AI and IoT-based Intelligent Health Care & Sanitation 9815136542, 9789815136548

Table of contents :
Cover
Title
Copyright
End User License Agreement
Contents
Preface
KEY FEATURES
List of Contributors
IoT Based Website for Identification of Acute Lymphoblastic Leukemia using DL
R. Ambika1,*, S. Thejaswini2,*, N. Ramesh Babu3,*, Tariq Hussain Sheikh4, Nagaraj Bhat5 and Zafaryab Rasool6
1. INTRODUCTION
2. LITERATURE SURVEY
3. MATERIALS AND METHODS
4. DATA COLLECTION
5. DATA PREPROCESSING
5.1. Resizing
5.2. Image Augmentation
6. DL – VGG 16
7. WEB DEVELOPMENT
8. RESULTS AND DISCUSSION
9. ADVANTAGES OF THE STUDY
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
AI and IoT-based Intelligent Management of Heart Rate Monitoring Systems
Vedanarayanan Venugopal1,*, Sujata V. Mallapur2, T.N.R. Kumar3, V. Shanmugasundaram4, M. Lakshminarayana5 and Ajit Kumar6
1. INTRODUCTION
2. LITERATURE SURVEY
3. PROPOSED SYSTEM
4. ARTIFICIAL NEURAL NETWORK
5. IMPLEMENTATION
6. RESULT AND DISCUSSION
7. STATE OF ART
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Deep Learning Applications for IoT in Healthcare Using Effects of Mobile Computing
Koteswara Rao Vaddempudi1,*, K.R. Shobha2, Ahmed Mateen Buttar3, Sonu Kumar4, C.R. Aditya5 and Ajit Kumar6
1. INTRODUCTION
2. LITERATURE SURVEY
3. MATERIALS AND METHODS
4. DATASET DESCRIPTION
5. ARTIFICIAL NEURAL NETWORK
6. IMPLEMENTATION
7. RESULT AND DISCUSSION
8. NOVELTY OF THE STUDY
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Innovative IoT-Based Wearable Sensors for the Prediction & Analysis of Diseases in the Healthcare Sector
Koteswara Rao Vaddempudi1, Abdul Rahman H Ali2, Abdullah Al-Shenqiti3, Christopher Francis Britto4, N. Krishnamoorthy5 and Aman Abidi6,*
1. INTRODUCTION
2. LITERATURE SURVEY
3. PROPOSED SYSTEM
3.1. Block Diagram
3.2. Flow Diagram
3.3. Hardware Implementation
3.4. IoT Cloud
4. RESULTS AND DISCUSSION
5. CONTRIBUTION TO THE HEALTH SECTOR
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Construction and Evaluation of Deep Neural Network-based Predictive Controller for Drug Preparation
K. Sheela Sobana Rani1, Dattathreya 2, Shubhi Jain3, Nayani Sateesh4, M. Lakshminarayana5 and Dimitrios Alexios Karras6,*
1. INTRODUCTION
2. MODEL DESCRIPTION
3. SYSTEM IDENTIFICATION
3.1. Material Balance Equation for Drug Preparation
3.2. Mass Flow Rate for Drug Preparation
4. PREDICTIVE CONTROLLER
5. RESULTS
5.1. Simulation Results of Drug Preparation for NN Controller and PID Controller
6. DISCUSSION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Machine Learning based Predictive Analysis and Algorithm for Analysis Severity of Breast Cancer
B. Radha1,*, Chandra Sekhar Kolli2, K R Prasanna Kumar3, Perumalraja Rengaraju4, S. Kamalesh4 and Ahmed Mateen Buttar5
1. INTRODUCTION
2. MATERIALS AND METHODS
3. FE
3.1. DB4
3.2. HAAR
4. CLASSIFICATION TECHNIQUES
4.1. SVM
4.2. RF
4.3. LDA
5. RESULT AND DISCUSSION
6. GAPS FILLED
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Glaucoma Detection Using Retinal Fundus Image by Employing Deep Learning Algorithm
K.T. Ilayarajaa1, M. Sugadev1, Shantala Devi Patil1, V. Vani2, H. Roopa2 and Sachin Kumar2,*
1. INTRODUCTION
2. LITERATURE SURVEY
3. MATERIALS AND METHODS
4. PREPROCESSING TECHNIQUES
5. DL MODEL
5.1. Transfer Learning (VGG-19)
5.2. CNN
6. RESULT ANALYSIS
7. DISCUSSION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Texture Analysis-based Features Extraction & Classification of Lung Cancer Using Machine Learning
Korla Swaroopa1,*, N. Chaitanya Kumar2, Christopher Francis Britto3, M. Malathi4, Karthika Ganesan5 and Sachin Kumar6
1. INTRODUCTION
2. METHODOLOGY
3. FE
3.1. GLCM
3.2. GLRM
4. CLASSIFICATION METHODS
4.1. SVM
4.2. KNN
5. RESULTS AND DISCUSSION
6. GAPS FILLED
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Implementation of the Deep Learning-based Website For Pneumonia Detection & Classification
V. Vedanarayanan1, Nagaraj G. Cholli2, Merin Meleet2, Bharat Maurya3, G. Appasami4 and Madhu Khurana5,*
1. INTRODUCTION
2. MATERIALS AND METHODS
3. PREPROCESSING TECHNIQUES
3.1. Data Resizing
3.2. Data Augmentation
4. DL TECHNIQUES
4.1. VGG-16
4.2. RESNET-50
5. WEB DEVELOPMENT
6. RESULT AND DISCUSSION
7. STATE OF ART
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Design and Development of Deep Learning Model For Predicting Skin Cancer and Deployed Using a Mobile App
Shweta M Madiwal1,*, M. Sudhakar2, Muthukumar Subramanian3, B. Venkata Srinivasulu4, S. Nagaprasad5 and Madhu Khurana6
1. INTRODUCTION
2. METHODOLOGY
3. PREPROCESSING
4. PREDICTION METHODS
5. DEPLOYMENT
6. RESULT
7. DISCUSSION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Feature Extraction and Diagnosis of Dementia using Magnetic Resonance Imaging
Praveen Gupta1,*, Nagendra Kumar2, Ajad2, N. Arulkumar3 and Muthukumar Subramanian4
1. INTRODUCTION
2. MATERIALS AND METHODS
3. FE TECHNIQUES
3.1. GLCM
3.2. GLRM
4. CLASSIFICATION TECHNIQUES
4.1. Support Vector Machine (SVM)
4.2. K-Nearest Neighbor (KNN)
4.3. Random Forest (RF)
4.4. Proposed Method
5. RESULT AND DISCUSSION
6. PROPOSED IDEA
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Deep Learning-Based Regulation of Healthcare Efficiency and Medical Services
T. Vamshi Mohana1,*, Mrunalini U. Buradkar2, Kamal Alaskar3, Tariq Hussain Sheikh4 and Makhan Kumbhkar5
1. INTRODUCTION
2. IOT IN MEDICAL CARE SERVICES
3. RELATED WORKS
4. PROPOSED SYSTEM
4.1. Interaction of RNN with LSTM
5. RESULTS AND DISCUSSION
6. NOVELTY OF THE PROPOSED WORK
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
An Efficient Design and Comparison of Machine Learning Model for Diagnosis of Cardiovascular Disease
Dillip Narayan Sahu1,*, G. Sudhakar2, Chandrakala G Raju3, Hemlata Joshi4 and Makhan Kumbhkar5
1. INTRODUCTION
1.1. Literature Survey
2. ML
3. METHODOLOGY
3.1. Source of Data
3.2. Data Pre-processing
4. ML ALGORITHM
4.1. NB Classifier
4.2. SVM
4.3 KNN
5. RESULT AND ANALYSIS
5.1. Performance Measures
5.2. Confusion Matrix: NB Classifier
5.3. Confusion Matrix: SVM
5.4. Confusion Matrix: KNN
5.5. ML Algorithm Comparison
6. IMPORTANCE OF THE STUDY
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Deep Learning Based Object Detection Using Mask R-CNN
Vinit Gupta1,*, Aditya Mandloi1, Santosh Pawar2, T.V Aravinda3 and K.R Krishnareddy3
1. INTRODUCTION
2. LITERATURE SURVEY
3. METHODOLOGIES
3.1. Data Collection and Preprocessing
3.2. Construction of Mask R-CNN
3.3. Training of Mask R-CNN
4. RESULT ANALYSIS
5. DISCUSSION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Design and Comparison Of Deep Learning Architecture For Image-based Detection of Plant Diseases
Makarand Upadhyaya1,*, Naveen Nagendrappa Malvade2, Arvind Kumar Shukla3, Ranjan Walia4 and K Nirmala Devi5
1. INTRODUCTION
1.1. Literature Survey
2. METHODOLOGY
3. DATA COLLECTION AND PREPARATION
3.1. Data Collection
3.2. Data Preprocessing
3.3. Data Augmentation
4. DEEP LEARNING NETWORKS
4.1. Convolution Neural Network
4.2. CNN-Long Short-Term Memory
5. RESULT AND DISCUSSION
5.1. CNN Performance Measure
5.2. CNN- LSTM
6. NOVELTY
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Discernment of Paddy Crop Disease by Employing CNN and Transfer Learning Methods of Deep Learning
Arvind Kumar Shukla1,*, Naveen Nagendrappa Malvade2, Girish Saunshi3, P. Rajasekar4 and S.V. Vijaya Karthik4
1. INTRODUCTION
2. LITERATURE SURVEY
3. METHODOLOGIES
4. DISEASE AND PREPROCESSING
4.1. Rescaling
4.2. Image Shearing
4.3. Zooming
4.4. Horizontal Flip
5. DL METHODS
5.1. CNN
5.2. Transfer Learning
6. RESULTS AND DISCUSSION
7. IDENTIFICATION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Deploying Deep Learning Model on the Google Cloud Platform For Disease Prediction
C.R. Aditya1,*, Chandra Sekhar Kolli2, Korla Swaroopa3, S. Hemavathi4 and Santosh Karajgi5
1. INTRODUCTION
2. LITERATURE SURVEY
3. METHODOLOGIES
3.1. Brain Tumor Data
3.2. Image Resizing
3.3. Image Rescaling
3.4. Image Data Generator
3.5. Construct VGG-16
3.6. Train VGG-16
3.7. Validate VGG-16
3.8. Deploy in GCP
4. Results and Discussion
5. REAL-TIME IMPLEMENTATIONS
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Classification and Diagnosis of Alzheimer’s Disease using Magnetic Resonance Imaging
K.R. Shobha1,*, Vaishali Gajendra Shende2, Anuradha Patil2, Jagadeesh Kumar Ega3 and Kaushalendra Kumar4
1. INTRODUCTION
2. METHODS AND MATERIALS
3. PREPROCESSING TECHNIQUES
3.1. Image Resizing
3.2. Smoothing using the Gaussian Method
4. FEATURE EXTRACTION
4.1. GLCM
4.2. HAAR
5. CLASSIFICATION TECHNIQUES
5.1. SVM
5.2. LDA
6. RESULT COMPARISON
7. DISCUSSION
CONCLUSION
CONSENT FOR PUBLICATON
CONFLICT OF INTEREST
ACKNOWLEDGEMENT
REFERENCES
Subject Index

Citation preview

AI and IoT-based intelligent Health Care & Sanitation Edited by Shashank Awasthi

Department of Computer Science and Engineering G.L. Bajaj Institute of Technology and Management (GLBITM), Affiliated to AKTU India

Mahaveer Singh Naruka

G.L. Bajaj Institute of Technology and Management (GLBITM) Affiliated to AKTU India

Satya Prakash Yadav

Department of Computer Science and Engineering G.L. Bajaj Institute of Technology and Management (GLBITM) Affiliated to AKTU, Greater Noida- 201306, U.P. India Graduate Program in Telecommunications Engineering (PPGET), Federal Institute of Education, Science, and Technology of Ceará (IFCE) Brazil

& Victor Hugo C. de Albuquerque

Department of Teleinformatics Engineering (DETI) Federal University of Ceará Brazil

AI and IoT-based intelligent Health Care & Sanitation Editors: Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav & Victor Hugo C. de Albuquerque ISBN (Online): 978-981-5136-53-1 ISBN (Print): 978-981-5136-54-8 ISBN (Paperback): 978-981-5136-55-5 © 2023, Bentham Books imprint. Published by Bentham Science Publishers Pte. Ltd. Singapore. All Rights Reserved. First published in 2023.

BSP-EB-PRO-9789815136531-TP-290-TC-18-PD-20230413

BENTHAM SCIENCE PUBLISHERS LTD.

End User License Agreement (for non-institutional, personal use) This is an agreement between you and Bentham Science Publishers Ltd. Please read this License Agreement carefully before using the book/echapter/ejournal (“Work”). Your use of the Work constitutes your agreement to the terms and conditions set forth in this License Agreement. If you do not agree to these terms and conditions then you should not use the Work. Bentham Science Publishers agrees to grant you a non-exclusive, non-transferable limited license to use the Work subject to and in accordance with the following terms and conditions. This License Agreement is for non-library, personal use only. For a library / institutional / multi user license in respect of the Work, please contact: [email protected].

Usage Rules: 1. All rights reserved: The Work is the subject of copyright and Bentham Science Publishers either owns the Work (and the copyright in it) or is licensed to distribute the Work. You shall not copy, reproduce, modify, remove, delete, augment, add to, publish, transmit, sell, resell, create derivative works from, or in any way exploit the Work or make the Work available for others to do any of the same, in any form or by any means, in whole or in part, in each case without the prior written permission of Bentham Science Publishers, unless stated otherwise in this License Agreement. 2. You may download a copy of the Work on one occasion to one personal computer (including tablet, laptop, desktop, or other such devices). You may make one back-up copy of the Work to avoid losing it. 3. The unauthorised use or distribution of copyrighted or other proprietary content is illegal and could subject you to liability for substantial money damages. You will be liable for any damage resulting from your misuse of the Work or any violation of this License Agreement, including any infringement by you of copyrights or proprietary rights.

Disclaimer: Bentham Science Publishers does not guarantee that the information in the Work is error-free, or warrant that it will meet your requirements or that access to the Work will be uninterrupted or error-free. The Work is provided "as is" without warranty of any kind, either express or implied or statutory, including, without limitation, implied warranties of merchantability and fitness for a particular purpose. The entire risk as to the results and performance of the Work is assumed by you. No responsibility is assumed by Bentham Science Publishers, its staff, editors and/or authors for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products instruction, advertisements or ideas contained in the Work.

Limitation of Liability: In no event will Bentham Science Publishers, its staff, editors and/or authors, be liable for any damages, including, without limitation, special, incidental and/or consequential damages and/or damages for lost data and/or profits arising out of (whether directly or indirectly) the use or inability to use the Work. The entire liability of Bentham Science Publishers shall be limited to the amount actually paid by you for the Work.

General: 1. Any dispute or claim arising out of or in connection with this License Agreement or the Work (including non-contractual disputes or claims) will be governed by and construed in accordance with the laws of Singapore. Each party agrees that the courts of the state of Singapore shall have exclusive jurisdiction to settle any dispute or claim arising out of or in connection with this License Agreement or the Work (including non-contractual disputes or claims). 2. Your rights under this License Agreement will automatically terminate without notice and without the

need for a court order if at any point you breach any terms of this License Agreement. In no event will any delay or failure by Bentham Science Publishers in enforcing your compliance with this License Agreement constitute a waiver of any of its rights. 3. You acknowledge that you have read this License Agreement, and agree to be bound by its terms and conditions. To the extent that any other terms and conditions presented on any website of Bentham Science Publishers conflict with, or are inconsistent with, the terms and conditions set out in this License Agreement, you acknowledge that the terms and conditions set out in this License Agreement shall prevail. Bentham Science Publishers Pte. Ltd. 80 Robinson Road #02-00 Singapore 068898 Singapore Email: [email protected]

BSP-EB-PRO-9789815136531-TP-290-TC-18-PD-20230413

CONTENTS PREFACE ................................................................................................................................................ i KEY FEATURES ........................................................................................................................... i LIST OF CONTRIBUTORS .................................................................................................................. iii CHAPTER 1 IOT BASED WEBSITE FOR IDENTIFICATION OF ACUTE LYMPHOBLASTIC LEUKEMIA USING DL .................................................................................... R. Ambika1, S. Thejaswini2, N. Ramesh Babu, Tariq Hussain Sheikh, Nagaraj Bhat and Zafaryab Rasool 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. MATERIALS AND METHODS ............................................................................................... 4. DATA COLLECTION ............................................................................................................... 5. DATA PREPROCESSING ........................................................................................................ 5.1. Resizing ............................................................................................................................ 5.2. Image Augmentation ........................................................................................................ 6. DL – VGG 16 ............................................................................................................................... 7. WEB DEVELOPMENT ............................................................................................................. 8. RESULTS AND DISCUSSION ................................................................................................. 9. ADVANTAGES OF THE STUDY ............................................................................................ CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 2 AI AND IOT-BASED INTELLIGENT MANAGEMENT OF HEART RATE MONITORING SYSTEMS .................................................................................................................... .Vedanarayanan Venugopal1, Sujata V. Mallapur, T.N.R. Kumar, V Shanmugasundaram, M. Lakshminarayana and Ajit Kumar 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. PROPOSED SYSTEM ............................................................................................................... 4. ARTIFICIAL NEURAL NETWORK ...................................................................................... 5. IMPLEMENTATION ................................................................................................................ 6. RESULT AND DISCUSSION ................................................................................................... 7. STATE OF ART ......................................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 3 DEEP LEARNING APPLICATIONS FOR IOT IN HEALTHCARE USING EFFECTS OF MOBILE COMPUTING ............................................................................................... ,Koteswara Rao Vaddempudi1, K.R. Shobha, Ahmed Mateen Buttar, Sonu Kumar C.R. Aditya and Ajit Kumar 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. MATERIALS AND METHODS ............................................................................................... 4. DATASET DESCRIPTION .......................................................................................................

1 2 2 4 5 5 6 6 6 9 9 13 13 14 14 14 14 16 17 18 20 23 26 27 31 31 31 31 31 32 33 34 34 36 37

5. ARTIFICIAL NEURAL NETWORK ...................................................................................... 6. IMPLEMENTATION ................................................................................................................ 7. RESULT AND DISCUSSION ................................................................................................... 8. NOVELTY OF THE STUDY .................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 4 INNOVATIVE IOT-BASED WEARABLE SENSORS FOR THE PREDICTION & ANALYSIS OF DISEASES IN THE HEALTHCARE SECTOR .................................................. Koteswara Rao Vaddempudi, Abdul Rahman H Ali, Abdullah Al-Shenqiti, Christopher, Francis Britto, N. Krishnamoorthy and Aman Abidi 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. PROPOSED SYSTEM ............................................................................................................... 3.1. Block Diagram ................................................................................................................. 3.2. Flow Diagram .................................................................................................................. 3.3. Hardware Implementation ............................................................................................... 3.4. IoT Cloud ......................................................................................................................... 4. RESULTS AND DISCUSSION ................................................................................................. 5. CONTRIBUTION TO THE HEALTH SECTOR ................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 5 CONSTRUCTION AND EVALUATION OF DEEP NEURAL NETWORKBASED PREDICTIVE CONTROLLER FOR DRUG PREPARATION .......................................... .K. Sheela Sobana Rani, Dattathreya , Shubhi Jain, Nayani Sateesh, M, Lakshminarayana and Dimitrios Alexios Karras 1. INTRODUCTION ...................................................................................................................... 2. MODEL DESCRIPTION .......................................................................................................... 3. SYSTEM IDENTIFICATION ................................................................................................... 3.1. Material Balance Equation for Drug Preparation ............................................................ 3.2. Mass Flow Rate for Drug Preparation ............................................................................. 4. PREDICTIVE CONTROLLER ................................................................................................ 5. RESULTS .................................................................................................................................... 5.1. Simulation Results of Drug Preparation for NN Controller and PID Controller ............. 6. DISCUSSION .............................................................................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ...............................................................................................................................

38 40 41 46 47 48 48 48 48 50 51 52 53 54 55 56 57 58 62 63 63 63 63 63 65 66 68 70 70 72 74 77 78 80 81 81 81 81 81

CHAPTER 6 MACHINE LEARNING BASED PREDICTIVE ANALYSIS AND ALGORITHM FOR ANALYSIS SEVERITY OF BREAST CANCER ...................................................................... 83 .B. Radha1, Chandra Sekhar Kolli, K R Prasanna Kumar, Perumalraja Rengaraju, S Kamalesh and Ahmed Mateen Buttar

1. INTRODUCTION ...................................................................................................................... 2. MATERIALS AND METHODS ............................................................................................... 3. FE ................................................................................................................................................. 3.1. DB4 .................................................................................................................................. 3.2. HAAR ........................................................................................................................................ 4. CLASSIFICATION TECHNIQUES ........................................................................................ 4.1. SVM ................................................................................................................................. 4.2. RF ..................................................................................................................................... 4.3. LDA ................................................................................................................................. 5. RESULT AND DISCUSSION ................................................................................................... 6. GAPS FILLED ............................................................................................................................ CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 7 GLAUCOMA DETECTION USING RETINAL FUNDUS IMAGE BY EMPLOYING DEEP LEARNING ALGORITHM ............................................................................. K.T. Ilayarajaa, M. Sugadev, Shantala Devi Patil, V. Vani, H. Roopa and Sachin ,Kumar 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. MATERIALS AND METHODS ............................................................................................... 4. PREPROCESSING TECHNIQUES ......................................................................................... 5. DL MODEL ................................................................................................................................. 5.1. Transfer Learning (VGG-19) ........................................................................................... 5.2. CNN ................................................................................................................................. 6. RESULT ANALYSIS ................................................................................................................. 7. DISCUSSION .............................................................................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 8 TEXTURE ANALYSIS-BASED FEATURES EXTRACTION & CLASSIFICATION OF LUNG CANCER USING MACHINE LEARNING .................................. ,Korla Swaroopa1, N. Chaitanya Kumar, Christopher Francis Britto, M. Malathi Karthika Ganesan and Sachin Kumar 1. INTRODUCTION ...................................................................................................................... 2. METHODOLOGY ..................................................................................................................... 3. FE ................................................................................................................................................. 3.1. GLCM .............................................................................................................................. 3.2. GLRM .............................................................................................................................. 4. CLASSIFICATION METHODS .............................................................................................. 4.1. SVM ................................................................................................................................. 4.2. KNN ................................................................................................................................. 5. RESULTS AND DISCUSSION ................................................................................................. 6. GAPS FILLED ............................................................................................................................ CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON ..................................................................................................

84 85 86 87 87 88 88 89 89 90 94 95 95 95 95 95 98 98 99 100 101 102 102 103 104 111 111 112 112 112 112 114 115 116 117 118 120 121 121 122 123 127 127 127

CONFLICT OF INTEREST ......................................................................................................... 127 ACKNOWLEDGEMENT ............................................................................................................. 127 REFERENCES ............................................................................................................................... 127 CHAPTER 9 IMPLEMENTATION OF THE DEEP LEARNING-BASED WEBSITE FOR PNEUMONIA DETECTION & CLASSIFICATION ......................................................................... V. Vedanarayanan, Nagaraj G. Cholli, Merin Meleet, Bharat Maurya, G. Appasami ,and Madhu Khurana 1. INTRODUCTION ...................................................................................................................... 2. MATERIALS AND METHODS ............................................................................................... 3. PREPROCESSING TECHNIQUES ......................................................................................... 3.1. Data Resizing ................................................................................................................... 3.2. Data Augmentation .......................................................................................................... 4. DL TECHNIQUES ..................................................................................................................... 4.1. VGG-16 ............................................................................................................................ 4.2. RESNET-50 ..................................................................................................................... 5. WEB DEVELOPMENT ............................................................................................................. 6. RESULT AND DISCUSSION ................................................................................................... 7. STATE OF ART ......................................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 10 DESIGN AND DEVELOPMENT OF DEEP LEARNING MODEL FOR PREDICTING SKIN CANCER AND DEPLOYED USING A MOBILE APP ................................ Shweta M Madiwal, M. Sudhakar, Muthukumar Subramanian, B. Venkata Srinivasulu, S. Nagaprasad and Madhu Khurana 1. INTRODUCTION ...................................................................................................................... 2. METHODOLOGY ..................................................................................................................... 3. PREPROCESSING .................................................................................................................... 4. PREDICTION METHODS ....................................................................................................... 5. DEPLOYMENT .......................................................................................................................... 6. RESULT ...................................................................................................................................... 7. DISCUSSION .............................................................................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 11 FEATURE EXTRACTION AND DIAGNOSIS OF DEMENTIA USING MAGNETIC RESONANCE IMAGING ............................................................................................... Praveen Gupta1, Nagendra Kumar, Ajad, N. Arulkumar and Muthukumar Subramanian 1. INTRODUCTION ...................................................................................................................... 2. MATERIALS AND METHODS ............................................................................................... 3. FE TECHNIQUES ...................................................................................................................... 3.1. GLCM .............................................................................................................................. 3.2. GLRM .............................................................................................................................. 4. CLASSIFICATION TECHNIQUES ........................................................................................

129 130 131 132 132 133 133 133 133 135 136 141 141 142 142 142 142 144 145 146 147 148 149 150 156 157 157 157 157 157 159 160 162 163 163 164 165

4.1. Support Vector Machine (SVM) ...................................................................................... 4.2. K-Nearest Neighbor (KNN) ............................................................................................. 4.3. Random Forest (RF) ........................................................................................................ 4.4. Proposed Method ............................................................................................................. 5. RESULT AND DISCUSSION ................................................................................................... 6. PROPOSED IDEA ..................................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 12 DEEP LEARNING-BASED REGULATION OF HEALTHCARE EFFICIENCY AND MEDICAL SERVICES ................................................................................................................. T. Vamshi Mohana1, Mrunalini U. Buradkar, Kamal Alaskar, Tariq Hussain Sheikh and Makhan Kumbhkar 1. INTRODUCTION ...................................................................................................................... 2. IOT IN MEDICAL CARE SERVICES .................................................................................... 3. RELATED WORKS ................................................................................................................... 4. PROPOSED SYSTEM ............................................................................................................... 4.1. Interaction of RNN with LSTM ....................................................................................... 5. RESULTS AND DISCUSSION ................................................................................................. 6. NOVELTY OF THE PROPOSED WORK .............................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 13 AN EFFICIENT DESIGN AND COMPARISON OF MACHINE LEARNING MODEL FOR DIAGNOSIS OF CARDIOVASCULAR DISEASE ................................................... Dillip Narayan Sahu1, G. Sudhakar, Chandrakala G Raju, Hemlata Joshi and Makhan Kumbhkar 1. INTRODUCTION ...................................................................................................................... 1.1. Literature Survey ............................................................................................................. 2. ML ................................................................................................................................................ 3. METHODOLOGY ..................................................................................................................... 3.1. Source of Data .................................................................................................................. 3.2. Data Pre-processing ......................................................................................................... 4. ML ALGORITHM ..................................................................................................................... 4.1. NB Classifier .................................................................................................................... 4.2. SVM ................................................................................................................................. 4.3 KNN .................................................................................................................................. 5. RESULT AND ANALYSIS ....................................................................................................... 5.1. Performance Measures ..................................................................................................... 5.2. Confusion Matrix: NB Classifier ..................................................................................... 5.3. Confusion Matrix: SVM .................................................................................................. 5.4. Confusion Matrix: KNN .................................................................................................. 5.5. ML Algorithm Comparison ............................................................................................. 6. IMPORTANCE OF THE STUDY ............................................................................................ CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON ..................................................................................................

165 167 168 169 169 173 173 174 174 174 174 176 177 178 180 182 183 185 188 188 188 189 189 189 191 192 192 193 194 194 195 195 195 196 196 197 198 198 199 200 203 204 204 205

CONFLICT OF INTEREST ......................................................................................................... 205 ACKNOWLEDGEMENT ............................................................................................................. 205 REFERENCES ............................................................................................................................... 205 CHAPTER 14 DEEP LEARNING BASED OBJECT DETECTION USING MASK R-CNN ..... Vinit Gupta, Aditya Mandloi, Santosh Pawar, T.V Aravinda and K.R Krishnareddy 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. METHODOLOGIES .................................................................................................................. 3.1. Data Collection and Preprocessing .................................................................................. 3.2. Construction of Mask R-CNN ......................................................................................... 3.3. Training of Mask R-CNN ................................................................................................ 4. RESULT ANALYSIS ................................................................................................................. 5. DISCUSSION .............................................................................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 15 DESIGN AND COMPARISON OF DEEP LEARNING ARCHITECTURE FOR IMAGE-BASED DETECTION OF PLANT DISEASES .................................................................... ,Makarand Upadhyaya1, Naveen Nagendrappa Malvade, Arvind Kumar Shukla Ranjan Walia and K Nirmala Devi 1. INTRODUCTION ...................................................................................................................... 1.1. Literature Survey ............................................................................................................. 2. METHODOLOGY ..................................................................................................................... 3. DATA COLLECTION AND PREPARATION ....................................................................... 3.1. Data Collection ................................................................................................................ 3.2. Data Preprocessing ........................................................................................................... 3.3. Data Augmentation .......................................................................................................... 4. DEEP LEARNING NETWORKS ............................................................................................. 4.1. Convolution Neural Network ........................................................................................... 4.2. CNN-Long Short-Term Memory ..................................................................................... 5. RESULT AND DISCUSSION ................................................................................................... 5.1. CNN Performance Measure ............................................................................................. 5.2. CNN- LSTM .................................................................................................................... 6. NOVELTY ................................................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 16 DISCERNMENT OF PADDY CROP DISEASE BY EMPLOYING CNN AND TRANSFER LEARNING METHODS OF DEEP LEARNING ......................................................... Arvind Kumar Shukla1, Naveen Nagendrappa Malvade, Girish Saunshi, P. Rajasekar and S.V. Vijaya Karthik 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. METHODOLOGIES .................................................................................................................. 4. DISEASE AND PREPROCESSING .........................................................................................

207 208 209 211 212 213 214 215 219 220 220 220 220 220 222 223 224 225 227 227 227 228 228 228 229 230 230 234 237 237 237 237 237 237 240 241 241 243 244

4.1. Rescaling .......................................................................................................................... 4.2. Image Shearing ................................................................................................................ 4.3. Zooming ........................................................................................................................... 4.4. Horizontal Flip ................................................................................................................. 5. DL METHODS ........................................................................................................................... 5.1. CNN ................................................................................................................................. 5.2. Transfer Learning ............................................................................................................. 6. RESULTS AND DISCUSSION ................................................................................................. 7. IDENTIFICATION .................................................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 17 DEPLOYING DEEP LEARNING MODEL ON THE GOOGLE CLOUD PLATFORM FOR DISEASE PREDICTION ...................................................................................... C.R. Aditya, Chandra Sekhar Kolli, Korla Swaroopa, S. Hemavathi and Santosh Karajgi 1. INTRODUCTION ...................................................................................................................... 2. LITERATURE SURVEY .......................................................................................................... 3. METHODOLOGIES .................................................................................................................. 3.1. Brain Tumor Data ............................................................................................................ 3.2. Image Resizing ................................................................................................................. 3.3. Image Rescaling ............................................................................................................... 3.4. Image Data Generator ...................................................................................................... 3.5. Construct VGG-16 ........................................................................................................... 3.6. Train VGG-16 .................................................................................................................. 3.7. Validate VGG-16 ............................................................................................................. 3.8. Deploy in GCP ................................................................................................................. 4. Results and Discussion ....................................................................................................... 5. REAL-TIME IMPLEMENTATIONS ...................................................................................... CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ............................................................................................................................... CHAPTER 18 CLASSIFICATION AND DIAGNOSIS OF ALZHEIMER’S DISEASE USING MAGNETIC RESONANCE IMAGING ............................................................................................... K.R. Shobha, Vaishali Gajendra Shende, Anuradha Patil, Jagadeesh Kumar Ega and Kaushalendra Kumar 1. INTRODUCTION ...................................................................................................................... 2. METHODS AND MATERIALS ............................................................................................... 3. PREPROCESSING TECHNIQUES ......................................................................................... 3.1. Image Resizing ................................................................................................................ 3.2. Smoothing using the Gaussian Method .......................................................................... 4. FEATURE EXTRACTION ....................................................................................................... 4.1. GLCM ............................................................................................................................. 4.2. HAAR .............................................................................................................................. 5. CLASSIFICATION TECHNIQUES ........................................................................................ 5.1. SVM .................................................................................................................................

245 245 245 245 246 246 246 248 252 252 253 253 253 253 255 256 256 257 258 259 259 260 261 262 263 263 265 266 267 267 267 267 267 269 270 271 272 273 273 273 273 275 275 275

5.2. LDA ................................................................................................................................. 6. RESULT COMPARISON .......................................................................................................... 7. DISCUSSION .............................................................................................................................. CONCLUSION ............................................................................................................................... CONSENT FOR PUBLICATON .................................................................................................. CONFLICT OF INTEREST ......................................................................................................... ACKNOWLEDGEMENT ............................................................................................................. REFERENCES ...............................................................................................................................

276 277 281 282 283 283 283 283

SUBJECT INDEX .................................................................................................................................... 2

i

PREFACE The book aims to provide a deeper understanding of the relevant aspects of AI and IoT impacting each other’s efficacy for better output. Readers may discover a reliable and accessible one-stop resource; An introduction to Artificial Intelligence presents the first full examination of applications of AI, as well as IoT presents the smart objects, sensors or actuators using a secure protocol to the Artificial Intelligence approaches. They are designed in a way to provide an understanding of the foundations of artificial intelligence. It examines Education powered by AI, Entertainment and Artificial Intelligence, Home and Service Robots, Healthcare re-imagined, Predictive Policing, Space Exploration with AI, and weaponry in the world of AI. Through the volume, the authors provide detailed, wellillustrated treatments of each topic with abundant examples and exercises.

KEY FEATURES ●

●

●

This book contains different topics about the most important areas and challenges in the Internet of Things. In this book, you will be able to read about the different subparts that compose the Internet of Things and the different ways to create a better IoT network or platform. This book will contain the possible architecture and middleware that you can use to create a platform and how to manage the data that you obtain from different objects (Smart Objects, sensors, or actuators) using a secure protocol or messages when you send the data through insecure protocols. These improvements can be applied in different ways. We can use different technologies and create different applications, or we can use the research of other fields like Big Data to process the massive data of the devices, Artificial Intelligence to create smarter objects or algorithms, Model-Driven Engineering to facilitate the use of any people, Cloud Computing to send the computation and the applications to the cloud, and Cloud Robotics to manage and interconnect the Robots between themselves and with other objects.

ii ●

This book will bring forth the details in a precise and concise form about all the right and relevant technologies and tools to simplify and streamline the concepts, processes, model development, and budget development for readers of respective backgrounds. Shashank Awasthi Department of Computer Science and Engineering G.L. Bajaj Institute of Technology and Management (GLBITM), Affiliated to AKTU Greater Noida- 201306, U.P. India Mahaveer Singh Naruka G.L. Bajaj Institute of Technology and Management (GLBITM) Affiliated to AKTU Greater Noida- 201306, U.P. India Satya Prakash Yadav Department of Computer Science and Engineering G.L. Bajaj Institute of Technology and Management (GLBITM) Affiliated to AKTU, Greater Noida- 201306, U.P. India Graduate Program in Telecommunications Engineering (PPGET), Federal Institute of Education, Science, and Technology of Ceará (IFCE) Fortaleza-CE Brazil & Victor Hugo C. de Albuquerque Department of Teleinformatics Engineering (DETI) Federal University of Ceará Brazil

iii

List of Contributors Aman Abidi

Swinburne University of Technology, Victoria, Australia

Abdullah Al-Shenqiti

College of Medical Rehabilitation Sciences, Taibah University, Madinah, Saudi Arabia

Ajit Kumar

School of Computer Science and Engineering, Soongsil University, Seoul, South Korea

Anuradha Patil

Department of ECE, Faculty of Engineering and Technology Exclusive for Women Sharn Basva, University Kalburagi, Anuradha Patil, ECE Department, SB Campus, Vidya Nagar, Sharn Basva University Kalburagi, Karnataka, India

Ahmed Mateen Buttar

Department of Computer Science, University of Agriculture Faisalabad, Faisalabad, Punjab, Pakistan

Ahmed Mateen Buttar

Department of Computer Science, University of Agriculture Faisalabad, Faisalabad, Punjab, Pakistan

Ajit Kumar

School of Computer Science and Engineering, Soongsil University, Seoul, South Korea

Arvind Kumar Shukla

Department of Computer Application, , IFTM University, Moradabad, Uttar Pradesh, India

Ajad

Department of Electronics and Communication Engineering, S J C Institute of Technology, Chickballapur, Karnataka, India

Aditya Mandloi

Electronics Engineering Department, Medi-Caps University Indore, Madhya Pradesh, India

Abdul Rahman H Ali

Mahatma Gandhi University/Taibah University, Meghalaya, India

B. Radha

ICT & Cognitive Systems, Sri Krishna Arts and Science College, Coimbatore, Tamil Nadu, India

Bharat Maurya

Department of Political Science, Chaudhary Bansi Lal University, Bhiwani Haryana, India

B Venkata Srinivasulu

Department of CSE, JJTU, Rajasthan, India

Chandrakala G Raju

BMS College of Engineering, Bangalore, Karnataka, India

C.R. Aditya

Department of Computer Science & Engineering, Vidyavardhaka College of Engineering, Mysuru, Karnataka, India

Christopher Francis Britto

Information Technology & Computer Sciences, Mahatma Gandhi University, Meghalaya, India

Chandra Sekhar Kolli

Aditya College of Engineering & Technology, Surampalem, East Godavari, India

Dattathreya

Alva's Institute of Engineering and Technolog, Mijar, Moodbidri, Karnataka, India

Dimitrios Alexios Karras

National and Kapodistrian University of Athens (NKUA), Hellas, Greece

iv Dillip Narayan Sahu

Department of MCA, School of Computer Science, Gangadhar Meher University, Sambalpur, Odisha, India

G Appasami

Department of Computer Science, Central University of Tamil Nādu, Tiruvarur, Tamil Nādu, India

G. Sudhakar

Computer Science Engineering, School of Information Technology, Jawaharlal Nehru Technological University, Hyderabad, Telangana, India

Girish Saunshi

Department of Computer Science and Engineering, KLE Institute of Technology, KLEIT, Karnataka, India

Hemlata Joshi

Department of Statistics, CHRIST (Deemed to be University), Bangalore, Karnataka, India

H. Roopa

Information Science and Engineering, Bangalore Institute of Technology, Bangalore, Karnataka, India-560079

Jagadeesh Kumar Ega

Chaitanya (Deemed to be University), Hanamkonda, Telangana State, India

K. Sheela Sobana Rani

Department of Electrical and Electronics Engineering, Sri Ramakrishna Institute of Technology, Coimbatore, Tamil Nadu, India

Korla Swaroopa

Department of CSE, Aditya Engineering College, Surampalem- East Godavari, Andhra Pradesh, India

K.T. layarajaa

Department of Electronics and Communication Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu, India

Kaushalendra Kumar

Department of Bio-Science, Galgotias University, Greater Noida, Uttar Pradesh, India

K R Krishnareddy

SJM Institute of Technology, Chitradurga, Karnataka, India

Koteswara Rao Vaddempudi

ECE Department, Prakasam Engineering College, Kandukur, Prakasam, Andhra Pradesh, India

Karthika Ganesan

Sri Vidya College of Engineering and Technology, Sivakasi Main Road, Virudhunagar, Tamilnadu, India

Kamal Alaskar

Bharati Vidyapeeth (Deemed to Be University) Institute of Management Kolhapur, Maharashtra, India

K Nirmala Devi

Department of Computer Science and Engineering, Kongu Engineering College Perundurai, Erode, Tamilnadu, India

K R Prasanna Kumar

Siddaganga Institute of Technology, Tumkur, Karnataka, India

K.R. Shobha

Department of Electronics and Telecommunication Engineering, Ramaiah Institute of Technology, MSR Nagar, MSRIT Post, Bangalore, Karnataka, India

Muthukumar Subramanian

SRM Institute of Science & Technology, Tiruchirappalli Campus, (Deemed to be University), Trichy, Tamilnadu, India

Makarand Upadhyaya

Department of Management & Marketing, College of Business, University of Bahrain, Administration, Bahrain

v M. Malathi

Department of ECE, Vivekananda College of Engineering for Women (Autonomous), Elayampalayam, Namakka, Tamilnadu, India

Makhan Kumbhkar

Christian Eminent College, Indore, Madhyapradesh, India

M. Sugadev

Department of Electronics and Communication Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu, India

Makhan Kumbhka

Computer Science & Elex, Christian Eminent College, Indore, Madhyapradesh, India

Mrunalini U. Buradkar

Department of Electronics and Telecommunication Engineering, St.Vincent Pallotti College of Engineering & Technology, Nagpur, Maharashtra, India

Madhu Khurana

University of Gloucestershire, UK

M. Sudhakar

Department of Mechanical Engineering, Sri Sairam Engineering College, Chennai, Tamil Nadu, India

Muthukumar Subramanian

SRM Institute of Science & Technology, Tiruchirappalli Campus, (Deemed to be University), Trichy, Tamilnadu, India

Merin Meleet

Department of Information Science and Engineering, RV College of Engineering, Bengaluru, Karnataka, India

M. Lakshminarayana

Department of ECE, SJB Institute of Technology, Bengaluru, Karnataka, India-560060

N. Arulkumar

Department of Statistics and Data Science, CHRIST (Deemed to be University) Bangalore, Karnataka, India

Nagaraj Bhat

RV College of Engineering, Bengaluru, Karnataka, India-560059

Nayani Sateesh

Department of Information Technology, CVR College of Engineering, Vastunagar, Mangalpalli (V), Ibrahimpatnam (M), Rangareddy, Telangana, India-501510

Nagendra Kumar

Department of Electronics and Communication Engineering, S J C Institute of Technology, Chickballapur, Karnataka, India

N. Krishnamoorthy

Department of Software Systems and Engineering, Vellore Institute of Technology, Vellore Campus, Thiruvalam Road, Vellore, School of Information Technology and Engineering, Tamilnadu, India

Naveen Nagendrappa Malvade

Department of Information Science and Engineering, SKSVM Agadi College of Engineering and Technology, Lakshmeshwa, Karnataka, India

N. Chaitanya Kumar

Department of CSE, Sri Venkateswara Engineering College, Karakambadi Road, Tirupati, Andhra Pradesh, India

Nagaraj G Cholli

Department of Information Science and Engineering, RV College of Engineering, Bengaluru, Karnataka, India

N. Ramesh Babu

Department of Electronics and Communication Engineering, BMS Institute of Technology and Management, Bengaluru, Karnataka, India560064

vi Praveen Gupta

Department of Computer Science and Engineering, GITAM School of Technology, GITAM (Deemed to be University), Visakhapatnam, Andhra Pradesh, India

Perumalraja Rengaraju

Department of Information Technology, Velammal College of Engineering and Technology, Madura, Tamil Nādu, India

R. Ambika

Departmnt of Electronics and Communication Engineering, BMS Institute of Technology and Management, Bengaluru, Karnataka, India560064

Ranjan Walia

Department of Electrical Engineering, Model Institute of Engineering and Technology, Jammu & Kashmir, India

S. Hemavathi

CSIR - Central Electrochemical Research Institute (CECRI), CSIRMadras Complex, Chennai, Tamil Nadu, India-600113

S V Vijaya Karthik

Department of Computer Science and Engineering, KLE Institute of Technology, KLEIT, Dharwad, Karnataka, India

Shubhi Jain

Swami Keshvanand Institute of Technology Management & Gramothan, Ramnagaria, Jagatpura, Jaipur, Rajasthan, India

Shantala Devi Patil

Department of Electronics and Communication Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu, India

S. Kamalesh

Department of Information Technology, Velammal College of Engineering and Technology, Madura, Tamil Nādu, India

Sujata V. Mallapur

Sharnbasva University, Kalaburagi, Karnataka, India-585103

Sonu Kumar

Department of Electronics and Communication Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Guntur, Andhra Pradesh, India

S. Thejaswini

BMS Institute of Technology and Management, Bengaluru, Karnataka, India-560064

Sachin Kumar

College of IT Engineering, Kyungpook National University, Daegu, South Korea

Shweta M Madiwal

Department of CSE, Faculty of Engineering and Technology, Exclusively for Women's) Sharnbasva University Karnataka, Kalaburagi, India

S. Nagaprasad

Department of Computer Science, Tara Government College (A), Sangareddy, Telangana, India

Santosh Pawar

Electronics Engineering Department, Dr. APJ Kalam University, Indore, Madhya Pradesh, India

Santosh Karajgi

BLDEA's SSM College of Pharmacy and Research Centre Vijayapur, Karnataka, India

T V Aravinda

SJM Institute of Technology, Chitradurga, Karnataka, India

Tariq Hussain Sheikh

Government Degree College Poonch J&K UT, Poonch, J&K UT, India185101

T.N.R. Kumar

Ramaiah Institute of Technology, Bangalore, Karnataka, India-560054

vii T. Vamshi Mohana

Department of Computer Science, RBVRR Women's College, Hyderabad, Telangana, India

Vedanarayanan Venugopal

Electronics and Communication Engineering (ECE), Sathyabama Institute of Science and Technology (Deemed to be University), Chennai, Tamil Nadu, India-604119

V. Shanmugasundaram

Department of Electrical and Electronics Engineering, Sona College of Technology, Salem, Tamil Nadu, India-636005

V. Vani

Information Science and Engineering, Bangalore Institute of Technology, Bangalore, Karnataka, India

Vinit Gupta

Electronics Engineering Department, Medi-Caps University Indore, Madhya Pradesh, India

Vaishali Gajendra Shende

Department of ECE, Faculty of Engineering and Technology Exclusive for Women Sharn Basva, University Kalburagi, Anuradha Patil, ECE Department, SB Campus, Vidya Nagar, Sharn Basva University Kalburagi, Karnataka, India

Zafaryab Rasool

Deakin University, Geelong, Victoria, Australia

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 1-15

1

CHAPTER 1

IoT Based Website for Identification of Acute Lymphoblastic Leukemia using DL R. Ambika1,*, S. Thejaswini2, , N. Ramesh Babu3, , Tariq Hussain Sheikh4, Nagaraj Bhat5 and Zafaryab Rasool6 Department of Electronics and Communication Engineering, BMS Institute of Technology and Management, Bengaluru, Karnataka, India-560064 2 BMS Institute of Technology and Management, Bengaluru, Karnataka, India-560064 3 Department of Computer Science & Engineering, Amruta Institute of Engineering & Management Sciences, Bidadi, Karnataka, India-562109 4 Government Degree College Poonch J&K UT, Poonch, J&K UT, India-185101 5 RV College of Engineering, Bengaluru, Karnataka, India-560059 6 Deakin University, Geelong, Victoria, Australia 1

Abstract: A form of cancer known as leukemia, attacks the body's blood cells and bone marrow. This happens when cancer cells multiply rapidly in the bone marrow. The uploaded image is analyzed by the website, and if leukemia is present, the user is notified—a collection of pictures depicting leukemia as well as healthy bones and blood. Once collected from Kaggle, the data is preprocessed using methods like image scaling and enhancement. To create a Deep Learning (DL) model, we use the VGG-16 model. The processed data is used to “train” the model until optimal results are achieved. A Hypertext Markup Language (HTML) based website is built to showcase the model. Using a DL model, this website returns a response indicating whether or not the user's uploaded photograph shows signs of leukemia. The primary aim of this site is to lessen the likelihood that cancer cells may multiply while the patient waits for test results or is otherwise unaware of their condition. Waiting for results after a leukemia test can cause further stress and even other health problems, even if the person is found to be leukemia-free. This problem can be fixed if this website is used as a screening tool for leukemia.

Keywords: Deep Learning, Image Augmentation, Image Processing, Leukemia, VGG-16, Web Development. Corresponding author R. Ambika: BMS Institute of Technology and Management, Bengaluru, Karnataka, India; E-mail: [email protected] *

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

2 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

1. INTRODUCTION Extremely high numbers of white blood cells and platelets in the blood are thought to be the root cause of leukemia and other forms of bone marrow malignancy. Leukemia's early symptoms are similar to those of other diseases, such as fever, weight loss, chills, etc. Furthermore, the patient may experience no outward signs of leukemia's presence or progress as it is slowly killing them. Needle biopsies and other forms of differential diagnosis are used to check for leukemia. This test may take a while to get reliable findings, and the patient's condition or stress level may worsen as a result. The consequences for the patient's life are negative either way. The prevalence of this problem can be mitigated with the use of self-administered, accurate screening tests. The problem can be solved by a website that, given a picture of the patient's blood, can determine whether or not they have leukemia and provide that information to the user. After a user uploads an image, it is processed with a variety of image preprocessing processes, including cropping, resizing, and enhancing. Once the image has been preprocessed, a DL model is built using a module called VGG-16 and applied to the data. When using transfer learning with this DL module, the user needs just to adjust the first and last layers to achieve the desired results rather than the full module itself. To show the final result to the user, a website built in HTML is created alongside the model. Modifications to the website's output can be made in response to the results returned by the DL model. 2. LITERATURE SURVEY The unexpectedly high number of new cancer cases harms the mental health of the elderly. The World Health Organization classifies leukemia as an uncommon disease. Even though over one million Indians were diagnosed with leukemia, the absence of a precise forecast of the disease's presence is the primary reason for the severity of the disease. The following scientists have conducted extensive studies on the prevalence and effects of leukemia, as well as studies on DL algorithms that may be useful in detecting leukemia. Research into leukemia and its causes has been conducted by scientists at St. Jude Children's Research Hospital and other medical institutions. The comprehensive analysis [1] highlights the impact of diagnostics, prognostics, and therapies on advances in our understanding of the molecular processes involved in Acute Lymphoblastic Leukemia. Researchers from Philadelphia Children's Hospital's Oncology Division and Childhood Cancer Research Center report that 90% of patients with the most common childhood cancer are now cured. The most molecularly resistant subsets of acute lymphoblastic leukemia are the current focus of drug development efforts. Researchers from the Departments of

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 3

Hematology and Oncology, Pharmaceutical Sciences, and Pathological Studies, St. Jude Children's Research Center, and the Colleagues of Medicine and Pharmacy, Tennessee Health Center, St. Jude Children's Research Centre, agree that treating leukemia can be complicated depending on its type [2]. However, the success of your treatment can be enhanced by the use of several tools and methods. Chemical therapy was also proposed by them as an effective treatment for leukemia [3]. Leukemia is a malignancy of the blood and bone marrow. In a nutshell, cancer is the excessive growth of cells. Leukemia differs from other cancers in that it does not typically form a mass (tumor) visible in imaging tests like x-rays, increasing the likelihood of a catastrophic outcome. You can find a lot of leukemia cases today. Some are more prevalent among children, while others are more often observed among adults. Leukemia is the first step in the production of blood cells in the bone marrow. The stem cells go through a series of stages before reaching their final, mature state. Acute lymphocytic leukemia is the most common type of leukemia in people less than 39 years old, including children, adolescents, and young adults. Some 57% of new cases occur in persons under the age of 19. Hispanics and Caucasians are disproportionately affected. New cases of leukemia are diagnosed in about 1,7 per 100,000 American adults each year. The Department of Hematology at the NYU Perlmutter Cancer Center and the NYU School of Medicine conducted a comprehensive evaluation of the many subtypes of leukemia and their potential triggers. They claim that the treatment of acute lymphoblastic leukemia with dose-intensification chemotherapy and Allo-SCT represents a breakthrough in pediatric oncological sciences. Given the high-risk nature of the disease and the significant toxicity of chemotherapeutic treatments in adults, the outcomes are far less promising. Current drugs that are well-tolerated for the treatment of other cancers are now being investigated in Acute Lymphoblastic Leukemia [4]. The IoT, or the Internet of Things, is an effective resource that has many potential applications, one of which is in the medical profession. Alongside Big Data and AI, the IoT is expected to become a game-changer in many fields. According to a recent editorial article on IoT written by scholars at Indiana University in the United States of America, in general, IoT is defined as the interconnection of common objects, most of which also feature pervasive computing capabilities. Thanks to the exponential growth of the underlying technology, IoT offers tremendous potential for a plethora of new applications that will enhance the standard of living for all people. Researchers and professionals from all around the world have been focusing on IoT over the past few years. In the research [6], a brand-new image of the most recent state-of-the-art research on the topic is shown. When given the appropriate data, parameters, and conditions, DL can be

4 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

utilized to accurately categorize photos. This method has great potential for use in image classification for determining the likelihood of leukemia. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton published an in-depth explanation of DL and its applications throughout all of the technology in 2015. They concluded that ConvNets performed very well when used to recognize and segment visual elements. Despite these successes, mainstream computer and machine-learning groups had mostly abandoned ConvNets until the ImageNet competition. A Convolutional Neural Network (CNN) is trained to “turn” high-level images into subtitles by retrieving samples from a Recurrent Neural Network (RNN) as an additional entry from a test picture. This method can examine the images quickly [7]. To categorize brain scans, Indian researchers used a DL model called VGG16. Traditional machine-learning methods necessitate individualized classification algorithms. In contrast, CNN can sort images based on visual characteristics by extracting them from raw images. Only in this specific implementation were the VGG-16 model's final few levels modified to accommodate different types of images. The transmission-learned, pre-trained model was tested on a dataset obtained from the Harvard Medical School repository, which consists of both normal and pathological MR images of various neurological conditions. With this method, you can classify images from beginning to end without having to manually extract attributes [8]. The images of diseased eggplants were classified by a different group of Indian researchers using VGG-16, a DL model. They determined that features on the different VGG16 layers were fed to the MSVM to evaluate classification performance. According to the results of this research, field pictures (RGB) are superior to other types of images for classifying the five diseases. In the real world, the Cercospora plate was inaccurate because of misclassification. Researchers in this study were able to use images of leaves taken from isolated leaf samples and a field setting that had not been available to researchers before [9]. Even though it was done correctly, the website still needs to be flawless to give the user a satisfying interactive experience. Researchers from the School of Computing and Mathematical Sciences at Liverpool John Moores University have explored many methods that could be implemented in the real world to provide a spectacular experience. That said, a website development project might involve a wide range of activities, both technical and business in nature. In particular, there appears to be no standardized procedure for determining the overarching purpose or business functions to be supported by the company's website. 3. MATERIALS AND METHODS Several tools, including a DL model, a website, etc., are used to diagnose leukemia. The processing of medical images is nothing new in the medical area

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 5

[10, 11]. The following flowchart explains how the model and website function as a whole. Fig. (1) shows how photos acquired from Kaggle are utilized to train and evaluate the DL model. However, the images are preprocessed via image scaling and image augmentation before being utilized in the model.

Fig. (1). Work Flow of Leukemia Detection.

4. DATA COLLECTION Images of blood samples are gathered via Kaggle. Altogether, 15,135 images were gathered. To improve accuracy, we fed 70% of the data we obtained into the DL model training process, or roughly 10,594 data. Once the model had been created and trained, it was put to the test using the remaining 70% of the dataset, or 4541 images. 5. DATA PREPROCESSING Data from Kaggle may provide everything needed to create and train a model, but the images may have different data and parameters that make the model's performance inconsistent. As a result, before feeding images into DL models, they must be processed using certain methods. The images were processed using a

6 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

couple of different methods. Two of these techniques are image scaling and image enhancement. 5.1. Resizing To ensure that all images in a dataset are of the same size, a technique called “image resizing” is carried out. It is possible to alter the size of an image by raising or decreasing its pixel count. Reducing the size of an image while keeping its quality is called “image scaling.” Image scaling also has the primary benefit of cropping out any undesirable areas. 5.2. Image Augmentation Image augmentation is a technique for enhancing existing images to produce new data for use in the training of models. It's a helpful method for constructing CNN with a larger training set without needing to take any new pictures. In other words, it's a method for training DL models with an artificially larger dataset. If you use augmented images, your model's performance will improve even if you don't feed it a massive dataset of real-world images. 6. DL – VGG 16 When it comes to image classification, the VGG 16 or Oxford Net architecture (both of which use a DL algorithm) are among the most popular images. CNN systems employ this structure for image classification. As a pre-trained model, VGG-16 saves time and effort because it doesn't need to be tested and trained from scratch. There are a total of 16 layers in the VGG-16 model, 13 of which are convolutional and 3 of which are completely linked. The only layers that can be altered to suit the needs of the user are the top and bottom ones. Fig. (2) depicts the working of the VGG-16 architecture in detail. VGG is a particular network meant to be classified and located. As well as several prominent networks such as Alex Net and Google net, etc., VGG16 is employed in many difficulties with classifying images, however smaller network topologies, such as SqueezeNet, GoogLeNet, etc., are typically preferred. But for learning purposes, it is an excellent building block since it is straightforward to execute. There are 41 layers in the VGG-16 network. There are sixteen layers with learning weights: 13 coevolutionary layers and three completely linked. The VGG-16 architecture used in this research to detect leukemia is given in Table 1.

IoT Based Website

Fig. (2). Architechture of VGG – 16. Table 1. VGG -16 Architecture.

AI and IoT-based Intelligent Health Care & Sanitation 7

8 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

From the above architecture, it can be inferred that each convolutional block is used to train the model. This model is designed using the following steps. ● ●

● ● ● ●

● ●

●

A CNN model of the above-said parameters is imported. The top and the bottom layer of the model are then modified to classify the images based on the presence of leukemia. Fig. (3) represents a normal blood cell. Fig. (4) represents leukemia-affected blood cells. The images are then used to train the architecture. The trained VGG-16 model is then tested by using thirty percent of the images that are obtained from Kaggle. It can be found that the accuracy of the model increases with every epoch. The loss percentage of the model is inversely proportional to the accuracy as it decreases with every single epoch. In the end, the maximum accuracy is gained, and the minimum loss is obtained.

Fig. (3). Normal blood cell.

Fig. (4). Leukemia-affected blood cell.

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 9

The generated model can be used to detect the presence of leukemia in the image that will be uploaded by the user. 7. WEB DEVELOPMENT The main aspect of this leukemia detection is the web page. This web page allows the user to enter an image, and it employs the DL architecture to detect the presence of leukemia and displays the result to the user. This website is designed using HTML. HTML is an information structuring language used to define the website structure. Every site, from social networks to music services, you access in your online browser, employs HTML [12]. A check underneath the hood of any website would disclose a basic HTML code page with an HTML structure editor that gives a structure to the components of all the page sections, including their headers, footers, and primary content elements. The home page of the website will be as shown in Fig. (5).

Fig. (5). Home page of the website.

From image 5, it can be found that the header of the webpage is nothing but a title card of the page. Below the header lies an image of a leukemia-affected child, and it also contains some facts and statistics about leukemia. The next part of the page allows the user to upload an image by dragging it into the box or by browsing from the available files. Then comes a button that is to be pressed after the image is uploaded. 8. RESULTS AND DISCUSSION A website is being developed that takes an image of the blood as input, verifies whether the patient has leukemia, and presents the results. After the user uploads an image, it is processed using image preprocessing techniques such as scaling

10 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

and image augmentation. The preprocessed image is then evaluated using a DL model built with the VGG-16 module. This DL module enables transfer learning, which allows the user to alter only the start and final layers based on the requirements, rather than the complete module, to achieve the desired outcome [13]. Along with the model, a website is created using HTML and utilized to display output to the end-user. The website's output can be updated based on the DL model's outputs. After training, the model is tested numerous times to improve its accuracy. Table 2 shows the model's accuracy and loss at each epoch. Table 2. Comparison of accuracy and loss of the model by every epoch. Epoch

Accuracy

Loss

1

0.7794

0.5104

2

0.8144

0.4256

3

0.8138

0.4309

4

0.8253

0.4051

5

0.8254

0.4135

6

0.8335

0.3912

7

0.8369

0.3799

8

0.8375

0.3787

9

0.8456

0.3709

10

0.839

0.3859

11

0.8416

0.3819

12

0.8479

0.3637

13

0.8399

0.3811

14

0.8549

0.3491

15

0.849

0.3639

16

0.8499

0.3599

17

0.8465

0.3723

18

0.851

0.3601

19

0.8508

0.3583

20

0.8484

0.3626

The data shown in the above table is then converted into a graph so that it can be analyzed more clearly further. The graph is shown in Fig. (6).

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 11

Fig. (6). Graphical representation of the accuracy and loss.

The model's accuracy for the first epoch is 0.7794. This accuracy value eventually rises to a maximum of 0.8484 throughout the 20th epoch. The loss value in this scenario is 0.5104, which is the maximum value. The loss value is then reduced until it reaches a minimum of 0.3491 on the 14th epoch. However, the value of loss rises again, reaching 0.3626 at the end of the 20th epoch. When the model has reached its maximum accuracy, it is outfitted with a webpage created with the HTML. When the website is deployed using a server, it requests an image from the user to detect the presence of leukemia. After uploading the image, it will be preprocessed by scaling and image augmentation. The processed images are then transferred to the VGG-16 architecture to be examined for leukemia. If the image contains leukemia, the website presents the result as well as the probability score obtained by combining the accuracy and loss. Fig. (7) depicts the appearance of the website. This is the webpage once the image has been uploaded and the predict button has been pressed.

12 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

Fig. (7). Website when the image is uploaded.

Fig. (8) represents the website when the DL model detects Leukemia in the uploaded image.

Fig. (8). Website when leukemia is detected.

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 13

Fig. (9) represents the website when the DL model does not detect Leukemia in the uploaded image. In this case, the patient is free of Leukemia.

Fig. (9). Website when leukemia not is detected.

Once the user sees the results, the procedure of leukemia detection is complete, as demonstrated above. 9. ADVANTAGES OF THE STUDY Many studies have used the technology of DL in the prediction of leukemia and other such diseases. Some have used machine learning and artificial intelligence. The one fact that made this study stand out from the rest is that it also includes the development of a website. This website allows laymen to use this model easily, whereas the implementation of a DL model can be complicated for the user. CONCLUSION While leukemia is very uncommon, it has a significant impact on the patient's ability to think clearly and move freely. Upon its timely discovery, it can be tested for its ability to treat cancer. While waiting for blood test results or an appointment with the doctor, the user can relax and enjoy themselves on this HTML and DL-powered website. In contrast to other similar sites, this one has a

14 AI and IoT-based Intelligent Health Care & Sanitation

Ambika et al.

straightforward layout that's easy to follow and interpret. While this website is only a screening tool, a negative result could significantly reduce the patient's anxiety about the possibility of leukemia. The results from a cancer specialist may take some time, and this helps them feel more confident in the meanwhile. People in rural and tribal communities, who may not have been exposed to modern medical diagnostic procedures, might utilize this website to learn more. There will be less of a chance of leukemia being as common as breast cancer or another sort of cancer, thanks to this. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

Ching-Hon Pui, Mary V. Relling, Pharm D., and James R., "Downing. Acute Lymphoblastic Leukemia. Review article on Mechanisms of disease", N Engl J Med, vol. 350, pp. 1535-1548, 2004. [http://dx.doi.org/10.1056/NEJMra023001]

[2]

S.P. Hunger, and C.G. Mullighan, "Acute Lymphoblastic Leukemia in Children", N. Engl. J. Med., vol. 373, no. 16, pp. 1541-1552, 2015. [http://dx.doi.org/10.1056/NEJMra1400972] [PMID: 26465987]

[3]

Ching-Hon Pui, and Mary V. Relling, "Treatment of Acute Lymphoblastic Leukemia", Review article on Mechanisms of disease, 2006. [http://dx.doi.org/10.1200/JCO.2014.59.1636]

[4]

T. Terwilliger, and M. Abdul-Hay, "Acute lymphoblastic leukemia: a comprehensive review and 2017 update", Blood Cancer J., vol. 7, no. 6, p. e577, 2017. [http://dx.doi.org/10.1038/bcj.2017.53] [PMID: 28665419]

[5]

F. Xia, L.T. Yang, L. Wang, and A. Vinel, "Internet of Things", Int. J. Commun. Syst., vol. 25, no. 9, pp. 1101-1102, 2012. [http://dx.doi.org/10.1002/dac.2417]

[6]

Yann LeCun, Yoshua Bengio, and Geoffrey Hinton, "DL", Nature review article, 2015. [http://dx.doi.org/10.1038/nature14539]

[7]

T. Kaur, and T.K. Gandhi, "Automated Brain Image Classification Based on VGG-16 and Transfer learning", International Conference on Information Technology, 2019. [http://dx.doi.org/10.1109/ICIT48102.2019.00023]

[8]

A. K. Rangarajan, and R. Purushothaman, "Disease classification in eggplant using pre-trained VGG16 and MSVM", Sci. Rep., vol. 10, no. 1, p. 2322, 2020. [http://dx.doi.org/10.1038/s41598-020-59108-x] [PMID: 32047172]

[9]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images",

IoT Based Website

AI and IoT-based Intelligent Health Care & Sanitation 15

Med. Biol. Eng. Comput., vol. 58, no. 4, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6] [PMID: 31993885] [10]

M.J. Taylor, J. McWilliam, H. Forsyth, and S. Wade, "Methodologies and website development: a survey of practice", Inf. Softw. Technol., vol. 44, no. 6, pp. 381-391, 2002. [http://dx.doi.org/10.1016/S0950-5849(02)00024-1]

[11]

N. Bouhmala, A. Viken, and J.B. Lønnum, "Enhanced Genetic Algorithm with K-Means for the Clustering Problem", Int. J. Model. Optim, vol. 5, no. 2, pp. 150-154, 2015. [http://dx.doi.org/10.7763/IJMO.2015.V5.452]

[12]

J.S.G. Syed, M. Awais, and I. Faye, "Fabrice Meriaudeau “Mammogram classification using DL features", IEEE International Conference on Signal and Image Processing Applications, 2017. [http://dx.doi.org/10.1109/ICSIPA.2017.8120660]

[13]

L. Alzubaidi, J. Zhang, A.J. Humaidi, A. Al-Dujaili, Y. Duan, O. Al-Shamma, J. Santamaría, M.A. Fadhel, M. Al-Amidie, and L. Farhan, "Review of deep learning: concepts, CNN architectures, challenges, applications, future directions", J. Big Data, vol. 8, no. 1, p. 53, 2021. [http://dx.doi.org/10.1186/s40537-021-00444-8] [PMID: 33816053]

16

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 16-32

CHAPTER 2

AI and IoT-based Intelligent Management of Heart Rate Monitoring Systems Vedanarayanan Venugopal1,*, Sujata V. Mallapur2, T.N.R. Kumar3, V. Shanmugasundaram4, M. Lakshminarayana5 and Ajit Kumar6 Electronics and Communication Engineering (ECE), Sathyabama Institute of Science and Technology (Deemed to be University), Chennai, Tamil Nadu, India-604119 2 Sharnbasva University, Kalaburagi, Karnataka, India-585103 3 Ramaiah Institute of Technology, Bangalore, Karnataka, India-560054 4 Department of Electrical and Electronics Engineering, Sona College of Technology, Salem, Tamil Nadu, India-636005 5 Department of Medical Electronics Engineering, M S Ramaiah Institute Of Technology, MSRIT Post, M S Ramaiah Nagar, MSR Nagar, Bengaluru, Karnataka, India - 560054 6 School of Computer Science and Engineering, Soongsil University, Seoul, South Korea 1

Abstract: The Heart Rate Monitoring system is developed using Artificial Intelligence (ANN) and Internet of Things technology to detect the heartbeat and SpO2 of the patient to monitor the risk of heart attack and also make the person do regular checkups. Health monitoring is very important to us to make sure our health is in which condition. The proposed framework examines the sources of information gathered from sensors fit with the patients to discover the crisis circumstance. The IoT technology is employed; it helps anyone monitor the patient's health from anywhere. A deep learning algorithm such as ANN is utilized to identify whether the person’s health is normal or abnormal. In case an abnormality in health is noticed, an alert will be popped out. The proposed framework with artificial intelligence and IoT can reduce the death occurring due to heart rate, and other related issues can also be avoided.

Keywords: Accuracy, Artificial Neural Network Algorithm, Heart Rate, SpO2, Thing Speak. * Corresponding author Vedanarayanan Venugopal: Electronics and Communication Engineering (ECE), Sathyabama Institute of Science and Technology (Deemed to be University), Chennai, Tamil Nadu, India-604119; E-mail: [email protected]

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 17

1. INTRODUCTION Heart disease is presently the leading cause of death in China, claiming the lives of more than 1,500,000 people each year, according to official statistics. The likelihood of becoming bedridden and requiring medical treatment from family and friends is higher among patients with heart issues, according to the American Heart Association. Even though they may not feel well until late in the disease, it is too late for them to undo the damage that has already occurred. Furthermore, the vast majority of people die before obtaining any form of medical attention or treatment. The transformation of passive healthcare into a general strategy is therefore essential for enhancing heart disease healthcare performance and decreasing mortality rates. The physical conditions of their patients must be monitored by their physicians, and choices regarding when to deliver healthcare services must be made based on this monitoring. An effective real-time monitoring system is essential in delivering universal healthcare services. As a result, the world's population is growing at an alarming rate. When it comes to accommodating an ever-increasing number of people, larger cities are under tremendous strain. However, even though medical resources and facilities in urban areas are continually being expanded, the essential level of care has not yet been achieved. As a consequence of the high demand for urban healthcare management, technical developments have occurred that have resulted in effective remedies to the expanding difficulties facing the city's population. Due to the increase in the number of people who want medical attention, remote healthcare is becoming increasingly frequent. Wearable sensors have gained in popularity in recent years, and gadgets like these are now more readily available on the market at a cheaper cost for self-care and activity tracking than they were previously. These cutting-edge devices have been investigated for medical applications such as data logging and management, as well as real-time health monitoring and monitoring of patients. Healthcare services will be able to reach new heights as a result of the development of the Internet of Things [1]. The use of low-cost, trustworthy, and handy devices by patients, as well as smooth networking between patients, medical equipment, and doctors is ensured by this method of delivery. The sensors are in charge of recording a continuous stream of signals [2]. It is the researchers who receive the signals captured by the sensors, which have been coupled with physiological data and supplied over a wireless network. It is necessary to link this new information with existing health data to make sense of it [3]. Using current data records and decision support systems, it is feasible for the doctor to create a more accurate forecast and recommend early treatment. Modern technology can predict health concerns even if a doctor is not readily available due to the analysis performed on the data. Other

18 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

than being able to make predictions, robots may be able to design pharmaceuticals by thoroughly researching medical databases. Due to the improvements in progressive technology, it will be possible to dramatically lower healthcare expenses while simultaneously enhancing the accuracy of illness projections in the future. In this article, we discuss a service model in terms of technology and economics for the benefit of patients, as well as the challenges that remain in integrating the Internet of Things into the real-world medical environment. 2. LITERATURE SURVEY The health of an individual is vital to their overall well-being. In the context of health, the absence of illness as well as a better physical and mental state, are seen as blessings. Society as a whole is placing an increasing emphasis on health and healthcare, and new technologies are being implemented to support this. A result of the economic effect of COVID-19 in recent years has been the implementation of smart healthcare systems around the world. A smart healthcare system keeps an eye on patients from a distance to avoid disease transmission and provides highquality treatment at an affordable price for patients and their families. In this context, healthcare solutions incorporating IoT and machine learning are considered the most optimal [4]. These solutions have become more successful as a result of the Internet of Things (IoT), artificial intelligence (AI) and machine learning technologies, among others. A breakthrough in microelectronics has permitted the development of compact and low-cost medical sensing devices, which have had a significant impact on the delivery of medical care [5]. As a result, in healthcare systems, these therapies are separated into two categories: symptomatic treatments and preventive treatments. Disease prevention and early identification, as well as discovering the most effective treatment choices for a wide range of chronic disorders, are now highly valued by the general public. Consequently, national healthcare monitoring systems are becoming an inescapable development in the contemporary era. In recent years, there has been a significant increase in interest in Internet of Things (IoT) devices and machine learning algorithms for remotely monitoring patients to provide more accurate diagnoses in telemedicine. Real-time healthcare systems that regulate people's vital signs are in high demand, and researchers are working hard to build systems that are energy-efficient, cost-effective and scalable [6]. While conventional wireless communication technology may be employed in healthcare, it does so at the expense of higher expenses and radiation exposure, among other things. Health monitoring systems that do not rely on radiation provide more flexible communication routes and may be used to monitor a wide range of circumstances and conditions. Using a wireless connection to the local router, for example, can be used to communicate and send data within the home environment. When people need to utilize portable healthcare equipment while they are out in the

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 19

field, a wireless connection is needed to guarantee that data transfer and transmission are successful. The algorithms detect whether the people are inside or outside the building by employing machine learning methods. The Internet of Things, or IoT, is a versatile technology that may be applied to a variety of sectors, including the medical. Along with Big Data and Artificial Intelligence, the Internet of Things is said to be one of the most effective tools [7]. A multimodal system based on the internet of things (IoT) was developed as a result of this, and it was used to collect data from body area sensors (BAS) to provide first-time information and early warnings about the possibility of cardiac arrest [8]. The major purpose of this proposal is to design and construct a lowpower Internet of Things system that can collect heart rate and body temperature data from smartphones while not interfering with daily activities and activities. According to the research, sensor data may be examined using signal processing and machine learning methodologies to identify and possibly forecast abrupt cardiac arrest with a high degree of accuracy. In the system that has been created, there are no unique performance measures for each particular patient. It was recently announced that a healthcare monitoring kit would be added to the Internet of Things ecosystem [9]. In addition to heart rate, electrocardiogram (ECG), body temperature, and breathing rate, the devised system also maintained track of other important human health indices. A pulse sensor, a temperature sensor, a blood pressure sensor, an electrocardiogram sensor, and a Raspberry Pi are just a few of the primary hardware components that have been employed in this experiment thus far. It was necessary to gather sensor data and send it to a Raspberry Pi for processing, following which it was transmitted back to the Internet of Things network for dissemination. The absence of data visualization interfaces is the most serious issue in the system. An Internet of Things (IoT)-based cardiac disease monitoring system has been described for pervasive healthcare services [10]. This system offers four different data transmission modes to help balance the demands on healthcare resources, such as communication and computer resources, with the monitoring of patient's physiological signs, such as BP, ECG, and oxygen. They also created a prototype to illustrate the functionality of the system. For patients who are experiencing an urgent situation that requires immediate attention from a doctor, the IoT and machine learning concepts will be used to provide an immediate solution by monitoring their health, determining if they have a specific disease, and recommending an appropriate treatment plan [11]. The Internet of Things develops safety monitoring device that responds to changing conditions [12]. Depending on your preferences, you may divide the framework's set-up into three different layers: control, devices, and transport. To monitor body temperature in the control section, a DS18B20 sensor was applied. To measure heart rate, a pulse

20 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

sensor was utilized. The Wi-Fi module and Ethernet shield were used to transfer data from the Arduino to the cloud storage system at the transport layer of the network. Finally, the framework layer got an opportunity to obtain information from the server! As a result, many sensing devices cannot be correctly processed due to the use of Arduino Uno. An online smart patient monitoring system that maintains track of vital indicators, such as heart rate, a saturation of oxygen in the blood, and body temperature, is the objective of this research study. ThingSpeak and a server will store and assess the data based on the specifications provided above. This paper uses the Python programming language to develop and discuss an algorithm for tracking the health of numerous people suffering from various ailments. This method is also useful for data categorization and the organization of pharmacological need information. If you wish to see the output on your phone, you may either use the ThingSpeak app viewer or a mobile app to do this. This approach produces excellent results, which are also given in this paper. The book chapter is taken in the following ways. Sections 1 and 2 discuss the importance of health parameter measurements and the research work so far done in this domain. Section 3 details the proposed block diagram and flow chart. Section 4 contains a brief explanation of ANN. Section 5 deals with the implementation of the proposed work. Section 6 shows the result obtained through ThingSpeak and suggestions given by the ANN. Finally, the conclusion of the research work is given in Section 7. 3. PROPOSED SYSTEM The proposed system integrates the two advanced technology, artificial intelligence and IoT, to improve healthcare in society. Personal health parameters are the essential ones in this research. The sensor is utilized for this purpose. Then the data has to be sent to the internet, which helps others view the person's health. Doctors can also use this technology to view their patient's health remotely. Deep learning is essential for predicting whether a person's health is in normal condition or not. The proposed system has five important phases, from data collection to alert messages. Each phase and its description are detailed with the help of Fig. (1). ● ●

●

Data Acquisition: Getting health parameters from a sensor attached to the person Data Transmission: The sensor data is passed through the cloud with the help of the ESP826 module. This acts as the bridge between hardware and the cloud. Data Visualization: The collected sensor data is plotted in the graph. The values are updated once in 120 seconds.

Monitoring Systems ●

●

AI and IoT-based Intelligent Health Care & Sanitation 21

Data Analyzation: The collected sensor data is given to the ANN model for predicting whether there is a chance of disease occurrence or not. Alert: Based on the ANN results, the message is given to the respective person.

Fig. (1). Work Flow of the proposed system.

The hardware components used in this work and their working are detailed in Table 1. Table 1. Hardware used in the proposed system. Sl. Hardware No 1

Working

MAX30100 This gadget is capable of measuring both blood oxygen levels and heart rate. The units of measurement for blood oxygen saturation (SpO2) are %, while the units of measurement for heart rate (BPM) are beats-per-minute (BPM). One of the device's LEDs, which is red, emits infrared light, which is visible to the human eye. When it comes to measuring pulse rate, infrared light is all that is necessary. The amount of oxygen in the blood may be determined by utilizing red and infrared light. More blood is pushed through the body with each beat, which increases the quantity of oxygenated blood available. Due to this, the amount of oxygenated blood in the body decreases as the heart beats more slowly. The time elapsed between a rise and a decrease in oxygenated blood may be used to measure the pulse rate of an individual. According to their findings, although red light absorbs more from deoxygenated blood, infrared light absorbs more from oxygenated blood when both processes occur in the same time frame. The MAX30100's major function is to read both light source absorption levels and save them in a buffer that can be accessed via the I2C communication protocol, which is the protocol used to communicate with the device.

2

Arduino

Using the Arduino development environment on a PC, the microcontroller board is linked to the computer through USB. Just write the code and upload it to the microcontroller using the Arduino IDE. In the following step, the microcontroller executes the code, interacting with sensors and a Wi-Fi module, among other components.

3

ESP8266

The module may work as a stand-alone system or as a slave device to a host MCU, minimizing the communication stack overhead on the primary application CPU. Using this board SPI, I2C, and UART interfaces can be used for establishing communication with various I/O peripherals, the ESP32 can communicate the data with the cloud.

22 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

Fig. (2). Work Flow of the proposed system.

The proposed prototype and the depicted flowchart in Fig. (2) are regarding heart health monitoring. The working flow of the developed prototype started with the

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 23

initialization of sensors which is annexed. For detection, the heart rate sensor and oxygen level sensor are adjoined along with the prototype model. Both the heart rate and oxygen level of the patient are getting measured by the two appended sensors. The appraised values from the sensors are sent directly to the cloud. In this, one of the deep learning algorithms called ANN is employed. After passing to the cloud, the measured values will be exploited as the input for chosen ANN model. After training and testing with these data, the model will furnish the output. This model will give the output of whether the patient is normal or abnormal. If the output of the ANN model is received as “Normal” – then an alert or indication message will be looming as “You are in a safe zone”. But if the output of the model has “Abnormal” as its state – eventually an alert or indication message will be looming as “Have a consult with your cardiologist”. The necessity of monitoring the heart rate is detailed below: 1. For those suffering from heart disease, for example, if there is a shift in heart rhythm that we are unable to feel in our body. 2. The individual may not be aware of the change in cardiac rhythm at first, but if the change gets severe, the person may experience discomfort and may have a cardiac arrest. 3. When a severe attack occurs, the victim will be in extreme anguish and will be unable to think coherently, which may result in their death. 4. ARTIFICIAL NEURAL NETWORK This technique, which mimics the human nervous system, allows students to learn from instances of physical events or decision-making processes by using representative data. “ANN” is a synonym for “Artificial Neural Network.” The ability of artificial neural networks to generate empirical connections among independent and dependent variables and extract sensitive information and complicated knowledge from representative data sets distinguishes them. It is not necessary to presume a mathematical description of the event under consideration in establishing correlations among variables. ANN models have several benefits over regression models, such as the ability to deal with noisy data. In ANNs, which are separated into two halves, one or more layers of hidden nodes connect the input and output nodes. The input layer nodes fire activate functions to convey information to the hidden layer. Based on the information provided, the hidden layer nodes decide whether to fire or remain dormant. Fig. (3) depicts the architecture of the ANN model. Nodes in the output layer send values to other hidden layers, which send values back to other hidden layers. This operation is

24 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

repeated until all nodes in the output layer have received values equal to or greater than a predefined threshold. The phrase “hidden layers” refers to how evidence in these layers is weighted using weighting algorithms. Before ANNs be considered successful, a large number of cases must be trained on them. It is impossible to employ ANN to anticipate uncommon or catastrophic events in the absence of sufficient data.

Fig. (3). Architecture of ANN.

When making decisions, the ANNs do not allow for the replacement of quantitative data with expert information by expert knowledge. An ANN takes into account uncertainty by calculating the chance that each output node will be discovered in the network and using that probability as input information. Before we can predict what will happen in the future, we must first make some educated guesses about what will happen in the future. As the nodes in the hidden layer have no true physical significance, the output cannot be traced directly to the process that produced it. Artificial neural networks must have an inference pathway, where the data to be processed (inputs) and how it will be classified in

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 25

the future are both established in advance of the network's use. ANNs have been utilized to construct prediction models for heart disease prediction. The parameter and their values used to construct the ANN model are given in Table 2. Table 2. Hardware used in the proposed system. S. No

Parameter

Values

1

Hidden Layer

5

2

Number of neurons in the output layer

1

3

Activation function-Hidden layer

Relu

4

Activation function-Output layer

Sigmoid

5

Batch

32

6

Epoch

15

7

Metrics

Accuracy

8

Optimizer

Adam

9

Error

Categorical entropy

The flow of ANN is given in the below procedure: 1. Initialize the weight of all the layers 2. The transfer function is calculated using the below equation:

(1)

Where, w →Weight b → Bias 3. Then passed through the activation function (2)

26 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

4. Compute the error

(3)

(4) 5. Backpropagate: Error gradients (5)

(6)

(7) 6. Update weights (8) 7. Go to step 2 and the same process is repeated until the error is tolerable 5. IMPLEMENTATION The research work is implemented in the following way. It helps the person to analyse their health. 1. The MAX30100 sensor, which is used to measure heart rate and oxygen level, is connected to the analog pin of the micro-controller.

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 27

2. The ESP8266 is connected to the serial port of the micro-controller for enabling the IoT in work. 3. The connection between the cloud and the hardware is made using the authentication code present in the API section of ThingSpeak. 4. The data is sent to and viewed at the ThingSpeak cloud through ESP8266. The health parameters can’t only be viewed by the respective persons and also by their friends, relatives, and doctors. 5. The health parameters are updated in the ThingSpeak every 120 seconds. 6. The ANN is constructed and trained until an error value is tolerable. 7. The ANN model is validated using the metric called accuracy. 8. The sensor data like heart rate and oxygen level are taken and given to the ANN model. It makes the prediction. 9. Based on the prediction, the alert message is given to the user. 6. RESULT AND DISCUSSION The work is carried out in the way included in Section 5. The picked deep learning algorithm called ANN has been employed in the data present in the cloud. The health data taken from the human body is sent to the cloud. The health parameters are displayed in ThingSpeak in the graphical format. The collected data in the cloud is dispatched as input to the ANN model. The ANN model is used to predict the health condition of the person. The accuracy obtained from the training phase is nearly equal to 99%. It proves that the prediction made by the model is trustable. During the training of the model, the accuracy of the ANN model is achieved, as shown in the above Fig. (4). Training of the model is done by keeping the x-axis as epochs against the y- axis as accuracy in percentage. During the first epoch, the percentage accuracy is obtained as 38%. While in the case of the 5th, 10th, and 15th epochs – the percentage accuracy is obtained as 69%, 82%, and 98.89%, respectively. From this trained model, it is clear that as epochs get increasing, the percentage accuracy is also getting increased, and the accuracy is also getting increased. Thus, this ANN model is chosen and employed due to its higher accuracy produced in the result of the ANN model.

28 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

Fig. (4). Accuracy of ANN model in training.

The data measured using the sensors are stored in the cloud with the help of ThingSpeak. In ThingSpeak, the update in the output or the changes in the graph will be only for two minutes once. In Fig. (5) , the SpO2 level of the patient is monitored and shown in the ThingSpeak. Here, the date is under the x-axis, and the SpO2 level is under the y-axis. The level of SpO2 is approximately getting maintained constantly. In Fig. (6), the heart rate of the patient is monitored and shown in the ThingSpeak. Here, the date is under the x-axis, and the heart rate (BPM) level is under the y-axis. The level of monitored heart rate is approximately getting varied. This measured data will be sent to the ANN model.

Monitoring Systems

Fig. (5). Monitoring SpO2 in ThingSpeak.

Fig. (6). Monitoring Heart Rate in ThingSpeak.

AI and IoT-based Intelligent Health Care & Sanitation 29

30 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

From the ANN model, these data will be analyzed properly, and suggestions will be given according to the output. If the patient’s data is very well good, the suggestion will be “No worry” and their condition will be depicted as “Normal,” as shown in Fig. (7). In case the patient’s data is not advisable, the suggestion will be as “Abnormal,” and their condition will be depicted as “Contact the Doctor immediately,” as shown in Fig. (8). These suggestions and showing of the patient’s condition will be helpful for them to take immediate measures accordingly.

Fig. (7). ANN suggestion for the person with a good heart condition.

Fig. (8). ANN suggestion for the person with an abnormal heart condition.

Monitoring Systems

AI and IoT-based Intelligent Health Care & Sanitation 31

7. STATE OF ART There are plenty of technologies that are used to monitor heartbeat rates. Though, the accuracy of those technologies is not stable and results in the ineffective performance of the tool. Whereas, the accuracy of the model sound in this study is observed to be consistent at all points. Also, the usage of the MAX30100 sensor measures the oxygen level along with the heartbeat rate. This research helps to monitor the patient's health condition from anywhere in the world. CONCLUSION To keep track and monitor the patient's physiological characteristics, diagnostic devices for patients, the elderly, and the disabled are required. Wireless technology has been used to provide cutting-edge applications that will open up new doors in health care. With the advent of wireless technology, time-consuming tasks have been eliminated, and patients are no longer confined to their beds and instruments. The goal of this study is to develop and implement a system for monitoring patient heart health and displaying that data in realtime on a website that is accessible over the network. By using this approach, we were able to determine the most accurate way to assess a person's health. The prediction of disease at an early stage is very much important. The deep learning method gives the best solution for this problem. The accuracy of prediction is high in deep learning when compared to other available techniques. The ANN model is constructed and trained with an accuracy level of 98.89%. After that, the model is deployed in real-time measurement of a person’s health. For getting the person's health parameters, the MAX30100 sensor is used to measure both heart rate and oxygen level. The health data is transferred to the cloud and then given as input for the ANN model. The ANN model predicts the person's health and gives suggestions to the person. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None.

32 AI and IoT-based Intelligent Health Care & Sanitation

Venugopal et al.

REFERENCES [1]

Y. Yin, Y. Zeng, X. Chen, and Y. Fan, "The internet of things in healthcare: An overview", J. Ind. Inf. Integr., vol. 1, pp. 3-13, 2016. [http://dx.doi.org/10.1016/j.jii.2016.03.004]

[2]

T.S.F. Al-Turjman, S.P. Yadav, M. Kumar, V. Yadav "Transforming Management with AI, Big-Data, and IoT" Springer Nature, 2021, ISBN: 303086748X. [http://dx.doi.org/10.1007/978-3-030-86749-2]

[3]

H.T. Sullivan, and S. Sahasrabudhe, "Envisioning inclusive futures: Technology-based assistive sensory and action substitution", Futures, vol. 87, pp. 140-148, 2017. [http://dx.doi.org/10.1016/j.futures.2016.06.006]

[4]

X. Liu, P. He, W. Chen, and J. Gao, "Multi-task deep neural networks for natural language understanding", Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019. [http://dx.doi.org/10.18653/v1/P19-1441]

[5]

J. Sun, F. Khan, J. Li, M.D. Alshehri, R. Alturki, and M. Wedyan, "Mutual authentication scheme for ensuring a secure device-to-server communication on the internet of medical things", IEEE Internet Things J., vol. 8, no. 21, pp. 15663-15671, 2021. [http://dx.doi.org/10.1109/JIOT.2021.3078702]

[6]

Q. Zhang, L.T. Yang, Z. Chen, and P. Li, "An improved deep computation model based on canonical polyadic decomposition", IEEE Trans. Syst. Man Cybern. Syst., vol. 48, no. 10, pp. 1657-1666, 2018. [http://dx.doi.org/10.1109/TSMC.2017.2701797]

[7]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images", Med. Biol. Eng. Comput., vol. 58, no. 4, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6] [PMID: 31993885]

[8]

A.K.M.J.A. Majumder, Y.A. ElSaadany, R. Young, and D.R. Ucci, "An Energy Efficient Wearable Smart IoT System to Predict Cardiac Arrest", Adv. Hum. Comput. Interact., vol. 2019, pp. 1-21, 2019. [http://dx.doi.org/10.1155/2019/1507465]

[9]

A.D. Acharya, and S.N. Patil, "IoT based health care monitoring kit", In: 2020 Fourth international conference on computing methodologies and communication (ICCMC) IEEE, 2020, pp. 363-8. [http://dx.doi.org/10.1109/ICCMC48092.2020.ICCMC-00068]

[10]

C. Li, X. Hu, and L. Zhang, "The IoT-based heart disease monitoring system for pervasive healthcare service", Procedia Comput. Sci., vol. 112, pp. 2328-2334, 2017. [http://dx.doi.org/10.1016/j.procs.2017.08.265]

[11]

K. Srivardhan Reddy, and R. Sidaarth, "Sai Aneesh Reddy, Dr. Rajashree Shettar. IoT based Health Monitoring System using Machine Learning", International Journal of Advance Research and Innovative Ideas in Education., vol. 5, 2019. [http://dx.doi.org/16.0415/IJARIIE-10244]

[12]

S.P. Kumar, V.R.R. Samson, U.B. Sai, P.L.S.D.M. Rao, and K.K. Eswar, "The smart health monitoring system of the patient through IoT", In: 2017 International Conference on I-SMAC (IoT in social, mobile, analytics and cloud) (I-SMAC). IEEE. 2017, 551–6. [http://dx.doi.org/10.1007/s42979-020-00195]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 33-49

33

CHAPTER 3

Deep Learning Applications for IoT in Healthcare Using Effects of Mobile Computing Koteswara Rao Vaddempudi1,*, K.R. Shobha2, Ahmed Mateen Buttar3, Sonu Kumar4, C.R. Aditya5 and Ajit Kumar6 ECE Department, Prakasam Engineering College, Kandukur, Prakasam, Andhra Pradesh, India Department of Electronics and Telecommunication Engineering, Ramaiah Institute of Technology, MSR Nagar, MSRIT Post, Bangalore, Karnataka, India-560054 3 Department of Computer Science, University of Agriculture Faisalabad, Faisalabad, Punjab, Pakistan-38000 4 Department of Electronics and Communication Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Guntur, Andhra Pradesh, India-522502 5 Department of Computer Science & Engineering, Vidyavardhaka College of Engineering, Mysuru, Karnataka, India-570002 6 School of Computer Science and Engineering, Soongsil University, Seoul, South Korea 1 2

Abstract: Diabetes is a chronic ailment characterized by abnormal blood glucose levels. Diabetes is caused by insufficient insulin synthesis or by cells' insensitivity to insulin activity. Glucose is essential to health since it is the primary source of energy for the cells that make up a person's muscles and tissues. On the condition that if a person has diabetes, his or her body either does not create enough insulin or cannot utilize the insulin that is produced. When there isn't enough insulin or cells stop responding to insulin, many dextroses accumulate in the person's vascular framework. As time passes, this could lead to diseases such as kidney disease, vision loss, and coronary disease. Although there is no cure for diabetes, losing weight, eating nutritious foods, being active, and closely monitoring the diabetes level can all assist. In this research, we used Artificial Neural Network to create a Deep Learning (DL) model for predicting Diabetes. Then it was validated using an accuracy of 92%. In addition, with the help of the MIT website, a mobile application was constructed. This project will now assist in predicting the effects of diabetes and deliver personalized warnings. Early detection of pre-diabetes can be extremely beneficial to patients since studies have shown that symptoms of early diabetic difficulties frequently exist at the time of diagnosis.

Keywords: Accuracy, ANN Algorithm, Deep Learning, Diabetes, Portable Mobile Application. Corresponding author Koteswara Rao Vaddempudi: ECE Department, Prakasam Engineering College, Kandukur, Prakasam, Andhra Pradesh, India; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

34 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

1. INTRODUCTION Diabetes is a group of hormonal disorders characterized by persistent high blood sugar levels caused by an inadequacy in insulin release, insulin activity, or both. The importance of insulin as an anabolic molecule causes hormonal dysfunction in carbohydrates, lipids, and proteins (hormones). The severity of the presentation is determined by the length and type of diabetes [1]. DL is a mathematical progression built on a neural network with multiple layers. During the 1940s, McCulloch and Pitts proposed the concept of artificial brain organization. The core idea of neural organizing for data processing was influenced by biological neural frameworks [2]. In 2014, 8.5% of adults over the age of 18 had type 2 diabetes. Diabetes was the cause of 1.5 million deaths in 2019. Diabetes-related premature death increased by 5% between 2000 and 2016. Currently, key medical care administrations in numerous countries are unable to analyze and manage diabetes. Increasing access to essential medications is an essential aspect of preventing and managing diabetes and other noncommunicable diseases. DL can be used to forecast the status of diabetes as technology advances in the medical care field. DL's Artificial Neural Network technique is critical in constructing a precise forecast and providing the correct assessment [3]. Smartphone technology supports patients in the developed administration of several chronic infections, such as psychological wellness concerns, diabetes, obesity, cancer, and chronic obstructive pulmonary disease [4]. As a result of the results obtained from the MIT site, a mobile application has been created. The dataset comes from Kaggle and has a variety of attributes as its features. Initially, the dataset is pre-processed (that is, cleaned and prepared for the DL model to perceive). The ANN algorithm is then used to create a model utilizing this new information. The accuracy of this algorithm is determined and analyzed to obtain the best-prepared model for the forecast. Following a real assessment, an advanced mobile phone application has been developed with a pre-planned model for new clients to receive personalized warnings on self-evaluation of Diabetes and provides lifestyle modifications to eliminate Diabetes. 2. LITERATURE SURVEY This section intends to survey the studies conducted on DL applications for IoT in the clinical sector. A study [5] investigated the use of medical images through image fusion technology. The researchers [6] present a rigorous outline of DL study to monitor the patient's well-being using gadgets. The problem with this review is that the size of the dataset limits the depth of the DL model. The primary goal of this paper is to evaluate and summarize the emerging DL research work [7] on patient well-being monitoring. This paper also looks at various recent

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 35

DL fads in patient monitoring. The study [8] looked at various DL processes in the portable and individual pursuit that are based on sensors and serviceable in a way that extricates the features and offers reliable data. Straightforwardly or by implication, the well-known models claim to address genuine issues by combining mind-boggling Internet of Things (IoT) factors. The primary goal of this review article is to quickly summarize the current state of the Internet of Things frameworks in many sectors. The study [9] focused on the hypothesis of the Internet of Medical Things (IoMT). In emergency clinics, exchanging items stimulates control of the patient's state of being by gathering and mixing information on all essential indications. IoMT is divided into four stages: collection, storing, processing, and presenting. The systematic review is used to evaluate DL for cloud applications in medical sector frameworks in this research. By surveying the relevant works, significant DL applications for the cloud in the medical sector and clinical care are introduced. The article [10] discusses information combinations for IoT, focusing on computational and numerical procedures such as strategies in probability, and AI, and IoT complexions such as spread, irregularity, heterogeneity, and conditions where objects are traced. Sharing and cooperating on information and various assets would be critical to empowering practical universal conditions such as cities and societies. The journal [11] investigated several provisions of IoT-based improvements in the medical services sector and proposed multiple clinical organization structures and stages that make the IoT establishment accessible. This viewpoint paper aims to provide an overview of current IoT innovation in medical care, outline how IoT gadgets are further improving the well-being administration conveyance, and outline how IoT innovation will influence and disrupt global medical care in the coming decades. The journal article [12] offered a precise writing audit depicting comprehensive writing, including the supported living solutions. The research is divided into four major sections: self-checking, wearable framework, clinical administration framework, surrounding upheld daily bread in the elderly population, and profound learning-dependent framework to analyze heart infection. This precise writing study gives precise writing on surrounding assisted living arrangements and aids in understanding how encompassing assisted living aides propels patients with cardiovascular ailments to self-administration to decrease related grimness and mortalities. Naive Bayesian and SVM classifier approaches were utilized for diabetic categorization [13]. The PIMA Diabetes dataset was used in this study. To improve accuracy, the researchers used a feature extraction-based technique as well as k-fold cross-validation. In the experiments, the SVM outperformed the naïve Bayes model. Unfortunately, along with achieved accuracy, a cutting-edge evaluation is missing. Conduct a comparison of diabetic classification strategies

36 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

[14]. They used both PIMA data and a local diabetes database. Researchers used AdaBoost, K-NN, and the radial basis function to classify individuals as diabetic or not. They also used LDA and PCA for feature extraction, which was demonstrated to be successful with categorization techniques for improving accuracy and removing undesirable traits. Researchers used a machine learning technique to categorize and forecast diabetes. They employed four algorithms for diabetic classification: NB, DT, AdaBoost, and RF [15]. In addition to the 20 studies, the researchers used three additional division strategies to improve the findings. The researchers examined data from the National Health and Nutrition Examination Survey on diabetic and non-diabetic persons in the United States and discovered that the recommended approach produced favorable data. A comparison study of multiple machine learning techniques, notably Nave Bayes, Decision Tree, and MLP, was undertaken using the PIMA dataset for diabetes categorization [16]. When compared to other classifications, the authors discovered MLP to be superior. Fine-tuning and effective feature extraction, according to experts, may boost MLP efficiency. Introduced a comprehensive investigation of the DL-based examination works for diabetes prediction [17]. This study presents a detailed summary that diabetes is often preventable and can be avoided by making lifestyle modifications. These developments can also reduce the likelihood of developing cardiovascular disease and malignant growth. As a result, there is a pressing need for anticipatory equipment that can help specialists with early diagnosis of the infection and then prescribe the lifestyle changes required to limit the spread of the destructive illness. Deals with the advancement of patient well-being data to a huge number of data assortments accessible on the database intended to utilize the clinical benefits of the medical sector and 50 billion (IoT) gadgets for utilizing the course of trained professional and major-focused; the clinical benefit sector has fostered an innovation-centered blend of human advancement and clinical consideration [18]. The fog environment is designed to handle all of the problems caused by dispersed virtual reality. It aims to cope with distributed computing to keep the devices close to the server farm and distributed cache. Diabetes mellitus is associated with an increase in cardiovascular-related health problems. Diabetes treatment is dependent on knowledge of the disease's pathophysiology [19]. This investigation seeks to apply the ANN algorithm of DL technology to massive data to analyze diabetes occurrence and provide a personalized alarm message. 3. MATERIALS AND METHODS The input data is obtained from Kaggle and preprocessed to remove duplicate data and acquire the completed dataset, which is then fed into the ANN model for

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 37

training. The data is divided into a 7:3 ratio. The trained model is assessed using parameters like Accuracy and Loss. The remaining data is submitted to the model at random for testing, yielding an accuracy of 92%. Finally, the model is deployed as a Mobile Application using the MIT website. Fig. (1) depicts the flowchart for predicting Diabetes using DL.

Fig. (1). Flowchart for Predicting Diabetes and implanting it through a mobile application.

4. DATASET DESCRIPTION The dataset was obtained from Kaggle and contained documentation of around 768 people's diabetes status. The dataset includes eight features and their labeled outcomes.

38 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

Pregnancies- This attribute is numerical data, and it is the number of times the patient got pregnant. Glucose- It is numerical data about the patient's blood glucose level. Blood pressure- It is numerical data about the patient's diastolic blood pressure. Skin Thickness- It is numerical data about the patient's thickness of the triceps skin fold. Insulin- It is numerical data about the patient's insulin level. BMI- It is numerical data about the patient's weight in kilogram to the square of height in meters square in ratio. Diabetes pedigree function- It is numerical data on the likelihood of diabetes based on family history. Age- This attribute is numerical data and it is the age of the patient. Outcome- It is the numerical data on the occurrence of diabetes. This is the decision-making attribute. Splitting data is the process of dividing a dataset into learning and validation stages to prepare and track down model parameters and evaluate the model prediction. When using the available data to produce predictions, the train-test split methodology is used to estimate the execution of the ANN algorithm. It is a quick and straightforward system to implement, with the results allowing for an analysis of the presentation of the ANN algorithm for predictive modeling problems. Even though it is simple to use and interpret, there are situations when the procedure methodology should not be used, such as when the user has a limited number of available data and when extra preparation is required, such as when it is used for arranging, and the dataset is not modified. 5. ARTIFICIAL NEURAL NETWORK The DL algorithm utilized to foresee the event of diabetes is Artificial Neural Network (ANN) algorithm. This network is a naturally roused organization of unnatural brain cells designed to perform explicit duty. These natural techniques for processing are known as the significant advancement in the Computation Industry. Artificial Neural Networks are an unprecedented sort of DL strategy that is planned by the human brain. The ANN can gain knowledge from past data and

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 39

give yields as assumptions or arrangements. ANNs are real nonlinear models which show a confounded association between the data sources and yield to track down other patterns. Numerous tasks, such as picture acknowledgment, speech acknowledgment, machine interpretation along with clinical diagnosis, utilize artificial neural organization. A critical advantage of ANN is the way that it gains information from the model instructive assortments. Random function approximation is the most general utilization of ANN. This is additionally equipped for taking representative information instead of the whole data set to furnish the required result. With ANN, the programmer can improve existing information examination strategies inferable from their high-level predictive abilities. In a neural organization, there are three fundamental layers●

●

●

Input Layer- This layer is the primary layer of an ANN that obtains the information data as different texts, numbers, audio, picture, and so on. Hidden Layer- This layer is the middle layer of ANN. This layer performs different kinds of numerical calculations on the information data set and perceives required patterns. Output Layer- In the output layer, we acquire the outcome that we obtain through diligent calculations performed by the center layer. The yield of ANN is mostly dependent on batch size, weights, learning rate, and biases. Every node in ANN has an approximate weight.

The ANN model used in this research for predicting diabetes is illustrated in Fig. (2).

Fig. (2). Architecture of ANN.

40 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

6. IMPLEMENTATION 1. Input Data- Kaggle data for approximately 768 participants were collected. Age, skin thickness, glucose, BMI, pregnancy, insulin, diabetes pedigree function, and blood pressure are the eight features. 2. Data Preprocessing- Data preparation is used to assess the dataset's errors and missing data. This is followed by a heatmap, as depicted in Fig. (3). It summarizes the data as a contribution to ongoing research and as a diagnostic for advanced analyses. The section on skin thickness and blood pressure has been removed because it has a weaker relationship with the outcome. Finally, the normalized data is obtained to proceed with additional processing.

Fig. (3). Heatmap for correlation identification.

3. Data Splitting- The updated dataset is parted into features and labels and further into the training and testing dataset in the proportion of 7:3 utilizing the Scikit learn library.

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 41

4. Algorithm- ANN algorithm has been utilized to train the set and validate the dataset to predict the most appropriate model. In the ANN algorithm, we have provided 6 neurons as inputs. This ANN algorithm has 4 Hidden Layer neurons such that the 1st hidden layer is provided with 10 neurons and Relu activation function,2nd layer is provided with 50 neurons and Relu activation function,3rd layer is provided with 10+Relu activation function, and the 4th layer with 5 neurons and Relu activation function. The output layer contains 1 neuron with a sigmoid activation function. Then the batch is fixed as 12 and epochs as 2000. 5. Confusion matrix - The confusion matrix is utilized to estimate the value of Accuracy, True Negative rate, False Positive Rate, False Negative Rate, True Positive rate, and Precision. The best model for diabetes prediction is acquired using the above metrics. 6. Mobile Application Development- MIT App Inventor is a web-based advancement stage that anybody can use to tackle real issues. It permits interlopers to PC programming for generating application software (apps) for the working frameworks (OS): Android, and iOS. It is open-source and free programming. It makes use of a visual interface (GUI) like the programming languages, which licenses customers to migrate perceptible things to make software that could deal with Android gadgets. App Inventor additionally upholds the utilization of cloud data through an experimental Firebase. The anticipated diabetes model is fed into the MIT app inventor utilizing code and the developers can utilize it to evaluate and alter the etiquette of their software progressively. 7. RESULT AND DISCUSSION The data is sent through the ANN network. This dataset has been subjected to 2000 epochs with a batch size of 12. The accuracy at the 0th epoch is 55%, which steadily climbs to 75% at the 250th epoch, 80% at the 500th epoch, which still rises to 85% at the 1750th epoch, and more than 90% at the 2000th epoch. Similarly, the loss at the 0th epoch is 14%, which has decreased to 5% at the 250th epoch and 4% at the 500th epoch, which continues to decline from 3% to 1% in the range of 750th to 2000th epochs, as shown in Figs. (4 and 5). The Kaggle dataset contains 6 attributes and approximately 768 diabetes patient records. The data is divided into two classes, the learning set and the evaluation set, in a 7:3 ratio. The ANN algorithm was used to develop an acceptable model for dissecting the diabetes incident. To deconstruct the model using performance measures, confusion matrix parameters such as True Negative, False Positive, True Positive, and False Negative have been used.

42 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

True Positive (TP) - The patient is influenced by diabetes continuously, and the model likewise predicts something very similar False Positive (FP) - Though, the patient isn't influenced by diabetes continuously, the model predicts that the patient is influenced by diabetes. True Negative (TN) – The patient isn't influenced by diabetes continuously and the model predicts something very similar. False Negative (FN) – Though, the patient is influenced by diabetes continuously, the model predicts that the patient is not influenced by diabetes. The confusion matrix of the designed ANN model is given in Fig. (6). This algorithm provides 98 True Positives, 45 True negatives, 7 False positives, and 4 False negative values.

Fig. (4). Accuracy of ANN model.

Mobile Computing

Fig. (5). Loss of ANN model.

Fig. (6). Confusion matrix of ANN model.

AI and IoT-based Intelligent Health Care & Sanitation 43

44 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

Table 1 gives the formula for performance metrics employed in the research work. Table 1. Formula of performance metrics. Performance Metrics

Formula

True Positive rate

TP/Actual positive=TP/(TP+FN)

True Negative rate

TN/Actual negative=TN/(TN+FP)

False Positive rate

FP/Actual negative=FP/(TN+FP)

False Negative rate

FN/Actual positive=FN/(TP+FN)

Precision

TP/(TP+FP)

Accuracy

(TP+TN)/(TP+TN+FP+FN)

The metrics used to evaluate the model are given below: Precision- It is the small portion of suitable illustration among the recovered illustrations. Accuracy- It is utilized to choose the best model. It is the level of right forecast in the test dataset. True Positive rate (TPR) -TPR is the feasibility that the actual positive will test positive. True Negative Rate (TNR) -The true negative rate is the probability that the real negative will test negative. False Positive Rate (FPR) - The false-positive rate is defined as the likelihood of falsely dismissing the null hypothesis. False Negative Rate (FNR) -The test outcome indicates that the person does not have a particular sickness or condition when the individual has the illness or condition. Table 2 compares the ANN algorithm's performance analysis. The resultant TNR is 86.53, TPR is 96.07, FNR is 3.92, and FPR is 13.46. This work's precision and accuracy are 93.33% and 92.85%, respectively. Fig. (7) depicts the performance analysis of the ANN algorithm using the Bar graph, which provides a better understanding of the DL algorithm in predicting the onset of diabetes with an accuracy of 92%.

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 45

Table 2. Performance of ANN algorithm. Performance Metrics

Performance Metrics in %

Accuracy

92.85714

TNR

86.53846

TPR

96.07843

FNR

3.921569

FPR

13.46154

Precision.

93.33333

Fig. (7). Bar graph depicting the ANN performance analysis for diabetes prediction.

Fig. (8) depicts the home page of the built mobile application. The user's health conditions, such as pregnancy, blood glucose level, age, Insulin level, Pedigree function, and Body Mass Index, are entered into the program and processed to provide a warning message about diabetic symptoms. Figs. (9a and 9b). show the final results supplied by the developed portable application for predicting the incidence of diabetes in two distinct patients. In Fig. (9a) the user has provided his/her personal information such as pregnancy as 6, blood glucose level as 148, insulin as 0, body mass index as 33.6, pedigree

46 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

function as 0.627, and age as 50. Fig. (9a) depicts the alert message “Diabetes Symptoms-Please Consult the Doctor Immediately” provided by the mobile application. In Fig. (9b), the user has input his/her personal information for pregnancy as 1, blood glucose level as 85, insulin as 0, body mass index as 26.6, pedigree function as 0.351, and age as 31. Fig. (9b) shows the alarm message “No Diabetes-You are safe” provided by the mobile application.

Fig. (8). Home page of diabetes predicting mobile application.

8. NOVELTY OF THE STUDY DL technology has been used in several studies to predict diabetes and other illnesses. This research is distinguished by the fact that it includes the

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 47

development of a website. This website makes it easy for ordinary individuals to use this approach. Because using a DL model or another similar technique might be tough for people with little or no experience with DL.

Fig. (9). Alert message for a person.

CONCLUSION This study demonstrates the development of a cloud-based, portable framework for diabetic patients and their medical care providers. It can aid with more effective patient screening and the early detection and prediction of many diseases. This study looks at a portable compact application for early diabetes detection and a DL-based application for biomedical information derived from

48 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

clinical data sets. Initially, the data was obtained from Kaggle and preprocessed using several ways. Following that, it is sent into the ANN model for additional analysis. After undergoing 2000 epochs, the best model with an accuracy of 92% was discovered and put in a portable mobile application for convenient use by patients in pandemics such as covid. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

A.T. Kharroubi, and H.M. Darwish, "Diabetes mellitus: The epidemic of the century", World J. Diabetes, vol. 6, no. 6, pp. 850-867, 2015. [http://dx.doi.org/10.4239/wjd.v6.i6.850] [PMID: 26131326]

[2]

H. Bolhasani, M. Mohseni, and A.M. Rahmani, "Deep learning applications for IoT in health care: A systematic review", Informatics in Medicine Unlocked, vol. 23, p. 100550, 2021. [http://dx.doi.org/10.1016/j.imu.2021.100550]

[3]

S. Gupta, H.K. Verma, and D. Bhardwaj, Classification of diabetes using Naïve Bayes and support vector machine as a technique. Operations Management and Systems Engineering. Springer: Singapore, 2021. [http://dx.doi.org/10.1007/978-981-15-6017-0_24]

[4]

U.U.R. Muhammed, M. Aleem, M.A. Islam, and S Ahmed, "Smart applications for diabetes management: A comprehensive survey and ranking", In: Health Informatics J vol. 26. , 2020, no. 2, pp. 1133-1151. [http://dx.doi.org/10.1177/1460458219869159]

[5]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images", Med. Biol. Eng. Comput., vol. 58, no. 4, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6] [PMID: 31993885]

[6]

R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R.X. Gao, "DL and its applications to machine health monitoring: A Survey", Mech. Syst. Signal Process., vol. 14, no. 8, 2015.

[7]

Thompson Stephan Fadi Al-Turjman, Satya Prakash Yadav, Manoj Kumar, Vibhash Yadav "Transforming Management with AI, Big-Data, and IoT"; Springer, Cham, Vol. 1, issue. 1, pp. 317, 20222. [http://dx.doi.org/10.1007/978-3-030-86749-2]

[8]

PP. Ray, "A survey on internet of Things architectures", Journal of King Saud University-Computer and Information Sciences, vol. 30, pp. 291-319, 2018. [http://dx.doi.org/10.1109/IEMCON.2018.8614931]

[9]

C.A. da Costa, C.F. Pasluosta, B. Eskofier, D.B. da Silva, and R. da Rosa Righi, "Internet of Health Things: Toward intelligent vital signs monitoring in hospital wards", Artif. Intell. Med., vol. 89, pp.

Mobile Computing

AI and IoT-based Intelligent Health Care & Sanitation 49

61-69, 2018. [http://dx.doi.org/10.1016/j.artmed.2018.05.005] [PMID: 29871778] [10]

F. Alam, R. Mehmood, I. Katib, N.N. Albogami, and A. Albeshri, "Data fusion and IoT for smart ubiquitous environments: A Survey", IEEE Access, vol. 5, pp. 9533-9554, 2017. [http://dx.doi.org/10.1109/ACCESS.2017.2697839]

[11]

S.M. Riazul Islam, M.D. Kwak, M.H. Kabir, M. Hossain, and K.S. Kwak, "The Internet of Things for Health Care: A Comprehensive Survey", IEEE Access, vol. 3, pp. 678-708, 2015. [http://dx.doi.org/10.1109/ACCESS.2015.2437951]

[12]

M.A. Qureshi, K.N. Qureshi, G. Jeon, and F. Piccialli, "DL-based ambient assisted living for selfmanagement of cardiovascular conditions", Neural Comput. Appl., vol. 434, no. 3, 2021. [http://dx.doi.org/10.1007/s00521-020-05678-w]

[13]

E.W. Evan, M Tissenbaum, and F Harunani, "MIT App Inventor: sObjectives, Design, and Development", Computational Thinking Education, 2019, pp. 31-49.

[14]

D.K. Choubey, M. Kumar, V. Shukla, S. Tripathi, and V.K. Dhandhania, "Comparative analysis of classification methods with PCA and LDA for diabetes", Curr. Diabetes Rev., vol. 16, no. 8, pp. 833850, 2020. [http://dx.doi.org/10.2174/18756417MTAz5ODYi2] [PMID: 31971112]

[15]

M. Maniruzzaman, M.J. Rahman, B. Ahammed, and M.M. Abedin, "Classification and prediction of diabetes disease using machine learning paradigm", Health Inf. Sci. Syst., vol. 8, no. 1, pp. 7-14, 2020. [http://dx.doi.org/10.1007/s13755-019-0095-z] [PMID: 31949894]

[16]

R. Ahuja, S.C. Sharma, and M. Ali, "A diabetic disease prediction model based on classification algorithms", Annals of Emerging Technologies in Computing, vol. 3, no. 3, pp. 44-52, 2019. [http://dx.doi.org/10.33166/AETiC.2019.03.005]

[17]

H. Naz, and S. Ahuja, "Deep learning approach for diabetes prediction using PIMA Indian dataset", J. Diabetes Metab. Disord., vol. 19, no. 1, pp. 391-403, 2020. [http://dx.doi.org/10.1007/s40200-020-00520-5] [PMID: 32550190]

[18]

Hariharan and K. Rajkumar.: “The importance of fog computing for healthcare 4.0-based IoT solutions.”; Fog Data Analytics for IoT Applications: Next Generation Process Model with State of the Art Technologies, 2020; pp.471-494.

[19]

R. Simó, and C. Hernández, "Treatment of diabetes mellitus: general goals, and clinical practice management", Rev. Esp. Cardiol., vol. 55, no. 8, pp. 845-860, 2002. [PMID: 12199981]

50

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 50-64

CHAPTER 4

Innovative IoT-Based Wearable Sensors for the Prediction & Analysis of Diseases in the Healthcare Sector Koteswara Rao Vaddempudi1, Abdul Rahman H Ali2, Abdullah Al-Shenqiti3, Christopher Francis Britto4, N. Krishnamoorthy5 and Aman Abidi6,* ECE Department, Prakasam Engineering College, Kandukur, Prakasam, Andhra Pradesh, India Mahatma Gandhi University/Taibah University, Meghalaya, India-793101 3 College of Medical Rehabilitation Sciences, Taibah University, Madinah, Saudi Arabia-41411 4 Mahatma Gandhi University, Meghalaya, India-793101 5 School of Information Technology and Engineering, Department of Software Systems and Engineering, Vellore Institute of Technology, Vellore Campus, Thiruvalam Road, Vellore, Tamilnadu, India - 632014 6 Swinburne University of Technology, Melbourne, Victoria, Australia 1 2

Abstract: Health monitoring may be required regularly in everyday life, which might help predict the significant health consequences. Accurate surveillance is required for effective health parameters like temperature, stress, heart rate and blood pressure (BP) in the medical and healthcare domains. The Ideal health-related characteristics for efficient persistent health monitoring are established in this study. The primary goal of the device is to monitor the health parameters of a person in everyday life, facilitating psycho-physiological supervision to examine the relationship between underlying emotional states, including changing stress levels, and the progression and prognosis of cardiovascular disease. Non-invasive sensors are employed here to observe the mentioned health-related variables. The observed data will be stored in the cloud for further processing. IoT technology has been used to process and store the measured parameters in the cloud. At the same time, the device will give a notification in the form of an alarm to the concerned person. The data can be frequently monitored by the guardian and the concerned doctor. This may help to keep an eye on the people even if they are far away from the person and the stored data can be viewed at any time from anywhere. Thus, the wearable device will record the health parameters of a person, which may assist them to know their mental and physical health, as well as give Corresponding author Aman Abidi: Swinburne University of Technology, Melbourne, Victoria, Australia; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 51

alerts in case of abnormalities. Implementation of this system will be helpful for the people to get an awareness about their health condition and also make them stay healthy.

Keywords: Blood Pressure, Cloud, Heart Rate, Display, Temperature. 1. INTRODUCTION In today's world, an IoT of improvement in the research area of the healthcare domain, and the new products under development mainly focused on Wireless transmission for the reason of comfortability and accessibility. Even these days, people are at risk of dying because of not getting medical support on time, and sometimes patients are undergoing unexpectedly become victims of a specific cause of heart issues and attacks, which may be caused by a lack of required medical care provided to patients at the appropriate time. The proposed idea is for precisely monitoring and alerting physicians and family members of elderly people. This research aims to merge IoT technology with health surveillance to render the issue even more customized and responsive by permitting devices to communicate with each other. This study will look into a variety of wearable health monitoring devices that consumers may use to track their health parameters. The wireless sensor has been used to gather data to construct a health inspection system. To accomplish real-time monitoring, the data is combined with the IoT for processing, connecting, and computation. ThinkSpeak cloud has been used to store the data collected from the device. The intimation will be given to the user if there is any abnormality observed by the device. The data is getting stockpiled in the cloud, which could be accessed by their doctor and the well-wisher of the user. Here, a wearable device is designed to provide efficient surveillance of the major healthcare variables such as temperature, BP, heart rate, and stress monitoring. In this pandemic situation, the temperature is an important parameter that needs to be monitored regularly; BP is a common health parameter that frequently varies according to the mental state of the person, heart rate is also an important health parameter that needs attention, and it plays a major role to monitor the stress level of a person. Nowadays, stress is a big problem in modern life. Everyone is focused on their job, and virtually everyone, including students and employees, is working frantically to achieve deadlines and objectives. People are typically aware that they are under a lot of job pressure and mental stress, yet they ignore their health. They also fail to take their medicine at the appropriate times, which might result in catastrophic consequences and even

52 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

death. If left unchecked, some sensor values, such as heart rate and BP, might be alarming. When the appropriate medication is prescribed at the right moment, this could hopefully avoid medical complications in the future and lower the risk of mortality. As a result, developing stress detection and health monitoring technologies that can assist individuals in understanding their mental and physical states is critical. The health parameters are monitored using sensors, and stores the collected information over the cloud using a WiFi-enabled controller; this wearable health monitoring system helps to monitor the person from anywhere by their guardians, and also patient health reports will be continuously shared with the doctor so they can easily identify if there is any abnormality. Implementation of this device will help people to lead a safe and healthy environment to live. 2. LITERATURE SURVEY According to Tamilselvi et al. [1], an IoT-based health monitoring system can measure vital signals like heart rate, oxygen saturation, core temperature, and eye movement. The heartbeat, SpO2, temperature, and eye blink sensors were used to record the data, which was then processed by an Arduino-UNO. There are no patient-specific performance measurements for the proposed device, even though it was used. Acharya et al. [2] predicted a healthcare inspection kit based on the Internet of Things (IoT). Other useful healthcare factors were examined by the installed method, including heartbeat, ECG, respiratory rate, and body temperature. Here, a pulse sensor, a digital temperature sensor, a BP sensor, an electrocardiogram sensor, and a Raspberry Pi are the primary hardware components. IoT sensors' data will now be processed on a Raspberry Pi before being transmitted back to the cloud for storage. The device's fundamental flaw is that it does not allow for the development of interconnections between modeling methodologies. Trivedi et al. [3] proposed a mobile phone-based healthcare parameter inspection system that uses microcontrollers to monitor the health of patients. The ATMega 328p microprocessor received the data from the analog sensors and processed it. Analog values are converted into digital data using the built-in analog-to-digital converter (ADC). The built-in gadget received the physical characteristics via Bluetooth. The Bluetooth gadget relied on a single component that didn't cover a large area of the spectrum. There is a new IoT safety monitoring device that can be moved around [4]. The platform's configuration is divided into three layers: control, device, and transport. Pulse oximeter and DS18B20 sensors were used to monitor the body's temperature and pulse rate, respectively. The Wi-Fi module of the transport layer was used to transfer data from the Arduino to the cloud. Finally, data were obtained from the server using the framework layer. However, several sensors

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 53

were unable to be adequately addressed due to the use of an Arduino Uno. An IoT-enabled personal health surveillance system was recommended by M. Al Yami et al. in a study paper. A person's BP, pulse, temperature, and other vital physiological data, as well as heart rate, may all be monitored in real time by the system. Wireless sensors are being used to store data for health monitoring in this research. Processing, networking, and calculation for real-time supervision are all done via the IoT. It was shown that the proposed system had both universal precision and trustworthy testability, both of which helped overcome the shortcomings of the current health monitoring platform [5, 6]. Using IoT technology, the author [7, 8] has developed a real-time remote analysis system that could aid medical practitioners in diagnosing patients more swiftly, integrating various modes such as online consultation with video conversation. Heart rate monitoring can be done via a smartphone, according to Gregoski et al. [9]. Light and a camera on a smartphone were used to assess blood flow in fingers and estimate cardiac output-specific blood flow. As a result of the system's development, people with smartphones may now check their heart rate without having to move their hands from their pockets or purses. It's an interesting idea, but it's not practical if you have to monitor your heart all the time. As of right now, the IoT is perceived in many different ways. One way that the IoT can collect data is through the use of various sensors, such as those found in smartphones and other connected devices like GPS and infrared sensors [10]. To monitor healthcare indicators such as pulse, respiration rate, BP, and body temperature, Rezaeibagha and Mu [11] developed an Agent system that uses portable wireless sensors. 3. PROPOSED SYSTEM Patients or elderly people are supposed to be supervised by a healthcare attendant or a family member under today's social health insurance framework, where patients stay at home after operations—most of the people who are supposed to work nowadays. So, the supervision of the patients is becoming a struggle to keep track of their family members, particularly elderly patients. To resolve this concern, the proposed device is an IoT-based patient health surveillance system. The device uses sensor technologies, a microcontroller, and a Wi-Fi module to assist users in trying to keep track of their family members. The proposed wearable device is planned to design in the form of a watch, and it can observe the health parameters like temperature, BP, heart rate, and stress. The observed data using these sensors will be sent to the cloud storage and display device. The observed data will be diagnosed by the controller, and once the parameter is sensed to increase the threshold value, then the indication will be

54 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

given to the user and also the notification regarding the abnormality in the variation of parameters like temperature, BP, heart rate, and stress. The notification can be given to the guardian of the person and the doctor. These data are supposed to store in the cloud that doctors can use to keep an eye on the health of the person in a continuous manner. The constant surveillance of the health parameter may help the person to know their physical health and also the mental health of a person. This may also lead them to have a precaution before suffering from any viral illnesses. 3.1. Block Diagram The block diagram of the proposed device is given below. The device comprises a heart rate sensor, temperature sensor, and BP sensor to measure the health care parameters. The sensed data will be given to the microcontroller, which will be processed and show the value in the display, and an indication will also be given. The ThingSpeak cloud has been used to store the data using the IoT. In the ThingSpeak dashboard, there will be an indicator to show the level of every health parameter considered. The graphs are employed there to have a view of the changing parameter according to the time. The observed data will be stored in the ThingSpeak cloud. Fig. (1) shows the block diagram of IoT-based wearable sensors for the prediction and analysis of disease in healthcare.

Fig. (1). Block diagram of IoT-based wearable sensors for the prediction and analysis of disease in healthcare.

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 55

The sensor used to measure the health parameters are detailed below: MLX90614 Sensor: The temperature sensor MLX90614 has been used in the device to observe the temperature of the person; it is a noninvasive sensor and the operative voltage required for this sensor is 3.3V. The MLX90614 is a noncontact infrared thermometer. Due to the small size of the TO-39 container, both the IR-sensitive thermopile detector chip and the signal conditioning ASIC can be placed in the same package. The MLX90614 thermometer is equipped with a low-noise amplifier, a 17-bit ADC, and a powerful DSP unit, providing exceptional accuracy and resolution. As the sensor is smaller, less expensive, and more accurate, it has the potential to be employed in wearable electronics applications. MAX30100 Sensor: The MAX30100 sensing package merges a pulse oximeter as well as a heart monitor into a single unit. This optical sensor monitors blood absorption by employing two LEDs, one infrared, and a photodetector that measures illumination from both wavelengths. This LED color combination is excellent for reading data with one's fingers. The digital signal information is recorded in a 16-deep FIFO inside the device utilizing program memories. It communicates with the host microcontroller through an I2C user device [12]. BMP180 Sensor: The BMP180 Sensor measures barometric pressure and temperature with exceptional precision at a reasonable cost. You may also use it as an altimeter because the pressure fluctuates with altitude. A 3.3V regulator, I2C, is attached to a PCB with the sensor. In terms of firmware and software, it is similar to the BMP085 3.2. Flow Diagram The proposed device will also help to find the mental state of the person. Fig. (2) flow diagram shows the working flow of the device. In this system, when the device is enabled, the sensor will start to sense the parameters of the human body. The sensed values will be processed by the microcontroller to check whether the sensed values are under the threshold value or not. If the sensed value is under the threshold value, then the device will conclude that the person is healthy. If it is not more or less than the threshold value, then the system says that the person is not healthy and needs to focus on health. Stress is one of the major issues faced by many working people nowadays. Due to heavy stress, the person may have a chance to get cardiovascular issues sometimes.

56 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

Fig. (2). Flow diagram of IoT-based wearable sensors for the prediction and analysis of disease in healthcare.

3.3. Hardware Implementation The implementation of this system has been done with the help of the microcontroller and the necessary input and output peripherals like sensors and display devices. The connection between the input and output peripherals with the microcontroller was made, and the sensed values were visualized using the OLED display. Fig. (3) reveals the heart rate, temperature, stress level, and health status of a person. From the above display, the value 72 is the heart rate, 94 is the temperature, and NORMAL states that the person is not stressed. The term healthy indicates the health of the person is good.

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 57

Fig. (3). Heart rate and stress display.

Fig. (4) shows the pressure sensed by the BP sensor. It shows the value of 106mmHg, which indicates the BP of the person is in an ideal state.

Fig. (4). BP display.

3.4. IoT Cloud ThingSpeak is an IoT cloud platform that allows you to communicate sensor data to the cloud [13]. The user can use a tool like MATLAB and other applications to analyze and visualize the data, as well as the user is allowed to create apps for the users. MathWorks supports the ThingSpeak platform [14]. The ThingSpeak cloud can be used to gather the data observed by the sensor, and also the dashboard of the platform shows the present value of each sensor. The indicators are also used to indicate the abnormalities observed by the sensor. Usually, it will be in an OFF state; once the observed values exceed the threshold

58 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

value, the indicator will turn ON. The data gathered from the sensors can be stored in the cloud, and those data can be accessed by their doctor or by their guardian at any time from anywhere. This helps to track the health condition of their family members even if they are not around them. 4. RESULTS AND DISCUSSION The wearable device is designed in the form of a watch. This watch can observe health-related variables like temperature, BP, heart rate, and stress of the person. The observed values will be given to the microcontroller, and the value will be displayed in the OLED display device connected to the controller. The gathered data will also be sent to the ThingSpeak cloud. Healthy adults may have a temperature between 97.8 degrees F to 99 degrees F. If the temperature exceeds the above-mentioned temperature, then it is considered hypothermia. LM 35 has been used in this device to observe the temperature of the person. The value observed using LM 35 has been visualized as a graph, and the graph is shown below. Fig. (5) indicates the variation in the temperature parameter of a human. The temperature for a normal human might be considered 98.6 degrees. The observed value of temperature for a person using ThingSpeak has been shown in the graph. The indicators are used in the ThingSpeak dashboard for showing the alert on abnormalities in the health parameter values. So, whenever the observed temperature value is higher than the threshold value, then the red indicator will show that the temperature is high; whenever the temperature is low, it shows the green indicator, and for normal temperature, a yellow indicator will get enabled. The results observed from the various sensors have been discussed below,

Fig. (5). Observed Temperature using ThingSpea a) High-Temperature Indicator and b) Normal Temperature Indicator.

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 59

Fig. (6) shows the low temperature, normal temperature, and high temperature of the person. According to the graph, the temperature value is going high to 110 degrees at 19.09. So, the indicators for high temperature have been enabled in Fig. (6a). The indicators will get enabled according to the observed temperature, and Fig. (6b) shows the normal temperature. According to the graph, the temperature value is in the normal state at the point before 19.06, which is observed in Fig. (5). The normal indicator is enabled as the temperature of the person is normal.

Fig. (6). Indicators of temperature in ThingSpeak.

The pressure of the blood pressing against the walls of the arteries has been recognized as BP. Whenever the heart beats, blood is impelled into the arteries. When the heart beats and pumps blood, the BP will be on its supreme. BP can be observed as two parameters in terms of millimeters of mercury (mmHg). In general, the BP value considers three levels low, ideal, and high pressure and its unit is mmHg. ● ● ●

Ideal - 90/60 and 120/80. High - 140/90 or more. Low - 90/60 or less.

Fig. (7) shows the graph of the systolic pressure, where the systolic pressure is exerted by your heart whenever it pumps blood out of the human body.

60 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

Fig. (7). Observed Systolic Pressure using ThingSpeak.

Fig. (8) shows the graph of the diastolic pressure, where the diastolic pressure is the pressure that the heart perceives when it is rested during pulses.

Fig. (8). Observed Diastolic Pressure using ThingSpeak.

The heart rate is the number of times the heart beats in a particular period. The arteries enlarge and narrow by the flow of blood employing the heart pumps blood over them. A healthy adult may have a heart rate between 60 to 100 beats every minute. Pulse rate is one of the major parameters to analyze the stress parameter. When the heart rate is considered more than 90 per minute, the person is supposed to be stressed.

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 61

Fig. (9) shows indicators like low BP, normal BP, and high BP of a person. The above image shows that the person has a high BP at the point of 19.07 with the reference of the graph and Fig. (9a). shows the indication of low BP of the person observed at the point of 19.05 by the sensor and represented in the systolic and diastolic graphs. So, the Low BP indicator has been enabled, as seen in Fig. (9b).

Fig. (9). Observed indicator of heart rate in ThingSpeak High BP indicator and b) Low BP indicator.

Fig. (10) shows the graph of the heart rate, where the heart rate is a measure of the number of bits in a particular period. The heart rate of a normal human is 60 to 100 beats per second. The variation of the heart rate is observed for a person and is visualized in the graph. Fig. (11) shows the normal heart rate of the person. It is considered that the stress parameter also depends on the heart rate parameter. In Fig. (11a), yellow indicator has enabled, so it says that the heart rate of the person is normal. So, at 19.7, the heart rate of the person is normal, which is shown in a graph in Fig. (10) and the yellow indicator has got enabled. Finally, it is concluded that the person may not be in a stressed state, and Fig. (11b), shows the green indicators, which means that the heart rate of the person is at a low level. And it is enabled with the reference point of 19.6 of the graphs shown in Fig. (10) where the heart rate of the person is going low in the graph so the green indicators getting enabled accordingly. It is one of the most important issues that need to be focused on for humans.

62 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

Fig. (10). Observed indicator of heart rate in ThingSpeak Normal heart rate indicator b) Low heart rate indicator.

Fig. 11. Observed heart rate indicator in ThingSpeak.

5. CONTRIBUTION TO THE HEALTH SECTOR The proposed device is a wearable type. The sensors embedded in the proposed model are noninvasive. Therefore, the person using the particular device will not undergo any pain or harm during the measurement of the health parameters. The existing model in the market measures a health parameter and displays it only in the device, but the proposed model is designed to send the data to the Thingspeak cloud. The data stored in the Thingspeak cloud will be stored in the cloud for later

Healthcare Sector

AI and IoT-based Intelligent Health Care & Sanitation 63

use. So, the data will be directly communicated to the health care personnel and intimate the abnormalities in the health parameters of the person to their guardian. CONCLUSION During the investigation of the health-related parameters, the various sensor was integrated with a controller with in-built wireless technology—monitoring of a person's health using the IoT technology in this study. The observed data of the sensors are assessed. A criterion for supportive approaches is the real-time risk for health problem recognition through wearable sensors. The above study presents an overview of various wearable IoT-based health and behavior-monitoring techniques. Second, it demonstrates a smart and unique health monitoring integrated development environment that allows for real-time patient monitoring or aged users and availability of information either through the web. Using the IoT, health-related data of a person will be gathered, and real-time monitoring and alarms will be provided if there are any abnormalities. Considered the healthcare parameter like temperature, heartbeat, BP and stress, the above-mentioned parameter will be monitored and sent to the controller and directed to the display device and the cloud for further use. The main advantage of this device is that the doctor and the guardian could retrieve the data stored in the cloud at any time. This may help the person to track their health as well as their guardian, and the doctor can keep an eye on the person’s health. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

V. Tamilselvi, S. Sribalaji, P. Vigneshwaran, and P. Vinu, GeethaRamani J. “IoT-based health monitoring system”. International Conference on Advanced computing and Communication Systems, p. 386–9, 2020.

[2]

A.D. Acharya, and S.N. Patil, “IoT-based health care monitoring kit”. Fourth International Conference on Computing Methodologies and Communication, p. 363–8, 2020. [http://dx.doi.org/10.1109/ICCMC48092.2020.ICCMC-00068]

[3]

S. Trivedi, and A.N. Cheeran, “Android-based health parameter monitoring”. International Conference on Intelligent Computing and Control Systems, pp. 1145–9, 2017.

64 AI and IoT-based Intelligent Health Care & Sanitation

Vaddempudi et al.

[http://dx.doi.org/10.1109/ICCONS.2017.8250646] [4]

S.P. Kumar, V.R.R. Samson, U.B. Sai, P.L.S.D.M. Rao, and K.K. Eswar, “The smart health monitoring system of patients through IoT”. International Conference on I-SMAC (IoT in social, mobile, analytics and cloud), pp. 551–6, 2017. [http://dx.doi.org/10.1109/I-SMAC.2017.8058240]

[5]

M. Al Yami, and D. Schaefer, "Fog computing as a complementary approach to cloud computing", International Conference on Computer and Information Sciences. pp. 1–5, 2019. [http://dx.doi.org/10.1109/ICCISci.2019.8716402]

[6]

A. Sharma, Sarishma, R. Tomar, N. Chilamkurti, and B-G. Kim, "Blockchain Based Smart Contracts for Internet of Medical Things in e-Healthcare", Electronics (Basel), vol. 9, no. 10, p. 1609, 2020. [http://dx.doi.org/10.3390/electronics9101609]

[7]

S.O. Agnisarman, K. Chalil Madathil, K. Smith, A. Ashok, B. Welch, and J.T. McElligott, "Lessons learned from the usability assessment of home-based telemedicine systems", Appl. Ergon., vol. 58, pp. 424-434, 2017. [http://dx.doi.org/10.1016/j.apergo.2016.08.003] [PMID: 27633239]

[8]

S. Banerjee, and S. Roy, "Design of pulse rate and body temperature monitoring system with Arduino via Wi-Fi and Android-based gadget", Int. J. Eng. Res. Technol.Studies, vol. 2, no. 5, pp. 302-306, 2016.

[9]

M.J. Gregoski, M. Mueller, A. Vertegel, A. Shaporev, B.B. Jackson, R.M. Frenzel, S.M. Sprehn, and F.A. Treiber, "Development and validation of a smartphone heart rate acquisition application for health promotion and wellness telehealth applications", Int. J. Telemed. Appl., vol. 2012, pp. 1-7, 2012. [http://dx.doi.org/10.1155/2012/696324] [PMID: 22272197]

[10]

M.M. Islam, A. Rahaman, and M.R. Islam, "Development of smart healthcare monitoring system in IoT environment", SN Computer Science, vol. 1, no. 3, p. 185, 2020. [http://dx.doi.org/10.1007/s42979-020-00195-y] [PMID: 33063046]

[11]

F. Rezaeibagha, and Y. Mu, "Practical and secure telemedicine systems for user mobility", J. Biomed. Inform., vol. 78, pp. 24-32, 2018. [http://dx.doi.org/10.1016/j.jbi.2017.12.011] [PMID: 29288816]

[12]

J.S. Paiva, D. Dias, and J.P.S. Cunha, "Beat-ID: Towards a computationally low-cost single heartbeat biometric identity check system based on electrocardiogram wave morphology", PLoS One, vol. 12, no. 7, p. e0180942, 2017. [http://dx.doi.org/10.1371/journal.pone.0180942] [PMID: 28719614.]

[13]

Satya Prakash Yadav, Bhoopesh Singh Bhati, Dharmendra Prasad Mahato, Sachin Kumar, “Federated Learning for IoT Applications”, Springer Innovations in Communication and Computing, 2022.

[14]

L.E. Romero, P. Chatterjee, and R.L. Armentano, "An IoT approach for integration of computational intelligence and wearable sensors for Parkinson’s disease diagnosis and monitoring", Health Technol. (Berl.), vol. 6, no. 3, pp. 167-172, 2016. [http://dx.doi.org/10.1007/s12553-016-0148-0]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 65-82

65

CHAPTER 5

Construction and Evaluation of Deep Neural Network-based Predictive Controller for Drug Preparation K. Sheela Sobana Rani1, Dattathreya 2, Shubhi Jain3, Nayani Sateesh4, M. Lakshminarayana5 and Dimitrios Alexios Karras6,* Department of Electrical and Electronics Engineering, Sri Ramakrishna Institute of Technology, Coimbatore, Tamil Nadu, India-641010 2 Alva's Institute of Engineering and Technology, Mijar, Moodbidri, Karnataka, India-574225 3 Swami Keshvanand Institute of Technology Management & Gramothan, Ramnagaria, Jagatpura, Jaipur, Rajasthan, India-302017 4 Department of Information Technology, CVR College of Engineering, Vastunagar, Mangalpalli (V), Ibrahimpatnam (M), Rangareddy, Telangana, India-501510 5 Department of Medical Electronics Engineering, M S Ramaiah Institute Of Technology, MSRIT Post, M S Ramaiah Nagar, MSR Nagar, Bengaluru, Karnataka, India - 560054 6 National and Kapodistrian University of Athens (NKUA), Hellas, Greece, 1

Abstract: The evaporator used in the pharmaceutical industry is for drug preparation. The purpose of the evaporator in drug manufacturing is to extract the water content in the material through the heating process. In this research, the SISO evaporator is taken, which contains temperature as input and dry matter content as output. The mathematical modelling of the drug preparation evaporator is done with the help of the system identification method. Controlling and maintaining the temperature inside an evaporator is a tedious process. In this regard, the Neural Network predictive controller (NNPC) is designed and implemented for drug preparation. It helps to predict the future performance of the evaporator and tune the control signal based on that. The setpoint tracking challenge is given to the designed controller. For analysing the performance of the controller, the error metrics, such as integral square error (ISE), integral absolute error (IAE), integral time square error (ITSE), and integral time, absolute error (ITAE), are employed. The time-domain specification, such as rise time, settling time, and overshoot, is also used to better understand controller performance. From the above two analyses, the conclusion is made that the predictive controller is performing well in comparison with the conventional PID controller in the drug preparation pharmaceutical industry. Corresponding author Dimitrios Alexios Karras: National and Kapodistrian University of Athens (NKUA), Hellas, Greece; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

66 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

Keywords: Error Metrics, PID Controller, Predictive Controller, Time-domain Specifications. 1. INTRODUCTION In this civilization, drugs and pharmaceutical engineering play a major role. In the initial twentieth century, the basics of contemporary Indian pharmaceuticals and pharmaceutical business were placed. India is now producing a bigger volume of diverse medicinal ingredients The pharmaceuticals and drugs manufacturing industry has emerged as one of the country's most important sectors. Approximately 82 percent of the majority of bulk pharmaceuticals are produced domestically, with the remainder imported [1, 2]. The collective process of distillation, evaporation, condenser, dryer crystallizer and other unit activities used in drug manufacture produces a large amount of effluent and causes environmental deterioration. Manufacturing is distinguished by a wide range of goods, developments, size of plants, and also the amount and value of wastewater produced. In truth, the pharmaceutical sector encompasses a wide variety of trades, each with its processes and procedures. As a result of this variability, describing a typical pharmaceutical effluent is nearly difficult. An evaporator is an element movement that separates the fluid from solids through heat interchange by earnings of evaporation or bubbling. The inspiration behind disappearing is to emphasize a response of a non-volatile solute (i.e., solids) and a dissoluble (i.e., fluid), which is commonly water. Evaporating a part of the dissoluble amasses the solute into a more adhesive fluid item. Evaporation is dependably operated in food preparation [3, 4], synthetic, and drug ventures. Concentrating aqueous solutions by evaporation is a common practice. It is necessary to bring the liquor to a boil in an evaporator and then remove the vapor. If the solution contains liquified components, the resulting strong liquor may become saturated, which will lead to the formation of rock crystals. The following are the numerous sorts of liquids that can be evaporated. 1. People who can resist high temperatures, such as those in the 330 K range, are free of decay. 2. Those that create solids once concentrated, where instance, crystal size and shape may be essential, and that do not. 3. These boil at about the same heat as waters at the given pressure, and those with a much more boiling point. The flush is evaporated by raising the temperature of the mixture. The temperature is largely offered to provide the latent heat of vaporization, and it can

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 67

accomplish significant efficiency in heat utilization by utilizing technologies for collecting heat from the vapor. While low-pressure drain steam from turbines is the most common heating medium, specific temperature transfer fluids are also utilized. The particle implementation of data on temperature transfer to boiling liquid, as well as a realization of what occurs to the fluid during concentration, are required for the construction of an evaporation unit. In addition to the three primary characteristics listed above, liquors with an inverse solubility curve and a high likelihood of depositing scale on the heating surface should be given different considerations. The solvent is vaporized to produce a thick, strong liquor solution during evaporation. Evaporation varies from drying in that the deposit is a liquid—occasionally a very sticky one—relatively than a solid; it changes from purification in that the vapor is a mixture; it differs from crystallization in that the importance is on extracting a solution relatively than establishing and structuring the crystal. A slurry of crystals can form during the evaporation process in the saturated mother liquor. Vapor is condensed and discarded in evaporation, with the abundant liquor being the prized by-product. Mineral-bearing water, for example, is routinely evaporated to create a liquid that does not have solid particles for boiler feed, particular process needs, or humanoid feeding. Even though it is commonly referred to as liquid distillation, the actual procedure is called desertion. Fig. (1) depicts how large-scale evaporation technologies are being used to recover potable water from seawater. The desired product is condensed water in this example. Only a small percentage of the total water in the feed is collected, with the rest being discharged into the sea [5]. Evaporation contrasts with the Dryness of the product, and drying results in the vanishing of concentrated fluid non-strong. Dissipation may be utilized as an underlying advance in delivering a dehydrated item on the off chance that the fluid concentrate, goes through a drying interaction, for example, shower aeration. The mix of dissipation and shower drying is regularly used to do crushed items, like crushed milk. This mix of cycles is financially appealing because higheffectiveness dissipation is essentially less expensive than drying and different strategies for eliminating water [6]. The evaporators are contained in 2 areas: a warming segment and a smoke/watery divider segment. Those segments can be positioned inside a private vessel, or the heating area may be outside the container that families the smoke/watery division segment [7].

68 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

Fig. (1). Heating and vapor/liquid separation sections of evaporator [8].

2. MODEL DESCRIPTION The modest technique of evaporation is a certain area provided with appropriate heat transfer, and the solution is fed to the evaporator. The direct contact condenser or a surface condenser helps to condense the vapor generated in the evaporator. Finally, at the bottom of the evaporator, the concentrated product is obtained. In the single effect evaporator, the liquid arrives at the upper of the evaporator frame and moves down on the one crosswise of the plate. The falling film single effect evaporator is modeled by pouring; the black liquor pours down one side of the plate after entering the upper part of the evaporator body. The dimpled plate structure encourages turbulent flow. When there are numerous plates, distribution devices ensure that the liquor is distributed evenly. On the other side of the plate, steam is introduced. The solution evaporates in the opposite direction as the liquor evaporates. A circulating pump forces the concentrated black liquor to exit the plate stack at the bottom of the effect [9]. The heat exchanger, the evaporating part, and the separator are the main part of the predictable evaporator. Separation, wherein vapor escapes from the liquid and drifts to the condensation or other apparatus, and evaporation, wherein liquid boils and evaporates. Numerous evaporators utilize a single cylinder to drive all three

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 69

portions. In the tube's centre, a vapor burner portion is placed, with pipes flowing across it and the evaporation of liquors rising. The top of the cylinder is covered with baffles that enable steam to escape while keeping liquid dewdrops from mixing with the vapours from the surface of the liquid. There are a few enigmas. The vapor condenses in the heat exchanger's outside jacket, referred to as a calandria in this form of evaporation. The fluid being evaporated is warmed on both the inside and outside of the tubes. The heat transfer resistance is governed by the coefficients of the vapor and fluid films, as well as the material of the tube wall. Although the rates and types of liquid circulation are exceedingly hard to distinguish, they have a major effect on evaporation [10, 11]. In the drug preparation industry, the following are significant real-world issues in evaporators: ● ●

●

●

The highest point of heating that can be tolerated, maybe well below 100°C. To achieve a reasonably high temperature, promote liquid circulation across the heat transfer surfaces. The thickness of the fluid, which will typically rise significantly as the concentration of liquified elements rises, heat transmission coefficients, and to avoid any limited overheating. The tendency to foam makes it difficult to distinguish between liquid and vapor.

As evaporation progresses, the residual liquors become more concentrated, and the boiling temperature rises as a result. Whenever the source of heat remains unchanged, raising the boiling temperature diminishes the temperature drop. Finally, the total rate of heat transmission will be lowered. Furthermore, when the concentration of the solute increases, the liquid viscous increases, often significantly, affecting motion and heat transfer ratios and resulting in reduced boiling rates. Another difficulty is that heat transfer rates have been observed to change with actual temperature decreases when regarded as a whole. The following are the primary features that influence the proportion of evaporation: ● ● ● ● ●

The proportion at which heat can be supplied to the fluid, The amount of heat necessary to evaporate each kilogram of liquid, The liquid's extreme permissible temperature, The pressure at which point evaporation occurs, and Variations in the foodstuff that can occur during the evaporation process

The calculations for heat transmission to boiling liquids, as well as the convection and conduction equations, govern temperature transfer in evaporators. The heat

70 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

must come from a cause at an appropriate temperature, which is condensing steam in most cases. Steam is obtained moreover straight from a boiler or a prior evaporation step in an additional evaporator. Other means of heating, such as straight fire or electric resistance heaters, face significant opposition due to the requirement to evade local high temperatures and the high costs of electricity. Condensing vapor temperatures might be too high for the product in some circumstances, thus, hot water may be utilized instead. 3. SYSTEM IDENTIFICATION The plant model is required to develop a controller for a plant. The system identification toolbox was used to identify a plant model. System identification creates a mathematical model from measurable data using the statistical method. The main goal is to predict a model using input-output data that has been measured. 3.1. Material Balance Equation for Drug Preparation Incorporating the principle of material preservation to a different capacity of size ∆Z gives: 𝑀|𝑡+∆𝑡 − 𝑀|𝑡 = 𝐺|𝑧 ∆𝑡 − 𝐺|𝑧+∆𝑧 ∆𝑡 − 𝑊∆𝑡

(1)

Where, M→ Mass of the liquor (kg), G→ Mass flow rate of black liquor (kg/s), Divide A t and take the limit as ∆t→ 0 yields.

In the circumstance of the drug liquor component, Equation 5 is converted as, 𝝉(𝛿𝐺/𝛿𝑡) = 𝐺|𝑧 − 𝐺|𝑍+∆𝑧 − 𝑊

Where, r→ Stagnant time in the different capacity of the rate VZ in the z-direction, i.e.

(2)

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 71

T=(∆Z/VZ) and Vz=u = average velocity (m/s) Solving, (∆𝑧⁄𝑉𝑧 ) 𝑋 (𝛿𝐺 ⁄𝛿𝑡) = 𝐺|𝑧 − 𝐺|𝑍+∆𝑧 − 𝑊

(3)

𝐺|𝑧 − 𝐺|𝑧+∆𝑧 − 𝑊 ⁄∆𝑧 ∆𝑧

(4)

(1⁄𝑉𝑧 )𝑋(𝛿𝐺 ⁄𝛿𝑡) =

Defining W = W/∆z gives 𝐺|𝑧 − 𝐺|𝑧+∆𝑧 −𝑊 ∆𝑧 (𝛿𝐺/𝛿𝑡) = 𝑉𝑧 𝑋 (𝐺|𝑧 − 𝐺|𝑍+∆𝑍 )/∆𝑧 − 𝑊 x 𝑉𝑧 (1⁄𝑉𝑧 )𝑋(𝛿𝐺 ⁄𝛿𝑡) =

(5) (6)

Taking the limit as ∆Z-→0 then yield (𝛿𝐺/𝛿𝑡) + 𝑉𝑧 (𝛿𝐺/𝛿𝑡) = −𝑊 𝑥 𝑉𝑧

(7)

Dry Solid balance for drug preparation At a given position z, 𝛿 (𝑀 𝑋𝑠 )/( 𝛿𝑡) = 𝑀 𝑋 (𝛿𝑋𝑠 / 𝛿𝑡) + 𝑋𝑠 𝑥 (𝛿𝑀/𝛿𝑡)

(8)

Where, M → Mass (kg) = Gτ Xs→drug liquor dry solids content (kg/kg) G →drug liquor mass flow rate (kg/s) τ →Residence time (𝛿𝑋𝑠/𝛿𝑡) + 𝑉𝑧(𝛿𝑋𝑠/𝛿𝑧) = (𝑋𝑠/𝐺) x 𝑊 𝑉𝑧

(9)

Equations 7 and 9 state that structure is considered quasilinear with the recognized. The dynamic behaviour of the evaporator is necessary to develop a set of differential equations: The equations are the following:

72 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

Overall mass balance equations 𝑑(𝐿 + 𝐺) =𝐹−𝑂−𝐵 𝑑𝑡

(10)

3.2. Mass Flow Rate for Drug Preparation To make the Equations 3 and 4 dimensionless, the following variables are introduced G* → (Drug liquor mass flow rate/ drug liquor mass flow rate at the initial stage) = (kg/s)/(kg/s) 𝐺 𝐺0

𝐺∗ =

Content for Dry Solids in drug X*→(Drug liquor dry solids content / Initial drug liquor dry solids content) = (kg/kg)/(kg/kg) 𝑋∗ =

𝑋𝑠 𝑋𝑆.0

Space Synchronize along the Plate Z* → (Planetary synchronize along the plate / Plate length) = meter/meter 𝑍∗ =

𝑍 𝐿

Plate Velocity in drug preparation V* → (Flow velocity along the plate / Friction velocity at initial conditions) = (m/s)/(m/s) ܸ‫  כ‬ൌ 

ܸ௓ ܸ௙Ǥ଴

‹Ž–Š‹…‡••ƒ–‹‹–‹ƒŽ…‘†‹–‹‘ ܸ௙௢ =ට

”ƒ˜‹–ƒ–‹‘ƒŽƒ……‡Ž‡”ƒ–‹‘

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 73

By using the above equations, the mathematical modelling for the system is designed by considering the differences in boiling point rise, total heat transfer coefficient, and physical-thermal coefficient of the liquid. The mathematical model has been designed to maintain the concentration of liquid concerning temperature. Consider a block diagram of the single-effect evaporator and the mass and energy balance equation over ith the effect is assumed for a single-effect evaporator. From Fig. (2), let us consider,

Fig. (2). Block diagram of the single-effect evaporator.

Where, T → Temperature of the Liquid in (° Celsius) V → Liquid in the Evaporator X →dry matter content L→Mass holds up of the liquid Vi+1 → Drug liquid that enters the evaporator Ti+1 → Temperature of the liquor in the evaporator

74 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

Ci-1 → Condensate liquid outlet Xi → Dry matter content in the evaporator. Mass balance in the evaporator section is, ‫ܮ‬௜ାଵ ൌ  ‫ܮ‬௜ ൅  ܸ௝

(11a)

Mass balance in the steam chest is, ܸ௜ିଵ ൌ  ‫ܥ‬௜ିଵ

(11b)

Overall mass balance is, ‫ܮ‬௜ାଵ ൌ ‫ܮ‬௜ ൅ ܸ௜

(12)

Partial mass balance for solids is. ‫ܮ‬௜ାଵ ൌ ‫ܮ‬௜ ൅ ܸ௜

(13)

Overall, energy balance for temperature inlet and dry matter content is: ‫ܮ‬௜ାଵ ܺ௜ାଵ ൌ ‫ܮ‬௜ ܺ௜ ൌ  ‫ܮ‬௙ ܺ௙

(14)

‫ܮ‬௜ାଵ ܺ௜ାଵ ൌ  ‫ܮ‬௜ ݄௅௜ ൅ ܸଵ ‫ܪ‬௩ଵ ൅  ο‫ܪ‬௜

(15)

Where,

Thus, the overall equation for a single effect evaporator concerning temperature as input and dry matter content as output. 4. PREDICTIVE CONTROLLER The preceding limit method is used in the NNPC method. The model reaction is forecast by a NN model on a limited time horizon [12]. A mathematical optimization algorithm practices the forecasts to find the controller output that minimizes the subsequent presentation principle over the quantified time limit.

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 75

ࡺ૛

ࡶ ൌ ෍ ሺ࢟࢘ ሺ࢚ ൅ ࢐ሻ െ ࢟࢓ ሺ࢚ ൅ ࢐ሻሻ૛ ൅ ࣋ሺ࢛ᇱ ሺ࢚ ൅ ࢐ െ ૚ሻ െ ࢛ᇱ ሺ࢚ ൅ ࢐ െ ૛ሻሻ૛

(16)

࢐ୀࡺ૚

A constant, SISO process step-response model may be stated as: ࡺି૚

࢟ሺ࢑ ൅ ૚ሻ ൌ  ࢅ૙ ൅  ෍ ࡿ࢏ ο࢛ሺ࢑ െ ࢏ ൅ ૚ሻ ൅ ࡿࡺ ࢛ሺ࢑ െ ࡺ ൅ ૚ሻ

(17)

࢏ୀ૚

Where y(k+1) → Output variable at (k+1) ∆uk-i+1 → Variation in the input Y and u → variable deviation The typical variables are coefficients, S1 to SN. Naturally, N is selected so that 30 is less than or equal to N greater than or equal to 120. The primary value, y(0), is represented by y0. For effortlessness, we will assume that y0 = 0. The objective of the Predictive controller: 1. To Avoid input and output limitations from being violated. 2. Achieve best set points for around output variables, although keeping others within prescribed ranges. 3. Avoid the input variables from moving too much. 4. Sometimes when the sensor or actuator is unattainable, the regulator has as many process variables as possible. The limits across the trailing mistake and controller additions evaluated are N1, N2, and Nu. The uncertain regulator signal is u′, the anticipated response is yr, and the network model response is ym. The value determines the proportion of the squared sum of controlling steps to the performance metric [13]. Consider a block diagram of the single-effect evaporator and the mass and energy balance equation over ith the effect is assumed for a single-effect evaporator.

76 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

From Fig. (3), the NNPC technique is depicted in the block diagram below. This plant model and the increase in efficiency block make up the controller. The optimization block selects the u′ values that minimize J, and then the plant receives the optimum u.

Fig. (3). NN controller.

Fig. (4) shows the optimized NN. For Real-time prediction and controller tasks, multi-layered NNs (MLNNs) are used. The NN predictor is used to forecast highly nonlinear gas network dynamics in a transient condition on the fly. The Real-time NN controller uses the forecast data to identify the near-optimal regulator inputs (compressor rotational speeds) to meet the new requirements. These may be accomplished by chasing the networks, which earlier determined best outlet pressures [14].

Fig. (4). Optimized NN model.

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 77

This plant model forecasts the output values of the model depending on previous inputs and model outputs. The plant model for the NN structure is depicted in Fig. (5).

Fig. (5). NN plant model structure [15].

Where u(t) → system input Yp(t) →output of the plant Ym(t) →output of NN model plant TDL → Previous input values of Tapped delay lines IWi,j →From input j to layer I weight matrix LWi,j →layer j to layer I weight matrix The NN training data signal is the forecast mistake between the model output and the NN output. The NN plant model forecasts future plant output values founded on preceding inputs and plant outputs [16]. The system outlet demand flow rates are modified in two separate scenarios to test the control scheme's robustness. Furthermore, the controller's performance is assessed in the face of noise and disturbances, confirming its efficiency and accuracy [17]. 5. RESULTS The evaporator's output for drug preparation, such as dry matter content, was collected using the Temperature as input. The acquired data were used to identify the system. For the linear SISO system, the Model-based NN was designed and

78 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

analysed. Performance criteria are used to compare the Model-based NN output to that of a traditional PID controller. 5.1. Simulation Results of Drug Preparation for NN Controller and PID Controller When the feed flow is given a step input of -1 at the sample moment of 10, the PID controller produces the most undershoot and error than the MPC to reach the set point. The vapor flow is then given a step input of 2 at sampling moment 60. The PID controller takes longer than the Model-based NN to get settled and the oscillation has also reduced compared with the PID controller. Fig. (6). shows the regulatory response of the Model-based NN controller and PID controller.

Fig. (6). The regulatory response of Model-based NN controller and PID controller

Table 1 shows the performance analysis of time-domain specification with the controllers. By analysing the time domain specification for the PID and NN controller, the rise time for the NN controller is 0.9454 Sec and for the PID controller, it is 0.4629 sec; the is a 0.5-sec delay in the NN controller. The settling time for the NN controller is 4.18094 and the PID controller is 14.7976 Sec. The settling Min value for the NN controller is 1.8094 and for the PID controller is

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 79

0.8374, the settling Max value for the NN controller is 2.2816 and for the PID controller is 3.0973. The overshoot for the NN controller is 13.9196 and for the PID controller is 54.9849. There is no undershoot for both the controller. The Peak value for the NN controller is 2.2816 and the PID controller is 3.0973. So, by comparing the overall time domain specification values, the Model-based NN controller has a better performance comparing the PID controller. So, the performance of the system for drug preparation is high in the NN controller Table 1. Time-domain analysis for PID controller and NN controller. Time Domain Specification/ Controllers

NN

PID

Rise Time

0.9456

0.4629

Settling Time

4.1874

14.7976

Settling Min

1.8094

0.8374

Settling Max

2.2816

3.0973

Overshoot

13.9196

54.9849

Undershoot

0

0

Peak

2.2816

3.0973

Table 2 lists the performance criteria for both PID and MPC controllers, including ITAE, ISE, and IAE. The above comparison illustrates that the Model-based NN controller produces an ISE (Integral square error) of 15.77 and the PID controller has an error value of 20.4, then the IAE (integral Absolute Error) value of the NN controller is 10.99 and PID is 17.31. For ITSE (Integral Time Square Error), the NN controller has 818.2 and the PID controller has 1093. The last error comparison is on ITAE (Integral Time Absolute Error) has an error metric of 590.1 and the PID controller value is 945.5. By comparing all the error metrics, the least amount of error is in the NN controller when compared to the PID controller and it is clear from Fig. (7). Table 2. Error metric analysis. Controller/Error

ISE

IAE

ITSE

ITAE

NN Controller

15.77

10.99

818.2

590.1

PID Controller

20.4

17.31

1093

945.5

80 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

Fig. (7). Graphical comparison of the controller using error metrics.

Fig. (8) shows the error response of drug preparation for the SISO system, which has a temperature as input and dry content as output when NN and PID controller is employed.

Fig. (8). Error comparison for NN and PID controller.

6. DISCUSSION The capacity of the NN controller to perform feature engineering on its own distinguishes it from other approaches. In this strategy, an algorithm evaluates the

Drug Preparation

AI and IoT-based Intelligent Health Care & Sanitation 81

input for correlated characteristics, then integrates those to boost faster learning without being explicitly told. PID controllers are complicated due to two factors: tuning techniques for evaluating the PID controller's parameters, which are utilized to regulate the temperature of the evaporator unit. PID tuning demands a greater understanding of computation or systems modeling. The performance of the SISO evaporator is particularly challenging to attain because it is a highly non-linear system. All these difficulties are overcomed by the NN controller, which also gives better performance. CONCLUSION In this paper, the SISO evaporator for drug preparation is considered, and the temperature is given as input and dry matter content as output. The input and output data are collected, and with transfer function for the system is identified. The Model-based NN was taken as acceptable to achieve a better presentation of the system. The NN controller is analysed and then compared with the presentation of the NN controller with the PID controller. The comparison for the controllers is made using error metrics and time-domain specification parameters. In the error metrics comparison, the values of all four error parameters are very low for the NN controller in comparison with the PID controller, and then in time domain analysis, the time domain specification responses are fast to attain stability for the drug preparation system. The time-domain specification improves the performance ratio of the Single effect evaporator system that is used in drug preparation. As a result of the comparison, it is shown that the NN controller has a better performance comparing the PID controller for drug preparation. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

M.A. Roa, and A.A. Vitali, Fruit Juice Concentration and Preservation. Handbook of Food Preservation., M.S. Rahman, Ed., Marcel Dekker: New York, NY, 1999.

[2]

N. Bouhmala, "A multilevel learning automata for MAX-SAT", Int. J. Mach. Learn. Cybern., vol. 6, no. 6, pp. 911-921, 2015.

82 AI and IoT-based Intelligent Health Care & Sanitation

Rani et al.

[http://dx.doi.org/10.1007/s13042-015-0355-4] [3]

G.D. Saravacos, Kluwer, “Handbook of Food Processing Equipment”. Academic/Plenum Publishers: New York, NY, 2002. [http://dx.doi.org/10.1007/978-1-4615-0725-3]

[4]

N. Bouhmala, A. Viken, and J.B. Lønnum, "Enhanced Genetic Algorithm with K-Means for the Clustering Problem", Int. J. Model. Optim, vol. 5, no. 2, pp. 150-154, 2015. [http://dx.doi.org/10.7763/IJMO.2015.V5.452]

[5]

G. Mehta., S. M. Prabhu. D, J. Kantawala, "Industrial wastewater treatment The Indian experience", Indian Assoc. Environ. Management, vol. 22, pp. 276-287, 1995.

[6]

P.G. Smith, Evaporation and Drying, Chapter 12 in Introduction to Food Engineering. Kluwer Academic/Plenum Publishers, New York, NY, 2003.

[7]

K.R. Morison, and R.W. Hartel, Evaporation and Freeze Concentration. Handbook of Food Engineering., D.R. Heldman, D.B. Lund, Eds., 2nd ed. CRC Press: Boca Raton, FL, 2007.

[8]

Hackett, Bryan W, PE.: The Essentials of Continuous Evaporation, Chemical Engineering Progress; New York, Vol. 114, Issue 5, pp. 24-28, 2018.

[9]

Z. Stefanov, and K.A. Hoo, "A distributed-parameter model of black liquor falling film evaporators. Part I. Modeling of a single plate", Ind. Eng. Chem. Res., vol. 42, no. 9, pp. 1925-1937, 2003. [http://dx.doi.org/10.1021/ie020483a]

[10]

J.A. Winchester, and C. Marsh, "Dynamics and control of falling film evaporators with mechanical vapor recompression", Chem. Eng. Res. Des., vol. 77, no. 5, pp. 357-371, 1999. [http://dx.doi.org/10.1205/026387699526340]

[11]

N.T. Russell, H.H.C. Bakker, and R.I. Chaplin, "A comparison of dynamic models for an evaporation process", Chem. Eng. Res. Des., vol. 78, no. 8, pp. 1120-1128, 2000. [http://dx.doi.org/10.1205/026387600528274]

[12]

P. Vaishnavi, K. Sneha, and K.M. Nandhini, "Design And Implementation of Model Predictive Controller for MIMO System", Int. J. Scient. Technol. Res, vol. 8, no. 08, pp. 1843-1846, 2012.

[13]

S. Bolognani, S. Bolognani, L. Peretti, and M. Zigliotto, "Design and Implementation of Model Predictive Control for Electrical Motor Drives", IEEE Trans. Ind. Electron., vol. 56, no. 6, pp. 19251936, 2009. [http://dx.doi.org/10.1109/TIE.2008.2007547]

[14]

S. Mohanty, "Artificial neural network based system identification and model predictive control of a flotation column", J. Process Contr., vol. 19, no. 6, pp. 991-999, 2009. [http://dx.doi.org/10.1016/j.jprocont.2009.01.001]

[15]

H. Demuth, and M. Beale, "NN Toolbox for Use with Matlab", Control Syst. (Tonbridge), pp. 205242, 2002.

[16]

M.G. Forbes, R.S. Patwardhan, H. Hamadah, and R.B. Gopaluni, "Model Predictive Control in Industry: Challenges and Opportunities", IFAC-PapersOnLine, vol. 48, no. 8, pp. 531-538, 2015. [http://dx.doi.org/10.1016/j.ifacol.2015.09.022]

[17]

S.P. Yadav, S. Zaidi, A. Mishra, and V. Yadav, "Survey on Machine Learning in Speech Emotion Recognition and Vision Systems Using a Recurrent Neural Network (RNN)", Arch. Comput. Methods Eng., vol. 29, no. 3, pp. 1753-1770, 2022. [http://dx.doi.org/10.1007/s11831-021-09647-x]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 83-97

83

CHAPTER 6

Machine Learning based Predictive Analysis and Algorithm for Analysis Severity of Breast Cancer B. Radha1,*, Chandra Sekhar Kolli2, K R Prasanna Kumar3, Perumalraja Rengaraju4, S. Kamalesh4 and Ahmed Mateen Buttar5 ICT & Cognitive Systems, Sri Krishna Arts and Science College, Coimbatore, Tamil Nadu, India-641008 2 Aditya College of Engineering and Technology, Surampalem, Andhra Pradesh, India-533291 3 Department of Computer Science & Engineering, Siddaganga Institute of Technology, Tumkur, Karnataka, India-572103 4 Department of Information Technology, Velammal College of Engineering and Technology, Madurai, Tamil Nādu, India-625009 5 Department of Computer Science, University of Agriculture Faisalabad, Faisalabad, Punjab, Pakistan-38000 1

Abstract: Breast cancer is the 2nd frequent occurrence of cancer among women, after skin cancer, according to the American Cancer Society. By using mammography, it is possible to detect breast cancer before it has spread to other parts of the body. It primarily affects females, though males can be affected as well. Early identification of breast cancer improves survival chances significantly, however, the detection procedure remains difficult in clinical studies. To solve this problem, a Machine Learning (ML) algorithm is used to detect breast cancer in mammogram images. In this study, 100 images from the mini-MIAS mammogram database were used, 50 of which were malignant and 50 of which were benign breast cancer mammograms. Before training the model, the sample image datasets are pre-processed using numerous techniques. The required features are then extracted from the sample images using Feature Extraction (FE) techniques, such as Daubechies (DB4) and HAAR. Finally, the extracted features are fed into ML classifiers such as Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Random Forest (RF) to create a model. Several performance metrics are used to evaluate FE and classification. According to the results of the analysis, the HAAR FE with the RF model is the ideal combination, with an accuracy level of 91%.

Keywords: Accuracy, Breast Cancer, Confusion Matrix, Diagnosis, Feature, Metrics. * Corresponding author B. Radha: ICT & Cognitive Systems, Sri Krishna Arts and Science College, Coimbatore, Tamil Nadu, India-641008; E-mail: [email protected]

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

84 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

1. INTRODUCTION Breast cancer is one of the three most frequent types of cancer in the world, and it affects women of all ages. Breast cancer is thought to be curable if detected early [1]. Breast cancer exhibits a broad spectrum of clinical characteristics, with certain tumors steadily growing and having a better prognosis, while some are combative. It is usually characterized by abnormal growth in the breast, which then aggregates and forms lumps or masses. These cells can spread from the breast to the lymphatic system or even other parts of the body. According to scientists, 5-10 percent of the total cases of breast cancer are prompted by a mutated gene or by family heredity [2]. The impacts of global invasive breast cancer are projected to reach an estimation of 3 million cases per year by 2050. These statistics emphasize the shocking pervasiveness of breast cancer and the critical need for treatment and prevention measures. Breast carcinoma is currently classified based on morphological characteristics, with cancerous cells classified into types such as infiltrating ductal carcinoma and a large number of distinct types such as mucinous, tubular, medullary, adenoid cystic carcinoma, and infiltrating lobular carcinoma, among others [3]. According to epidemiological studies, addressing socioeconomic factors is critical in ensuring that almost all females have sufficient access to primary treatment, from monitoring to an effective remedy, and only drastic execution can reduce the global risk of developing breast carcinoma. A computer-assisted diagnostic approach is in greater demand as a result of this pressing issue. The accuracy of an optimized SVM classifier utilized by Ilias Maglogiannis et al. for automated detection of the Wisconsin Diagnostic Breast Cancer datasets is around 97% [4]. Hussain et al. published the findings of a comparison analysis on SVM with various kernels for breast carcinoma detection, concluding that with all features included, the accuracy is 94.6% [5]. Sri Hari Nallamala et al. proposed the detection of breast cancer using three different classification methods such as SVM, Logistic Regression Algorithm, and K-Nearest Neighbor (KNN) Algorithm and concluded with a precision of 98.50% [6]. Furthermore, M. Muthu Rama Krishnan et al. claim that by combining the kernel function with the training-test partition, an SVM-based classifier may achieve an accuracy of 99.385%and 93.726% for two separate datasets [7]. Lal Hussain et al. used different resilient classifiers, like Bayesian strategy, Decision Tree, and SVM kernels, to identify carcinoma mammograms from the subjects [8]. Furthermore, Chelvian Aroef et al. conducted a comparison study for classifying breast cancer using ML models like RF and SVM. According to the experimental result, RF obtained the best acc-

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 85

uracy of 90.90% using 80% training data, while SVM obtained 95.45% using 80% [9]. Saima Rathore et al. proposed a combination feature set comprising conventional statistical features that relied on characteristics and variants of statistical moments and Haralick feature and used an SVM classifier to classify the biopsy images, which resulted in an accuracy of 98.85% [10]. According to Bazazeh et al. analysis, SVMs perform best in terms of accuracy, specificity, and precision, but RF offers the best chance of properly diagnosing cancers [11]. Further Ricvan Dana Nindrea et al. used four main ML algorithms for their study and concluded by saying that the SVM has a good accuracy of 97.13% with high precision and low error when compared to other ML algorithms [12]. Khourdifi et al. did a study on breast cancer detection using 5 different ML algorithms and compared the result and concluded that SVM had proven to give the best performance at several levels compared to other ML algorithms [13]. The journal is systemized as follows: Section 2 discusses the materials used and the procedure used to detect Breast cancer. DB4 and HAAR FE techniques are covered in section 3. SVM, RF, and LDA are among the classification algorithms covered in section 4. Section 5 examines the performance of several FE and classification algorithms in combination. Finally, Section 6 suggests that for Breast cancer diagnosis, it is best to aggregate FE with different classifier methods. 2. MATERIALS AND METHODS The breast mammograms for this study were taken from the mini-MIAS mammography database, which has a total of 100 mammogram images. 50 of the 100 photos are malignant, while the other 50 are benign. ML Algorithms are employed to train and test these mammography images; an 8:2 split is used for the image data, where 80% of it goes to training, and the remaining 20% goes to testing. For training and testing processes, the complete image collection is divided into 8:2 ratios (i.e., 80:20 in this case). The images collected initially are raw RGB images. To analyze and enhance these images further, some preprocessing methods to increase the pixel brightness, filtering, segmenting, resaturation methods, etc., are used. Fig. (1) below illustrates the wavelet-based FE and classification of breast mammography image data for breast cancer in a flow chart.

86 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

Fig. (1). Proposed research flow.

3. FE The vital stage in ML is FE which aims to reduce the total number of features in image data by developing new ones from old ones. Because the acquired dataset is likely to contain a large number of samples, processing it will necessitate a large amount of computing power. This massive amount of data is sorted into

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 87

smaller alliances through the use of FE, and as a result, the required attributes are selected and combined, reducing the volume of the original data. This condensed set articulates the majority of the information in the original dataset. If the acquired data is too large, the FE technique is very functional for reducing the original data without losing any fundamental data. To retrieve useful information from the mammogram breast images, the texture-based extraction procedure is used. DB4 and HAAR methods are used in this research, and it is detailed in the below section. 3.1. DB4 The DB4 technique is an image compression technique that uses DB4 wavelets, and it was found by Ingrid DB4, a Belgian Physicist, and Mathematician. It is a new class of wavelet that is not only orthogonal but could also be implemented using simple filters like Short Digital filters. The wavelets are almost similar to HAAR wavelets, but these are smoother without the jumps, unlike HAAR wavelets. As the resulting representation, such as an image, was a redundant expansion with complex-valued coefficients, the image processing field did not pay much attention to the symmetrical DB4 multi-resolution studies. The actual condition is quite simple and creates a simple authenticity between the Fourier modes. In comparative analysis with normal real DB4 analyses, the current study investigates the type of connection between complex wavelet coefficients, demonstrating the DB4 technique's natural redundancy [14]. The Multiresolution analysis is used in the DB4 technique to find the sequence of the closed subspaces. 3.2. HAAR The Haar wavelet is a series of rescaled square-shaped functions, equivalent to Fourier analysis. Alfred Haar postulated it, and they are similar to convolutional kernels. Haar technique can detect edges quite efficiently and also can detect lines efficiently. Haar technique can represent most of the relevant features. The Haar transform has a lot of advantages in terms of computational and memory requirements, therefore, it's worthwhile considering. The least number of required arithmetic operations is another benefit of the Harr spectrum in this and comparable applications. The method is used as an alternative for checking the equivalence of functions that are difficult to fully represent [15]. The Haar technique is implemented by adding all the intensities of white pixels and the value of some of the intensities of the black pixels. The difference between the sum of the white and black pixel intensities determines the optimal intensity of the Haar feature to be implemented.

88 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

4. CLASSIFICATION TECHNIQUES To distinguish between malignant and benign breast cancer mammographic images, several classification algorithms are applied once the essential information is extracted using FE methods [16]. In the section below, the classification algorithms used in this study are described in depth. 4.1. SVM SVM is regarded as a cutting-edge classification technique due to its reliability and efficiency. Currently, it is considered one of the most well-known classification algorithms, offering computational advantages over its contenders. It is a feasible technique that requires only a few examples to train and is unchanged by the number of dimensions. In addition, effective methods for training SVM are being developed at a rapid pace. The ultimate objective of SVM is to find the ideal classification function in the training data for differentiating between members of two classes. In a linearly separable image dataset, the hyperplane f(x) passes between two classes, malignant and benign cancer in this case, and is kept separate using the linear classification function. Since there are so many hyperplanes, the primary objective is to choose one with the maximum margin or proximity between data points from both classes. Attempting to increase the margin distance provides reinforcement, making resulting data points easier to classify. Only very few hyperplanes can be labeled as SVM solutions, even though there is an endless number of them. These models perform the best in terms of generalization, and they not only provide the greatest classifier outcomes but also provide a great deal of flexibility in terms of categorizing newly acquired data [17]. The following is the formula for calculating the final classifier:

𝑓(𝑥) =∑𝛼𝑖 𝑥𝑖𝑇 𝑥 + 𝑏𝑚

(1)

The main challenge with this categorization method is the dispersion of data, which is lengthy and makes linear separation difficult. SVM sets up the kernel function, which is a technique used to produce data as input and convert it into the feature needed for processing. The Gaussian RBF Kernel Function, a radial-based function, is used in this work [18]. This Kernel function is typically used when there is no previous experience with the data. The Gaussian RBF Kernel Function is denoted by the following equation:

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 89

𝑒𝑥𝑝(−𝛾||𝑥𝑖 − 𝑥𝑗 ||2 𝐾(𝑥𝑖 , 𝑥𝑗 ) = 2𝜎 2

(2)

RBF is a universal kernel function that performs well on smaller datasets, which is ideal for our research. 4.2. RF The RF is made up of several tree classifiers, and these classifiers are obtained from a random vector which is acquired individually from the input vector [19]. In this technique, the training set for each classifier is generated with N samples randomly and replacing them with the N number of the primary training set. The learning modal generates a classifier from the sample and then combines all the classifiers from each trial to generate the final classifier. Every classifier records a vote for the class to which an example belongs. If many classes get the greatest number of votes, the final classifier is chosen at random. There are many methodologies for extracting the features for the decision tree, and the majority of them effectively allocate a quality evaluation to the feature. In a decision tree, one of the most popularly used FE metrics is the Gain Ratio criterion and the Gini Index. The Gini Index is used as a feature screening method by the RF classifier, which analyses the imperfection of a characteristic of the classes. The Gini index is stated as follows for a given training dataset T: for a specified training set T, choosing a case (pixel) at random that belongs to a class Ci. The Gini Index is represented by the below equation: ∑ ∑ 𝑓(𝐶𝑖 , 𝑇)/|𝑇|) (𝑓(𝐶𝑗 , 𝑇)/|𝑇|) 𝑗≠𝑖

(3)

where f(Ci,T)/|T|) is a representation of the degree of randomness inherent in the selected instance of class Ci. RF classifiers have the following benefit over decision tree methods. 4.3. LDA LDA is a fundamental data analysis technique used to identify subsample lower dimension, in comparison with the actual data dimension, in which the fundamental problem's data points are separable. Separability is a way of measuring mean value and variance. The most significant aspect of LDA is that

90 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

the outcomes can be gained quickly by finding solutions to the basic eigenvalue system [20]. LDA helps greatly when the number of input parameters expands to the point where a computational modeling function becomes challenging. The LDA method searches the space for matrices that can demonstrate standards between classes. Decision analysis assists in learning about the class classifier Pr(G|Y=y). Assume fk(y)is the class density of Y in the class H=k [21]. Then the application of the Bayes theorem is given by the following equation:

Pr(𝐻 = 𝑘|𝑌 = 𝑦)

𝑓𝑘 (𝑦)𝜋𝑘 ∑𝑘 𝑙=1 𝑓𝑙 (𝑦)𝜋𝑙

(4)

If each implied density class is represented by a multinomial Gaussian model, then:

𝑓𝑥 (𝑦) =

1 1/2 (2𝜋)𝑝/2 || ∑−1 𝑘 ||

1

exp{- (y-𝜇𝑘 ) ∑−1 𝑘 ( 𝑦 − 𝜇𝑘 )} 2

(5)

LDA creates prediction by estimating the possibility that a particular set of inputs falls into each of the classes. The resulting class is the one with the greatest chance, and a prognosis is produced. 5. RESULT AND DISCUSSION A total of 100 mammography pictures of breasts were gathered from the miniMIAS mammogram database. The dataset is separated into two sections, one for training the classifier model and the other for testing the classifier model in an 8:2 ratio. The performance measures are used to evaluate the FE and the ML model. Accuracy, Precision, True/False Positive and Negative Rate are the performance metrics employed in this study. The performance metrics are determined using the confusion matrix items, which are discussed below with the help of Fig. (2). The model predicted the image as cancer; originally, the image was also cancer, then represented as True Positive in the confusion matrix. Next, the model prediction and actual image are healthy; it is called True Negative. The actual image is cancer, and the model prediction is healthy; it is called a False Negative, and vice versa, it is called a False Positive.

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 91

Fig. (2). Confusion matrix.

The percentage of correct predictions for the test data is known as accuracy. It is determined by dividing the total true prediction by the number of the total test sample, as shown in the equation below: Accuracy=

𝐶𝑜𝑟𝑟𝑒𝑐𝑡 𝑃𝑟𝑒𝑑𝑖 𝑇𝑜𝑡𝑎𝑙 𝑃𝑟𝑒𝑑𝑖𝑐𝑖𝑡𝑖𝑜

(6)

Precision can be calculated using the below formula. It can estimate by the ratio of truly predicted as cancer to the predicted as cancer which contains both true and false cancer predictions. Precision=

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒

(7)

True Positive Rate is the measure of the correct prediction of cancer from the total prediction of cancer. The formula used to calculate the True Positive Rate is as follows. True Positive Rate=

்௥௨௘௉௢௦௜௧௜௩௘ ்௥௨௘௉௢௦௜௧௜௩௘ାி௔௟௦௘ே௘௚௔௧௜௩௘

(8)

92 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

True Negative Rate is the measure of the correct prediction of healthy images from the total prediction of healthy images. The formula used to calculate the True Negative Rate is as follows. True ܰ݁݃ܽ‫ ݁ݒ݅ݐ‬Rate=

்௥௨௘ே௘௚௔௧௜௩௘

(9)

ி௔௟௦௘௉௢௦௜௧௜௩௘ା்௥௨௘ே௘௚௔௧௜௩௘

False Positive Rate is the measure of incorrect prediction of cancer from the total of correct healthy image prediction and wrong cancer image prediction. The formula used to calculate the False Positive Rate is as follows. False Positive Rate=

ி௔௟௦௘௉௢௦௜௧௜௩௘

(10)

்௥௨௘ே௘௚௔௧௜௩௘ାி௔௟௦௘௉௢௦௜௧௜௩௘

False Negative Rate is the measure of incorrect prediction of healthy images from the total of correct cancer image prediction and wrong healthy image prediction. The formula used to identify the False Negative Rate is given below. ி௔௟௦௘ே௘௚௔௧௜௩௘

False Negative Rate =

(11)

்௥௨௘௉௢௦௜௧௜௩௘ାி௔௟௦௘ே௘௚௔௧௜௩௘

These performance measures are used to determine which FE and classification model is the most effective for detecting breast cancer and healthy breasts using a mammographic image. Several ML algorithms, such as RF, SVM, and LDA, are assessed using the DB4 FE approach, as exposed in Table 1. The table shows that RF has an accuracy rate of approximately 85%, which is extremely high in this case. LDA has a lower accuracy level of 80%. RF and LDA give high and low precision, with 84% and 78%, respectively. Fig. (3) displays a visual depiction of the performance analysis of DB4 when it is combined with different classifiers to detect breast cancer. Table 1. Performance analysis of breast cancer prediction using DB4 and the ML model. Model\ Metrics

SVM

RF

LDA

TP

41

42

39

TN

42

43

41

FP

9

8

11

FN

8

7

9

ACCURACY

83

85

80

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 93

(Table 1) cont.....

Model\ Metrics

SVM

RF

LDA

PRECISION

82

84

78

TNR

82.35294

84.31373

78.84615

TPR

83.67347

85.71429

81.25

FNR

16.32653

14.28571

18.75

FPR

17.64706

15.68627

21.15385

Fig. (3). Performance evaluation of DB4 with ML model.

Table 2 shows the results of a comparison of the classifiers RF, SVM, and LDA after the HAAR FE method. The accuracy rate of RF is 91%, as shown in Table 2, and it has a high level of accuracy. For LDA, the accuracy level is only 87%, which is low by comparison. Precisions such as 86.4%, 85.4%, and 90.3%, are achieved using SVM, LDA, and RF, respectively. The table is converted into a graphical format for better visualization and easy understanding, which is shown in Fig. (4).

94 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

Table 2. Performance analysis of breast cancer prediction using HAAR and the ML model. Model\ Metrics

SVM

RF

LDA

TP

45

47

41

TN

43

44

46

FP

7

5

7

FN

5

4

6

ACCURACY

88

91

87

PRECISION

86.53846

90.38462

85.41667

TNR

86

89.79592

86.79245

TPR

90

92.15686

87.23404

FNR

10

7.843137

12.76596

FPR

14

10.20408

13.20755

Fig. (4). Performance evaluation of HAAR with ML model.

6. GAPS FILLED The identification of algorithms with the highest accuracy which is used to predict breast cancer is discussed. Being a sensitive issue, the prediction of breast cancer has to be as high as possible to decrease the number of causalities. The best FE and ML classifiers are identified separately. The best combination of wavelet features with ML for breast cancer is not discussed so far. The experimental

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 95

results in Section 6 help to identify that the HAAR with RF gives good prediction results. The model developed using this study can be helpful in the successful prediction of breast cancer. CONCLUSION The number of women diagnosed with breast cancer continues to rise daily. Generally speaking, the sooner an issue is identified and treated, the greater the likelihood of a complete recovery is. The MIAS mammography database is used to collect and process the dataset, which is a collection of mammographic images. Raw images are pre-processed to increase the categorization accuracy. And it is possible to extract meaningful information from an image using pre-processed data since FE may be performed on it. To identify breast cancer based on feature data received from a database, a variety of classifier models are used. Evaluation of the model and FE techniques are evaluated through the use of performance measurements. The RF classifier with the HAAR FE delivers a high accuracy of 91% with a precision of 90.3%, according to the data analysis. The SVM with HAAR then provides a second high level of accuracy of 88%. The LDA classifier with HAAR FE algorithms achieves the third high accuracy level. This analysis aids in determining the best FE and classification model for breast cancer detection. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

M. Baldonado, C.C.K. Chang, L. Gravano, and A. Paepcke, "The Stanford Digital Library metadata architecture", Int. J. Digit. Libr., vol. 1, no. 2, pp. 108-121, 1997. [http://dx.doi.org/10.1007/s007990050008]

[2]

K.B. Bruce, L. Cardelli, and B.C. Pierce, Comparing Object Encodings. Theoretical Aspects of Computer Software., M. Abadi, T. Ito, Eds., vol. 1281. Springer-Verlag: Berlin, Heidelberg, New York, 1997, pp. 415-438. Lecture Notes in Computer Science. [http://dx.doi.org/10.1007/BFb0014561]

[3]

Computer Science Today. Recent Trends and Developments. vol. 1000. Springer-Verlag: Berlin, Heidelberg, New York, 1995. J. Van Leeuwen, Ed., Lecture Notes in Computer Science.

96 AI and IoT-based Intelligent Health Care & Sanitation

Radha et al.

[4]

I. Maglogiannis, E. Zafiropoulos, and I. Anagnostopoulos, "An intelligent system for automated breast cancer diagnosis and prognosis using SVM based classifiers", Appl. Intell., vol. 30, no. 1, pp. 24-36, 2009. [http://dx.doi.org/10.1007/s10489-007-0073-z]

[5]

M. Hussain, S.K. Wajid, A. Elzaart, and M. Berbar, "A Comparison of SVM Kernel Functions for Breast Cancer Detection", In: 2011 Eighth International Conference Computer Graphics. Singapore, 2011. [http://dx.doi.org/10.1109/CGIV.2011.31]

[6]

S.H. Nallamala, P. Mishra, and S.V. Koneru, "Breast Cancer Detection using ML Way", Int J Recent Technol Eng, vol. 8, p. 2S3, 2019.

[7]

M.M.R. Krishnan, S Banerjee, and C Chakraborty, " Statistical analysis of mammographic features and its classification using SVM", In: Expert Syst Appl. vol. 37. , 2010, no. 1, pp. 470-478. [http://dx.doi.org/10.11648/j.wcmc.20190701.11]

[8]

L. Hussain, W. Aziz, S. Saeed, S. Rathore, and M. Rafique, “Automated Breast Cancer Detection Using ML Techniques by Extracting Different Feature Extracting Strategies.” 17th IEEE International Conference on Trust, Security and Privacy In Computing And Communications/ 12th IEEE International Conference On Big Data Science And Engineering (TrustCom/BigDataSE), pp. 327-331, 2018. [http://dx.doi.org/10.1109/TrustCom/BigDataSE.2018.00057]

[9]

C. Aroef, Y. Rivan, and Z. Rustam, "Comparing RF and SVMs for breast cancer classification", Telkomnika (Telecommunication Computing Electronics and Control), vol. 18, no. 2, pp. 815-821, 2020. [http://dx.doi.org/10.12928/telkomnika.v18i2.14785]

[10]

S. Rathore, M. Hussain, A. Ali, and A. Khan, "A recent survey on colon cancer detection techniques", IEEE/ACM Trans. Comput. Biol. Bioinformatics, vol. 10, no. 3, pp. 545-563, 2013. [http://dx.doi.org/10.1109/TCBB.2013.84] [PMID: 24091390]

[11]

D. Bazazeh, and R. Shubair, “Comparative study of ML algorithms for breast cancer detection and diagnosis”; 5th International Conference on Electronic Devices, Systems and Applications, pp. 1-4, 2016. [http://dx.doi.org/10.1109/ICEDSA.2016.7818560]

[12]

R.D. Nindrea, T. Aryandono, L. Lazuardi, and I. Dwiprahasto, "Diagnostic Accuracy of Different Machine Learning Algorithms for Breast Cancer Risk Calculation: a Meta-Analysis", Asian Pac. J. Cancer Prev., vol. 19, no. 7, pp. 1747-1752, 2018. [http://dx.doi.org/10.1016/j.tranon.2020.100907] [PMID: 30049182]

[13]

Y. Khourdifi, and M. Bahaj, "Applying Best ML Algorithms for Breast Cancer Prediction and Classification", In: 2018 International Conference on Electronics, Control, Optimization, and Computer Science (ICECOCS), 2018, pp. 1-5. [http://dx.doi.org/10.1109/ICECOCS.2018.8610632]

[14]

J.M. Lina, "Image Processing with Complex DB4 Wavelets", J. Math. Imaging Vis., vol. 7, no. 3, pp. 211-223, 1997. [http://dx.doi.org/10.1023/A:1008274210946]

[15]

Radomir S. Stanković; Bogdan J. Falkowski.: “The Haar wavelet transform its status and achievements.”; Vol. 29, Issue. 1, pp. 25–44, 2003. [http://dx.doi.org/10.1016/S0045-7906(01)00011-8]

[16]

Yadav, Satya & Mahato, Dharmendra Prasad & Linh, Nguyen, “Distributed Artificial Intelligence: A Modern Approach”. 2020. [http://dx.doi.org/10.13140/RG.2.2.19776.69128]

[17]

X. Wu, V. Kumar, J. Ross Quinlan, J. Ghosh, Q. Yang, H. Motoda, G.J. McLachlan, A. Ng, B. Liu,

Breast Cancer

AI and IoT-based Intelligent Health Care & Sanitation 97

P.S. Yu, Z-H. Zhou, M. Steinbach, D.J. Hand, and D. Steinberg, "Top 10 algorithms in data mining", Knowl. Inf. Syst., vol. 14, no. 1, pp. 1-37, 2008. [http://dx.doi.org/10.1007/s10115-007-0114-2] [18]

H. Rajaguru, K. Ganesan, and V.K. Bojan, "Earlier detection of cancer regions from MR image features and SVM classifiers", Int. J. Imaging Syst. Technol., vol. 26, no. 3, pp. 196-208, 2016. [http://dx.doi.org/10.1002/ima.22177]

[19]

M. Pal, "Random forest classifier for remote sensing classification", Int. J. Remote Sens., vol. 26, no. 1, pp. 217-222, 2005. [http://dx.doi.org/10.1080/01431160412331269698]

[20]

P. Xanthopoulos, P.M. Pardalos, and T.B. Trafalis, LDA.Robust Data Mining.”; SpringerBriefs in Optimization. Springer: New York, 2013. [http://dx.doi.org/10.1007/978-1-4419-9878-1]

[21]

H. Rajaguru., S. Kumar Prabhakar., “Time-Frequency Analysis (dB2 and dB4) for Epilepsy Classification with LDA Classifier.”; 2017 2nd International Conference on Communication and Electronics Systems (ICCES), pp. 708-711, 2017. [http://dx.doi.org/10.4236/cn.2020.124009]

98

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 98-113

CHAPTER 7

Glaucoma Detection Using Retinal Fundus Image by Employing Deep Learning Algorithm K.T. Ilayarajaa1, M. Sugadev1, Shantala Devi Patil1, V. Vani2, H. Roopa2 and Sachin Kumar2,* Department of Electronics and Communication Engineering, Sathyabama Institute of Science and Technology, Chennai, Tamilnadu, India-600119 2 Information Science and Engineering, Bangalore Institute of Technology, Bangalore, Karnataka, India-560079 1

Abstract: Glaucoma is an eye disease that can result in permanent blindness if not detected and treated in the early stages of the disease. The worst part of Glaucoma is that it does not come up with a lot of visible symptoms, instead, it can go from the preliminary stage to a serious issue quickly. A Deep Learning (DL) model is capable of detecting the presence of Glaucoma by analyzing the image of the retina which is uploaded by the user. In this research, two DL algorithms were used to detect the presence of Glaucoma in the uploaded image. The DL algorithms include the convolutional neural network or the CNN and the transfer learning algorithm. The transfer learning algorithm is implemented by the VGG-19 model. Once two DL models were developed using the above-mentioned algorithms, the models were trained and tested using the images of the retina. The trained models are tested to find the better model. The efficiency of the model is measured based on some metrics. These metrics include the True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). Using these metrics, the true positive rate, the true negative rate, the false-positive rate, and the false-negative rate are calculated. From the above values, the DL algorithm, which is more efficient than the other one in identifying Glaucoma, can be found.

Keywords: CNN, DL Algorithm, Glaucoma, Performance Metrics, VGG-19. 1. INTRODUCTION Glaucoma is an eye condition that is caused due damage to the optic nerve in the eye. This results in the development of a blind spot due to the elevated blood pressure in the eyes and even results in permanent blindness. There are different Corresponding author Sachin Kumar: Information Science and Engineering, Bangalore Institute of Technology, Bangalore, Karnataka, India-560079; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 99

types of glaucoma, such as open-angle glaucoma, angle-closure glaucoma, normal-tension glaucoma, pigmentary glaucoma, etc. There are not a lot of preventive measures present for glaucoma as it can be considered a hereditary disease. In India, about 12 million people whose age is above 40 were diagnosed to be affected by glaucoma [1]. Sometimes it becomes tougher for glaucoma-affected people to live a normal life like others. One of the most important and dreadful facts about glaucoma is that it cannot be diagnosed in its initial stages due to the lack of any visible symptoms [2]. Hence, it becomes necessary for people who are born in a family with the traits of glaucoma to undergo some simple processes for the detection of the presence of glaucoma to prevent themselves from being affected by glaucoma in the premature stage itself. This research works in the development of two DL methods using two algorithms named the transfer learning algorithm and the CNN algorithm [3]. Both the models are then trained and validated to achieve higher accuracy and lower loss level. Once the models are trained, they are tested to find the most effective algorithm that can be used for glaucoma detection. By evaluating the test results, the best DL algorithm for the detection of glaucoma can be determined. 2. LITERATURE SURVEY Glaucoma is one of the most threatening eye diseases in the world; many researchers have used various techniques to reduce the worst consequences caused by glaucoma. These techniques include both earlier detection and advanced treatment methods. The researchers of the Federal University of Brazil researched glaucoma detection using machine learning algorithms. The call detail record metric is the most significant information in the detection of glaucoma, according to the methodologies they utilized, which included certain machine learning approaches. Furthermore, all of the studies found that adopting an automated system for ocular structural evaluations has the benefit of reducing variance in medical competence assessments. They concluded that DL is one of the most efficient algorithms in terms of computer vision and processing of retinal fundus images. They also said that one downside of DL is that it requires an adequate outcome to create a large database and fast response time [4]. Tomaz Ribeiro Viana Bisneto et al. of Brazil discovered the usage of a Generative adversarial network in the detection of glaucoma. The findings they obtained indicated the promise of texture extraction approaches since the suggested taxonomic diversity indices proved to be effective in OD region characterization. They got positive findings, with substantial values in all outcome indicators and a 100% success rate using several classifiers. Their discovery might help professionals with auxiliary

100 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

analyses during glaucoma identification, ensuring a better quality of life for patients and a better prognosis for glaucoma therapy [5]. Soheila Gheisari et al. from the University of Technology, Australia, came up with the idea of combining the CNN algorithm and the RCNN in the detection of glaucoma with higher efficiency. They claimed that by using a retinal picture sequence to train and validate a combined CNN and RNN, they were able to extract both spatial and temporal data, resulting in a much better glaucoma diagnosis. In differentiating glaucoma from healthy eyes, the VGG19 + LSTM combination outperformed all other networks, with all the parameters like sensitivity, specificity, and F-measure above ninety-five percent. In the end, they found that Deep CNNs were an effective AI-based technique for finding clinically relevant characteristics in retinal fundus pictures [6]. CNN is not a new technology that is used for object detection. Yanan Sun et al. of the Sichuan University, China designed an automatic CNN model which can be used for image classification. An new encoding strategy was developed by them to encode the in-depth layers of the CNNing algorithms, incorporating skip connections to encourage deeper CNN algorithms to be produced during evolution, and the construction of the new encoding technique of CNN algorithms, all while working with limited computational resources. On two difficult benchmark datasets, the suggested approach was compared against numerous current CNN models, including 8 manually produced CNNs, 6 mechanized + manually adjusted CNN algorithms, and four mechanical algorithms that found CNN structures [7]. Like the CNN algorithm, the transfer learning algorithm can also be used for object detection. Taranjit Kaur et al. from the Indian Institute of Technology, India, used CNN in the classification of the images of the brain. They evaluated a lot of models to find the one with the results. They concluded that among all the models they tested, the Alexnet model performed the best, with classification results of 100 percent, 94 percent, and 95.92 percent for all three different datasets. The results they achieved are relatively higher than that of the many existing image classification algorithms. They also stated that in the current DL-based studies, the transferred Alexnet model achieves greater performance metrics [8]. 3. MATERIALS AND METHODS A dataset consisting of 705 retinal fundus images is collected from the ACRIMA database. Out of the 705 images, 396 images contain a retina with glaucoma infection and 309 normal images of the retina. The images are then divided into three parts for three different purposes. 493 images from the total dataset are used to train both DL models. Another 141 samples of the retina are used to validate the models. The final 71 models are used to test the efficiency of the models. Two DL models were developed using transfer learning and the CNN algorithm. The transfer learning algorithm is implemented using the VGG-19 architecture. Once

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 101

the models are developed, they are trained, validated, and tested. The test results are then analyzed based on some performance metrics. These metrics include the true positive, true negative, false positive, and false negative values of the models. The below diagram explains the steps involved in the prediction of glaucoma using the DL algorithms pictographically. The workflow of the research is shown in Fig. (1).

Fig. (1). Block diagram of Glaucoma Detection.

4. PREPROCESSING TECHNIQUES The images obtained from the ACRIMA database are used in almost every step of Glaucoma detection. Hence, the images must be perfect in every aspect to develop the model with higher efficiency. To achieve this, certain pre-processing techniques can be used. These techniques include image resizing and image augmentation. ●

●

The process of adjusting the dimensions of all the photos in a dataset so that they are all similar dimensions is known as image reshaping. When reshaping an image, the total number of pixels can be raised or lowered. Image reshaping ensures that image quality is maintained when an image is zoomed or enlarged. Image augmentation is important. In this technique, the existing image is taken,

102 AI and IoT-based Intelligent Health Care & Sanitation

●

Ilayarajaa et al.

and geometrical transformations like flip, zoom, rotate, etc. for making the model train on various data. The processed images are then divided into three parts. Both DL models are trained using 493 photos from the complete dataset. A total of 141 retinal samples are utilized to verify the models. The final 71 models are used to assess the models' efficacy.

5. DL MODEL A DL model is a computer-generated classification system for photos, texts, and even sounds. Various DL methods are employed to create solutions for diagnosing the disease. CNN's, radial basis function networks, recurrent neural networks, transfer learning, and other algorithms are among them [9, 10]. The DL model in this study is developed using two algorithms: CNNs and the transfer learning method. The transfer learning method is implemented using the VGG-19 architecture. Both the DL algorithms are more clearly discussed further. 5.1. Transfer Learning (VGG-19) The VGG-19 architecture is used to construct the transfer learning algorithm. The Oxford Net architecture, also known as the VGG 19 architecture, is a DL architecture that uses a DL algorithm to identify pictures. This design is used in CNNing systems for image categorization. The VGG-19 model includes 19 layers, 13 of which are convolutional and three of which are fully connected. The only two levels that may be modified to match the user's demands are the topmost and bottommost layers [11]. Transfer learning is a strategy or approach to train models, not a specific machine learning algorithm. The knowledge gained during earlier training is reused to aid in the completion of a new activity. VGG-19 offers several benefits, including the notion that it is a pre-trained model that does not require extensive training, testing, and validation [12]. VGG is a network type that may be classified and tracked. There are 41 layers in the VGG-19 network. There are 19 learningweighted layers, 13 of which are collaborative and supportive and three of which are completely connected [13]. Fig. (2) consists of the architecture of the VGG-19 architecture when it is completely trained based on the requirement. In this case, the requirement will be the detection of the presence of glaucoma from the retinal fundal images.

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 103

Fig. (2). VGG-19 Architecture.

Each convolutional block is utilized to train the model, as shown in the above Fig. The developed model may be used to detect the presence of glaucoma in a picture that will be tested or verified. 5.2. CNN The CNN is a DL system that can take an image as input, assign meaning to distinct objects or elements of the picture, and then discriminate between them. When compared to other DL-based classification algorithms, one of the key advantages of the CNN approach is that it does not require much preprocessing [14]. The design of the CNN algorithm is based on the organization of the Visual Cortex and is comparable to the connective theory of neurons in the human brain.

104 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

The CNN technique may also conduct image classification in several convolutional layers, reducing the number of DL models required to complete a single job. The CNN algorithm is claimed to be one of the finest algorithms for extracting high-level properties from a picture, such as edges [15]. The sequential order which was followed during the construction of the CNN model for the process of glaucoma detection is shown in Fig. (3).

Fig. (3). CNN Architecture.

In the above Figure, it can be seen that there are four layers: an activation layer, pooling layer, a convolutional layer, and a fully connected layer, all of which are coupled so that CNNs may process and interpret input to categorize pictures. The photographs are then analyzed to check whether there are any symptoms of plant disease. 6. RESULT ANALYSIS Two DL models were developed using the transfer learning algorithm and the CNN algorithm. The transfer learning algorithm is implemented using the VGG19 architecture. Once the models are developed, they are trained, validated, and tested. The test results are then analyzed based on some performance metrics. The performance metrics include the true positive, true negative, false positive, and false negative values of the models. This step is mandatory to compare the accuracy and loss value of both algorithms. When glaucoma is present in a picture, and the model detects it properly, the real output is shown as positive. True Negative (TN): When glaucoma is not present in

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 105

a picture, and the model detects it accurately, the outcome is shown as negative. False Positive (FP): When glaucoma is present in a picture, and the model fails to detect it, the real outcome is shown as positive. If glaucoma is not present in a picture and the model detects it incorrectly, the real output is shown as negative. Based on the above performance metrics, the accuracy and the loss value of both algorithms are tabulated. Table 1 contains the accuracy value and the loss value of the transfer learning algorithm during training. Table 1. Training accuracy and loss of transfer learning algorithm. Epoch

Training Accuracy

Training Loss

0

0.8698630332946777

0.28365132212638855

1

0.8972602486610413

0.26046496629714966

2

0.9212328791918347

0.20296837389469147

3

0.9190959124565125

0.2252890199620514

4

0.9006849527359009

0.2259654998779297

5

0.9452054500579834

0.18439199030399323

6

0.9503424763679504

0.17865334451198578

7

0.9400684833526611

0.17603550851345062

8

0.9417808055877686

0.1971869158744812

9

0.9571917653083801

0.15202458202838898

10

0.9589040875434875

0.1455105096101761

11

0.965753436088562

0.14151352643966675

12

0.9417808055877686

0.14090575277805328

13

0.9452054500579834

0.15004700422286987

14

0.9520547986030579

0.1367044299840927

The training accuracy of the transfer learning method is 86 percent in the first epoch, as shown in the table above, and the accuracy increases as the number of epochs increases. The transfer learning algorithm's accuracy has grown to 95% after the fourteenth period. It can be seen that the algorithm's loss at the first epoch is about equal to 28%, and it continues to decrease as the epochs rise. Finally, the transfer algorithm suffers a 13 percent loss. The tabulated values are explained graphically in Fig. (4).

106 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

Fig. (4). Training accuracy and loss of VGG-19.

Table 2 contains the accuracy value and the loss value of the transfer learning algorithm during the validation. Table 2. Validation accuracy and loss of transfer learning algorithm. Epoch

Validation Accuracy

Validation Loss

0

0.96875

0.19947955012321472

1

0.96875

0.1885835826396942

2

0.96875

0.1722395122051239

3

0.921875

0.20042726397514343

4

0.953125

0.19374321281909943

5

0.96875

0.12109236419200897

6

0.96875

0.1255747675895691

7

0.921875

0.1390729546546936

8

0.984375

0.10914437472820282

9

0.984375

0.10123242437839508

10

0.984375

0.1156655102968219

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 107

(Table 2) cont.....

Epoch

Validation Accuracy

Validation Loss

11

0.953125

0.11944677817821503

12

0.984375

0.09958772361278534

13

0.984375

0.10017640888690948

14

0.984375

0.09955428540706635

As indicated in the table above, the transfer learning method's training accuracy is 96 percent in the first epoch, and it improves as the number of epochs grows. At the end of the fourteenth epoch, the accuracy of the transfer learning algorithm had increased to 98%. It's also worth noting that the algorithm's loss in the first epoch is roughly equivalent to 19%, and it gradually decreases as the number of epochs grows. In the end, the transfer algorithm loses 9% of the time. The loss value is so small that it can be neglected. In Fig. (5), the tabular numbers are visually explained.

Fig. (5). Validation accuracy and loss of transfer learning.

108 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

Table 3 contains the accuracy value and the loss value of the CNN algorithm during training. Table 3. Training accuracy and loss of CNN algorithm. Epoch

Validation Accuracy

Validation Loss

0

0.6304348111152649

0.7321705222129822

1

0.8224637508392334

0.41351139545440674

2

0.8768119235733032

0.3125768005847931

3

0.8713768124580383

0.31538867950439453

4

0.9196666865348819

0.20803412795066833

5

0.9149305820465088

0.20139507949352264

6

0.9196666865348819

0.20246319492557526

7

0.9347826242446899

0.19619191286830902

8

0.9420289993286133

0.15422913432121277

9

0.9196666865348819

0.22149819155433655

10

0.9275362491907666

0.17931151390075684

11

0.9402173757553101

0.15204143524199922

12

0.9347826242446899

0.15529650449752808

13

0.9655796885490417

0.10744067281484604

14

0.95652174949646

0.12333853542804718

Fig. (6). Training accuracy and loss of CNN.

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 109

The training accuracy of the transfer learning method is 63 percent in the first epoch, as shown in the table above, and the accuracy increases as the number of epochs increases. The transfer learning algorithm's accuracy has grown to 95% after the fourteenth period. It can also be observed that the algorithm's loss at the first epoch is about equal to 73%, and it continues to decrease as the number of epochs increases. In the end, the transfer algorithm suffers a 12 percent loss. Though it may seem like the CNN algorithm is less efficient than the transfer learning algorithm, it is important to notice the drop in the loss value during every epoch. This fact makes CNN one of the best DL algorithms for glaucoma detection. The values tabulated are shown as a graph in Fig. (6). Table 4 contains the validation results of the CNN algorithm. Table 4. Validation accuracy and loss of CNN algorithm. Epoch

Validation Accuracy

Validation Loss

0

0.8020833134651184

0.47919759154319763

1

0.8958333134651184

0.28837230801582336

2

0.8645833134651184

0.3435284197330475

3

0.90625

0.19661729037761988

4

0.9444444179534912

0.14713825285434723

5

0.9196666865348819

0.21173448860645294

6

0.8645833134651184

0.2580515444278717

7

0.9375

0.14425787329673767

8

0.9375

0.2010868787765503

9

0.9479196865348819

0.19525928676128387

10

0.9444444179534912

0.12180391999075699

11

0.9196666865348819

0.11969588088989258

12

0.9583333134651184

0.10521870851519724

13

0.9895833134651184

0.06276941299438477

14

0.96875

0.09335782378911972

As shown in the table above, the transfer learning method's validation accuracy is 80 percent in the first epoch, and it improves as the number of epochs grows. At the end of the fourteenth epoch, the accuracy of the transfer learning algorithm had increased to 96%. It's also worth noting that the algorithm's loss in the first epoch is roughly equivalent to 47%, and it gradually decreases as the number of epochs grows. In the end, the transfer algorithm loses 9% of the time. The loss

110 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

value is so small that it can be neglected. Again, the CNN algorithm is the best in terms of little loss values. The values tabulated are shown as a graph in Fig. (7).

Fig. (7). Validation accuracy and loss of CNN.

However, it is mandatory to analyze the performance of both algorithms using the performance index. The 71-glaucoma sample is used to test the designed algorithm A confusion matrix is generated based on the values obtained from both algorithms, and the results are tabulated in Table 5 below. Table 5. Confusion matrix of CNN. Model/Confusion Matrix

TP

TN

FP

FN

CNN

34

30

4

3

VGG-19

33

32

2

4

The VGG-19 and CNN are compared in terms of performance. Consequently, the true positive, true negative, false positive, and false-negative rates have all been estimated using the data. Table 6 shows that the transfer learning approach outperforms the CNN algorithm when dealing with true rates. As can be seen in Fig. (8), the metrics that were calculated are represented in a graphical format.

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 111

Table 6. Metrics of CNN. Model/Metrics

TNR

TPR

FNR

FPR

CNN

88.23529

91.89189

8.108108

11.76471

VGG-19

94.11765

89.18919

10.81081

5.882353

Fig. (8). Performance Comparison.

7. DISCUSSION The study aims to attain maximum accuracy in the DL algorithm. As glaucoma is such a delicate issue, the accuracy of glaucoma prediction must be as high as feasible to reduce the number of causalities. Two DL models were developed using transfer learning and the CNN algorithm. The transfer learning algorithm is implemented using the VGG-19 architecture. Once the models are developed, they are trained, validated, and tested. The test results are then analyzed based on some performance metrics. The correct prediction rate attained by both models is above 90%, the model established in this research can be extremely useful in glaucoma prediction. CONCLUSION The ACRIMA database provided a dataset with 705 photographs of the retinal fundus images, including 396 glaucomatous and 309 normal images. After that, image augmentation is used to preprocess the photos. Picture rescaling, image

112 AI and IoT-based Intelligent Health Care & Sanitation

Ilayarajaa et al.

sharing, image zooming, and horizontal flipping are all part of this image augmentation approach. Two separate DL algorithms were used to create a DL model capable of detecting the presence of glaucoma. CNN and transfer learning algorithms are among the DL algorithms. The constructed models are then trained repeatedly until they attain the maximum and lowest accuracy values. The values of both training and the validation of the models are tabulated and examined. By calculating the difference in every performance metric and comparing it to the latter algorithm, it can be found that the transfer learning algorithm is better than the CNN algorithm in glaucoma detection. As the accuracy percentage of the developed DL model is very high, it would be a powerful preliminary-level test that can be performed in the process of prediction of glaucoma for the patient. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

George, Ronnie MS; Ve, Ramesh S. MPhil; Vijaya, Lingam MS:, "Estimated Burden of Disease", J. Glaucoma, vol. 19, no. 6, pp. 391-397, 2010. [http://dx.doi.org/10.1097/IJG.0b013e3181c4ac5b] [PMID: 20711029]

[2]

N. Sengar, M.K. Dutta, R. Burget, and M. Ranjoha, "Automated detection of suspected Glaucoma in digital fundus images", In: 40th International Conference on Telecommunications and Signal Processing, 2017, pp. 749-752. [http://dx.doi.org/10.1109/TSP.2017.8076088]

[3]

A. Dey, and K.N. Dey, "Automated Glaucoma Detection from Fundus Images of Eye Using Statistical Feature Extraction Methods and Support Vector Machine Classification", In: Proceedings of the International Conference, 2019. [http://dx.doi.org/10.1117/12.2524215]

[4]

D.M.S. Barros, J.C.C. Moura, C.R. Freire, A.C. Taleb, R.A.M. Valentim, and P.S.G. Morais, "Machine learning applied to retinal image processing for glaucoma detection: review and perspective", Biomed. Eng. Online, vol. 19, no. 1, p. 20, 2020. [http://dx.doi.org/10.1186/s12938-020-00767-2] [PMID: 32293466]

[5]

T.R.V. Bisneto, A.O. de Carvalho Filho, D.M.V. Magalhães, “Generative adversarial network and texture features applied to automatic glaucoma detection,”; Applied Soft Computing, vol. 90, pp. 106195, ISSN 1568-4946, 2020.

[6]

S. Gheisari, S. Shariflou, J. Phu, P.J. Kennedy, A. Agar, M. Kalloniatis, and S.M. Golzan, "A combined convolutional and recurrent neural network for enhanced glaucoma detection", Sci. Rep., vol. 11, no. 1, p. 1945, 2021.

Learning Algorithm

AI and IoT-based Intelligent Health Care & Sanitation 113

[http://dx.doi.org/10.1038/s41598-021-81554-4] [PMID: 33479405] [7]

Y. Sun, B. Xue, M. Zhang, G.G. Yen, and J. Lv, "Automatically Designing CNN Architectures Using the Genetic Algorithm for Image Classification", IEEE Trans. Cybern., vol. 50, no. 9, pp. 3840-3854, 2020. [http://dx.doi.org/10.1109/TCYB.2020.2983860] [PMID: 32324588]

[8]

T. Kaur, and T.K. Gandhi, "Deep CNNs with transfer learning for automated brain image classification", Mach. Vis. Appl., vol. 31, no. 20, 2020. [http://dx.doi.org/10.1007/s00138-020-01069-2]

[9]

Y. LeCun, Y. Bengio, G. Hinton.: "Deep Learning"., Nature, Vol. 521, no. 7553, pp.436-44, 2015.

[10]

S. Yadav, F. Al-Turjman, V. Yadav, M. Kumar, and T. Stephan, "Transforming Management with AI", In: Big-Data, and IoT, 2021.

[11]

S.J. Pan, and Q. Yang, "A Survey on Transfer Learning", IEEE Trans. Knowl. Data Eng., vol. 22, no. 10, pp. 1345-1359, 2010. [http://dx.doi.org/10.1109/TKDE.2009.191]

[12]

S. Akçay, M.E. Kundegorski, M. Devereux, and T.P. Breckon, "Transfer learning using CNNs for object classification within X-ray baggage security imagery", IEEE International Conference on Image Processing, 2019 [http://dx.doi.org/10.1109/ICIP.2016.7532519]

[13]

R. Chauhan, K.K. Ghanshala, and R.C. Joshi, "CNN for Image Detection and Recognition", First International Conference on Secure Cyber Computing and Communication, pp. 278-282. [http://dx.doi.org/10.1016/j.icte.2020.08.002]

[14]

U.R. Acharya, S.L. Oh, Y. Hagiwara, J.H. Tan, M. Adam, A. Gertych, and S.T. Ru, "A deep CNN model to classify heartbeats", Comput. Biol. Med., vol. 89, 2017. [http://dx.doi.org/10.1016/j.compbiomed.2017.08.022] [PMID: 28869899]

[15]

Y. Zhu, Y. Chen, Z. Lu, S.J. Pan, G-R. Xue, Y. Yu, and Q. Yang, "Heterogeneous Transfer Learning for Image Classification", In: Twenty-Fifth AAAI Conference on Artificial Intelligence, 2011.

114

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 114-128

CHAPTER 8

Texture Analysis-based Features Extraction & Classification of Lung Cancer Using Machine Learning Korla Swaroopa1,*, N. Chaitanya Kumar2, Christopher Francis Britto3, M. Malathi4, Karthika Ganesan5 and Sachin Kumar6 Department of CSE, Aditya Engineering College, Surampalem- East Godavari, Andhra Pradesh, India-533437 2 Department of CSE, Sri Venkateswara Engineering College, Karakambadi Road, Tirupati, Andhra Pradesh, India 3 Information Technology & Computer Services, Mahatma Gandhi University, Meghalaya, 793101, India 4 Department of ECE, Vivekananda College of Engineering for Women (Autonomous), Elayampalayam, Namakkal, Tamilnadu, India-637205 5 Sri Vidya College of Engineering and Technology, Sivakasi Main Road, Virudhunagar, Tamilnadu, India-626005 6 College of IT Engineering, Kyungpook National University, Daegu, South Korea 1

Abstract: Lung cancer is a form of carcinoma that develops as a result of aberrant cell growth or mutation in the lungs. Most of the time, this occurs due to daily exposure to hazardous chemicals. However, this is not the only cause of lung cancer; additional factors include smoking, indirect smoke exposure, family medical history, and so on. Cancer cells, unlike normal cells, proliferate inexorably and cluster together to create masses or tumors. The symptoms of this disease do not appear until cancer cells have moved to other parts of the body and are interfering with the healthy functioning of other organs. As a solution to this problem, Machine Learning (ML) algorithms are used to diagnose lung cancer. The image datasets for this study were obtained from Kaggle. The images are preprocessed using various approaches before being used to train the image model. Texture-based Feature Extraction (FE) algorithms such as Generalized Low-Rank Models (GLRM) and Gray-level co-occurrence matrix (GLCM) are then used to extract the essential characteristics from the image dataset. To develop a model, the collected features are given into ML classifiers like the Support Vector Machine (SVM) and the k-nearest neighbor's algorithm (k-NN). Corresponding author Korla Swaroopa: Department of CSE, Aditya Engineering College, Surampalem- East Godavari, Andhra Pradesh, India-533437; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 115

To evaluate FE and classification, several performance metrics are used, such as accuracy, error rate, sensitivity specificity, and so on.

Keywords: Classification, CT scan, Lung Adenocarcinoma, Performance Metrics, Texture. 1. INTRODUCTION The primary cause of all forms of cancer in humans is a bodily mutation in cellular DNA. Adenocarcinoma is cancer that develops in the glandular tissues surrounding the organs and produces biological fluids like phlegm, digestive fluids, etc. It can be found in many different organs, including the colon, breasts, lungs, prostate, and esophagus. Adenocarcinoma of the lung is one of the most prevalent kinds of lung cancer when compared to squamous and small carcinoma. It is responsible for about 40% of all non-small carcinoma cases and 30% of all cancer cases overall. This form of carcinoma is located in the peripheral region of the lung and, in certain cases, in the areas of inflammation or scars. It is one of the most prevalent cancers detected in people who have never smoked [1]. Smoking nicotine is by far the most serious reason for lung cancer in individuals, but other factors such as family genetic heritage and environmental exposures to smoke and other carcinogenic substances also play a role. The most prevalent cause of lung tumor formation is a genetic abnormality in the p53 gene, which functions as a tumor suppressor and is accountable for cell division and cell death in roughly 52% of people [2]. Lung cancer is a kind of carcinoma that affects a vast number of people around the world, with a 5-year survival chance of less than 12% to 15%. A low-dose CT scan is often advised for high-risk patients, such as former and present heavy smokers. If the CT scan reveals mediastinal nodes, the patient will be advised to undergo various tests to stage cancer, and treatment options will be discussed based on the stage of the malignancy. This shows that in recent years, a computer-aided diagnostic tool for screening lung cancer at a preliminary phase with great precision has become quite desirable. Though early detection of lung cancer is critical, some studies have used ML algorithms to analyze data and diagnose lung cancer. Singh, G.A.P., et al. in their study of the direction and classification of lung cancer using various ML techniques, propose the seven classifiers trained using the GLCM FE method that contains all 14 features and conclude that MLP neural network has the maximum classification accuracy [3]. Bayanati et al. created a methodology for separating cancer from benign mediastinal lymph. Their research combined texture-based FE methods (GLCM and GLRL) and structure attributes with regression models and SVM Classifiers [4]. Alves, A.F.F., et al. developed an approach that included 15

116 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

characteristics with three ML techniques and five distinct operators, yielding an accuracy of less than 5% [5]. Makaju et al. proposed an approach where they employed a watershed segmentation technique for identification and a classical SVM classifier to classify the carcinoma in a Lung cancer CT scan image. The proposed approach gave an accuracy of 92%.[6]. Tyagi and colleagues advocated for a complete examination of a wide range of image segmentation approaches to detect and diagnose lung cancer. These methodologies make use of edge detection, thresholding methods, and PDE-dependent algorithms, among other techniques. One of these methods is Marker Controlled Watershed Segmentation. Various classification methodologies for the detection of lung carcinoma lesions using image classification can be combined with numerous segmentation techniques [7]. Faisal et al. attempted to determine the excluding value of several classifiers to enhance the accuracy of lung cancer prediction. To achieve maximum accuracy, they used Multi-Layer Perceptron, SVM, Nave Bayes, and other classifiers to analyze the datasets retrieved from the UCI repository [8]. Nasser and Abu-Naser (2019) used ANN to improve a lung cancer diagnosis computer-aided model. Anxiety, yellow fingers, and other symptoms are used to identify lung cancer. It is proposed that an ANN model be developed, trained, and evaluated utilizing data from a “survey lung cancer” dataset, as well as additional patient-specific information [9]. The following is how the article is built: Section 2 discusses the materials and processes for detecting lung cancer. Section 3 discusses the texture-based FE approaches employed in this work, such as GLCM and GLRM. SVM and k-NN are two of the classification algorithms discussed in Section 4. In Section 5, the performance of a variety of FE and classification approaches is compared. Finally, for lung cancer detection, Section 6 suggests integrating FE with several classifier methods. 2. METHODOLOGY This study makes use of a Lung CT image dataset from the Kaggle database, which totals 430 images. To improve the image quality, the raw photos are normalized utilizing a multitude of preprocessing procedures. The FE process follows, with the texture-based FE method being used in this research. Finally, to get the best level of accuracy, the images are classified using ideal classifiers. The complete images are divided into two parts in this work, and the above-mentioned ML techniques are used to train and test these CT images. In this study, 80% of the images (i.e., 344 images) are utilized as training sets, whereas 20% of the images (i.e., 86 images) are used as testing sets. Figs. (1 and 2) show sample lung

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 117

CT images, whereas Fig. (3) shows a block diagram of the stages required in employing ML approaches to detect lung cancer.

Fig. (1). Sample Normal Lung CT image.

Fig. (2). Sample Abnormal Lung CT image.

3. FE Choosing a FE approach is the most important factor in achieving great diagnostic accuracy. FE is a more involved method that entails attempting to turn the input data into low-dimensional data while retaining the bulk of the significant information. To improve output accuracy, precision, and other performance measures, the FE approach is commonly employed in conjunction with classifiers [10]. The important characteristics of the Lung CT scan images were extracted using texture-based FE techniques such as GLCM and GLRM in this work.

118 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

Fig. (3). Block diagram of Lung Cancer detection model using ML.

3.1. GLCM When it comes to images, texture can be represented by the average spatial relationship between the grey tones in the image. This was one of the first techniques to be developed based on this notion. Earlier work by Julesz can be linked back to his notion that human texture discrimination requires just secondorder probabilities, which laid the groundwork for our current formulation of texture discrimination. The GLCM is one of the easiest methods for extracting texture-based data. The GLCM is a matrix whose size is dictated by the number of

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 119

grey levels in the image. In a grey image, the GLCM stores information about the likelihood of two neighboring pixel pairings [11]. GLCM extracts four unique texture-based characteristics, which are described in detail using the following equations: Energy: σ௨ǡ௩ሺܲሺ‫ݑ‬ǡ ‫ݒ‬ሻሻଶ

(1)

Entropy: െ ෍ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ݈‫ܲ݃݋‬ሺ‫ݑ‬ǡ ‫ݒ‬ሻ

(2)

௨ǡ௩

Correlation: ሺ‫ ݑ‬െ ߤሻሺ‫ ݒ‬െ ߤሻ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ߪଶ

(3)

ͳ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ͳ ൅ ሺ‫ ݑ‬െ ‫ݒ‬ሻଶ

(4)

െ෍ ௨ǡ௩

Homogeneity: ෍ ௨ǡ௩

Inertia: ෍ሺ‫ ݑ‬െ ‫ݒ‬ሻଶ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ

(5)

௨ǡ௩

Variance: ෍ሺ‫ ݑ‬െ ߤሻଶ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ

(6)

௨ǡ௩

Where, u,v → matrix indices ߪ ൌ ෍ሺ‫ ݒ‬െ ߤ௬ ሻଶ ෍ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ௩

௨

ߤ ൌ ෍ ‫ ݑ‬෍ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ൌ ෍ ‫ ݒ‬෍ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ௨

௩

௩

௨

(7) (8)

120 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

3.2. GLRM The Grey Level Run-Length Matrix (GLRM) is an approach used to minimize the dimensionality of the pre-processed dataset. The texture is defined as a pattern of grey-intensity pixels that are oriented in a given direction relative to the reference pixels in the context of texture. GLRM can be used to extract texture features from the grey-level run-length matrix, which can then be used to analyze the texture. Essentially, it is an approach for searching for groups of pixels with the same grey level value that always moves in the same direction across an image. When viewed from a certain angle, the length of a run is the number of pixels in a row that have the same grey value as one another. In GLRM, a matrix is derived for each of the 11 textural indexes, as well as for the four distinct orientations. The GLRLM element (x,y) represents the number of homogenous runs of ‘y’ pixels in an image with the intensity ‘x’. GLAM extracts 11 distinct texture-based features [12], a few of which are detailed using the equations below: Short-run: ͳ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ෍ ݊ ‫ݒ‬ଶ

(9)

ͳ ෍ ݆ ଶ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ݊

(10)

ͳ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ෍ ݊ ‫ݑ‬ଶ

(11)

ͳ ෍ ‫ݑ‬ଶ ܲሺ‫ݑ‬ǡ ‫ݒ‬ሻ ݊

(12)

௨ǡ௩

Long-run: ௨ǡ௩

Low Gray level run: ௨ǡ௩

High gray level run: ௨ǡ௩

Gray level non-uniformity: ͳ ෍ ൭෍ ‫݌‬ሺ‫ݑ‬ǡ ‫ݒ‬ሻ൱ ݊ ௨

Run-length non-uniformity:

௩

ଶ

(13)

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 121

ͳ ෍ ൭෍ ‫݌‬ሺ‫ݑ‬ǡ ‫ݒ‬ሻ൱ ݊ ௨

ଶ

(14)

௨

Run percentage: ෍ ௨ǡ௩

݊ ‫݌‬ሺ‫ݑ‬ǡ ‫ݒ‬ሻ‫ݒ‬

(15)

4. CLASSIFICATION METHODS .

Once the essential features are collected using FE methods [13], different classification algorithms are used to diagnose lung cancer from CT scans of lung cancer. The categorization methods utilized in this study are detailed in the section below. 4.1. SVM A collection of approaches known as supervised learning methods (SVMs) may be used for both classification and regression are known as SVMs. They are a classification system that is a member of the generalized linear classification family of systems. SVM is unique in that it can reduce classification error while simultaneously increasing the geometric margin. SVM classifiers are frequently referred to as “Maximum Margin Classifiers,” which refers to classifiers that maximize their margin of error. SVM is founded on the concept of reducing structural risk to the greatest extent possible (SRM). The most efficient SVM may be constructed by shifting the input vector to a higher-dimensional space and then generating a hyperplane with the greatest possible separation between the two dimensions. Each of the two parallel hyperplanes is divided into two sections, with one on either side, of which two parallel hyperplanes are generated, and each of these parallel hyperplanes is divided into two parts. The separating hyperplane is defined as the hyperplane that maximizes the distance between two parallel hyperplanes while also connecting the two hyperplanes. For the classifier to be considered to have reduced generalization error, it must have a wider margin or distance between the parallel hyperplanes than a certain threshold value. Generally, the use of SVM classifier and biomarkers in cancer prognosis is very significant. The precision of this classifier is much greater. SVM is used to determine a Hyperplane that provides the largest possible difference between two groups of data. SVM translates data into a higher-dimensional feature space to create a Hyperplane. A kernel function is used by SVM to achieve this nonlinear

122 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

mapping [14]. RBF is the optimal kernel function employed with an SVM classifier in cancer detection, according to research work. The following is the final classifier equation: ே

݂ሺ‫ݖ‬௜ ሻ ൌ ෍ ‫ן‬௡ ‫ݕ‬௡ ‫ܭ‬൫‫ݖ‬௡ǡ ‫ݖ‬௜ ൯ ൅ ܾ

(16)

௡ୀଵ

Where, yn→ Output n→ Multiplier γ → Optimizing unit zn → Vector 4.2. KNN The KNN technique can be used to classify objects based on a set of training examples that are provided. Lazy learning is a type of KNN in which the function is only approximated locally and the full calculation is postponed until the classification is completed. When the distribution of the data is not known, the KNN classification technique is the most fundamental and straightforward classification procedure to use in this situation. As a result of applying this rule, each query is given the majority label of the training set neighbors that are knearest to it, and each query is assigned to one of three classes: The KNN technique for non-parametric classification is straightforward, but it has proven to be effective in a variety of situations. An effective strategy for both regression and classification in this classifier is to apply values to the inputs of the neighbors, such that the closer neighbors give more to the median than the farther neighbors. The k-NN of a database 't' are retrieved to categorize it, resulting in a 'k' neighborhood. An appropriate value for k must be fixed to apply the k-NN method, and the classification's performance is strongly contingent on this k number. The simplest approach to determine an appropriate number for ‘k’ is to run the algorithm numerous times with various k values and select the one with the best accuracy [15].

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 123

The implementation technique for the KNN algorithm is as follows: ௡

‫ ܦ‬ൌ ඩ෍ሺ‫ݔ‬௡ െ ‫ݕ‬௡ ሻଶ

(17)

௜ୀଵ

● ● ● ● ●

Estimate the Euclidean distance D. Using ascending order arrange Euclidean distance D. Based on the k value set the neighbors are taken If more neighbors are belonging to Malignant, the given image is Malignant Else, the given image is Benign

5. RESULTS AND DISCUSSION The Kaggle database was used to acquire 430 lung CT scan images for this study. In an 8:2 ratio, the dataset is split into two halves, one for testing and the other for training the classification algorithm (i.e., 80% and 20%). Performance measures are used to assess the FE methods and the classification algorithm that results. Accuracy, Specificity, Sensitivity, Precision, False Negative and False Positive Rates, and Error Rates are among the performance indicators employed in this study. The contingency table's elements are used to examine and evaluate the classification model. These performance evaluations aid in the development of the most optimized FE and classification algorithm for lung cancer detection. Texture-based FE algorithms like GLRM and GLCM are compared to classifiers like k-NN and SVM in Table 1. Table 1 reveals that the degree of accuracy for GLCM with an SVM classifier is quite high, at 97%. When the GLCM approach is tested with a k-NN classifier, the accuracy output is 94%, which is the secondhighest. The accuracy of the GLRM FE approach using SVM and k-NN classifier is 93% and 91%, respectively. When using the GLRM technique with SVM, the precision level is as high as 95%, and when using the GLRM technique with the k-NN classifier, the precision level is only 92%. When GLCM is combined with SVM, the error rate is as low as 2.3%, while when GLRM is combined with a kNN classifier, the error rate is as high as 8.1%, which makes GLCM with SVM an ideal combination of the algorithm with high accuracy and low error rate. Figs. (4-7) show a graphical representation of the performance study of lung cancer detection using GLCM and GLRM with several classifiers.

124 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

Table 1. Analysis of Lung cancer CT scan images using texture-based FE with various classifier. FE + Model

GLCM+SVM

GLCM+KNN

GLRM+SVM

GLRM+KNN

ACCURACY

97.67442

94.18605

93.02326

91.86047

ERROR RATE

2.3256

5.814

6.9767

8.1395

PRECISION

95.34884

93.02326

95.45455

92.68293

F-MEASURE

97.61905

94.11765

93.33333

91.56627

Specificity

95.55556

93.18182

95

93.18182

Sensitivity

100

95.2381

91.30435

90.47619

FNR

0

4.761905

8.695652

9.52381

FPR

4.444444

6.818182

5

6.818182

The graphical depiction of a Lung CT image using a combination of GLCM FE and SVM classifier is shown in Fig. (4). The graph illustrates that when the GLCM is used with the SVM classifier, the accuracy is 97% with a low error rate of only 2%. It also reveals that the accuracy is around 95%, with a specificity of 95%. This model has a % F-measure rate and a flawless % sensitivity. The model has a 0% FNR and a 4% FPR.

Fig. (4). Pictorial interpretation of Lung CT image for the conjunction of GLCM FE with SVM classifier.

Fig. (5) interpreted below, gives the graphical depiction of the Lung CT image for the conjunction of GLCM FE and k-NN classifier. The graph illustrates that when GLCM is combined with a k-NN classifier, the accuracy is 94% with a 5% error

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 125

rate. It also demonstrates that the accuracy is around 93%, with a specificity of 93%. This model has a 94 percent F-measure rate and a % sensitivity. The model has a 4% FNR and a 6% FPR.

Fig. (5). Pictorial interpretation of Lung CT image for the conjunction of GLCM FE with k-NN classifier.

Fig. (6) mentioned below gives the pictorial representation of the Lung CT image for the conjunction of GLRM FE with the SVM classifier. The graph illustrates that when GLCM is combined with an SVM classifier, the accuracy is 93% with a 6% error rate. It also reveals that the accuracy is around 95%, which is the greatest among all the models, and that the specificity is likewise 95%. This model has a % F-measure rate and a 91% sensitivity. The model has an 8% FNR and a 5% FPR. The graphical depiction of a Lung CT image for a combination of GLRM FE and k-NN classifier is shown in Fig. (7). The graph reveals that when GLRM is combined with a k-NN classifier, the accuracy is 91%, the lowest of all the models, with an error rate of up to 8%. It also demonstrates that the accuracy is around 92%, with a % specificity. This model has an F-measure rate of 91% and a sensitivity of 90%. The model has an FNR of 9% and an FPR of 6%, respectively.

126 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

Fig. (6). Pictorial interpretation of Lung CT image for the conjunction of GLRM FE with SVM classifier.

Fig. (7). Pictorial interpretation of Lung CT image for the conjunction of GLRM FE with k-NN classifier.

Machine Learning

AI and IoT-based Intelligent Health Care & Sanitation 127

6. GAPS FILLED Many studies have been undertaken for the prediction of lung cancer. But the problem with their studies is the identification of the best ML model with texturebased FE, whereas there are three ML algorithms and two FE used in this study. The usage of various ML and FE results in the comparison of result prediction with the actual input. This comparison leads us to decide on the best ML algorithm and texture-based FE techniques combination that can be employed in the prediction. CONCLUSION Lung cancer can be deadly, but new therapies have made it possible for many individuals to live and recover from it, especially if they are diagnosed early. The Kaggle database is used to collect and analyze the Lung CT image dataset in this study. The raw images are pre-processed to enable them to be used for classification. FE identifies the image's crucial data using pre-processed data. Finally, the acquired feature data is fed into several lung cancer detection classification algorithms. The model is evaluated using performance measures as well as FE techniques. This study shows that the GLCM FE approach paired with the SVM classifier is the best algorithm combination for diagnosing lung cancer from CT scan images after reviewing the image data, with a 97%, % precision, and a low error rate of only 2%. This study backs up choosing the best FE and ML classifier model for detecting lung cancer-affected CT scan images. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

D.J. Myers, and J.M. Wallen, Lung Adenocarcinoma. StatPearls. StatPearls: Treasure Island, FL, 2021.

[2]

J. Gariani, S.P. Martin, A.L. Hachulla, W. Karenovics, D. Adler, P.M. Soccal, C.D. Becker, and X. Montet, "Noninvasive pulmonary nodule characterization using transcutaneous bioconductance", Medicine (Baltimore), vol. 97, no. 34, p. e11924, 2018.

128 AI and IoT-based Intelligent Health Care & Sanitation

Swaroopa et al.

[http://dx.doi.org/10.1097/MD.0000000000011924] [PMID: 30142805] [3]

G.A.P. Singh, and P.K. Gupta, "Performance analysis of various machine learning-based approaches for detection and classification of lung cancer in humans", Neural Comput. Appl., vol. 31, no. 10, pp. 6863-6877, 2019. [http://dx.doi.org/10.1007/s00521-018-3518-x]

[4]

H. Bayanati, R.E. Thornhill, C.A. Souza, V. Sethi-Virmani, A. Gupta, D. Maziak, K. Amjadi, and C. Dennie, Quantitative CT texture and shape analysis: can it differentiate benign and malignant mediastinal lymph nodes in patients with primary lung cancer., 2015. [http://dx.doi.org/10.1007/s00330-014-3420-6]

[5]

A.F.F. Alves, S.A. Souza, R.L. Ruiz Jr, T.A. Reis, A.M.G. Ximenes, E.N. Hasimoto, R.P.S. Lima, J.R.A. Miranda, and D.R. Pina, "Combining machine learning and texture analysis to differentiate mediastinal lymph nodes in lung cancer patients", Physical and Engineering Sciences in Medicine, vol. 44, no. 2, pp. 387-394, 2021. [http://dx.doi.org/10.1007/s13246-021-00988-2] [PMID: 33730292]

[6]

S. Makaju, P.W.C. Prasad, A. Alsadoon, A.K. Singh, and A. Elchouemi, "Lung Cancer Detection using CT Scan Images", Procedia Comput. Sci., vol. 125, pp. 107-114, 2018. [http://dx.doi.org/10.1016/j.procs.2017.12.016]

[7]

P. Tripathi, S. Tyagi, and M. Nath, "A comparative analysis of segmentation techniques for lung cancer detection", Pattern Recognit. Image Anal., vol. 29, no. 1, pp. 167-173, 2019. [http://dx.doi.org/10.1134/S105466181901019X]

[8]

M.I. Faisal, S. Bashir, Z.S. Khan, and F.H. Khan, “An evaluation of ML classifiers and ensembles for early-stage prediction of lung cancer.”; In 2018 3rd International Conference on Emerging Trends in Engineering, Sciences, and Technology (ICEEST), IEEE, pp. 1-4, 2018. [http://dx.doi.org/10.48161/qaj.v1n2a58]

[9]

I.M. Nasser, and S.S. Abu-Naser, "Lung Cancer Detection Using Artificial Neural Network", Int. J. Eng. Inform. Sys., vol. 3, no. 3, pp. 17-23, 2019.

[10]

N. Chumerin, and M.M. Van Hulle, “Comparison of Two FE Methods Based on Maximization of Mutual Information.”; 16th IEEE Signal Processing Society Workshop on ML for Signal Processing, pp. 343-348, 2006.

[11]

R. Asery, R.K. Sunkaria, L.D. Sharma, and A. Kumar, "Fog detection using GLCM based features and SVM", In: 2016 Conference on Advances in Signal Processing, 2016, pp. 72-76. [http://dx.doi.org/10.1109/CASP.2016.7746140]

[12]

M. Khojastehnazhand, and H. Ramezani, "Machine vision system for classification of bulk raisins using texture features", J. Food Eng., vol. 271, p. 109864, 2020. [http://dx.doi.org/10.1016/j.jfoodeng.2019.109864]

[13]

S. Prakash Yadav, and S. Yadav, "Fusion of Medical Images in Wavelet Domain: A Hybrid Implementation", Comput. Model. Eng. Sci., vol. 122, no. 1, pp. 303-321, 2020. [http://dx.doi.org/10.32604/cmes.2020.08459]

[14]

H. Wang, and G. Huang, "Application of support vector machine in cancer diagnosis", In: Med Oncol. vol. 28. , 2011, pp. 613-618. [http://dx.doi.org/10.1007/s12032-010-9663-4]

[15]

Guo G., Wang H., Bell D., Bi Y., Greer K.: “KNN Model-Based Approach in Classification.”; Lecture Notes in Computer Science, Springer, Berlin, Heidelberg, Vol 2888, 2003. [http://dx.doi.org/10.1007/978-3-540-39964-3_62]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 129-143

129

CHAPTER 9

Implementation of the Deep Learning-based Website For Pneumonia Detection & Classification V. Vedanarayanan1, Nagaraj G. Cholli2, Merin Meleet2, Bharat Maurya3, G. Appasami4 and Madhu Khurana5,* Electronics and Communication Engineering (ECE), Sathyabama Institute of Science and Technology (Deemed to be University), Chennai, Tamil Nadu, India-604119 2 Department of Information Science and Engineering, RV College of Engineering, Bengaluru, Karnataka, India-560059 3 Department of Political Science, Chaudhary Bansi Lal University, Bhiwani Haryana, India, 127021 4 Department of Computer Science, Central University of Tamil Nādu, Tiruvarur, Tamil Nādu, India-610005 5 University of Gloucestershire, UK 1

Abstract: It is often difficult to diagnose several lung illnesses, such as atelectasis and cardiomegaly, as well as Pneumonia, in hospitals due to a scarcity of radiologists who are educated in diagnostic imaging. If pneumonia is diagnosed early enough, the survival rate of pulmonary patients suffering from the disease can be improved. Most of the time, chest X-ray (CXR) pictures are used to detect and diagnose pneumonia. When it comes to detecting pneumonia on CXR images, even an experienced radiologist may have difficulty. It is vital to have an automated diagnostic system to improve the accuracy of diagnostic results. It is estimated that automated pneumonia detection in energy-efficient medical systems has a substantial impact on the quality and cost of healthcare, as well as on response time. To detect pneumonia, we employed deep transfer learning techniques such as ResNet-18 and VGG-16. Each of the model's four standard metrics, namely accuracy, precision, recall, and f1-score, are used to evaluate. The best model is established by the use of metrics. To make pneumonia detection simple, the website is designed by employing the best model.

Keywords: Deep Learning, Performance Metrics, Pneumonia, ResNet-18, Streptococcus Pneumoniae. Corresponding author Madhu Khurana: Lecturer in cyber-Security, University of Gloucestershire, UK; E-mail: [email protected] *

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

130 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

1. INTRODUCTION Pneumonia is a prevalent lung disease that is the main cause of death in children and the elderly. Though various factors such as age, underlying illness, and environment have a role in the development of pneumonia in babies and small children, Streptococcus pneumoniae is the most prevalent bacterial cause of pneumonia. Pneumonia affects children of all ages, however, it is more common in children under the age of 5 and adults above the age of 75. Chlamydia pneumoniae and Mycoplasma pneumoniae have recently been identified as the microorganisms that cause pneumonia in children over the age of five [1]. According to WHO estimates, 450 million cases of pneumonia are registered each year, resulting in roughly 4 million fatalities, accounting for 7 percent of the global mortality rate of 57 million people [2]. Cough, respiratory pain, fever, and abnormalities in X-rays are considered as the analysis of Pneumonia. However, no clinical approach can properly define the pathogenesis of pneumonia. Doctors can identify pneumonia in hospitalized patients using a medical examination, medical records, diagnostic tests such as phlegm or blood work, pulmonary X-rays, and other imaging techniques. The most frequent tool for detecting lung infections such as pneumonia is chest X-rays, which are getting less expensive because of technological developments in biomedical instruments [3]. The competence to distinguish between bacterial and viral pneumonia might have significant consequences for treatment. Despite progress, diagnostic testing in many afflicted patients still fails to identify the underlying pathogens. Computer-assisted diagnostics can overcome the problem of professional availability. Artificial intelligence research is progressing at a rapid pace, which can be extremely beneficial. Stephen et al. [4] proposed using a Deep Learning (DL) model to diagnose pneumonia. They created a DL model from scratch using a set of chest X-ray scans. Meanwhile, researchers were interested in such pre-trained systems' ability to execute X-ray image processing tasks, and they sought to grasp more about the potential of various DL models. For example, Loey et al. [5] employed AlexNet, GoogleNet, and ResNet-18 to describe and categorize X-ray images for corona diagnosis via a transferable learning technique. Rajpurkar et al. [6] constructed a DL model for diagnosing pneumonia in chest x-ray images using the dataset ChestX-ray14. Using CheXNet, a hundred-twenty-one layer CNN, they categorized chest X-rays with a good rate that exceeded expert clinicians. Their system also discovered 14 more ailments in addition to pneumonia. Szegedy et al. [7] built 3 new unique deep CNN models that are combinations of Inception and ResNet systems. Their technique generated good outcomes. On the ImageNet classification challenge testing dataset, they had a top 5 error rate of 3.08 percent.

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 131

The structure of the article is as follows: The materials and techniques for detecting Pneumonia are discussed in Section 2. The preprocessing techniques used in this study, such as image scaling and image augmentation, are discussed in Section 3. Two of the DL algorithms covered in Section 4 are VGG-16 and ResNet-50. Web development is described in Section 5. Finally, in Section 6, the results are examined, and in Section 7, the general conclusion of the study is presented. 2. MATERIALS AND METHODS With the use of an image dataset taken from the Kaggle website, this study has been able to create and assess an image-based DL model capable of accurately diagnosing patients with Pneumonia through the use of a user-friendly humaninterfaced website. The process of using DL models to diagnose pneumonia is shown schematically in Fig. (1).

Fig. (1). Block diagram of Pneumonia detection model using DL Algorithms.

132 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

Anterior-posterior X-ray scans of the chest were collected from a prior group of young children at the Guangzhou Women and Children's Medical Center in Guangzhou, aged one to five years. All of the patients' thoracic X-ray exams were part of their usual medical care. To analyze chest x-ray images, all thoracic radiographs were first checked for quality checks, with any images that resulted in poor quality or illegible being discarded. Once the image had been diagnosed, they were assessed by two separate professionals before being given the green light to be included in the AI system. A third expert reviewed the assessment set to confirm that there were no errors in the findings. Firstly, 5863 JPEG X-ray images are acquired and divided into two groups: Pneumonia-affected X-ray images and normal X-ray images. X-ray images (JPEGs) were used to create 3166 balanced data sets from an unbalanced image data set. For training, testing, and validation, the chest x-ray image data was divided into three equal halves with a ratio of seven to two to one (i.e., 70 percent, 20 percent, and 10 percent of the overall data set) (i.e., 70 percent, 20 percent, and 10 percent of the total data set). Each of the three sets has a different number of images: 2216 for the training set, 316 for testing, and 634 for validation. 3. PREPROCESSING TECHNIQUES Preprocessing stage is the process of transforming unprocessed data into something that can be used in a learning model. It is the initial and most crucial stage in the development of a DL model. Actual data is frequently distorted, absent, and in an improper format, making it impossible to employ directly in DL models. Data preprocessing is required to clean the data and make it suitable for a DL model, which enhances the model's efficacy and precision. The obtained data is preprocessed in this study utilizing image scaling and image augmentation techniques. 3.1. Data Resizing Resizing the raw image data set is a very important step to be followed while creating a DL model. Since most of the DL model architectures require the image data set to be the same size, the collected raw image data set is resized to achieve this. Resizing is the process of changing the dimension of a picture without leaving anything out. When scaling an image, just the file size changes. For this study, before training the DL model, the raw picture data set was scaled to a size of 224x224.

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 133

3.2. Data Augmentation To attain high accuracy, DL models often require a huge quantity of training data. Image data augmentation is commonly used to improve the DL model's performance. Image augmentation deliberately creates training images using several processing methods or a combination of methods such as shifts, random rotation, flips, and shear, among others. The picture augmentation methods involved in this research were image rotation, image flipping, and image zooming. 4. DL TECHNIQUES It is becoming increasingly popular to employ DL methods in the field of medical image analysis [8, 9]. DL has made it possible to detect the presence or absence of disease in medical imaging, which has grown increasingly prevalent in recent years. The neural networks are investigated in-depth and have the potential to revolutionize the disease using medical imaging. In this work, two transfer learning method is employed using VGG-16 and RESNET-50. Both models are detailed below. 4.1. VGG-16 The Oxford Net architecture, commonly known as VGG-16, is a 16-layer convolutional neural network built by University of Oxford professors K. Simonyan and A. Zisserman. The pre-trained network of this DL model has been trained on over a million photographs from the ImageNet database. ImageNet is a database of images of various resolutions. Images were reduced to 256x256 pixel resolutions in this way. The network was pre-trained to classify pictures into 1000 discrete item categories. A range of feature representations for various forms of images has been trained by the network as a result. The VGG-16 algorithm has 16 layers, 13 of which are convolutional and 3 of which will be completely linked. There are also five max-pooling layers in it [10]. The topmost layer and the lowermost layer are the only two levels that may be adjusted to fulfill the user's requests. Given below, Fig. (2) is the building of the VGG-16 model. 4.2. RESNET-50 The residual network, usually known as ResNet-50, is a multi-layer transfer training system with a depth of 50 layers. ResNet is mainly used to classify pictures. The network's convolutional layers feature a 3X3 filter, as well as convolutional layers that explicitly downsample at a rate of 2 downsample. It is possible to create a 256-channel network utilizing the ReLU and softmax activation algorithms as the last layers of the network. The network's learning rate

134 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

is 0.000001. After that, a 34-layer basic network architecture based on VGG-19 is employed to build the bypass connection. As a result of the bypass linkages, the design is altered into a residual network [11]. The ResNet-50 model's design is displayed in Fig. (3).

Fig. (2). Architecture of VGG – 16.

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 135

Fig. (3). Architecture of ResNet-50.

5. WEB DEVELOPMENT The user-friendly webpage is the final purpose of this study in diagnosing Pneumonia. The website was built with Django, a high-level Python web platform

136 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

that allows for the quick creation of safe and stable websites. It can work virtually with any client architecture and offer content in almost any format (including HTML, JSON, XML, etc.). This web tool allows users to upload a chest X-ray image and uses a DL algorithm to determine whether or not they have Pneumonia before alerting them. Fig. (4) shows an image of the website's home page.

Fig. 4. Homepage of the website.

6. RESULT AND DISCUSSION For this study, 5863 chest X-rays were obtained from the Kaggle database. The collected X-ray is parted into three groups, namely training, validating, and testing the DL method, in a 7:2:1 ratio (i.e., 70%, 20%, and 10%). Training Accuracy, Training Loss, Validation Accuracy, and Validation Loss are some of the performance metrics utilized in this study to evaluate the DL algorithm. These performance assessments contribute to the creation of the best DL system for detecting Pneumonia. The preprocessed image is evaluated using a DL model created with the VGG-16 and ResNet-50 modules. The performance

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 137

characteristics of the training set of the VGG-16 model, such as accuracy and loss, are shown in Table 1 for each epoch. Table 1. Accuracy and Loss analysis of VGG-16. Epochs

Training Phase

Validation Phase

Accuracy

Loss

Accuracy

Loss

0

0.8727466172276

0.33041855686364

0.8943217673145

0.23801852762127

1

0.9341552903638

0.18388415875069

0.9621450903494

0.12272572364426

2

0.9431047676086

0.1565302696075

0.8722398050964

0.31005041461792

3

0.9377256804749

0.16126086427307

0.9731861292532

0.10358668187286

4

0.9494584812866

0.13629573667084

0.8635329140302

0.3427799401253

5

0.9476541289087

0.13765722131988

0.9558359456743

0.11294174133938

6

0.9539711437314

0.11239003073776

0.9747633934096

0.08748178929095

7

0.9580324822937

0.11361615756454

0.9731861220532

0.08885020062901

According to Table 1 , the training set's accuracy using the VGG-16 algorithm at the zeroth epoch is around 0.87%, with a loss of 0.18%. With each epoch increment, the accuracy improves, and the loss lowers, reaching a maximum accuracy equal to 0.95% and a loss percentage of 0.11% at the seventh epoch. Then, using the VGG-16 method at the zeroth epoch, the validation set's accuracy is about 0.89%, with a loss of 0.23%. The accuracy improves, and the loss decreases with each epoch increment, achieving a maximum accuracy of 0.97% and a loss percentage of 0.08% at the sixth epoch. Figs. (5 and 6) show a graphical representation of the performance study of Pneumonia detection using VGG-16 architecture.

Fig. (5). Accuracy analysis of VGG-16.

138 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

Fig. (6). Loss analysis of VGG-16.

Table 2 shows that using the ResNet-50 method, the training set's accuracy at the zeroth epoch is roughly 0.82%, with a loss of 3.2%. The accuracy increases, and the loss decreases with each epoch increment, achieving a maximum accuracy of 0.89% at the fourth epoch and a minimum loss percentage of 0.21% at the sixth epoch. The validation set's accuracy is roughly 0.5% using the ResNet-50 approach at the zeroth epoch, with a loss of 2.5%. With each epoch increment, the accuracy increases and the loss reduces, reaching a maximum accuracy of 0.71% and a loss of 0.34% at the seventh epoch. The performance analysis of Pneumonia detection using ResNet-50 architecture is depicted graphically in Figs. (7 and 8). Table 2. Accuracy and Loss analysis of VGG-16. Epochs

Training Phase

Validation Phase

Accuracy

Loss

Accuracy

Loss

0

0.826714813709259

3.263126373290156

0.5

2.52524845712

1

0.852888107298047

0.964512765075623

0.5

0.82412342442

2

0.837545098742371

0.51476312523

0.5

1.85194386145

3

0.870938599109497

0.89934966813202

0.5

0.70359876693

4

0.893501818180842

0.280218986569977

0.5

0.3781881851

5

0.830776154994946

0.344854348583984

0.517350137

0.57685217412

6

0.829422354698812

0.219899284396301

0.647523204

0.4238426822

7

0.87770760059669

0.283429718175781

0.71293377876

0.3458746874

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 139

Fig. (7). Accuracy analysis of ResNet-50.

Fig. (8). Loss analysis of ResNet-50. Table 3. Test data performance. Model

Accuracy

Recall

Precision

F-Measure

VGG-16

94.62025316

96.02649

92.948718

94.46254

RESNET-50

74.36708861

71.59763

78.571429

74.9226

Table 3 shows that the VGG-16 has the highest accuracy of 94.6% to ResNet-50, which has a 74.3%. VGG-16 architecture has a precision of 92.9%, whereas ResNet-50 architecture has a precision of 78.5%. VGG-16 and ResNet-50 have recall values of 96.6% and 71.55%, respectively. VGG-16 has a 94.4% F-measure value, while ResNet-50 has a 74.9% F-measure value. Fig. (9) given below,

140 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

shows the graphical representation of performance comparison of VGG-16 and ResNet-50 on Pneumonia Detection. The model is integrated with a Django-developed webpage after it has attained its maximum accuracy. The user then uploads the chest X-ray image preprocessed with image scaling. The preprocessed image is then transmitted to the VGG-16 architecture, where it is examined for Pneumonia detection. Figs. (10 and 11) are given below to illustrate the front-end of the website with Pneumonia-affected and unaffected chest x-rays.

Fig. (9). Graphical representation of performance comparison of VGG-16 and ResNet-50 on Pneumonia Detection.

Fig. 10. Front-end of the website when Pneumonia is not affected.

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 141

Fig. 11. shows that if the probability score is less than 0.5, the individual's uploaded chest x-ray is classified as normal.

7. STATE OF ART Several studies have used DL technologies to predict pneumonia and other diseases. The fact that this study includes the creation of a website sets it unique from the others. This website simplifies the use of this paradigm by ordinary people. Because using a DL model or other tools might be challenging for folks who have little or no experience with DL. CONCLUSION Pneumonia can range in severity from a simple ailment to a serious or lifethreatening condition, and it can even be fatal. The Kaggle database was used to gather and analyze the chest X-ray images in this study. To achieve uniform proportions across all images, the raw images are pre-processed with image scaling and image augmentation. Using two separate techniques, two DL models were created. The VGG-16 and the ResNet-50 method are two of the algorithms. The generated models were then trained using the database's processed images. To improve accuracy and lower the loss value, the trained model is validated. After the models have been trained and validated, the performance of both models constantly improves as the number of epochs increases. As the number of epochs rises, the loss is the inverse of the accuracy. However, the model's total efficiency can only be assessed once it has been tested. When examined, the transfer learning algorithm (VGG-16) was shown to be the most efficient for pneumonia identification, with a % accuracy rate. In addition, the transfer learning method

142 AI and IoT-based Intelligent Health Care & Sanitation

Vedanarayanan et al.

outperforms other performance criteria, including recall, accuracy, and the Fmeasure. A website is developed using Django, and it is integrated with the DL model. The website successfully provides accurate results when tested in realtime. This website can be used as a preliminary test in the detection of pneumonia. This process can reduce the time wasted during the waiting of test results for pneumonia. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

A. Prayle, M. Atkinson, A. Smyth: “Pneumonia in the developed world.”; vol. 12, no. 1, pp. 60–69, 2011. [http://dx.doi.org/10.1016/j.prrv.2010.09.012]

[2]

WHO. “Revised global burden of disease (GBD) 2002 estimates." 2002.

[3]

A. Tilve, S. Nayak, S. Vernekar, D. Turi, P.R. Shetgaonkar, and S. Aswale, "Pneumonia detection using DL approaches", In: International Conference on Emerging Trends in Information Technology and Engineering IEEE, 2020, pp. 1-8. [http://dx.doi.org/10.1155/2021/3514821]

[4]

O. Stephen, M. Sain, U.J. Maduh, and D.U. Jeong, "An efficient DL approach to pneumonia classification in healthcare", J. Healthc. Eng., vol. 2019, pp. 1-7, 2019. [http://dx.doi.org/10.1155/2019/4180949] [PMID: 31049186]

[5]

M. Loey, and F. Smarandache, "M Khalifa N E: “Within the lack of chest covid-19 x-ray dataset: a novel detection model based on GAN and deep transfer learning", Symmetry (Basel), vol. 12, no. 4, p. 651, 2020. [http://dx.doi.org/10.3390/sym12040651]

[6]

P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, M. P. Lungren, Andrew Y. Ng "Chexnet: Radiologist-level pneumonia detection on chest x-rays with DL.”; 2017.

[7]

C. Szegedy, S. Ioffe, V. Vanhoucke, A. A. Alemi: “Inception-v4, inception-resnet and the impact of residual connections on learning.”; 31st AAAI Conference on Artificial Intelligence, 2017. [http://dx.doi.org/10.1609/aaai.v31i1.11231]

[8]

S. P. Yadav, and S. Yadav, "Mathematical implementation of fusion of medical images in continuous wavelet domain." J. Advan. Res. Dynamic. Cont. Syst., vol. 10, no. 10, pp. 45-54, 2019.

[9]

F. Al-Turjman, S.P. Yadav, M. Kumar, V. Yadav, T. Stephan, Ed., “Transforming Management with AI, Big Data, and IoT”. Springer Nature; 2022. [http://dx.doi.org/10.1007/978-3-030-86749-2]

Pneumonia Detection & Classification

AI and IoT-based Intelligent Health Care & Sanitation 143

[10]

M.K. Pandit, S.A. Banday, R. Naaz, and M.A. Chishti, "Automatic detection of COVID-19 from chest radiographs using deep learning", Radiography, vol. 27, no. 2, pp. 483-489, 2021. [http://dx.doi.org/10.1016/j.radi.2020.10.018] [PMID: 33223418]

[11]

R. Jain, P. Nagrath, G. Kataria, V. Sirish Kaushik, and D. Jude Hemanth, "Pneumonia detection in chest X-ray images using convolutional neural networks and transfer learning", Measurement, vol. 165, p. 108046, 2020. [http://dx.doi.org/10.1016/j.measurement.2020.108046]

144

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 144-158

CHAPTER 10

Design and Development of Deep Learning Model For Predicting Skin Cancer and Deployed Using a Mobile App Shweta M Madiwal1,*, M. Sudhakar2, Muthukumar Subramanian3, B. Venkata Srinivasulu4, S. Nagaprasad5 and Madhu Khurana6 Department of CSE, Faculty of Engineering and Technology (Exclusively for Women's) Sharnbasva University, Kalaburagi. Karnataka, India-585101 2 Department of Mechanical Engineering, Sri Sairam Engineering College, Chennai, Tamil Nadu, India-600044 3 SRM Institute of Science & Technology, Tiruchirappalli Campus, (Deemed to be University), Trichy, Tamilnadu, India - 621105 4 Department of CSE, JJTU, Rajasthan, India-333010 5 Department of Computer Science, Tara Government College (A), Sangareddy, Telangana, India500021 6 University of Gloucestershire, UK 1

Abstract: Melanoma, the deadliest form of skin cancer, is becoming more common every year, according to the American Cancer Society. As a result of the artifacts, low contrast, and similarity to other lesions, such as moles and scars on the skin, diagnosing skin cancer from the lesions might be difficult. Skin cancer can be diagnosed using a variety of techniques, including dermatology, dermoscopic examination, biopsy and histological testing. Even though the vast majority of skin cancers are non-cancerous and do not constitute a threat to survival, certain more malignant tumors can be fatal if not detected and treated on time. In reality, it is not feasible for every patient to have a dermatologist do a complete examination of his or her skin at every visit to the doctor's office or clinic. To solve this challenge, numerous investigations are being conducted to provide computer-aided diagnoses. In this work, skin cancer can be predicted from an image of the skin using deep learning techniques such as convolutional neural networks. The accuracy and loss functions of the model are used to evaluate its overall performance. The mobile app is created to detect skin cancer using the developed model. As soon as the images have been submitted, the app can communicate with the user about their progress.

Corresponding author Shweta M Madiwal: Department of CSE, Faculty of Engineering and Technology (Exclusively for Women's) Sharnbasva University, Kalaburagi. Karnataka, India-585101; E-mail: [email protected] *

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 145

Keywords: Convolutional Neural Network, Deep Learning Algorithm, Melanoma, Web Development.

1. INTRODUCTION Melanoma is a kind of cancer growing increasingly widespread worldwide, particularly in Western countries. Melanoma is still thought to be mostly influenced by sunlight exposure. Other factors, such as certain genetic changes, have a role in melanoma development in addition to UV exposure. Melanoma is considered one of the deadliest and most aggressive forms of skin cancer, especially cutaneous malignant melanoma. Surgical removal, radiotherapy, photodynamic therapy, immunotherapy, chemotherapy and laser surgery are some of the modern treatment options. Based on the patient's condition, tumor stage, and locale, the treatment approach may comprise standalone drugs or combination therapy [1]. Melanoma of the skin kills 55’500 people a year. The disease's incidence and fatality rates vary greatly over the world, based on the availability of rapid diagnosis and basic healthcare. Melanoma research has progressed significantly during the last ten years. Immunotherapy with anti-CTLA4 monoclonal antibodies and targeted treatment with BRAF inhibitors have emerged as two independent ways to try and extend life in patients with advanced melanoma cancer. However, due to pre-existing or developed tolerance, virtually all of them relapse. As a result, malignant melanoma has a dismal long-term prognosis [2]. Despite improvements, many affected patients' screening test still misses pinpointing the underlying disorders. Professional dependability can be addressed using computer-assisted diagnosis. Artificial intelligence technology is moving at an incredible speed, which has the potential to be tremendously useful. The problem of melanoma detection through dermoscopy images has been studied for decades. Reviews have been published that analyze a selection of papers published in the last decade [3]. Deep learning techniques for melanoma identification have been used in several studies recently. To identify a melanoma, Codella et al. presented an integrated technique that used convolutional neural networks (CNNs), sparse coding, and SVM [4]. To collect multi-scale characteristics for melanoma identification, Kawahara et al. used a deep convolutional network [5]. Ayan (2018) investigated the efficiency of a CNN model for the detection of skin problems by comparing it to a dataset that had not been upgraded and a dataset that had been improved. According to the authors of this work, preprocessing approaches may be effective in building powerful classifiers when there is insufficient data. Performance between networks that used updated datasets compared to networks that used unenhanced datasets was

146 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

significantly different [6]. Esteva et al. reported the outcomes of using Convolutional neural networks, training a large sample size (129450 clinical images) from 18 distinct healthcare professionals, open-access internet archives (including ISIC) and Stanford Medical Center. The testing findings for three class recognition problems showed average reliability of 72.1 percent, which was superior to 21 board-certified dermatologists' findings [7]. Many computer-aided diagnostic applications, such as classification [8], detection [9], and segmentation, have recently seen successes through convolutional neural networks (CNNs) with high-level feature learning capability. The raw picture is fed into the deep neural networks, eliminating the timeconsuming pre-processing, categorization, and development of customized features. Convolutional Neural Networks (CNN) have been acknowledged in the scientific area for a few years now, but the ImageNet classification challenge brought them to the spotlight. The following is the article's structure: Section 2 discusses the materials and techniques for diagnosing Melanoma. Section 3 discusses the preprocessing steps employed in this work, such as image scaling and image augmentation. An algorithm for deep learning in Section 4, for example, CNN, is discussed. Section 5 discusses web development. Finally, in Part 6, the results are discussed, and in Section 7, the study's overall conclusion is stated. 2. METHODOLOGY The study makes use of image data from the Kaggle mobile app. Initially, 2900 sample images are gathered to train and test the image using deep learning methods to create a user-friendly human-interfaced mobile app that accurately detects Melanoma in individuals. Fig. (1a and 1b) depict benign and melanoma skin images. The obtained skin image data set is separated into three groups with a 7:2:1 ratio: training, testing and validation (i.e., 70%, 20%, and 10% of the total data set). It is estimated that the training dataset contains 2030 images, the validation dataset has 580 images, and the testing dataset contains 290 images. To assure the image dataset's quality, pre-processing methods such as image scaling and image augmentation are used to examine the images for quality assurance. After the images have been preprocessed, deep learning models are used to determine if the image is melanoma-affected or not. Fig. (2) depicts a block diagram of the phases involved in using deep learning to diagnose melanoma.

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 147

Fig. (1). Sample skin image.

Fig. (2). Working model of detection of melanoma.

3. PREPROCESSING Almost every phase of the melanoma detection process uses images from the Kaggle database. As a result, the images must be flawless in every way to

148 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

construct the model more efficiently. Certain pre-processing techniques can be utilized to accomplish this [10, 11]. Image augmentation and scaling are two of these strategies. Almost every phase of the melanoma detection process uses images from the Kaggle database. As a result, the images must be flawless in every way to construct the model more efficiently. Certain pre-processing techniques can be utilized to accomplish this. Reshape, rescale, and image augmentation are three of these strategies and are detailed in Table 1. Table 1. Pre-processing Techniques. Pre-processing

Definition

Output

Reshaping

Process of altering the dimensions of all the images in a dataset so that they are all the same size. The total number of pixels in an image can be increased or decreased when scaling it.

Size = 256*256 dimension.

Rescale

Done to convert all the pixel values between 0 and 1. The value of the pixel is between the value of 0 to 255. For doing resizing, all the pixel values are divided by 255.

Pixel value = 0 to 1

Augmentation Process of creating more images using the existing ones. Done to Flip in both horizontal make the model more effective. In this work, augmentation and vertical directions, techniques are used. rotation by random degrees.

The total of 2900 images is then divided into three parts. 70% of the total images, i.e., 2030 images, are used to train the deep learning model. Twenty percent of the image, or 580 images, are used to validate the model. The rest of the images, i.e., the 290 images, are used to test the accuracy, and the loss of the model developed using the convolutional neural network algorithm. 4. PREDICTION METHODS A computer-generated categorization system for photographs, texts, and even sounds is known as a deep learning model. This deep learning model may be created using a variety of deep learning approaches. Among the algorithms are convolutional neural networks, radial basis function networks, recurrent neural networks and transfer learning. Convolutional neural networks were used to create the deep learning model in this study. This is because the CNN algorithm produces more accuracy than many other deep learning algorithms. An image is sent into a Convolutional Neural Network, which assigns meaning to various objects or characteristics of the image and subsequently distinguishes between them [12]. One of the primary advantages of the CNN technique over

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 149

other deep learning-based classification algorithms is that it does not require extensive preprocessing. The CNN algorithm is based on the Visual Cortex's arrangement, which is similar to the connective theory of neurons in the human brain. The CNN method may also classify images in several convolutional layers, lowering the number of deep learning models needed to perform a single task. One of the best methods for extracting high-level characteristics from an image, such as edges, is the CNN algorithm [13 - 15]. The CNN structure shown in Fig. (3) is the sequence in which the CNN model was built for melanoma detection.

Fig. (3). CNN Architecture.

There are three layers in the above design. They are the pooling layer, the convolutional layer, and a flattened layer, which are all linked together so that convolutional neural networks can process and interpret input to categorize photos. The architecture additionally includes layers such as dense and output in addition to the previously stated layers. The images are then evaluated to see whether there are any signs or symptoms of melanoma. From Fig. (3), it can be seen that the image first passes through the convolutional layer, which contains about 64 filters. The output of the first convolutional layer is then passed through the second convolutional layer, which contains 256 filters. The process is then repeated some other times. Once the image passes all the convolutional layers, it enters the flattened layer and then the dense layer. The combination of both layers then classifies the images based on the requirement. In this research, the images are classified into two types – Benign and Malignant. 5. DEPLOYMENT The mobile app is the most important part of this melanoma detection [16]. This mobile app allows the user to upload an image, and it uses deep learning

150 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

architecture to identify the existence of melanoma and provide the user with the results. The MIT App Inventor was used to create this webpage. A header and accompanying text describe the guidelines for uploading an image on the mobile app. The user may also explore and select the appropriate image. When no picture is uploaded, the home page of the mobile app appears, as illustrated in Fig. (4).

Fig. (4). Mobile app: Home page.

6. RESULT The convolutional neural network technique was used to create a deep learning model. The model is trained, verified, and tested once it is created. After that, the test results are examined using various performance measures. Based on the

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 151

above-mentioned performance measures, a confusion matrix is also created. This phase is required to assess the algorithm's accuracy and loss value. The true result is indicated as affirmative when melanoma is present in a photo, and the model correctly recognizes it. True Negative (TN): When there is no melanoma in a photo, and the model correctly recognizes it, the result is reported as negative. False Positive (FP): When melanoma is present in a photograph, but the model fails to identify it, the true result is displayed as positive. The genuine output is indicated as negative if melanoma is not present in a photo and the program identifies it mistakenly. The accuracy and loss value of both methods are calculated using the given performance criteria. The accuracy curve of the convolutional neural network method during training and validation is shown in Fig. (5).

Fig. (5). Accuracy plot of the CNN algorithm.

From the above graph, it can be inferred that the accuracy of the model is very low, which is almost equal to 60% during the first epoch of training. But the accuracy value increases as the epochs rise and reach the maximum accuracy of almost 95% during the last epoch of training. In the case of validation, the accuracy of the model during the first epoch is somewhere between 75% and 80%. The efficiency of the model also depends on the loss value along with the

152 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

accuracy. Below, Fig. (6) explains the change in the loss value of the CNN algorithm with every epoch of training and validation.

Fig. (6). Loss plot of CNN algorithm.

The loss of the model is quite large, approximately equal to 68% during the first period of training, as seen in the graph above. However, as the number of epochs declines, the loss value lowers until it achieves a minimum of about 18% during the final epoch of training. During validation, the loss value is 50% during the first epoch, and it gradually decreases to 10% during the last epoch of validation. This loss value is so small that it can be neglected, making the model more efficient. A confusion matrix is generated based on the results obtained from testing the model. The confusion matrix consists of four performance metrics named True Positive, False Positive, True Negative, and False Negative. The confusion matrix which was developed for this research is depicted in Fig. (7).

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 153

Fig. (7). Confusion matrix.

From the values of the performance metrics, values of some other parameters are calculated. The CNN model correctly predicted 127 samples as Malignant and 138 samples as Benign. The model wrongly predicted 16 samples as Benign, but it is Malignant. Then 9 samples are predicted as Malignant by the model, but originally it was Benign. The parameters include the accuracy, error rate, True Negative Rate (TNR), True Positive Rate (TPR), False-Negative Rate (FNR), And False-Positive Rate (FPR). The formulae that are used to calculate the following values are as follows. Accuracy = (TP+TN)/(TP+TN+FP+FN)

(1)

Error Rate = (FP+FN)/(TP+TN+FP+FN)

(2)

FPR = FP/FP+TN

(3)

FNR = FN/FN+TP

(4)

TPR = TP/TP+FN

(5)

TNR = TN/TN+FP

(6)

Where FP is the False Positive, TN is the True Negative, FN is the False Negative, and TP is the True Positive value. The correctly predicted disease by the developed CNN model is indicated by the matrix TPR and TNR. The wrongly predicted disease by the developed model is FPR and FNR. The values of the above metrics are tabulated in Table 2.

154 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

Table 2. Performance metrics of CNN algorithm. Metrics

Values

ACCURACY

91.37931

ERROR RATE

8.6207

TNR

93.87755

TPR

88.81119

FNR

11.18881

FPR

6.122449

The accuracy of the developed deep learning model is 91% which is good enough. Also, the error rate is so small that it can be ignored. The data shown in the above table is represented as a graph in Fig. (8).

Fig. (8). Graphical representation of performance metrics.

As the model reached a higher accuracy, a mobile app was developed using the MIT App Inventor by integrating the deep learning model. The mobile app is used as a bridge between the user and the deep learning model. The screenshot of the mobile app when the mobile app detects melanoma in its serious stage, i.e., the malignant stage, is shown in Fig. (9) .

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 155

Fig. (9). Mobile app: Upload malignant melanoma image.

Whether the mobile app detects melanoma in its initial stage, i.e., the benign stage or not, for doing this verification, the benign image is uploaded onto the mobile app, and the predict button is pressed to make the model classify the skin image. The result obtained during this stage is shown in Fig. (10).

156 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

Fig. (10). Mobile app: Upload benign melanoma.

The image uploaded in the app and the result obtained through the app are benign; all the things are understood by looking at the Fig. . From the above Fig., it can be found that the mobile app is capable of detecting benign. Now it is clear that the mobile app is good for skin cancer prediction. 7. DISCUSSION Web applications and software applications are something that users use in daily life. When an application is capable of predicting skin cancer, it will be really

Deep Learning Model

AI and IoT-based Intelligent Health Care & Sanitation 157

helpful for patients. Also, the accuracy of the constructed model is really high resulting in perfect prediction. CONCLUSION From the Kaggle database, a dataset of 2900 images was collected. The images are then preprocessed using image augmentation. This picture augmentation method includes image rescaling, image sharing, image zooming, and horizontal flipping. The convolutional neural networks technique was used to create a deep learning model. The created model is then trained until it achieves the highest accuracy and lowest error values. The values of the models' training and validation are reviewed. Finally, the model produces a greater accuracy of 91.3 percent and a very low loss value of 8.6 percent in the testing phase. The constructed deep learning model's accuracy percentage is so high, it might be a useful initial-level test in the process of melanoma prediction. To make this research accessible to a normal person, the mobile app is created using the MIT App Inventor. The working of the mobile app is verified by uploading various samples of skin images. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

B. Domingues, J. Lopes, P. Soares, and H. Pópulo, "Melanoma treatment in review", ImmunoTargets Ther., vol. 7, pp. 35-49, 2018. [http://dx.doi.org/10.2147/ITT.S134842] [PMID: 29922629]

[2]

A.M.M. Eggermont, A. Spatz, and C. Robert, "Cutaneous melanoma", Lancet, vol. 383, no. 9919, pp. 816-827, 2014. [http://dx.doi.org/10.1016/S0140-6736(13)60802-8] [PMID: 24054424]

[3]

N. Codella, J. Cai, M. Abedini, R. Garnavi, A. Halpern, and J.R. Smith, "Deep learning, sparse coding, and SVM for melanoma recognition in dermoscopy images", In: International Workshop on Machine Learning in Medical Imaging Springer: Cham, Switzerland, 2015, pp. 118-126. [http://dx.doi.org/10.1007/978-3-319-24888-2_15]

[4]

J. Kawahara, A. Bentley, and G. Hamarneh, “Deep features to classify skin lesions.”; In Proceedings of the 2016 IEEE 13th International Symposium on pp. Biomedical Imaging (ISBI), Prague, Czech

158 AI and IoT-based Intelligent Health Care & Sanitation

Madiwal et al.

Republic, pp. 1397–1400, 2016. [http://dx.doi.org/10.1109/ISBI.2016.7493528] [5]

E. Ayan, and H.M. Ünver, “Data augmentation is important for the classification of skin lesions via deep learning.”; In 2018 Electric Electronics, Computer Science, Biomedical Engineerings’ Meeting. IEEE, 2018, pp. 1-4. [http://dx.doi.org/10.1155/2022/1797471]

[6]

A. Esteva, B. Kuprel, R.A. Novoa, J. Ko, S.M. Swetter, H.M. Blau, and S. Thrun, "Dermatologistlevel classification of skin cancer with deep neural networks", Nature, vol. 542, no. 7639, pp. 115-118, 2017. [http://dx.doi.org/10.1038/nature21056] [PMID: 28117445]

[7]

A.A.A. Setio, F. Ciompi, G. Litjens, P. Gerke, C. Jacobs, S.J. van Riel, M.M.W. Wille, M. Naqibullah, C.I. Sanchez, and B. van Ginneken, "Pulmonary nodule detection in ct images: false positive reduction using multi-view convolutional networks", IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1160-1169, 2016. [http://dx.doi.org/10.1109/TMI.2016.2536809] [PMID: 26955024]

[8]

H.C. Shin, H.R. Roth, M. Gao, L. Lu, Z. Xu, I. Nogues, J. Yao, D. Mollura, and R.M. Summers, "Deep convolutional neural networks for computer-aided detection: Cnn architectures, dataset characteristics, and transfer learning", IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1285-1298, 2016. [http://dx.doi.org/10.1109/TMI.2016.2528162] [PMID: 26886976]

[9]

Ronneberger, P. Fischer, and T. Brox.: U-net: “Convolutional networks for biomedical image segmentation., International Conference on Medical Image Computing and Computer-Assisted Intervention Springer, 2015, pp. 234-241. [http://dx.doi.org/10.1007/978-3-319-24574-4_28]

[10]

S. P. Yadav, and S. Yadav, "Mathematical implementation of fusion of medical images in continuous wavelet domain." J. Adv. Res. Dynam. Contro. Sys, vol. 10, no. 10, pp. 45-54, 2019.

[11]

S. Prakash Yadav, and S. Yadav, "Fusion of Medical Images in Wavelet Domain: A Hybrid Implementation", Comput. Model. Eng. Sci., vol. 122, no. 1, pp. 303-321, 2020. [http://dx.doi.org/10.32604/cmes.2020.08459]

[12]

Y. Sun, B. Xue, M. Zhang, G.G. Yen, and J. Lv, "Automatically Designing CNN Architectures Using the Genetic Algorithm for Image Classification", IEEE Trans. Cybern., vol. 50, no. 9, pp. 3840-3854, 2020. [http://dx.doi.org/10.1109/TCYB.2020.2983860] [PMID: 32324588]

[13]

T. Kaur, and T.K. Gandhi, "Deep convolutional neural networks with transfer learning for automated brain image classification", Mach. Vis. Appl., vol. 31, no. 3, p. 20, 2020. [http://dx.doi.org/10.1007/s00138-020-01069-2]

[14]

C. Affonso, A.L.D Rossi, F.H.A Vieira, and A.C.P.D.L.F.D Carvalho, "Deep learning for biologicalimage classification", Expert Syst. Appl., vol. 85, pp. 114-122, 2017. [http://dx.doi.org/10.1016/j.eswa.2017.05.039]

[15]

M. Abdolshah, M. Teimouri, and R. Rahmani, "Classification of X-Ray images of shipping containers", Expert Syst. Appl., vol. 77, pp. 57-65, 2017. [http://dx.doi.org/10.1016/j.eswa.2017.01.030]

[16]

M.J. Taylor, J. McWilliam, H. Forsyth, and S. Wade, "Methodologies and website development: a survey of practice", Inf. Softw. Technol., vol. 44, no. 6, pp. 381-391, 2002. [http://dx.doi.org/10.1016/S0950-5849(02)00024-1]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 159-175

159

CHAPTER 11

Feature Extraction and Diagnosis of Dementia using Magnetic Resonance Imaging Praveen Gupta1,*, Nagendra Muthukumar Subramanian4

Kumar2,

Ajad2,

N.

Arulkumar3

and

Department of Computer Science and Engineering, GITAM School of Technology, GITAM (Deemed to be University), Visakhapatnam (Andhra Pradesh)- India-530045 2 Department of Electronics and Communication Engineering, S J C Institute of Technology, Chickballapur, Karnataka-562101, India 3 Department of Statistics and Data Science, CHRIST (Deemed to be University) Bangalore, Karnataka, India - 560029 4 SRM Institute of Science & Technology, Tiruchirappalli Campus, (Deemed to be University), Trichy, Tamilnadu, India - 621105 1

Abstract: Dementia is a state of mind in which the sufferer tends to forget important data like memories, language, etc.. This is caused due to the brain cells that are damaged. The damaged brain cells and the intensity of the damage can be detected by using Magnetic Resonance Imaging. In this process, two extraction techniques, Gray Level Co-Occurrence Matrix (GLCM) and the Gray Level Run-Length matrix (GLRM), are used for the clear extraction of data from the image of the brain. Then the data obtained from the extraction techniques are further analyzed using four machine learning classifiers named Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), and the combination of two classifiers (SVM+KNN). The results are further analyzed using a confusion matrix to find accuracy, precision, TPR/FPR – True and False Positive Rate, and TNR/FNR – True and False Negative Rate. The maximum accuracy of 93.53% is obtained using the GLRM Feature Extraction (FE) technique with the combination of the SVM and KNN algorithm.

Keywords: Confusion matrix, Dementia, Extraction techniques, Magnetic resonance imaging.

Corresponding author Praveen Gupta: Department of Computer Science and Engineering, GITAM School of Technology, GITAM (Deemed to be University), Visakhapatnam (Andhra Pradesh)- India-530045; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

160 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

1. INTRODUCTION Dementia is a brain disorder that is caused due damage to brain cells. When the brain cells are damaged, it becomes uneasy for them to communicate with each other, resulting in mental disorders such as memory loss and forgetting the way to communicate with other people. Sometimes, it also leads to a lack of problemsolving and quick-thinking abilities. In India, almost 5% of people above the age of 60 suffer from dementia or even the severe stage of dementia, i.e., Alzheimer’s [1]. Dementia is a chronic disorder that can only be treated, and there is no cure. It also may last for a lifetime, even if it is treated properly. Studies state that there are chances for the risk of dementia to double in the upcoming twenty years. By 2040, about 80 million people will be affected by dementia. Especially in western countries, dementia is the worst-case, i.e., Alzheimer’s disease will be one of the most common diseases, with about 60% of the cases [2]. The main cause of dementia is damage to the brain cells. The brain cells are damaged due to the lack of supply of oxygen for a long period. Though people don’t realize the impact of this brain cell damage, when this condition is prolonged for a long time, it leads to dementia and Alzheimer’s. The initial stages of dementia include a lack of interest in performing everyday tasks, frequent memory issues, confused behavior, etc.. Normally, dementia is not a hereditary disease as it does not affect the genetic parts of the human brain. But some rare mutations of dementia can be inherited by the offspring of the patient. Even in that case, dementia will not be as worse as it was with the successor. Dementia is a disorder that often affects people aged more than 65. Due to the increase in population in the past few decades, scientists have already predicted that the population will rapidly increase in the next few decades. The increase in population also leads to an increase in people whose age is above 65. By the end of 1950, 8% of the whole population of the United States consisted of people aged above 65; by 1995, this 8% increased to 12.8%. With this data, it can be predicted that the percentage of people whose age is greater than 65 will be 20% of the total population. In that 20% of the people, there are chances for 35% of the total elderly people to get affected by dementia and Alzheimer’s. About 0.6 percent of the total population has been affected by dementia in the last decade, and this percentage is again increased to 0.9 percent [3]. Dementia is also linked with some other chronic disorders like Traumatic, and Parkinson's. Dementia is an irreversible condition that should be diagnosed in the initial stages. The following tests are done in the initial stages of dementia to diagnose the intensity of the disorder. First, cognitive and neurological tests are taken to evaluate the thinking and problem-solving abilities of the person. If the person is diagnosed with the illness of fundamental activities from the cognitive and neurological test, then several brain scans are done to check whether there is any damage in the brain cells. These brain scans involve the CT scan, which is an

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 161

updated version of an X-Ray; the MRI scan, which uses magnetic resonant images, and the PET scan, which uses a radioactive tracer to detect damage in the brain cells. If any brain cells are found damaged in the above-mentioned scans, then some genetic blood tests are taken to ensure the presence of dementia. The process of magnetic resonance imaging can be done by various combinations of extraction techniques and machine learning classifier algorithms. The extraction techniques include the GLCM and the GLRM. Both methods will be further analyzed throughout the article and the method whose performance index is high will be found. The data obtained from the extraction technique is then classified using some machine learning classifying algorithms. Those algorithms are the SVM algorithm, KNN algorithm, RF algorithm, and the combination of SVM and KNN. These four algorithms are also implemented, and the best classifier algorithm is found. The usage of machine learning algorithms and extraction techniques is not new in the medical field. The volumetric usage of magnetic resonance imaging in the study of patients with dementia, vascular dementia, and in the worst case, Alzheimer’s. According to them, the MRI scan can be used to detect the presence of a pathological body called the Lewy body. The Lewy bodies are present in the striatum part of the brain. The final finding of the research states that the usage of additional algorithms like neuroimaging techniques can be used in the detection of dementia, Alzheimer’s, and vascular dementia [4]. The extraction of data is done using texture-based extraction techniques. The support vector machine integrated with some other pasting votes techniques can be used to detect the presence of dementia in its early stages. At first, the MR images of the patient’s brain are taken from the database, and the wavelet features are extracted from the image using SVM. The diagnosis of dementia using this method includes three steps. They are feature selection, extraction, and classification. The Morphometry technique that uses Voxel’s method is used in this research for FE. Finally, the classification is done using the SVM technique [5]. Another research states the usage of another classification method, i.e., the K-Nearest Neighbor method, for classification purposes. In that research, they state that the KNN algorithm is one of the easiest and simplest algorithms. The distance formula by Minkowski is used for classification in this research. Along with the KNN algorithm, the Support Vector Machine Algorithm and the Linear Discriminant Analysis method are used in the classification process. The following journal is explained in the following way: The second part of the journal explains the materials and methods used to detect dementia using MRI, along with the block diagram and its explanation. The third part explains the extraction techniques that are used in the process. The fourth portion consists of the machine learning classification techniques, formulae, and expressions used to classify. The fifth part analyses the outputs obtained from different combinations

162 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

of extraction techniques and classification algorithms. The last part, the sixth part, states the best combination to detect dementia with higher precision and accuracy. 2. MATERIALS AND METHODS The magnetic resonance images are obtained from the dementia dataset from Kaggle with a total of 1002 images. Out of these, 501 images are images of the brains of dementia-affected people and 501 images of the brains of normal people. All the total images are divided into two parts in the ratio of 8:2 for learning and testing. Of all the total images, 801 MRI scan images are used for training, and 201 MRI scan images are used for testing purposes. The MRI scan images are further processed using some prescreening techniques like resizing, conversion and filtering. The further processing and classification of the image are explained below in Fig. (1).

Fig. (1). Flow Diagram of detection of dementia using MRI.

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 163

3. FE TECHNIQUES FE is a process in which certain linear combinations are constructed using continuous features. The features should have a good discriminatory power between classes. This FE is used to convert complicated machine learning systems to simpler ones. The original input variable is represented as every variable in feature space in a linear combination. This is also utilized to decrease the number of existing features in the older dataset by manipulating the features of the dataset by combining or deleting a few features for easier analysis of the dataset. On the whole, the main task of feature selection is to find the potential set of features from the database containing a lot of unnecessary data [7]. There are many FE techniques, such as independent component analysis, kernel principal component analysis, multifactor dimensionality reduction, partial least squares, etc.. In this journal, two extraction techniques are named the GLCM method and the GLRM method. Both extraction techniques are explained further below. 3.1. GLCM The Gray Level Co-Occurrence matrix method uses a histography image to extract the main features from the magnetic resonance image. A GLCM histogram image is created when two grayscale value occurs at the same time on the image. The GLCM works by determining how often a pixel with the given intensity appears next to another pixel with the same direction and orientation. To determine how often pairs with pixels with certain values and a defined spatial relationship occurs in a picture, the GLCM functions calculate the frequency of these pairing of pixels, create a GLCM, and then extract statistical measures from the matrix. Those matrix entries include the second-order statistical probabilities depending on the gray value of both the rows and columns. If the intensity ranges are wide, the size of the transition matrix will be large. This statistical approach generates the confusion matrix of each picture in the database [8]. Using the GLCM technique, about 28 types of different texture descriptors can be formed [9]. In this research, about 19 texture descriptors of the GLCM are used. They are some average, sum variance, homogeneity, contrast, inverse difference, cluster prominence, sum entropy, information measure of correlation, maximum probability, the sum of squares, normalized inverse different moment, difference variance, autocorrelation, dissimilarity energy, normalized inverse difference, correlation, cluster shade, entropy, and difference entropy. By the efficient usage of these texture extraction features of GLCM, an image with the necessary features only can be obtained. An image’s (ac,bc) coordinates are put in the middle of the neighbor window N (ac,bc). This matrix is defined by the total number of gray level pixels at the

164 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

direction in the GLCMp,θac,bc matrix. Here p is the distance between the total occurrences in the matrix. ݂ሺܽͳ ǡ ܾͳ ሻ ൌ ‫ݔ‬ǡ݂ሺܽʹ ǡ ܾʹ ሻ ൌ ‫ݕ‬ ‫ۓ‬ ‫݁ݎ݁ܪ‬ ۖ ሺܽʹ ǡ ܾʹ ሻ ൌ ሺܽͳ ǡ ܾͳ ሻ ൅ ሺ‫ߠݏ݋ܿ݌‬ǡ ‫ߠ݊݅ݏ݌‬ሻ ‫ ۔‬ሺܽ ǡ ܾ ሻܽ݊݀ሺܽ ǡ ܾ ሻܹ߳ሺܽ ǡ ܾ ሻ ͳ ͳ ʹ ʹ ܿ ܿ ۖ ‫ە‬ 

(1)

Once [1] is done, the pixel’s feature vector is the sum of all weighted chosen features across all weighted frames that compromise that pixel. The total pixel value is given by: ௡

݂ሺ‫݌‬ሻ ൌ ෍ ‫ݓ‬ሺ‫݌‬ሻǤ ݂௫ ሺ‫ܥ‬௔ ሻ

(2)

௔ୀଵ

In the above expression, the values represent fx(.) → The xth element of the feature vector Ca → The pixel at the center wp→The weight allied to pixel p n → Total number of overlapping windows. 3.2. GLRM The GLRM is expanded as a gray-level Run-Length Matrix method. This method uses a two-dimensional histography image of the initial image whose features are to be extracted. In texture analysis, the gray-level run-length matrix may be used to extract the texture characteristics. For instance, the GLRM technique can be used to extract higher-order texture statistics. As an example, the gray-level runlength matrix contains a stripe of pixels that are traveling in a definite direction. With the same level of intensity, the run time of the GLRM technique is determined by how many pixels are there in the image. It’s termed the run-length rate. The run length is the number of pixels in a similar definite direction in which the adjacent pixels take a similar gray intensity [10]. In this research, about seven types of FE techniques of the GLRM method are used. They include Run Length Nonuniformity, Low Short and Long Run Emphasis, High Gray-Level Run

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 165

Emphasis, Run Percentage, Short Run Low and High Gray-Level Emphasis, Gray-Level Nonuniformity, Long Run Low and High Gray-Level Emphasis. 4. CLASSIFICATION TECHNIQUES Once the required features are retrieved from the magnetic resonance images, certain machine-learning classification techniques are applied to the featured images containing dementia and the normal images. Several techniques can be used for this purpose, and four classification techniques are used for this research. Those four algorithms are more clearly explained below. 4.1. Support Vector Machine (SVM) The SVM algorithm is the most popular classification method that can be used to solve big data categorization challenges in less time. This traditional machine learning method is a very useful technique today. The multidomain applications in the big data environment can be solved using SVM in particular. An SVM training method creates a model that forecasts the class of a novel sample based on a set of training samples, each of which has been designated as fitting to one of the many classes. When it comes to statistical learning, SVM is better at generalizing problems. Using the statistical learning theory, making predictions and decisions is easier with the SVM algorithm. When it comes to image categorization, the support vector machine is one of the simplest algorithms to implement. In the SVM method, there are two kinds of image classification. They are linear and non-linear patterns. Linear patterns are those that are easily identifiable or that can be simply divided into images of small dimensions. Nonlinear image patterns are those that are not easily distinguishable or cannot be separated easily [11]. The formula that is used for SVM modeling is as follows. ௟

ͳ ௡ ோ ௡ ሺ߱ ሻ ߱ ൅ ܲ ෍ ߦ௜௡  ʹ

(3)

௜ୀଵ

ሺ߱௡ ሻோ ߶ሺ‫ܣ‬௜ ሻ ൅ ܾ ௡ ൒ ͳ െ ߦ௜௡ ǡ݂݅‫ݕ‬௜ ൌ ݊ ሺ߱௡ ሻோ

௡

߶ሺ‫ܣ‬௜ ሻ ൅ ܾ ൑ ͳ െ

ߦ௜௡ ǡ݂݅‫ݕ‬௜

ߦ௜௡ ൒ Ͳǡ ݅ ൌ ͳǡ ǥ Ǥ Ǥ ǡ ݈

്݊

(4) (5) (6)

௟

ͳ ௡ ோ ௡ ሺ߱ ሻ ߱ ൅ ܲ ෍ ߦ௜௡  ʹ ௜ୀଵ

(7)

166 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

In the expression, Ai - Attribute data ϕ,P - Penalty value It becomes difficult for the images to be separated linearly because of the scattering of the data. To overcome this, SVM adds a kernel function K(An,A), which converts the actual data into a new space and returns it to the application. The transformation technique is combined with the dot product ØA. To this, the data must be transcribed into a higher dimension. The kernel can be expressed as the below equation. ൫୬ǡ ୧ ൯ ൌ ‫׎‬ሺ୬ ሻ‫׎‬ሺ୧ ሻ

(8)

Using the equation of the kernel that is shown above, the formula for the hyperplane can be found. The formula for the hyperplane is given: ‫ܭ‬൫‫ܣ‬௡ǡ ‫ܣ‬௜ ൯ ൌ ‡š’ሺെߛԡ‫ܣ‬௡ െ ‫ܣ‬௜ ԡଶ ൅ ‫݋‬ሻ

(9)

In the above expression, o and γ → Parameters that are used for the optimization An → Support vector data Finally, the categorization of the features using the SVM is done using the expression given below. ெ

݂ሺ‫ݖ‬௜ ሻ ൌ ෍ ‫ן‬௠ ‫ܮ‬௠ ‫ܭ‬൫‫ݖ‬௠ǡ ‫ݖ‬௜ ൯ ൅ ܾ ௠ୀଵ

In the above expression, Ln= Data that is labeled m= Legrange multiplier

(10)

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 167

4.2. K-Nearest Neighbor (KNN) The KNN algorithm is again an easy categorization algorithm that is utilized for data mining. It has been used for a long time. The main flaw of this algorithm is that it is inefficient and imprecise when dealing with large amounts of data. For a long time, data preparation approaches like dimension reduction and missing data replacement have been aimed at the KNN method's effectiveness. As a result, the k-nearest neighbor approach is currently used to detect and repair faulty data, such as eliminating noise and duplicate samples or imputing incomplete data. KNN converts large data into small required data, which is data of good quality to yield excellent results from any other technique in machine learning [12]. With the increasing complexity of the dataset, the efficiency of most advanced algorithms that use data structures like trees rapidly, and tree structures have diverse complexity levels across diverse datasets [13]. The first research direction against the imputation of the k-nearest neighbor algorithm improved all the efforts in a specific direction. An actual selection of the k-nearest neighbor method is used to determine the direction. Different distance functions or weighing techniques or both can produce k nearest neighbors that are different from each other. In any case, the goal is to find a machine that emphasizes some attributes while minimizing the impact of the rest on the available dataset [14]. These K training samples represent the K nearest neighbors of the unidentified collections. Certain distance measures, such as Euclidean, were used to quantify the resemblance between the collections. The KNN categorization assigns the second most common class among these unknown collections. In the case of k=n, the unidentified collection is assigned to the pattern field category of the training dataset that grows in size as the number of input values increases. Hence the K value increases in size as well [15]. The K-Nearest Neighbor algorithm is implemented using the following steps.

௡

‫ ܦ‬ൌ ඩ෍ሺܽ௡ െ ܾ௡ ሻଶ

(11)

௜ୀଵ

i. The Euclidean distance is calculated using the given expression ii. The found distance values are arranged in increasing order. iii. Use the preceding arrangement of the distance values to calculate the distance starting from the very first one. iv. Next, estimate the k points using the k distances. v. Find the images in which dementia is detected.

168 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

vi. Evaluate the other normal images to check whether they don’t fit the condition of dementia. 4.3. Random Forest (RF) The random forest algorithm is a machine-learning classification algorithm that uses several decision trees to find the optimal result. Initially, the random forest algorithm is the collection or ensemble of classification and regression trees (CART) that have been trained on data sets of the same as the training sets. In other words, it is the generalized prediction of the error. A study by Breiman in the year 1996 proved that the out-of-bag error is the same for a test set of the same size as the training set. This eliminates the need for a separate test set when using the OOB estimate. A CART tree consists of different classes and the forest predicts which of those classes will get the most votes [16]. The random forest algorithm is the more efficient algorithm for predicting a task. This is because linear regression is based on the linearity assumption. This assumption simplifies the model’s interpretation, but it’s not often flexible enough to make accurate forecasts it. Data non-linearities are easily accommodated by random decision forests, which tend to perform better than linear regression. Machine learning algorithms such as random forests, on the other hand, are very much effective in handling large datasets [17]. Not all predictor values are used at the same time. Each tree is replaced by a sample that is drawn from the sampling set in a forest building process, about 1/3 of the original instances are left out. Out-of-bag data refers to this set of instances [18]. Using OOB error estimation, each tree in the forest has its OOB dataset. The expression for the generalization error is as follows. GE * = P x,y (mg(A,B)) < 0

(12)

A margin function measures how much the right classes (x,y) the average number of votes exceeds the average vote of any other class. The expected value of the margin function is used to determine the strength of the random forest S = E A, B (mg (A, B))

(13)

If ρ is the average value of the distance between the trees, then the maximum value of the generalization error is given as, GE* ≤ ρ (1 – s 2) / s2

(14)

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 169

To improve the accuracy of this algorithm, the decision trees of the algorithm, the trees must be accurate and diverse to provide better results. 4.4. Proposed Method The proposed technique starts with the KNN classifier and then moves on to the SVM classifier. The procedure includes using KNN Classification on samples with the same class as the first 30 neighbors. Following that, samples with different first 30 neighbors are sent to the SVM classifier. The SVM classifier is trained with its first 30 neighbors in the neighborhood of the total training sample for the SVM classifier of each observation. Precomputed kernels are utilized for the SVMs to improve and speed up categorization. The SVM classifiers learned using this method, on the other hand, are only trained on the first 30 neighbors, which is a tiny quantity. Furthermore, the first 30 neighbors’ class labels may be randomized. Increase the number of training samples of its SVM. It effectively enhances the performance of the discriminative nearest neighbor classification approach. For SVM training, the first 30 neighbors of each category instead of the first 30 neighbors of the whole training set are utilized. This approach consistently gives superior outcomes. 5. RESULT AND DISCUSSION The images of the normal brain and dementia-affected brain are first obtained from the Kaggle database. A total of 1002 images are taken for this research. The total images are then divided into two in the ratio of 8:2. The 8 parts of the total image, i.e., 801 images, are used for training the combination of extraction techniques and classifier models. The rest of the images (201) are used to test the same combination. Once all the images undergo the FE and the classification process, it is necessary to compare the efficiency of all the combinations to find the better one. This comparison is made using a confusion matrix. Certain performance metrics like accuracy, TNR, TPR, FNR, FPR, and precision. Those parameters are more clearly explained below. True Positive or TP: If dementia is present in an image and the model senses it correctly and shows the actual output as positive. True Negative or TN: If dementia is not present in an image and the model senses it correctly and shows the actual output as negative. False Positive or FP: If dementia is present in an image and the model does not

170 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

sense it wrong and shows the actual output as positive. False Negative or FN: If dementia is not present in an image and the model senses it wrong and shows the actual output as negative. Accuracy: The proportion of right dementia predictions to the total of correct and incorrect predictions. The statement is expressed below the expression.

‫ ݕܿܽݎݑܿܿܣ‬ൌ

ܶܲ ൅ ܶܰ ‫ͲͲͳ כ‬ ܶܲ ൅ ܶܰ ൅ ‫ ܲܨ‬൅ ‫ܰܨ‬

(15)

True Positive Rate: The proportion of exact dementia forecast to the total of right dementia and incorrect healthy projection, typically stated in the formula. ܴܶܲ ൌ

ܶܲ ‫ͲͲͳ כ‬ ‫ ܰܨ‬൅ ܶܲ

(16)

True Negative Rate: There is a formula to express the proportion of correct normal forecasts to all correct normal forecasts and incorrect dementia forecasts by the model. ܴܶܰ ൌ

ܶܰ ‫ͲͲͳ כ‬ ܶܰ ൅ ‫ܲܨ‬

(17)

False Positive rate: The proportion of incorrect dementia diagnosis to the sum of correct healthy diagnosis and incorrect dementia diagnosis. This is stated as follows: ‫ ܴܲܨ‬ൌ

‫ܲܨ‬ ‫ͲͲͳ כ‬ ܶܰ ൅ ‫ܲܨ‬

(18)

False Negative Rate: The proportion of incorrect healthy diagnoses to total right dementia and incorrect healthy forecast by ML model, which can be stated as: ‫ ܴܰܨ‬ൌ

‫ܰܨ‬ ‫ͲͲͳ כ‬ ‫ ܰܨ‬൅ ܶܲ

(19)

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 171

Precision: The proportion of appropriate dementia forecast to overall correct and incorrect dementia forecast. The formula for calculating the model's precision is given here. ܲ‫ ݊݋݅ݏ݅ܿ݁ݎ‬ൌ

ܶܲ ‫ͲͲͳ כ‬ ‫ ܲܨ‬൅ ܶܲ

(20)

Table 1 analyses the performance metrics from the confusion matrix for the classification algorithms like SVM, KNN, RF, and the combination of SVM and KNN for the gray-level co-occurrence matrix extraction technique. Table 1 represents the performance metrics for the same classification algorithms but a different extraction technique, i.e., the GLRM method. Table 1. Evaluation of dementia images using the GLCM FE with various ML models. Classifier

Accuracy

TNR

TPR

FNR

FPR

Precision

SVM

91

87.962

94.565

5.434

12.037

87.000

KNN

87.562

84.259

91.397

8.602

15.740

8.333

RF

89.054

87.037

91.397

8.602

12.962

85.858

SVM+ KNN

92.537

88.888

96.774

3.225

11.111

88.235

Table 2. Evaluation of images with dementia using GLRM FE with the various ML model. Classifier

Accuracy

TNR

TPR

FNR

FPR

Precision

SVM

90.046

91.111

89.189

10.810

8.888

92.523

KNN

88.557

88.888

88.288

11.711

11.111

90.740

RF

89.552

87.777

90.990

9.009

12.222

90.178

SVM+ KNN

93.532

92.222

94.594

5.405

7.777

93.75

From Table 1 and Table 2, it can be inferred that the gray-level run-length matrix method is more efficient in detecting images with dementia. In the GLRM, the maximum accuracy of 93% is obtained by the SVM +KNN algorithm. The least accuracy is obtained by the KNN algorithm. This statistic is explained pictographically using a bar graph in Figs. (2 and 3).

172 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

Fig. (2). Graphical analysis of images with dementia identification using GLCM FE with the various ML model.

From Table 1 and Fig. (2), it can be found that the combination of the SVM and the KNN algorithm provides the best results for almost every performance metric. The maximum accuracy of 92.5% is obtained from this algorithm. The SVM algorithm is the next efficient algorithm, with an accuracy percentage of 91%. The third most efficient algorithm for the GLCM is the RF algorithm, with an accuracy percentage of 89.05%. The weakest algorithm is the KNN algorithm, with an accuracy of 87.5%. Both the accuracy and the precision of this combination are relatively higher than all the other classifying algorithms.

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 173

Fig. (3). Graphical analysis of images with dementia identification using GLRM FE with the various ML model.

From Table 2 and Fig. (3), it can be found that the combination of SVM and KNN algorithm is again the best algorithm for dementia detection. The accuracy rate of this algorithm is about 93.5%, and the precision is 93.75%. The second most efficient algorithm is the SVM algorithm with an accuracy percentage of 90.04%. Next to SVM, the RF algorithm is the most efficient, with an accuracy percentage of 89.5%. The KNN algorithm is said to be the weakest algorithm for dementia detection, with an accuracy rate of 88% and a precision range of 90.7% 6. PROPOSED IDEA The accuracy of the model developed in this research is typically higher than the pre-existing model, which makes this model more efficient and reliable. The different combinations of feature extraction techniques and classifiers were used in this research to identify the best model for detecting dementia. The hybrid classifier is employed in this work to get better accuracy. The hybrid technique starts with the KNN classifier and then moves on to the SVM classifier. CONCLUSION Though dementia is not a fatal disease, it is a chronic disease that can disturb the normal lives of human beings. Due to increased stress levels, people are more

174 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

likely to be affected by dementia in the next few decades. As wise people say, prevention is always better than cure. Rather than trying some treatments or remedies to cure dementia, it is always easier to regularly check the brain functions for such internal damage or its symptoms. In this research, about 1002 images with images containing dementia-affected brains and normal brains have been collected from the Kaggle database for training and testing. The main features of the images are then extracted from the images using two extraction techniques – GLCM and GLRM. The main features are then classified using four machine-learning classification algorithms. The four algorithms are SVM, KNN, RF, and the combination of SVM and KNN. The outputs of all the techniques are further analyzed using the confusion matrix. It also used the TPR, TNR, FPR, FNR, accuracy, and precision to find the best combination of extraction techniques and the classifying algorithm. In the end, it is found that the combination of GLRM with the SVM+KNN technique will provide the best results for detecting dementia images and normal ones, with an accuracy percentage of 93.5. The second-best accurate combination is the pair of SVM and KNN with the GLCM extraction technique. Using this theory, the best combination of extraction techniques and the classifying algorithm for the diagnosis of dementia can be found. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

K.S. Shaji, V.P. Jithu, and K.S. Jyothi, "Indian research on aging and dementia", Indian J. Psychiatry, vol. 52, no. 7, suppl. 1, p. 148, 2010. [http://dx.doi.org/10.4103/0019-5545.69227] [PMID: 21836672]

[2]

L. Rizzi, I. Rosset, and M. Roriz-Cruz, "Global epidemiology of dementia alzheimer’s and vascular types", In: Biomed Res Int, 2014. [http://dx.doi.org/10.1155/2014/908915]

[3]

M.F. Mendez, and J.L. Cummings, Dementia: a clinical approach. Butterworth-Heinemann, 2003.

[4]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images", Med. Biol. Eng. Comput., vol. 58, no. 4, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6] [PMID: 31993885]

[5]

R. Barber, I. McKeith, C. Ballard, and J. O’Brien, "Volumetric MRI study of the caudate nucleus in

Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 175

patients with dementia with Lewy bodies, Alzheimer’s disease, and vascular dementia", J. Neurol. Neurosurg. Psychiatry, vol. 72, no. 3, pp. 406-407, 2002. [http://dx.doi.org/10.1136/jnnp.72.3.406] [PMID: 11861709] [6]

D. Bansal, K. Khanna, R. Chhikara, R.K. Dua, and R. Malhotra, "Classification of magnetic resonance images using bag of features for detecting dementia", Procedia Comput. Sci., vol. 167, pp. 131-137, 2020. [http://dx.doi.org/10.1016/j.procs.2020.03.190]

[7]

N.M. Mathkunti, and S. Rangaswamy, "Machine learning techniques to identify dementia", SN Computer Science, vol. 1, no. 3, p. 118, 2020. [http://dx.doi.org/10.1007/s42979-020-0099-4]

[8]

S. Khalid, T. Khalil, and S. Nasreen, “A survey of feature selection and feature extraction techniques in machine learning: In 2014 Science and Information Conference IEEE., pp. 372-378, 2014. [http://dx.doi.org/10.1109/SAI.2014.6918213]

[9]

S. Jafarpour, Z. Sedghi, and M.C. Amirani, "A robust brain MRI classification with GLCM features", Int. J. Comput. Appl., vol. 37, no. 12, pp. 1-5, 2012.

[10]

Z. Xiao, Y. Ding, T. Lan, C. Zhang, C. Luo, and Z. Qin, "Brain MR image classification for Alzheimer’s disease diagnosis based on multifeature fusion", Comput. Math. Methods Med., vol. 2017, pp. 1-13, 2017. [http://dx.doi.org/10.1155/2017/1952373] [PMID: 28611848]

[11]

W.N. Hafsha, W. Kairuddin, and W.M. Hazifah, "W Mahmud: Texture feature Analysis for different resolution level of Kidney Ultrasound Images", In: IOP Conference Series: Material Science vol. 226. , 2017. [http://dx.doi.org/10.1088/1757-899X/226/1/012136]

[12]

A. Pradhan, "Support vector machine-a survey", Int. J. Emerg. Technol. Adv. Eng., vol. 2, no. 8, pp. 82-85, 2012.

[13]

I. Triguero, D. García-Gil, J. Maillo, J. Luengo, S. García, and F. Herrera, "Transforming big data into smart data: An insight on the use of the k-nearest neighbors algorithm to obtain quality data", Wiley Interdiscip. Rev. Data Min. Knowl. Discov., vol. 9, no. 2, p. e1289, 2019. [http://dx.doi.org/10.1002/widm.1289]

[14]

S. Xia, Z. Xiong, Y. Luo, L. Dong, and G. Zhang, "Location difference of multiple distances based knearest neighbors algorithm", Knowl. Base. Syst., vol. 90, pp. 99-110, 2015. [http://dx.doi.org/10.1016/j.knosys.2015.09.028]

[15]

S. Zhang, "Nearest neighbor selection for iteratively kNN imputation", J. Syst. Softw., vol. 85, no. 11, pp. 2541-2552, 2012. [http://dx.doi.org/10.1016/j.jss.2012.05.073]

[16]

J. Han, J. Pei, and M. Kamber, Data mining: concepts and techniques. Elsevier, 2011.

[17]

A. Sarica, A. Cerasa, and A. Quattrone, "Random Forest algorithm for the classification of neuroimaging data in Alzheimer’s disease: a systematic review", Front. Aging Neurosci., vol. 9, p. 329, 2017. [http://dx.doi.org/10.3389/fnagi.2017.00329] [PMID: 29056906]

[18]

M. Schonlau, and R.Y. Zou, "The random forest algorithm for statistical learning", Stata J., vol. 20, no. 1, pp. 3-29, 2020. [http://dx.doi.org/10.1177/1536867X20909688]

176

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 176-190

CHAPTER 12

Deep Learning-Based Regulation of Healthcare Efficiency and Medical Services T. Vamshi Mohana1,*, Mrunalini U. Buradkar2, Kamal Alaskar3, Tariq Hussain Sheikh4 and Makhan Kumbhkar5 Department of Computer Science, RBVRR Women's College, Hyderabad, Telangana- 500027, India 2 Department of Electronics and Telecommunication Engineering, St. Vincent Pallotti College of Engineering & Technology, Nagpur, Maharashtra - 441108 India 3 Bharati Vidyapeeth (Deemed to Be University), Institute of Management, Kolhapur, Maharashtra, India 4 Department of Computer Science, Government Degree College, Poonch, J&K UT-185101, India 5 Christian Eminent College, Indore, Madhyapradesh-452001, India 1

Abstract: There has been an increase in new diseases in recent years, which has had both economic and societal consequences. Patients in the modern environment require not only constant monitoring but also all-encompassing smart healthcare solutions. These systems keep track of the patient's health, store data, and send alerts when critical conditions arise. Healthcare may be considerably improved with the use of Artificial Intelligence and Machine Learning (ML) systems. These systems can help with earlier diagnosis of diseases, as well as more specific treatment plans. As big data, the Internet of Things with many more smart technologies grows more widespread; deep learning is becoming more popular. Due to the apparent rising complexity and volume of data in healthcare, artificial intelligence (AI) will be used more frequently. This work aims to develop a deep learning-based smart healthcare monitoring system. This system keeps track of patients' health, analyses numerous parameters, categorizes data, and organizes requirements. The algorithm using the python program is developed and discussed to track the health of several patients with various illnesses. This method also aids in the categorization of data, organization of pharmacological requirements. This approach yields satisfactory performance, and the results are also provided.

Keywords: Deep Learning, Healthcare, IoT, Regulation.

Corresponding author T. Vamshi Mohana: Department of Computer science, RBVRR Women's college, Hyderabad, Telangana- 500027, India; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 177

1. INTRODUCTION The Healthcare industry has grown tremendously in the last decades, not only in Research and Development but also in data mining and organization. Mobile technologies and smart devices made significant improvements in monitoring and assistance [1]. Due to the supplementary health issues, monitoring the health and organizing the previous data are more critical. Though wearable devices have been given continuous monitoring and organization, those data are not reliable in extreme cases since we require a sophisticated health monitoring and communication system that can also be used to organize the data of a patient and suggest the drug and surgical requirements [2]. The Internet of Things (IoT) improves the performance of medical equipment by reducing human error. In this technique, medical devices transmit patient data to a gateway that can be stored and analyzed. Monitoring all patients from different locations is one of the main impediments to implementing the Internet of Things for healthcare applications [3]. In recent years, Wireless Medical Sensor Network (WMSN) has been gaining attention in remote monitoring system. Implanted biosensors are used in this method to detect sensitive information about a patient that may then be transferred to remote hospitals for additional analysis. Medical professionals can access patient data via WMSN from anywhere in the world with the help of the Internet [4]. Data acquired from individuals and patients require authorization from the system because of the increased sensitivity due to the participation of people. Consumers are deterred from taking full advantage of the technology because of their privacy concerns. To protect the accuracy, security, and privacy of healthcare data, we must develop a security framework for real-time patient monitoring. Healthcare practitioners and rescuers have been using a variety of security frameworks and devices for some time now. However, these systems get increasingly absurd with time. The flow of data and its related elements are mentioned in the Fig. (1).

Fig. (1). Flow of data from the patient to healthcare system.

178 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

Recent decades have seen a shift in the world population's size and composition. For example, such demographic shifts have a profound impact on health and healthcare. Even in wealthy countries, people are living longer and healthier lives. This should be celebrated and seen as an opportunity for people to enjoy longer and healthier lives. As a result, healthcare and living circumstances will need to be significantly improved. Healthcare costs are increasing at an alarming rate. Paradoxically, all nations around the globe are seeing rapid economic growth. Our daily lives are impacted by deep learning to various extents. These are among the most popular applications, and they are constantly evolving [5]. This paper aims to present an online smart patient monitoring system that monitors vital parameters like Heart Rate, SpO2, and Body Temperature. These details are shared with ThingSpeak and a server to store and analyze the data. The algorithm using the python program is developed and discussed to track the health of several patients with various illnesses. This method also aids in the categorization of data, and organization of pharmacological requirements. The output can be viewed using a mobile application also using the ThingSpeak app viewer. This approach yields satisfactory performance, and the results are also provided. 2. IOT IN MEDICAL CARE SERVICES With the arrival of the “Internet of Things,” it is now feasible to distinguish “patient data” and “physical sensor data” in the analysis and diagnosis of a physician. The greatest benefit of the Internet of Things in healthcare is a reduction in maintenance costs, followed by an increased probability of receiving healthcare. The inclusion of a personal and digital healthcare network was a valuable learning experience, and cloud health services were projected to emerge as a result of mobile information and general technology killer applications. The Internet of Things (IoT) is already in use as a primary platform for evaluating neuronal attention. But even though adequate monitoring mechanisms are unavailable, significantly higher risks can be taken. The Internet of Things (IoT) and other innovations are used here. Such caution is in the best interests of the patient. Sensed data are employed to assess patient information. The caregiver is capable of providing sufficient medical advice. IoT devices, which are frequently used by people with disabilities, necessitate additional supervision. The deployment of sensors has been used to collect monitoring tactics to maintain a continuous material movement by the patients referred to the facility for caregivers. As either a result, the treatment quality improves. This leads to higher hospital bills. Fig. (2) shows the impact of the Internet of Things (IoT) on healthcare.

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 179

Fig. (2). Framework of IoT in Health care Medical Services.

There are a variety of wearable gadgets that can provide natural statistics such as calories burnt, steps taken as well as information on heart rate and blood pressure. Incredible and obvious, these devices have a great impact on our users' overall fitness. Some notable devices and their applications are tabulated in a Table 1. Table 1. Wearable Devices and their Work description. Device Name

Working in brief

Electroencephalography Electroencephalography (EEG) measures the brain's activity in real-time. An EEG data gathering system, wireless communication, and analog-to-digital synchronization applications, as well as low-level, real-time signal processing, are all possible with the Wireless Sensors. Pulse Oximetry

Connecting the pump-oximeter, which has an image detector and LEDs, is as simple as sticking a finger or an ear to it. For example, the human body's infrastructure is tested by sending or returning red light to it. Measurement of oxygen saturation was aided by the differentiation between the installation level and the amount of deoxygenated hemoglobin present in the bloodstream. A photoplethysmograph is used to calculate heart rate (PPG).

Electrocardiography

As the heart continues to beat, a waveform monitors it and sends time information. It's also difficult to automate ECG calculations using wireless sensor devices because the technology isn't ready yet.

Blood Pressure

To quantify it, we can look at the energy expended by the blood pumping into the blood arteries. To calculate these types of sensors for hand frame and systolic measurements, oscillometric methods must be applied.

180 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

(Table 1) cont.....

Device Name Electromyography

Working in brief The electrical signals of the muscle are studied in muscle research. In this approach, the EMG signal is a useful tool for observing human muscle activity.

Fig. (3). Global Wearable Devices Revenue-Source-Global data, Technology Center.

The graphical data year-wise from 2018-2023 in Fig. (3) represents the global data on Wearable devices, and it can be seen that its revenue is increasing drastically every year. 3. RELATED WORKS ●

●

Professionals can now check and monitor a patient's disease with simple and good treatment, even in village areas, due to advancements in healthcare. Patients can benefit from health monitoring services that gather and analyze health data to treat a wide range of complex health issues on a big number. Health care monitoring is a valid clinical trial development approach for confirming that participants' health is monitored in compliance with certain handouts. Many strategies [6] for healthcare monitoring may be considered and carried out by specialists in the field. As a result, the proposed research examines the existing deep learning and transfer learning procedures, tactics, and methods in health monitoring. This in-depth investigation will be used to support the creation of new health monitoring systems. The Mask-RNN approach is used to extract several key areas of the patient's body [7]. Using data mining association rules, these critical points are subsequently translated into six main biological organs. Moreover, the data from

Medical Services

●

●

●

●

●

AI and IoT-based Intelligent Health Care & Sanitation 181

the discovered key point coordinates are analyzed to determine the patient's level of discomfort. Finally, normal and unpleasant movements are classified using range and temporal criteria. These important factors are also used to assess a person's bedside posture. The amount of comfort and discomfort is measured by regularly checking the patient's body positions. IoT By merging and splitting technology and resources, health wearable gadgets are addressing new concerns. Health wearable devices may be utilized to check the health of in- and out-patients on a regular and periodic basis. This research proposes the E-Healthcare Monitoring System [8], Internet of Things technological assets coupled to Deep - learning techniques to construct an enhanced automation system. For correct diagnosis, this system will connect, monitor, and make judgments. This paper intends to provide a more efficient method, and it requires all the views of the nutrients to be dissected. Following that, it will use Faster RCNN to identify the nutrients and change items. As the system design [9 - 13] was created and developed in Android, one of the Feature Sets, the random forest calculation method, was to evaluate and recognize the data in the frame to determine the shape of the supplements. The producers can eventually compute the number of calories after examining the volume of nourishment. Identifying fruit and veggies using object identification and deep learning neural network models is proposed in this paper as an effective way of identifying food objects and assessing calorie consumption. The use of strong AI techniques and machine learning algorithms has made it possible to provide individualized care that focuses on offering emotional support targeted to a specific individual. The methods and parameters used in various systems for mental health monitoring [14 - 17], such as virtual consulting, targeted treatment, and diagnostic methods, are reviewed in this article. Finally, we offer a system that combines the aforementioned methodologies to provide customized mental health care. There are numerous flaws in most of the studies on the subject, such as the lack of machine learning, sensor security over the network, and so on. Some of these difficulties have been discussed in this work, along with the system's output. Additionally, based on a combination of machine learning and the existing system, the paper [18 - 22] provides a model for forecasting future disorders in the patient or that remote site. As part of this research [23 - 25], a low-cost method for communicating patient vital signs during an emergency has been proposed. To measure the patient's condition, sensors are connected to a wireless network. Through a Wi-Fi module on the controller, data from the sensors is collected and sent up into the cloud to be processed. These studies are carried out remotely by a physician, who has access to the study data via the internet and input operations. Meanwhile,

182 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

doctors' burden is reduced as a result of the use of remote viewing, and patients receive accurate health status information. If a patient requires immediate medical attention, the doctor is notified of the need for immediate attention. 4. PROPOSED SYSTEM The block diagram of the LSTM-RNN algorithm-based Smart and Intelligent Health care monitoring system is depicted in Fig. (4). This model consists of a Sensor Unit, Arduino Uno board, and a Data management unit. Body temperature, Heart Rate, and Blood Oxygen level Sensors are utilized to measure the patient’s body and health condition. The outputs of these Sensing units were given to the Arduino Uno board. The output of the Arduino board is displayed in ThingSpeak through the internet.

Fig. (4). Block Diagram of the Proposed System.

Long-Short Term Memory (LSTM) algorithm is used to read the patient's condition by analyzing the Heart rate, Oxygen level, and Temperature values driven from the ThingSpeak server and give pharmacological requirements. Python program is used to create the LSTM model and produce suggestions.

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 183

4.1. Interaction of RNN with LSTM During the 1980s, the Recurrent Neural Network (RNN) was initially constructed. An input layer, one or more hidden layers, and an output layer, make up the structure. To store critical information from past processing steps, the RNNs have a chain-like structure made up of repeating units. Instead of relying on a single input, RNNs use a feedback loop. This means that the output of the previous step is transmitted back into the network to impact the outcome of the current step, and each following step. In this way, RNNs have proved successful at learning sequences. RNN uses a backpropagation algorithm to update its weights and calculate gradients. Due to the vanishing and expanding gradient difficulties, this algorithm is not efficient enough to learn a pattern from long-term dependency. An LSTM, or long-short-term memory network, can learn long-term dependencies. Hochreiter & Schmidhuber (1997) introduced them, to address problems of the drawbacks of the RNN by adding additional interactions per cell. In addition, they have a wide range of applications and are now widely used. To avoid the issue of long-term tolerance, LSTMs are explicitly designed. Not only does it come naturally to them, but it's not something they struggle to learn! In all recurrent neural networks, the modules of the neural network are repeated. In conventional RNNs, this repeating module will have a relatively simple structure, such as a single tan-h layer. Fig. (5) shows the architectural layer of LSTM followed by the fundamental equations of LSTM given by (1)-(6). ~

Ct = tanh (xtUg+at-1Wg+B) ~

Ct = σ (ft*Ct-1+itCt)

at = σ (tanh(Ct)*Ot) It = σ (xtUi+at-1Wi) ft = σ (xtUf+at-1Wf) Ot = σ (xtUo+at-1Wo)

Here, i, f, o are called the input, forget, and output gates, respectively.

(1) (2) (3) (4) (5) (6)

184 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

Fig. (5). Long-Short Term Memory (LSTM) Architecture.

In a conventional LSTM network, cells are used to store data. As a result of this, two different states are passed to the next cell: the cell state and a hidden state. Information can pass forward in a largely unaltered manner because of the cell state, which is the fundamental data chain. Some linear changes, on the other hand, are possible. Cell state data can be added or withdrawn using sigmoid gates. Each weight in the gate is equivalent to the weight in a layer or a set of matric operations in a computer program. Because LSTMs use gates to manage the memorization process, the long-term reliance problem is avoided.

Fig. (6). Layer Structure of LSTM.

Apart from the input, hidden and output layers of normal Neural networks, the Recurrent layer LSTM, made up of memory units, is called the LSTM (). The

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 185

fully connected layer that commonly precedes LSTM layers and is used to predict the output is called Dense (). As an alternative, we can create an array of levels and provide it to the function Object () of Sequential parameters. Fig. (6) shows the layer structure of the LSTM model. 5. RESULTS AND DISCUSSION Fig. (7) depicts the project's flow chart, in which the sensors' data is read, sent to the ThingSpeak server, displayed, and stored in the cloud for future use. If the sensors detect abnormal conditions, a warning message is delivered to the patient immediately, advising them to see a doctor. LSTM model is used to analyze the data received from the ThingSpeak server, and the model predicts the conditions and provides general suggestions to be carried out.

Fig. (7). Flowchart of Smart Health Monitoring System.

186 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

Fig. (8) shows the real-time SpO2 value of a patient that is taken through the Arduino board. SpO2 value is normal at 96% for a certain period and from 96% to 94% after that. The heart rate bpm values are shown in Fig. (9) for the same period, and the values are steady. The body temperature value is shown in Fig. (10), and the value is also steady at 100°F. All three values are monitored, and if any values are considered critical, past and present values are analyzed by the LSTM model to give suggestions or alarms about the situation.

Fig. 8. SpO2 Live Monitoring Output.

Fig. (9). Heart Rate Live Data.

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 187

Fig. (10). Body Temperature Live Monitoring.

Fig. (11) shows the conditions received from the ThingSpeak server. All three values are analyzed by the LSTM model and the general suggestions are provided to the patient and physician. All the values are normal for a certain period, the suggestion is provided accordingly. Since there is a sudden drop in SpO2 value and the temperature is also increasing gradually, the LSTM model considers this situation as an abnormal condition as per the previous records, a warning was provided, as shown in Fig. (12), to the patient and physician that, helps to organize the next steps.

Fig. (11). Suggestions were given by the LSTM model at normal condition.

188 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

Fig. (12). Suggestions were given by the LSTM model at abnormal condition.

6. NOVELTY OF THE PROPOSED WORK This endeavor attempts to develop a deep learning-based smart healthcare monitoring system. This system monitors patient health and analyses, categorizes, and organizes data. A python script is used to follow several individuals with varying conditions. This technique would aid in the categorization of data and the organization of pharmaceutical requirements. This strategy is effective, as seen by the reported findings. In Python, the LSTM algorithm was developed to follow the health of a variety of patients. This strategy is effective, as seen by the findings obtained. CONCLUSION This research presents a smart health monitoring system that focuses on the LongTerm Memory (LSTM) algorithm. Three vital parameters, like Heart Rate, Body Temperature, and Blood Oxygen Level, were calculated using various sensors. These values are displayed to the ThingSpeak Server through an Arduino Board. This system also analyzed numerous past parameters, categorized the data, and organized the requirements. The LSTM algorithm was developed using the python program and discussed to track the health of several patients with various illnesses. The hardware results are also provided with real-time data output and the suggestions provided by the model. This approach yields satisfactory performance, and the results are also provided. CONSENT FOR PUBLICATON Not applicable.

Medical Services

AI and IoT-based Intelligent Health Care & Sanitation 189

CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

J. Gómez, B. Oviedo, and E. Zhuma, "Patient monitoring system based on internet of things", Procedia Comput. Sci., vol. 83, pp. 90-97, 2016. [http://dx.doi.org/10.1016/j.procs.2016.04.103]

[2]

P. Várady, Z. Benyó, and B. Benyó, "An open architecture patient monitoring system using standard technologies", IEEE Trans. Inf. Technol. Biomed., vol. 6, no. 1, pp. 95-98, 2002. [http://dx.doi.org/10.1109/4233.992168] [PMID: 11936602]

[3]

R. Kumar, and M.P. Rajasekaran, "An IoT-based patient monitoring system using Raspberry Pi", In: International Conference on Computing Technologies and Intelligent Data Engineering, 2016, pp. 14. [http://dx.doi.org/10.1109/ICCTIDE.2016.7725378]

[4]

R. Amin, S.K.H. Islam, G.P. Biswas, M.K. Khan, and N. Kumar, "A robust and anonymous patient monitoring system using wireless medical sensor networks", Future Gener. Comput. Syst., vol. 80, pp. 483-495, 2018. [http://dx.doi.org/10.1016/j.future.2016.05.032]

[5]

Y. Wang, S. Nazir, and M. Shafiq, "An Overview on Analyzing Deep Learning and Transfer Learning Approaches for Health Monitoring", Comput. Math. Methods Med., vol. 2021, pp. 1-10, 2021. [http://dx.doi.org/10.1155/2021/5552743]

[6]

I. Ahmed, G. Jeon, and F. Piccialli, "A Deep-Learning-Based Smart Healthcare System for Patient’s Discomfort Detection at the Edge of Internet of Things", IEEE Internet Things J., vol. 8, no. 13, pp. 10318-10326, 2021. [http://dx.doi.org/10.1109/JIOT.2021.3052067]

[7]

B. Godi, S. Viswanadham, A.S. Muttipati, O.P. Samantray, and S.R. Gadiraju, "E-“Healthcare Monitoring System using IoT with Machine Learning Approaches", In: International Conference on Computer Science, Engineering and Applications, 2020, pp. 1-5. [http://dx.doi.org/10.1109/ICCSEA49143.2020.9132937]

[8]

E.S. Muralidhar, T.S. Gowtham, A. Jain, and K. Padmaveni, "Development of Health Monitoring Application using Machine Learning on Android Platform", In: 5th International Conference on Communication and Electronics Systems, 2020, pp. 1076-1085. [http://dx.doi.org/10.1109/ICCES48766.2020.9137969]

[9]

J. Riyazulla Rahman, S. Sanshi, and N.N. Ahamed, “Health Monitoring and Predicting System using Internet of Things & Machine Learning,” 7th International Conference on Advanced Computing and Communication Systems, pp. 223-226, 2021. [http://dx.doi.org/10.1109/ICACCS51430.2021.9441856]

[10]

S. Gupta, D. Dahiya, and G. Raj, "Remote Health Monitoring System Using IoT", In: International Conference on Advances in Computing and Communication Engineering, 2018, pp. 300-305. [http://dx.doi.org/10.1109/ICACCE.2018.8441731]

[11]

V. Yeri, and D.C. Shubhangi, "IoT based Real-Time Health Monitoring", In: 2020 Second International Conference on Inventive Research in Computing Applications, 2020, pp. 980-984. [http://dx.doi.org/10.1109/ICIRCA48905.2020.9183194]

190 AI and IoT-based Intelligent Health Care & Sanitation

Vamshi Mohana et al.

[12]

F. Firouzi, F. Ye, K. Chakrabarty, and M.B. Tahoori, “Chip Health Monitoring Using Machine Learning”. IEEE Computer Society Annual Symposium on VLSI, pp. 280-283, 2014. [http://dx.doi.org/10.1109/ISVLSI.2014.119]

[13]

H. Pandey, and S. Prabha, "Smart Health Monitoring System using IOT and Machine Learning Techniques", In: 2020 Sixth International Conference on Bio Signals, Images, and Instrumentation, 2020, pp. 1-4. [http://dx.doi.org/10.1109/ICBSII49132.2020.9167660]

[14]

S.S. Sarmah, "An Efficient IoT-Based Patient Monitoring and Heart Disease Prediction System Using Deep Learning Modified Neural Network", IEEE Access, vol. 8, pp. 135784-135797, 2020. [http://dx.doi.org/10.1109/ACCESS.2020.3007561]

[15]

J. Azimi, "Takalo-Mattila, A. Anzanpour, A. M. Rahmani, J. Soininen and P. Liljeberg, “Empowering Healthcare IoT Systems with Hierarchical Edge-Based Deep Learning", In: IEEE/ACM International Conference on Connected Health: Applications, Systems and Engineering Technologies, 2018, pp. 6368. [http://dx.doi.org/10.1145/3278576.3278597]

[16]

D.S.R. Krishnan, S.C. Gupta, and T. Choudhury, “An IoT based Patient Health Monitoring System”, 2018 International Conference on Advances in Computing and Communication Engineering, pp. 0107, 2018. [http://dx.doi.org/10.1109/ICACCE.2018.8441708]

[17]

M.R. Ruman, A. Barua, W. Rahman, K.R. Jahan, M. Jamil Roni, and M.F. Rahman, "IoT Based Emergency Health Monitoring System", 2020 International Conference on Industry 4.0 Technology, pp. 159-162, 2020. [http://dx.doi.org/10.1109/I4Tech48345.2020.9102647]

[18]

V. Tamilselvi, S. Sribalaji, P. Vigneshwaran, P. Vinu, and J. GeethaRamani,"IoT Based Health Monitoring System", 2020 6th International Conference on Advanced Computing and Communication Systems, pp. 386-389, 2020. [http://dx.doi.org/10.1109/ICACCS48705.2020.9074192]

[19]

H. Bolhasani, M. Mohseni, and A.M. Rahmani, "Deep learning applications for IoT in health care: A systematic review", Informatics in Medicine Unlocked, vol. 23, 2021.100550. [http://dx.doi.org/10.1016/j.imu.2021.100550]

[20]

V. Mody, and V. Mody, Mental Health Monitoring System using Artificial Intelligence: A Review, 2019 IEEE 5th International Conference for Convergence in Technology (I2CT), pp. 1-6, 2019. [http://dx.doi.org/10.1109/I2CT45611.2019.9033652]

[21]

V. Ravindran, R. Ponraj, C. Krishnakumar, S. Ragunathan, V. Ramkumar, and K. Swaminathan, "IoTBased Smart Transformer Monitoring System with Raspberry Pi," 2021 Innovations in Power and Advanced Computing Technologies (i-PACT), pp. 1-7, 2021. [http://dx.doi.org/10.1109/i-PACT52855.2021.9696779]

[22]

V. Ravindran, and C. Vennila, "An energy-efficient clustering protocol for iot wireless sensor networks based on cluster supervisor management", Dokl. Bulg. Akad. Nauk., vol. 74, no. 12, 2021.

[23]

Ravindran, Vijay and Vennila, C. ‘Energy Consumption in Cluster Communication Using Mcsbch Approach in WSN’. pp. 1 – 11, 2022.

[24]

S.P. Yadav, D.P. Mahato, and N.T. Linh, Distributed Artificial Intelligence: A Modern Approach. CRC Press, 2020. [http://dx.doi.org/10.1201/9781003038467]

[25]

S.P. Yadav, S. Zaidi, A. Mishra, and V. Yadav, "Survey on machine learning in speech emotion recognition and vision systems using a recurrent neural network (RNN)", Arch. Comput. Methods Eng., vol. 29, no. 3, pp. 1753-1770, 2022. [http://dx.doi.org/10.1007/s11831-021-09647-x]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 191-206

191

CHAPTER 13

An Efficient Design and Comparison of Machine Learning Model for Diagnosis of Cardiovascular Disease Dillip Narayan Sahu1,*, G. Sudhakar2, Chandrakala G Raju3, Hemlata Joshi4 and Makhan Kumbhkar5 Department of MCA, School of Computer Science, Gangadhar Meher University, Sambalpur, Odisha-768001, India 2 School of Information Technology, Jawaharlal Nehru Technological University, Hyderabad, Telangana- 500085, India, 3 BMS College of Engineering, Bangalore, Karnataka-560019, India 4 Department of Statistics, CHRIST (Deemed to be University), Bangalore, Karnataka-560029, India 5 Christian Eminent College, Indore, Madhyapradesh-452001, India 1

Abstract: Cardiovascular disease has a significant global impact. Cardiovascular disease is the primary cause of disability and mortality in most developed countries. Cardiovascular disease is a condition that disturbs the structures and functionality of the heart and can also be called heart disease. Cardiovascular diseases require more precise, accurate, and reliable detection and forecasting because even a small inaccuracy might lead to fatigue or mortality. There are very few death occurrences related to cardio sickness, and the amount is expanding rapidly. Predicting this disease at its early stage can be done by employing Machine Learning (ML) algorithms, which may help reduce the number of deaths. Data pre-processing can be employed here to eliminate randomness in data, replace missing data, fill in default values if appropriate, and categorize features for forecasting and making decisions at various levels. This research investigates various parameters that are related to the cause of heart disease. Several models discussed here will come under the supervised learning type of algorithms like Support Vector Machine (SVM), K-nearest neighbor (KNN), and Naïve Bayes (NB) algorithm. The existing dataset of heart disease patients from the Kaggle has been used for the analysis. The dataset includes 300 instances and 13 parameters and 1 label are used for prediction and testing the performance of various algorithms. Predicting the likelihood that a given patient will be affected by the cardiac disease is Corresponding author Dillip Narayan Sahu: Department of MCA, School of Computer Science, Gangadhar Meher University (GMU), Sambalpur, Odisha-768001, India; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

192 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

the goal of this research. The most important purpose of the study is to make better efficiency and precision for the detection of cardiovascular disease in which the target output ultimately matters whether or not a person has heart disease.

Keywords: Bayes, Evaluation Metrics, Prediction, Pre-process, Supervised. 1. INTRODUCTION The major cause of mortality in the modern world is heart disease, also known as cardiovascular disease. A WHO study estimates that over 17.9 million people die each year as a result of cardiovascular disease. Middle to low economies has a substantial number of fatalities. Variables such as individual and career actions, as well as an inherited propensity, can all contribute to heart disease. The reasons for cardiovascular disease may include nicotine, excess drinking, anxiety, lack of physical activity, and also physiologic illness. The capacity to swiftly and effectively detect heart disease is crucial to taking preventative measures to avoid death. The process of data preprocessing is to extract essential information from massive datasets in a variety of disciplines, including medicine, business, and education. These algorithms can analyze huge amounts of data from several sectors, which may include the medical field. It is a computer-assisted alternative method over a traditional way of predicting models to reduce the errors in projected and actual results. The dataset is supposed to be analyzed using various ML algorithms. These data are analyzed by healthcare professionals for them to make efficient diagnostic decisions. Analyzing the health care domain dataset using classification algorithms gives clinical assistance. It includes a classification algorithm to test whether they can predict patients' cardiac problems. Data mining is the process of getting valid data or information from large datasets. Many ML methods, including regression, grouping, classification methods, and a few more segmentation techniques, can be used to do classification. A comparison analysis is used to analyse the classification approaches. This study will make use of data available on Kaggle. A classifier model was constructed using classification techniques to aid in the prediction of diseases. This research compares current technologies, as well as various methods that might be used to forecast heart illness. 1.1. Literature Survey Coronary disease was one of the major reasons for death in a few countries, such as India and the United States. Computations incorporating ML techniques such

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 193

as logistic regression, irregular woodlands, angle boosting, and SVM. For example, NB experiences a variety of heart-related difficulties. These computations can be used to improve the storing of data for valid and important reasons. Most analysis carries a variety of classifiers, for example, SVM, KNN, Neural Network and twofold discretization with Decision Tree [1]. It is assumed in the paper [2] that even using NB and Decision tree with data to start picking up counts produces superior results and greater precision when compared to conventional techniques. Since the nature of the heart is complicated, it must be handled with care. Heart disease severity is classified using a variety of approaches, including NB, Generic Algorithms, Decision Trees, and KNN. Mohan et al. [3] outline the procedure to combine two distinct ways to create a hybrid strategy with an accuracy of 88.4 percent, which is higher than the accuracy of any previous approach. Several researchers have researched ML algorithms that may be used for the prediction of cardiovascular illnesses. Kaur et al. [4] investigated this and defined the process by which interesting patterns and knowledge are extracted from a large dataset. They compare the accuracy of multiple ML and data preprocessing algorithms to determine the best and conclude the benefit of the SVM classifier. ML algorithms were utilized here to identify different types of diseases or illnesses, and many researchers like Kohali et al. [5], have started working on this for the prediction of heart disease using various ML models and concluded that the logistic regression algorithm has 87.1 percent of accuracy, SVM seems to have an accuracy of 87.1 percent, and the AdaBoost classifier has an 87.1 percent of accuracy. 2. ML Artificial intelligence is a sub-discipline of ML and is a technology widely used today. The main goal is to create systems that really can learn from their experiences and make predictions. It may use the training dataset to create a model that trains ML algorithms. To forecast cardiac disease, the approach integrates the new input data. The models may be generated by detecting hidden patterns in the input dataset using ML algorithms. For new datasets, it makes accurate predictions. To design an effective model, the dataset needs to be cleaned, and any missing values have to be filled with either zeros or nearby values. The model is instead evaluated for accuracy using the new set of input data to predict the illness of the heart. The model has been trained using a labeled dataset. It has input data as well as output data. A training data set can be used in supervised learning to train models to produce the preferred output. This may be prepared with the dataset containing both correct and incorrect data, which may allow the model to improve its

194 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

efficiency. The loss function used here is to assess the correctness of an algorithm, and it is adjusted until the error is suitably lessened. This type of algorithm may help to classify data and assign it to certain groups. It may use certain objects in the dataset and make the model educated to the presumption that those objects should be categorized or specified. 3. METHODOLOGY This study attempts to forecast the probability of having cardiovascular disease using computer-aided prediction of heart disease. To attain this purpose, we discussed the use of multiple supervised learning models on the data set, and this research paper will discuss a set of data analyses. The dataset is allowed to have only the highly important data, which may avoid multiple trials on a patient, as not many of the attributes may contribute significantly to the expected outcome. In Fig. (1), the flow of the proposed work has been given.

Fig. (1). Process Flow of Proposed work.

3.1. Source of Data To conduct this study, we used data that was obtained from Kaggle. It includes a true dataset containing 300 data samples with 14 multiple parameters (13 forecasters; 1 tag). To get the most realistic model of heart disease, we tested three different algorithms in this study.

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 195

3.2. Data Pre-processing Substantial volumes of data missing or noisy data may present in real-life data or raw data sets. These data are supposed to be pre-processed to reduce false predictions. The processed data is important for the ML model. 4. ML ALGORITHM Classification is a type of task that requires supervision. An ML problem involves creating a formula that converts input to output using sample input-output combinations. It derives a generative model consisting of a sequence of training instances. The input dataset is separated into two sections: train and test. There is an outcome variable in the test dataset that should be predicted or categorized. All algorithms extract features from the training phase to predict or classify the test set. Numerous studies have been conducted on the most well-known supervised ML algorithms [6 - 8]. 4.1. NB Classifier One of the supervised learning algorithms is the NB classifier. This type is a straightforward classification method that follows Bayes' theorem. It assumes strong (naive) independence from the features. The Bayes theorem works in the concept of probability. The predictors have no relationship with one another and do not correlate with one another [9]. NB classifiers are used in many complex real-world scenarios. The Bayesian formula is given, the b in the formula represents the features, and the y represents the label. ‫ܤ‬ ܲ ቀ ቁ ܲሺ‫ݕ‬ሻ ‫ݕ‬ ‫ݕ‬ ܲቀ ቁ ൌ ‫ܤ‬ ܲሺ‫ܤ‬ሻ ܾͳ ܾʹ ܾ݊ ܲ ቀ ቁ ܲ ቀ ቁ ǥ ܲ ቀ ቁ ܲሺ‫ݕ‬ሻ ‫ݕ‬ ‫ݕ‬ ‫ݕ‬ ‫ݕ‬ ൰ൌ ܲ൬ ܾͳǡ ܾʹǡ ǥ ǥ ܾ݊ ܲሺ‫ܤ‬ሻ

(1)

(2)

The classification of NB is based on the given formula ௡

ܾͳ ‫ ݕ‬ൌ ܽ‫ݔܽ݉݃ݎ‬௬ ܲሺ‫ݕ‬ሻ ෑ ܲ ൬ ൰ ‫ݕ‬ ௜ୀଵ

(3)

196 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

4.2. SVM Most researchers may use the SVM classifier because it may achieve excellent accuracy even with minimal computing resources. SVM stands for SVM, and this algorithm can use cases of both classification and regression types of problems [10]. It is, nevertheless, commonly used in the implementation of categorization types of problems. The purpose of the SVM algorithm is to predict the proper hyperplane in N-dimensional that should differentiate the data points. The primary goal is to identify a plane with the greatest margin between data points of the available types. The hyperplane is represented by: ்݊ ܾ ൅ ܽ ൌ Ͳǡ ȁȁ݊ȁȁ ൌ ͳ

(4)

Where, n → Unit vector The classification of SVM is based on the given formula, and b in the formula represents the features, and y represents the label. ‫ݕ‬ො ൌ ݄ሺܾሻ ൌ ‫݊݃݅ݏ‬ሺ்݊ ܾ ൅ ܽሻ

(5)

The model predicts the given data contain cardiovascular disease if the given condition satisfies ்݊ ܾ ൅ ܽ ൒ Ͳ݂݅‫ ݕ‬ൌ ͳ

(6)

The model predicts the given data don’t have cardiovascular disease if the given condition satisfies ்݊ ܾ ൅ ܽ ൑ Ͳ݂݅‫ ݕ‬ൌ െͳ

(7)

4.3 KNN ML models may provide the output values according to the compilation of input values. KNN is one of the commonly used classification algorithms. It may classify the data point based on the classification of its neighbors. The data in a dataset may be classified using KNN uses the resemblance measure of the beforehand stored data [11]. The constant value K in KNN stands for the number of nearest neighbors we utilized to classify the fresh data points- pretend that a new data point has to be placed in the appropriate category. The number of neighbors, which is k=5, needs to be assumed initially. The Euclidean distance is

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 197

the distance between two points that we learned about in geometry class. The nearest neighbor can be calculated with the help of Euclidean distance. The formula will be like this, ୬

†ሺ‡™†ƒ–ƒǡ ’”‡˜‹‘—•†ƒ–ƒሻ ൌ ඩ෍ሺ‡™†ƒ–ƒ െ ’”‡˜‹‘—•†ƒ–ƒሻଶ

(8)

୧ୀଵ

The classification of KNN is based on the given formula, and the b in the formula represents the features and y represents the label. The KNN model predicts the given data show cardiovascular disease if the given condition satisfies ‫ ݕ‬ൌ ͳǡ ݂݅݊‫̷ܾ݂݋ݎܾ݁݉ݑ‬ሾ‫ ݕ‬ൌ ͳሿ ൒ ݊‫̷ܾ݂݋ݎܾ݁݉ݑ‬ሾ‫ ݕ‬ൌ െͳሿ

(9)

The KNN model predicts the given data have the cardiovascular disease if the given condition satisfies ‫ ݕ‬ൌ െͳǡ ݂݅݊‫̷ܾ݂݋ݎܾ݁݉ݑ‬ሾ‫ ݕ‬ൌ െͳሿ ൒ ݊‫̷ܾ݂݋ݎܾ݁݉ݑ‬ሾ‫ ݕ‬ൌ ͳሿ

(10)

5. RESULT AND ANALYSIS A confusion matrix may help to estimate the performance of a classification model using a table structure. An N x N matrix is used to represent the confusion matrix. This helps evaluate the performance of a classification model planning to implement for the particular dataset, where N is considered several target classes. It may compare the actual goal values to the predicted values of ML models. For a binary classification issue, we'd use a 2 x 2 matrix with four values, as illustrated below: There will be two possible conditions for predicting the classes. They are 'Yes' or 'No'. If the model is supposed to forecast the existence of an illness, “yes” would indicate that they have the condition, while “no” would indicate that they do not have a disease. Here, the considered dataset has made 60 predictions.

198 AI and IoT-based Intelligent Health Care & Sanitation ●

●

●

●

Sahu et al.

True Positive (TP) - Correct Positive Prediction. The data represents the diseaseaffected person, and the model predicted it positively. True Negative (TN) - Correct Negative Prediction. The data represents the disease, not the affected person, and the model predicted it positively. False Positive (FP) - Incorrect Positive Prediction. The data represents the disease, not the affected person, and the model predicted it negatively. False Negative (FN) - Incorrect Negative Prediction. The data represents the disease-affected person, and the model predicted it negatively.

5.1. Performance Measures The performance of the ML model can be validated using some terms [12] like accuracy, precision, recall, and specificity. F1 score is detailed in Table 1. Table 1. Performance metrics and their formula. Performance Metrics Accuracy

Definition

Formula

The widely used metric to estimate a model is Accuracy=TP+TNTP+FP+TN+FN certainly not a clear measure of its effectiveness.

True Positive Rate (TPR)

The TPR is the possibility that a true positive will test positive. Indicates the high chance of the presence of the disease.

TPR=TP TP+FN

True Negative Rate (TNR)

The TNR is the possible ratio of true negative will test as negative. Value helps to know the chance of not having the disease.

TNR=TN TN+FP

False Positive Rate (FPR)

The ratio of false positives compared to all positive predictions. Determined by the number of real negatives predicted wrongly by the model.

FPR=FP TN+FP

False Negative Rate (FNR)

A ratio of false negatives is when the model predicts the negative category erroneously. Measured by the number of real positives predicted wrongly as negatives by the model.

FNR=FNTP+FN

Precision

It measures how often the prediction of the particular model is correct. The formula for measuring the precision is given below:

Precision=TPTP+FP

5.2. Confusion Matrix: NB Classifier The confusion matrix [13] obtained for the considered algorithm has illustrated below. The NB algorithm has been employed to predict the presence of heart disease in a data set. The confusion matrix obtained for the NB classifier is shown in Fig. (2).

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 199

Fig. (2). Confusion matrix of NB Classifier.

Here, the prediction done by the model designed using the NB classifier has been discussed. The model has predicted 28 values as true positive where the actual value and predicted value are the same and states that the person has been affected by the disease as well the model has predicted it correctly. It has predicted 30 true negatives where the actual value and predicted value are the same and states that the person has not been affected by the disease as well as the model has predicted it correctly. It has predicted 2 false-positive where the actual value and predicted value are different, and it states that the person has not been affected by the disease, but the model has predicted it wrongly. It has predicted 1 false-positive where the actual value and predicted value are different, and it states that the personal data has been affected by the disease, but the model has predicted it wrongly. 5.3. Confusion Matrix: SVM The SVM algorithm has been employed to predict heart disease. The confusion matrix obtained for the SVM has shown in Fig. (3).

200 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

Fig. (3). Confusion matrix of SVM Classifier.

The prediction made by the model created with an SVM classifier is detailed below. The SVM classifier predicted 28 true positive values, indicating that the real and projected values are the same, indicating that the individual has been afflicted by the disease and that the model accurately predicted it. It accurately predicted 29 true negatives, in which the actual and predicted values are the same, indicating that the person was not afflicted by the disease as the model anticipated. It has predicted two false positives in which the actual value and values which have been predicted are different, implying that the person is not affected by the condition although the model has predicted it as affected. It has predicted 1 false-positive where the actual value and predicted value are different, and it states that the personal data has been affected by the disease, but the model has predicted it wrongly. 5.4. Confusion Matrix: KNN The KNN algorithm has been employed to predict heart disease. The confusion matrix obtained for the K Nearest Neighbor algorithm has shown in Fig. (4).

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 201

Fig. (4). Confusion matrix of KNN Classifier.

The model built with the KNN classifier has been used to make a prediction, which was mentioned here. The KNN classifier accurately predicted 29 true positive values where the actual and forecasted values are the same, indicating that the person has been afflicted by the disease. It accurately predicted 27 true negatives, in which the actual and projected values are the same, indicating that the person was not impacted by the illness. It has predicted one false positive, in which the real and projected values are different, and it indicates that the person is not affected by the disease, yet the model has predicted it wrongly. It has predicted 3 false positives, in which the real and predicted values are different, and it claims that the disease has damaged personal data, but the model predicted it incorrectly. Performance of Classifier The NB, SVM, and K Nearest Neighbor algorithms have been employed to predict heart disease in this research work. The performance metrics obtained for the above-mentioned algorithm has shown in the following figure. The performance of the NB classifier is shown in Fig. (5), the SVM in Fig. (6), and the KNN classifier are given in Fig. (7).

202 AI and IoT-based Intelligent Health Care & Sanitation

Fig. (5). The efficiency of NB Classifier.

Fig. (6). The efficiency of the SVM Classifier.

Sahu et al.

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 203

Fig. (7). The efficiency of K Nearest Neighbor.

5.5. ML Algorithm Comparison The ML evaluation is very important for finalizing the model [14]. The designed models were evaluated using certain measures, and those measures are listed in the below-given Table 2. The table shows the accuracy, precision, F1 score, and performance Index of the above-considered algorithm to classify the dataset. The performance of each considered model has been evaluated and tabulated. Table 2 represents the accuracy, precision, FNR, FPR, TNR, and TPR measured for the NB (NB) classifier, SVM (SVM), and K nearest neighbor (KNN) classifier algorithms. From the considered models' NB has given a better accuracy of about 96.67%. Next to that is SVM which gives an accuracy of 95%, and the KNN classifier gives 93.3% of accuracy. So, it is concluded that NB will suit well for this prediction from the performance analysis. Table 2. ML Algorithm comparison. Accuracy

TNR

TPR

FNR

FPR

Precision

NB

96.6

93.7

100

0

6.25

93.3

SVM

95

93.3

96.5

3.44

6.45

93.3

KNN

93.3

96.4

90.6

3.27

3.57

96.6

204 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

The overall accuracy of the model has been compared and represented using a line graph shown in Fig. (8).

Fig. (8). Accuracy Comparison.

6. IMPORTANCE OF THE STUDY Using ML algorithms to predict this disease in its early stages may assist to reduce the mortality rate. Data pre-processing can be used to minimize randomness in data, recover missing data, bring up default values if necessary, and extract features for forecasting and decision-making on cardiovascular disease. Several models mentioned here will fall under the category of supervised learning algorithms. The study's primary goal is to improve detection accuracy for cardiovascular illness, where the target output ultimately determines whether the patient has heart disease or not. CONCLUSION The ultimate goal is to develop several data mining approaches that may be used to accurately forecast cardiac disease. Our objective is to provide efficient and

Cardiovascular Disease

AI and IoT-based Intelligent Health Care & Sanitation 205

reliable predictions with a smaller number of characteristics and tests. Only 14 key qualities are considered in this study. KNN, NB, and SVM were the four data mining Classification algorithms used in this study. The dataset was pre-processed before being utilized in the model. It is found that SVM's accuracy is comparatively better than the other two algorithms. In the future, disease prediction will be implemented using combined models to improve early heart disease diagnosis accuracy. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

J. Patel, S. Tejalupadhyay, and S. Patel, Heart Disease prediction using ML and Data Mining Technique. Journal – IJCSC. vol.7, 2016.

[2]

G. Parthiban, Int. J. Appl. Inf. Syst., vol. 3, pp. 25-30, 2012. [http://dx.doi.org/10.5120/ijais12-450593]

[3]

S. Mohan, C. Thirumalai, and G. Srivastava, "Effective Heart Disease Prediction Using Hybrid Machine Learning Techniques", IEEE Access, vol. 7, pp. 81542-81554, 2019. [http://dx.doi.org/10.1109/ACCESS.2019.2923707]

[4]

A. Kaur, and J. Arora, "Heart disease prediction using data mining techniques: A survey", Int J Adv. Res. Comp. Sci., vol. 9, no. 2, pp. 569-572, 2018. [http://dx.doi.org/10.26483/ijarcs.v9i2.5872]

[5]

P. Kohli, and S. Arora, “Application of ML in Disease Prediction,” pp. 1-4, 2018. [http://dx.doi.org/10.1109/CCAA.2018.8777449]

[6]

D. Abd, J.K. Alwan, M. Ibrahim, and M.B. Naeem, "The utilisation of ML approaches for medical data classification and personal care system management for sickle cell disease", In: 2017 Annual Conference on New Trends in Information & Communications Technology Applications (NTICT), 2017, pp. 213-218. [http://dx.doi.org/10.1109/NTICT.2017.7976147]

[7]

M. Fatima, and M. Pasha, "Survey of Machine Learning Algorithms for Disease Diagnostic", Journal of Intelligent Learning Systems and Applications, vol. 9, no. 1, pp. 1-16, 2017. [http://dx.doi.org/10.4236/jilsa.2017.91001]

[8]

A. Gavhane, G. Kokkula, I. Pandya, and P. Devadkar, "Prediction of Heart Disease Using ML", In: 2018 Second International Conference on Electronics, Communication and Aerospace Technology (ICECA), 2018, pp. 1275-1278. [http://dx.doi.org/10.1109/ICECA.2018.8474922]

[9]

A.H. Jahromi, and M. Taheri, "A non-parametric mixture of Gaussian NB classifiers based on local

206 AI and IoT-based Intelligent Health Care & Sanitation

Sahu et al.

independent features", In: 2017 Artificial Intelligence and Signal Processing Conference (AISP), 2017, pp. 209-212. [http://dx.doi.org/10.1109/AISP.2017.8324083] [10]

S. Alam, "Moonsoo Kang, Jae-Young Pyun and G. -R. Kwon, “Performance of classification based on PCA, linear SVM, and Multi-kernel SVM", In: 2016 Eighth International Conference on Ubiquitous and Future Networks (ICUFN), 2016, pp. 987-989. [http://dx.doi.org/10.1109/ICUFN.2016.7536945]

[11]

M.A. jabbar, B.L. Deekshatulu, and P. Chandra, "Classification of Heart Disease Using K- Nearest Neighbor and Genetic Algorithm", Procedia Technol., vol. 10, pp. 85-94, 2013. [http://dx.doi.org/10.1016/j.protcy.2013.12.340]

[12]

S.P. Yadav, "Emotion recognition model based on facial expressions", Multimedia Tools Appl., vol. 80, no. 17, pp. 26357-26379, 2021. [http://dx.doi.org/10.1007/s11042-021-10962-5]

[13]

I. Bakagiannis, V.C. Gerogiannis, G. Kakarontzas, and A. Karageorgos, "ML product key performance indicators and alignment to model evaluation," 2021 3rd International Conference on Advances in Computer Technology, Information Science and Communication (CTISC), pp. 172-177, 2021. [http://dx.doi.org/10.1109/CTISC52352.2021.00039]

[14]

Y.S. Al-Turjman, F. Yadav, V. Kumar, M.S. Thompson, “Transforming Management with AI, BigData, and IoT”, 2021.

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 207-221

207

CHAPTER 14

Deep Learning Based Object Detection Using Mask R-CNN Vinit Gupta1,*, Aditya Mandloi1, Santosh Pawar2, T.V Aravinda3 and K.R Krishnareddy3 Electronics Engineering Department, Medi-Caps University, Indore, Madhya Pradesh, India452001 2 Electronics Engineering Department, Dr. APJ Kalam University, Indore, Madhya Pradesh, India-452001 3 SJM Institute of Technology, Chitradurga, Karnataka, India-577502 1

Abstract: Nowadays, an image must be verified or analyzed keenly for further processing methods. Few numbers of images can be analyzed manually for a particular object or a human being. But when millions and millions of images are present in a dataset, and every single one of them must be verified and classified based on the objects present in the image, it is necessary to find an algorithm or a technique to assist this process. Of many types and applications of object detection, this research aims at pedestrian detection. Pedestrian detection is a special way that only aims to detect the human beings in the uploaded image. The Mask R-CNN algorithm is used in pedestrian detection. The Mask R-CNN model is a Deep Learning (DL) model that can detect a certain object from an image and can also be used for image augmentation. The Mask R-CNN stands for Mask Regional Convolutional Neural Network (CNN). This is one of the classification techniques which can be designed using the Convolutional Network theories. The Mask R-CNN algorithm is the updated version of the Faster RCNN. The main difference between the faster R-CNN and the Mask R-CNN is that the mask R-CNN can bind the borders of the object detected while the faster R-CNN uses a box to identify the object. One of the major advantages of mask R-CNN is that it can provide high-quality image augmentation and is also one of the fastest image segmentation algorithms. This model can be easily implemented when compared to other object detection techniques. Python contains various inbuilt dependencies which can be installed with the help of repositories. These dependencies are used to design the model. The model is then trained and tested with various inputs for better accuracy. Image segmentation is used in various fields like medical imaging, video surveillance, traffic control, etc., so the mask R-CNN technique would be an extremely efficient DL algorithm.

Keywords: Data, DL, Mask R-CNN, MaxDets, Object Detection, Pedestrian. * Corresponding author Vinit Gupta: Electronics Engineering Department, Medi-Caps University, Indore, Madhya Pradesh, India-452001; E-mail: [email protected]

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

208 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

1. INTRODUCTION Images and videos are some forms of data that can be used for various purposes in our day-to-day life. Image processing and object detection are used in various fields like medical analysis, face detection, security systems, automated vehicles, and agriculture. Object detection is a computer vision task used to classify and detect the presence of individual objects or even human beings from an image. Object detection can be further classified into various tasks like face detection, animal detection, vehicle identification, security systems, picture retrieval, etc. Object detection can be employed using either machine learning algorithms or DL algorithms. Yet, DL algorithms used for object detection are far more effective than machine learning algorithms. By the year 2014 and after, object detection was effectively done using DL algorithms. Many DL algorithms can be used in object detection. Some of the most commonly used DL algorithms for object detection include You Only Look Once algorithm or the YOLO algorithm, the single-shot detector algorithm, SqueezeDet, and the mask R-CNN. The YOLO algorithm is one of the fastest DL algorithms that can bind a box regression just by scanning the image once. The single-shot detector is the most commonly used DL algorithm for object detection. The SqueezeDet algorithm is specifically designed for human driving and is also capable of bounding a box regression in one scan, just like the YOLO algorithm. The Mask R-CNN is an upgraded version of the Fast R-CNN. This algorithm is capable of detecting the object along with its border, and unlike the other object detection algorithms, the mask R-CNN does not use box regression bounding. Thus, the mask R-CNN algorithm can be claimed as the simplest and the most efficient DL algorithm. The process of object detection using the mask R-CNN is explained. Computer vision for real-time applications, such as driving assistance and video surveillance, is expanding, and pedestrian identification is one of the most promising areas of research. For the past ten years, the topic of pedestrian detection has piqued the interest of many academics in the research community. Pedestrian detection aims to identify the obvious objects in the information gathered from the labeled pedestrian and non-pedestrian examples in the information collection process. Due to its practical application in a variety of computer vision applications, such as video surveillance and robotics interaction, tremendous progress has been made in the last decade. It has become possible to deal with pedestrian detection by installing a variety of sensor systems on the autos and even integrating the data given by some of the sensors. Some of the sensors used to accomplish this goal include cameras and stereoscopic machine vision systems, laser sensors, time-of-flight cameras, and three-dimensional laser sensors (three-dimensional laser sensors). Despite all of the significant

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 209

advancements in the identification of pedestrians, this task continues to provide significant challenges. For example, one of them is the achievement of dependable performance under highly fluctuating lighting conditions, such as those found in real-world driving situations. 2. LITERATURE SURVEY Many types of research and projects have been developed based on the mask RCNN algorithm and its usage in object detection. For instance, by the end of the year 2019, Phil Ammirato et al. developed Mask-RCNN Baseline for Probabilistic Object Detection. They used a two-stage pipeline to develop a DL model using the mask R-CNN algorithm. They also used instance segmentation and the MSCOCO class to train the data. They concluded that there is a lot of possibility for development in future work with a final score of 21.432, albeit the lack of training data is a limiting problem. Most of the improvements in this paper are based on basic heuristics that use the PDQ calculation's structure [1]. Researchers from the National Authority for Remote Sensing and Space Science, Cairo, Egypt, and the 2 Faculty of Computers and Information, Cairo University, Giza, Egypt, including Amira S. Mahmoud et al., researched the usage of Adaptive Mask RCNN in object detection. They used the usage of the FPN and RPN to construct the DL model. They also used a loss function to increase the efficiency of the algorithm. The model is then optimized using stochastic gradient descent. According to them, the suggested method beats current baseline object detection methods, and they found that the performance of the algorithm is increased by six percent. By employing the switch between optimizers SWATS in the training phase, the proposed adaptive Mask RCNN is found to be the most efficient among other DL. They measured the efficiency using some metrics named PRC and IOU when compared to other methods that used the default optimizer [2]. According to Jeremiah W. Johnson, Mask-RCNN is an efficient approach for object detection, localization, and object instance segmentation in image features that was recently proposed. A cyclic learning rate regime is shown to allow effective Mask-RCNN model training without the need to refine the information gained, eliminating a laborious and time-consuming step in the training process. He concluded that the Mask-RCNN had been demonstrated to be capable of performing exceptionally effective and efficient automatic segmentation of a wide range of microscopy pictures of cell nuclei for a range of cells obtained under various conditions [3]. Again, by 2019, Evi Kopelowitz et al. used the threedimension mask R-CNN to detect lung nodules. They tweaked MaskRCNN's 2D implementation to work with 3D pictures and account for small item detection.

210 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

Appendix A and Appendix B contain information on how to fully implement the model. With overlapping sliding windows, we scan each image. Non-Max Suppression is used to filter out overlapping boxes (NMS). They only preserved the boxes whose segmentation value was greater than zero. This process helps to reduce False Positives (FPs). In the end, they found a significant correlation between anticipated volumes and GT, implying that the volume of nodules assessed and forecasted by their model is a trustworthy metric of nodule size and could eventually reduce the usage of manual segmentation [4]. A crew of research scholars from Queensland, including Haoyang Zhang, used the s Stochastic Weights Averaging in the process of object detection. According to them, the SWA is more effective than the mask R-CNN for object detection. They concluded that training an object detector for additional 12 epochs with cycle learning rates and averaging these 12 checkpoints as the final model can improve one of the effects of using SWA to object detection and instance segmentation. SWA, they concluded, increases object localization and categorization accuracy, leading to fewer false positives and higher recall rates [6]. Object detection was no new topic to researchers. However, a pre-trainable system that can be used for object detection was initially introduced by Constantine Papageorgiou et al. designed a generalized trainable system for object detection. They used various wavelets like HAAR wavelets. This model is said to be one of the most efficient object detection algorithms, as it can detect and analyze faces, people, cars, etc. They also used a classifier algorithm named Support Vector Machine for classification. The results are applicable on face, people an d car detection With the use of a support vector machine, they were successful in developing a detection system that met their objectives of high accuracy and low false-positive rates when used in conjunction with the classification engine. In an overfitted dictionary of Haar wavelets, Haar wavelets are used to represent the elements of the object class in an overfitted dictionary of Haar wavelets [7]. Zhengxia Zou et al. reviewed many machine learning and DL algorithms and summarized their research. According to them, the Viola-Jones detectors were one of the earliest object detection techniques. The VJ detector employs a straightforward form of identification known as moving windows, which entails scanning a picture for all possible positions and scales to see whether any of the windows contain a human face. The Viola-Jones detector is considerably updated by the usage of incorporating three main techniques that can help increase the detecting speed: “integral picture, “detection cascades” and “feature selection”. Following the Viola-Jones detecting technique, several techniques, such as the HOG detector, Deformable Part Based Model, and so on, are used. But the usage of CNN-based two-stage detectors is considered a milestone in the history of

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 211

object detection. The invention of various CNN algorithms like R-CNN, SPP net, Fast R-CNN, and mask R-CNN boosted the efficiency of all the existing object detection algorithms. Followed by the two-stage detectors, the one-stage detectors also paved the way for the most efficient object detection algorithm as it only needs to scan the image or a file once to predict the presence of an object. The one-stage detectors include the You Only Look Once (YOLO) algorithm, singleshot multi-box detector (SSD), retina net, etc. They concluded that the future of object detection relies on the invention of algorithms that are capable of performing tasks like lightweight object detection, domain adaption, and detection of relatively smaller objects [8]. The usage of DL algorithms in object detection is one of the smart moves in the history of object detection. Researchers from India summarized the usage of DL algorithms in the field of object detection. As a result of recent achievements in the fields of image classification, object identification, and natural language processing, the term “DL” has become a catchphrase, according to the authors. CNNs have the advantage of not requiring feature extractors or filters to be manually developed. Alternatively, they retrain from the pixel level before the end object categories. They concluded that object detection is the most important phase in the implementation of self-driving automobiles and robotics. Due to the infeasibility of surplus human power, the future breadth of DL algorithms in use will continue to expand. It is important to integrate processing into a centralized, high-performance GPU to process data from multiple servers at the same time and perform identification analysis in near real-time [9]. 3. METHODOLOGIES The implementation of the Mask R-CNN algorithm for object detection includes a few steps that are simple and easy to implement. The steps to be followed to complete the designing and evaluation of the DL model are pictorially represented as a flow chart in Fig. (1) . The data for training the DL model is first collected. The data is then preprocessed to achieve better results. The image resizing and augmentation is done in pre-processing steps. The Mask RCNN is chosen as a DL model. The collected data is split and used in particular parts for training the model and the remaining for testing. The output of testing data is evaluated to check whether the designed model is good or not.

212 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

Fig. (1). Workflow of Mask R-CNN.

3.1. Data Collection and Preprocessing To design and test the efficiency of the DL model, a set of data must be used. This data is used to both test and train the model. In this research, a dataset containing 170 images is used. These images are specifically taken for pedestrian detection. There will be a total of 345 pedestrians in the whole dataset. Out of the 170 images, 96 images were taken in and around the University of Pennsylvania, and the rest of the images were taken in and around Fudan University. Hence, the dataset is also known as the Penn-Fudan database for pedestrian detection. The obtained images must go through some preprocessing techniques so that they can be used to train and test the DL model. The image preprocessing is mostly done in two steps, namely image resizing and image augmentation [10]. Image resizing is the process of changing the dimensions of all the images in a dataset so that they are all the same size. The total number of pixels in an image can be increased or decreased when scaling it. When an image is zoomed or enlarged, picture resizing guarantees that the image quality is maintained. A further significant benefit of image scaling is that it can trim out unnecessary areas of the image [11].

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 213

Image augmentation is a method of modifying existing photos to increase the amount of data available for model training. When creating CNNs, it's a straightforward technique to enlarge the training set without having to gather fresh photos. In other words, it's a way to artificially increase the dataset accessible for DL model training. The usage of augmented photographs will increase the model's performance without necessitating the inclusion of a huge number of real images [12]. The preprocessed images are then divided into two parts. One part includes about 120 images, which are used to train the DL model; the second part includes 50 images which will be used to test the efficiency of the model along with the average precision and recall. 3.2. Construction of Mask R-CNN The mask R-CNN algorithm is implemented using the development of a DL model. The mask R-CNN can be classified into two categories. They are semantic segmentation and instance segmentation. Semantic segmentation detects the presence of objects as a cluster if one object is present on top of the other, and it cannot separate every single instance [13]. On the other hand, instance segmentation can precisely identify the presence of every instance. This research uses the instance segmentation of object detection as it is more efficient than semantic segmentation [14]. The mask R-CNN algorithm can be constructed using various methods. They include the development of the algorithm using matterport, TensorFlow, and Keras. This model can also be implemented using python. A DL model using the mask R-CNN algorithm is developed using TensorFlow and Keras. The steps to be followed while developing the DL model are as follows. ●

●

●

●

A backbone model is developed using the pretrained FPN. This backbone model is a simple classifier model which is used to develop the inner layer of the model and to train the mask R-CNN. Next to the backbone model, the region proposal network is developed. This is where the ROI or the region of interest is determined. This region of interest is then used to find the value of IoU, which is then used to determine the average precision and the average recall value. The region proposal network is developed using the anchors. The Anchors are precon Figured rectangles of various sizes that cover the whole input image. We choose the sizes and scales of the anchors based on our items. Anchors are selected based on the overlap of objects to be detected, and the RPN's duty will be to move them around to make them overlap even better, as well as shrink them to fit better.

214 AI and IoT-based Intelligent Health Care & Sanitation ●

●

Gupta et al.

The final step in the development of the mask R-CNN is to train the “head” models, which can be done in parallel. Only the foreground proposals contribute to the box loss, just as they do in the RPN, and an arbitrarily defined number of proposals contribute to the loss value. So it will not only forecast if a region is a foreground or background, but it will also classify it if it is foreground. Once the model is constructed, the model is then class-predicted. The corresponding filter can be used for class prediction. This can also be done using the mask heads which can be developed using the CNNs.

3.3. Training of Mask R-CNN The DL model, which was already constructed, is trained using the 120 images obtained from the Penn-Fudan dataset. The processed images are then uploaded into the DL model. The model is then activated, and some parameters like the loss, the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg are noted to validate the improvement in the performance of the model [15]. The below table tabulates the values of the loss, the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg for every epoch of the training of the DL model. Table 1. Loss of details for every epoch. Epoch

Loss

Loss Classifier

Loss box_reg

Loss Mask

Loss Objectness

Loss rpn_box_reg

0

5.8707

0.5411

0.3162

4.9738

0.0344

0.0053

1

0.2788

0.0257

0.0783

0.1664

0.0004

0.0079

2

0.3295

0.0367

0.0710

0.2169

0.0008

0.0041

3

0.1979

0.0317

0.0558

0.1077

0.0003

0.0024

4

0.2589

0.0431

0.0604

0.1527

0.0003

0.0024

5

0.2356

0.0399

0.0541

0.1366

0.0007

0.0043

6

0.1384

0.0201

0.0176

0.0986

0.0001

0.0020

7

0.2075

0.0224

0.0441

0.1371

0.0018

0.0022

8

0.2068

0.0243

0.0340

0.1433

0.0004

0.0049

9

0.1194

0.0064

0.0119

0.0999

0.0004

0.0007

From the above Table 1, it can be inferred that the loss percentage of the model is gradually decreasing as the number of epochs is increasing. The value of the loss is 5.8707 at the first epoch, which is a greater value for an object detection algorithm. This big loss value leads to the failure of the algorithm as it may fail to detect the presence of certain objects as needed. Along with the loss value, it is

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 215

also noted that the other parameters like the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg are also very high during the first epoch. It is mandatory to reduce the values of the loss, the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg for the DL algorithm to work with higher efficiency and performance. It is attained by training the model again and again and by increasing the number of epochs. When done correctly, it can be seen that the values of the loss, the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg are constantly decreasing as the number of epochs in increased. The model acquires the minimum loss value of 0.0064 by the end of the tenth epoch. The values of the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg are also decreased to 0.0119, 0.0999, 0.0004, and 0.0007, respectively. These values of the loss, the loss classifiers, the loss box reg, the loss mask, the objectness, and the loss rpn box reg are the minimum values that can be obtained for the DL model used for object detection. The fact that the minimum value can be attained in the tenth epoch itself reveals the efficiency of the Mask R-CNN algorithm. 4. RESULT ANALYSIS A DL model which is capable of detecting objects from the images uploaded with a maximum accuracy level, is developed using the Mask R-CNN algorithm. This model is then trained a lot of times until it achieves maximum accuracy. It could be observed that the loss, the loss classifier value, the loss box reg value, the loss mask value, the loss objectness value, and the loss rpn box reg value decreases gradually as the number of epochs is increased and reaches the minimum possible value. The model is then tested using real-time images that were obtained from the Penn-Fudan database. Fig. (2) is the first example of the accuracy of object detection. From the above image, it can be seen that the model detects the presence of pedestrians with an accuracy score of 0.966, which is great for an object detection algorithm. Also, the model can predict the presence of both pedestrians with the same accuracy level.

216 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

Fig. (3) is again an example of the output which is being tested by the model.

Fig. (2). Output for test image -1.

Fig. (3). Output for the test image -2.

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 217

The above image is slightly more complicated than the first image as it contains many pedestrians, some of them even overlap each other. Yet the model can detect almost six members from the image. Out of all the members detected, it could be seen that the minimum accuracy score is 0.919, which is again a good accuracy percentage for such a complicated image. It should also be noted that this model is capable of instance segmentation as it can differentiate the presence of the pedestrians whose image is hidden behind the other one, and it can detect them individually without grouping them. Fig. (4) is the last example of the output of the DL algorithm that uses mask RCNN for object detection.

Fig. (4). Output for the test image – 3.

The above image is again more complicated than the last one as it contains a pedestrian who lies in the corner of the image. In the DL model, the man in the corner has a larger risk of being neglected than the man in the middle. In terms of identifying the pedestrian in the upper right of the image, CNN's Mask R-model is more accurate than the other models in the dataset. Along with the man in the corner, the model is also capable of detecting the pedestrians who are present in the center of the images with an accuracy score of 0.997 for two pedestrians and 0.966 for another one. Following the preceding examples, it is obvious that the model is capable of detecting pedestrians with a higher accuracy score regardless of whether or not the

218 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

pedestrians overlap or where in the image the pedestrian appears. Though the image only contains pedestrians and not objects like vehicles, trees, or animals, it can be observed that the model maintains the maximum accuracy percentage of ninety percent throughout the process. The model can be slightly modified in such a way that it can detect other objects as well. Below, Fig. (5) tabulates the outputs obtained from all the epochs and includes the average precision value.

Fig. (5). Average precision for given conditions.

From the above Fig., it can be inferred that the average precision of the DL model is maximum when the IoU of the model is 0.50. At the same time, it is necessary for the area to cover all of the graphs and for the maxDets to remain at a hundred to attain this maximum precision. The second maximum precision is 0.948, which is slightly lesser than the maximum precision. This value is reached when the IoU is 0.75, the area covers the entire graph and the maxDets value is 100. The minimum value of the average precision is 0.692. This precision is reached when the IoU lies between the range of 0.5 to 0.95. But only the medium of the area is covered by the graph to attain the minimum precision. However, the maxDets value will remain at 100 during the minimum precision as well. The below Fig. (6) contains the data of the same conditions of the IoU, area, and Maxdets to achieve the maximum average recall value and the minimum average recall value.

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 219

Fig. (6). Average recall for given conditions.

When the mask R-CNN algorithm is used to create a DL model, the average recall value ranges from 0.50 to 0.95 for a DL model built using this algorithm. The maximum average recall value is 0.886. At the same time, it is necessary for the area to cover a large portion of the graph and for the maxDets to remain at a hundred to attain this maximum precision. The second maximum precision is 0.870, which is slightly lesser than the maximum precision. This value is reached twice. The first time is when the IoU lies between the range of 0.5 to 0.95, the area covers the entire graph, and the maxDets value is 10. The second time is when the IoU lies between the range of 0.5 to 0.95, the area covers the entire graph and the maxDets value is 100. The minimum value of the average precision is 0.392. This precision is reached when the IoU lies between the range of 0.5 to 0.95. But only all of the area is covered by the graph to attain the minimum precision. However, the maxDets value will remain as 1 during the minimum precision of the model. 5. DISCUSSION A developed Mask R-CNN DL model capable of accurately identifying objects in uploaded photos. A drop in the loss value can be seen when the number of epochs is increased. Additionally, this model is capable of instance segmentation since it can distinguish the existence of the pedestrians whose picture is obscured by the other one, and it can identify them individually without grouping them.

220 AI and IoT-based Intelligent Health Care & Sanitation

Gupta et al.

CONCLUSION There are a lot of machine learning and DL algorithms available for object detection; the mask R-CNN is found to be one of the most efficient algorithms available for the above-said process. Another important advantage of the mask RCNN is the fact that it was one of the simplest DL algorithms to implement. The below-mentioned steps were conducted to develop an efficient DL algorithm that can provide maximum accuracy. A dataset of 170 images of pedestrians is collected. The collected images are then preprocessed using a certain step. The steps include image resizing and image augmentation. A DL model which is capable of detecting objects from the images uploaded with a maximum accuracy level is successfully developed using the Mask R-CNN algorithm. The processed images are then used to train the model. This model is then trained a lot of times until it achieves maximum accuracy. It could be observed that the loss, the loss classifier value, the loss box reg value, the loss mask value, the loss objectness value, and the loss rpn box reg value decreases gradually as the number of epochs is increased and reaches the most possible minimum value. Once the model is trained sufficient times, the model is then tested for the average precision and average recall value. The model is then tested with some real-time images, and it is observed that the model is capable of detecting the presence of pedestrians with a higher accuracy score irrespective of the overlapping of the pedestrians or the position of the pedestrian in the image. From the above inferences, it can be concluded that the DL model, which was designed using the Mask R-CNN, is undoubtedly one of the most efficient algorithms in object detection. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

Phil Ammirato, Alexander C. Berg, “A Mask-RCNN Baseline for Probabilistic Object Detection”. arXiv:1908.03621, 2019.

[2]

A. Mahmoud, S. Mohamed, R. El-Khoribi, and H. AbdelSalam, "Object Detection Using Adaptive Mask RCNN in Optical Remote Sensing Images", Int. J. Intell. Eng. Syst., vol. 13, no. 1, pp. 65-76, 2020.

Deep Learning

AI and IoT-based Intelligent Health Care & Sanitation 221

[http://dx.doi.org/10.22266/ijies2020.0229.07] [3]

J.W. Johnson, "Automatic Nucleus Segmentation with Mask-RCNN", In: Science and Information Conference, 2019. [http://dx.doi.org/10.1007/978-3-030-17798-0]

[4]

E. Kopelowitz, G. Engelhard, “Lung Nodules Detection and Segmentation Using 3D Mask-RCNN” arXiv:1907.07676, 2019. [http://dx.doi.org/10.48550/arXiv.1907.07676]

[5]

M. Guillermo, "Implementation of Automated Annotation through Mask RCNN Object Detection model in CVAT using AWS EC2 Instance", In: 2020 IEEE Region 10 Conference. Tencon, 2020, pp. 708-713. [http://dx.doi.org/10.1109/TENCON50793.2020.9293906]

[6]

H. Zhang, Y. Wang, F. Dayoub, N. Sunderhauf. “SWA Object Detection”. arXiv:2012.12645, 2021.

[7]

C. Papageorgiou, and T. Poggio, "A Trainable System for Object Detection", Int. J. Comput. Vis., vol. 38, no. 1, pp. 15-33, 2000. [http://dx.doi.org/10.1023/A:1008162616689]

[8]

Z. Zou, Z. Shi, Y. Guo, and J. Ye, "Object Detection in 20 Years", Survey (Lond.), 2019.

[9]

A.R. Pathak, M. Pandey, and S. Rautaray, "Application of DL for Object Detection", In: International Conference on Computational Intelligence and Data Science vol. 132. , 2018, pp. 1706-1717. [http://dx.doi.org/10.1016/j.procs.2018.05.144]

[10]

Q. Wu, C. Shen, P. Wang, A. Dick, and A. van den Hengel, "Image captioning and visual question answering based on attributes and external knowledge", IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 6, pp. 1367-1381, 2018. [http://dx.doi.org/10.1109/TPAMI.2017.2708709] [PMID: 28574341]

[11]

P. Dollár, C. Wojek, B. Schiele, and P. Perona, "Pedestrian detection: an evaluation of the state of the art", IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 4, pp. 743-761, 2012. [http://dx.doi.org/10.1109/TPAMI.2011.155] [PMID: 21808091]

[12]

M.A. Fischler, and R.A. Elschlager, "The representation and matching of pictorial structures", IEEE Trans. Comput., vol. C-22, no. 1, pp. 67-92, 1973. [http://dx.doi.org/10.1109/T-C.1973.223602]

[13]

S.P. Yadav, "Emotion recognition model based on facial expressions", Multimedia Tools Appl., vol. 80, no. 17, pp. 26357-26379, 2021. [http://dx.doi.org/10.1007/s11042-021-10962-5]

[14]

S.P. Yadav, S. Zaidi, A. Mishra, and V. Yadav, "Survey on Machine Learning in Speech Emotion Recognition and Vision Systems Using a Recurrent Neural Network (RNN)", Arch. Comput. Methods Eng., vol. 29, no. 3, pp. 1753-1770, 2022. [http://dx.doi.org/10.1007/s11831-021-09647-x]

[15]

S.P. Yadav, "Vision-based Detection, Tracking, and Classification of Vehicles", IEIE Transactions on Smart Processing & Computing, vol. 9, no. 6, pp. 427-434, 2020. [http://dx.doi.org/10.5573/IEIESPC.2020.9.6.427]

222

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 222-239

CHAPTER 15

Design and Comparison Of Deep Learning Architecture For Image-based Detection of Plant Diseases Makarand Upadhyaya1,*, Naveen Nagendrappa Malvade2, Arvind Kumar Shukla3, Ranjan Walia4 and K Nirmala Devi5 University of Bahrain, Department of Management & Marketing, College of Business Administration, Bahrain 2 Department of Information Science and Engineering, SKSVM Agadi College of Engineering and Technology, Lakshmeshwar, Karnataka, India 582116 3 Department of Computer Application, IFTM University, Moradabad, Uttar Pradesh, India 4 Department of Electrical Engineering, Model Institute of Engineering and Technology, Jammu, Jammu & Kashmir-181122, India 5 Department of Computer Science and Engineering, Kongu Engineering College, Perundurai, Erode, Tamilnadu– 638060, India 1

Abstract: Agriculture provides a living for half of India's people. The infection in crops poses a danger to food security, but quick detection is hard due to a lack of facilities. Nowadays, Deep learning will automatically diagnose plant diseases from raw image data. It assists the farmer in determining plant health, increasing productivity, deciding whether pesticides are necessary, and so on. The potato leaf is used in this study for analysis. Among the most devastating crop diseases is potato leaf blight, which reduces the quantity and quality of potato yields, significantly influencing both farmers and the agricultural industry as a whole. Potato leaves taken in the research contain three categories, such as healthy, early blight, and late blight. Convolution Neural Network (CNN), and Convolution Neural Network- Long Short Term Memory(CNN-LSTM) are two neural network models employed to classify plant diseases. Various performance evaluation approaches are utilized to determine the best model.

Keywords: Accuracy, Blight, CNN, Loss, LSTM. Corresponding author Makarand Upadhyaya: University of Bahrain, Department of Management & Marketing, College of Business Administration, Bahrain; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 223

1. INTRODUCTION Agriculture made a tremendous contribution to the development of today’s modern primary food crops. Food insecurity [1, 2], which would be a significant problem arising from plant infections, is among the most pressing global issues confronting human civilization. Plant diseases, based on one research, contribute to approximately 16% of worldwide agricultural production loss [3]. Pesticide losses for wheat are expected to be over 50% and more than 25% for soya worldwide. Fast and effective identification of crop diseases is essential for both healthy and profitable cultivation as well as for avoiding financial loss of investments. Certain crops don't show any symptoms at all, which necessitates the employment of sophisticated diagnostic techniques. Though some plants have invisible signals, most diseases may be diagnosed by an expert plant pathologist through a visual inspection of an affected crop [4]. A crop pathologist's capability to track plants and recognize common symptoms of the disease are essential to their profession. Furthermore, due to a large number of plant species, alterations in the development of crop diseases, a consequence of global warming, and the quicker outbreak of infection to new locations, even expert pathologists are unable to detect some illnesses [5]. Agronomists benefit greatly from the existence of professional and sophisticated technologies which can properly identify crop diseases autonomously. However, providing such technology with a simplistic smartphone app even help non-expert farmers. This app gives significant profit to farmers who lack agronomists and technical support systems [6]. Plants that have been afflicted with a disease generally have visible scars or marks on the plant parts, such as leaves, branches, fruits, or flowers. In general, it is possible to use the out-of-the-ordinary appearances of many diseases as diagnostic tools. The leaves of crops are often vital sources for identifying plant illnesses, and the majority of infection signals can start solely on leaves [7]. While demand for processed potato products has increased, fresh potato consumption has dropped. As more people rely on potatoes as a staple diet, even minor differences in the nutritional profile of potatoes will have far-reaching effects on public health. As it is high in carbs, the potato delivers energy while being low in fat. Although potatoes are low in protein, they have a significant biological value (90-100). Potatoes are an excellent source of numerous critical nutrients, including vitamin C, potassium, and a variety of B vitamins. The skins contain a lot of fiber. Several compounds in potatoes contribute to their antioxidant effect, and there has recently been an increase in demand for cultivars with colored skins. It is critical to understand the health advantages of potatoes to improve the overall quality of our food supply.

224 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

1.1. Literature Survey The study [8] used the K-means clustering technique to extract lesions areas and merged the global color histogram (GCH), completed local binary pattern (CLBP), color coherence vector (CCV), and local binary pattern (LBP) to obtain the color and texture of an apple, it helps to detect the three types of infection in apple. The improved machine learning algorithm of the support vector machine (SVM) is used for classification, giving an accuracy level of 93%. Utilizing principal component analysis (PCA) and stepwise discriminant, scientists investigated multiple tomato diseases. The 18 main features, including shape, color, texture, and some information, are extracted from collected images [9]. To recover the important features, the above-mentioned model is employed. The two techniques' accuracy was 94.71 percent and 98.32 percent, correspondingly. The journal [10] discussed a variety of diagnostic approaches for identifying plant diseases, with an emphasis on imaging technology. KNN, K-means, and SVM are the main approaches given for crop diseases and categorization. In the next study, the researchers proposed hot-spot recognition using image processing techniques and for earlier detection of 3 categorizations of wheat diseases. They implemented their proposed technique on cell phones and tested it in a real-world setting [11]. A novel model for the autonomous identification of crop diseases called the Group Method of Data Handling logistics, is presented [12]. For identifying plant diseases, several techniques merge feature extraction with Neural Network Ensemble. It provides a greater generalization of learning ability by training a particular number of neural networks [13]. This approach was used only for detecting tea leaf diseases, with a testing accuracy of 91% [14]. Utilized Probabilistic Neural Network for disease detection [15]; it contains four layers, including an input, pattern, summation and output layer. Additionally, this neural net seems to have a quicker training time than Back Propagation Neural Network (BPNN). A three-layered BPNN was used for classification tasks, with the first layer containing 22 neurons trying to describe the characteristics of the image which is extracted in the feature extraction stage, the hidden layer containing comprising four neurons, and finally, the last layer containing just a single neuron during the learning process and the outcome to be 0 or 1 depending on whether the given input image was good or unhealthy [16]. Another strategy depends on imaging data and employing ANN in combination with the K-means clustering algorithm for automating recognition and categorization of infection present in plants [17]. ANN was made up of ten secret layers. The range of outcomes was six, corresponding to five types of disease leaf and a normal leaf. The performance measure of categorization obtained using this method was 94.67 percent of the total. The PlantVillage group has been gathering

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 225

10,000 pictures of normal and sick field crops [18] and making those available on the internet free of cost. Using a CNN method, researchers were able to classify 26 illnesses in 14 different crops utilizing 54,306 pictures. The capacity of the algorithms to forecast the proper crop diseases match, given 38 distinct categories, and evaluate their performance. The great model earns an F1 score of 99.34%, indicating the technical viability of the method. These findings pave the way for a cellphone-based plant disease diagnostic system. The authors discuss an autonomous wheat disease diagnostic structure based on a supervised deep learning architecture, that accomplishes incorporation of wheat infection classification and segmentation for disease areas by using image captioning for training datasets in wild environments [19]. Wheat Disease Directory 2017 is composed a fresh collection of wheat disease images, which is also gathered to examine the effectiveness of the approach. For 5-fold cross-validation, the algorithm obtains average classification accuracies of 95.12 percent and 97.95 percent for two distinct models, namely VGG-FCN-S and VGG-FCN- VD16. The chapter is discussed as follows. Sector 2 discusses the workflow of leaf disease detection using deep learning methods, including CNN and CNN-LSTM. Sector 3 details the data collection and preparation techniques. The deep algorithm is explained in Sector 4. The results and comparison are made in Sector 5. Finally, the Sector 6 concludes the better construction of a neural network for disease classification. 2. METHODOLOGY The step-by-step methodology used to detect potato disease using leaf images is shown in Fig. (1). The first step is collecting the data from the Kaggle website. The data contain healthy, early, and late blight leaf images. The collected images are passed to the pre-processing section, such as image resize and rescale. For a great model, more data are required. Therefore, the data augmentation is done on the preprocessed data. The data augmentation techniques used in this work are rotation and flipping. Then, the deep learning network is designed. Totally two networks are used named CNN and LSTM. The data is divided into three segments such as train, validate and test. The standard ratio is used for splitting the dataset. The model is finalized using the metrics like accuracy and loss. The number of epochs, filter, kernel size, and hidden layer is changed until the metrics reach the satisfaction level. After getting good accuracy and loss values, both models are finalized. Finally, both designed models are compared using the accuracy level.

226 AI and IoT-based Intelligent Health Care & Sanitation

Fig. (1). Workflow of potato leaf diseases detection using deep learning.

Upadhyaya et al.

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 227

3. DATA COLLECTION AND PREPARATION 3.1. Data Collection The first step inside the deep learning piping system is to gather data for constructing the deep learning classification model. Deep learning algorithms can only make excellent predictions based on the data on that they were educated. The healthy and diseased potato leaves include early blite, and late blight images collected from the internet (Kaggle Website). The sample image in each class is exposed in Fig. (2). The first image denotes the healthy potato leaf; the second one shows the potato leaf with early blite disease, and the last one displays the potato leaf with late blite disease.

Fig. (2). Representation of potato leaf with disease.

3.2. Data Preprocessing Data preprocessing may be defined as a data mining method that transforms raw data obtained from Kaggle sources into better information more suited for the operation. To look at it another way, it is a preparatory phase that organizes, sorts, and merges most of the accessible data. Data preprocessing is a significant stage in Deep Learning since it improves the data quality and simply derives great information. In Deep Learning, data preparation relates to the procedure of processing the raw data into usable data for the construction and learning of Deep Learning models. In this work, rescaling and reshaping techniques are used. The collected images are converted into standard size 256*256. That means the image contains 256 rows and 256 columns. The resizing is important because the neuron

228 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

in the input layers is based on the shape and type of image. In case any images collected from the web are of different sizes. This step helps to change the size of the image. 3.3. Data Augmentation The gathering and labeling of a significant quantity of disease images require a lot of labor, recourses, and economic ability in the identification of leaf diseases. Several plant diseases have a lower incubation period, making samples harder to obtain. Deep learning suffers from poor identification because of limited sample sizes and unequal datasets. To use a deep learning approach for leaf disease recognition, more data must be collected. The primary goal of using augmentation is to enhance the data and provide a little bit of imperfection to the images, thus aiding in preventing the fitting problem in the training phase. To have a comparable number of images in each category, data augmentation may assist. Rotating and flipping were used as two augmentation methods to make the training images more powerful and effective. The images were rotated in a clockwise and anticlockwise manner at an inclination of 20°. The flipping was accomplished by converting vertically-oriented images into right and left as well as up and down. 4. DEEP LEARNING NETWORKS 4.1. Convolution Neural Network The CNN is the famous deep learning structure that was motivated by live organisms' natural sensory perception capabilities. Hubel and Wiesel [20] discovered in 1959 that several units in the visual brain of animals are useful for detecting lighting in fields. Kunihiko Fukushima, who was encouraged by such an observation, suggested a recognition technique in 1980 [21] that might be considered the forerunner of CNN. CNN's, a group of image-processing techniques, were utilized in various applications, involving image categorization, object identification, and image semantic separation. CNN's can retrieve adequate information for image categorization due to their great capacity to autonomously learn high-level feature representation [22, 23]. CNNs are comprised of 3 layers: convolution, pooling, and fully connected (FC). The first two layers include convolution and pooling, which are responsible for extracting features, while the final, a fully linked layer, translates those features to the outcome, like categorization.

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 229

A convolution layer is a main share of CNN. It is made up of a sequence of numerical computations like convolution, which would be a specific form of linear action. The graphic shows the CNN that was used in this study. The image is given as input to the first convolution network, consisting of 32 filters, 3*3 kernels, and Relu, an activation function. Next, the pooling layer is used with filter size 2*2. Like that other five hidden layers are used with the combo of convolution and pooling. Next, the 6th pooling layer is converted to flatten layer which is connected to the dense layer. The neurons in the flattened layer consist of fully connected to the previous and succeeding output layer. The final output layer consists of 3 neurons, and the softmax activation function is used for multiple classifications. The architecture of CNN for potato disease classification is given in Fig. (3).

Fig. (3). CNN for potato leaf disease classification.

4.2. CNN-Long Short-Term Memory Deep learning employs a type of Recurrent Neural Network (RNN) known as Long Short-Term Memory (LSTM). It is important to note that, in contrast to other types of neural models, LSTM contains input connections. It is capable of processing not only single data points like pictures but whole streams. LSTM, for instance, relates to tasks including non-segmented handwritten text, speech synthesis, and the recognition of abnormalities in texture patterns. A typical LSTM unit has four gates: an entering, exiting and forgetting gate. The cell retains data for arbitrary lengths of time, and the three gates control the information flow in and out of the cell. As a result, LSTM networks are quite well for time series analysis, categorization, and prediction because there could be lags of undetermined duration between important incidents. LSTMs were created to solve the issues with standard RNN preparations, such as gradient vanishing [24].

230 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

The CNN-LSTM is the combination of CNN and LSTM. In this work, CNN is used to extract the required features of leaf, and those features are given as input to the LSTM model. As discussed in Sector 3 the same layer of CNN is used. The output of the flattened layer is taken and reshaped. The reshaped data is given as input to the LSTM layer. The LSTM layer is followed by a fully connected layer with a ‘relu’ activation function. The final layer contains 3 neurons with a ‘softmax’ activation function to predict the leaf conditions. The architecture of CNN-LSTM for potato disease classification is given in Fig. (4).

Fig. (4). CNN-LSTM for potato leaf disease classification.

5. RESULT AND DISCUSSION The image of potato disease is taken from the Kaggle website. Then the image is split in the ratio of 7:2:1. This ratio says that 70% of data aids for the training phase and 20% for the validation phase. Finally, the last 10% of data is taken for the testing phase. The model is finalized based on the performance measures such as accuracy and loss. 5.1. CNN Performance Measure Table 1 shows the accuracy and loss of training and validation data for 15 epochs. From the table, the training and validation accuracy increases, and at epoch 15, the accuracy value is nearly 98%. Similarly, the training and validation loss column in the table decrease as the number of epochs increases. At the final epoch, the training loss is 0.0436, and the validation loss is 0.0775. The two values are present at the tolerable level only. Therefore, the model is good enough to predict potato disease based on its leaf image.

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 231

Table 1. Accuracy and loss of training and validation data for 15 epoch - CNN algorithm. Epochs

Training Accuracy

Validation Accuracy

Training Loss

Validation Loss

1

0.5478

0.56

0.8898

0.9023

2

0.6875

0.7764

0.7372

0.6038

3

0.8071

0.9062

0.4571

0.3034

4

0.9095

0.9254

0.2517

0.2272

5

0.9109

0.9399

0.2265

0.1669

6

0.9401

0.9423

0.1554

0.1443

7

0.9494

0.923

0.1217

0.1843

8

0.9461

0.9519

0.1382

0.0967

9

0.966

0.9495

0.0907

0.1255

10

0.964

0.9639

0.0946

0.0969

11

0.976

0.9783

0.0662

0.0691

12

0.97

0.9519

0.0823

0.0963

13

0.984

0.9855

0.0499

0.0452

14

0.9833

0.9759

0.0567

0.0686

15

0.984

0.9807

0.0436

0.0775

For better understanding and visualization, the line plot is used. Fig. (5) shows the training and validation accuracy for each epoch. The magenta color plot in the graph represents the accuracy of training data, and the black color represents the accuracy of validation data. From analyzing the plot, the accuracy of both training and validation rapidly increases for the first 5 epochs. Then at epoch 15, both give a good accuracy level greater than 98%. Fig. (6) shows the train and validate loss for each epoch. Similar to the previous Fig. (5), the magenta color plot in the graph represents the loss in the training phase, and the black color represents the loss in the validation phase. From the figure, it is clear that the error between the actual and predicted data goes on decreasing.

232 AI and IoT-based Intelligent Health Care & Sanitation

Fig. (5). Training and Validation accuracy of CNN for 15 epochs.

Fig. (6). Training and Validation loss of CNN for 15 epochs.

Upadhyaya et al.

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 233

The confusion matrix of the CNN model for potato leaf disease is demonstrated in Fig. (7). The figure shows the number of true and false predicted diseases using the image by deploying the CNN model. The 125 values in the figure show that the number correctly predicted early blight leaf. Like that, 113 and 12 represent the correctly estimated late blight and healthy leaf. Numbers 3 and 2 represent that the model wrongly predicted the leaf has late blight, but it is an early and healthy leaf. The 1 represents the actual late blight leaf is presdicted wrongly as early blight. Finally, the sum of truly predicted leaf conditions is 250, and the sum of falsely predicted leaf conditions is 6. These values are highly helpful for calculating metrics.

Fig. (7). Confusion Matrix of CNN.

5.2. CNN- LSTM Training and validation data performance, such as the accuracy and loss concerning epochs, are shown in Table 2. According to the table, training and validation accuracy pt rising until epoch 15, and it achieved about 99 percent. A higher epoch count results in a reduction a training and validation loss. The training loss was 0.0224, and the validation loss was 0.0302 in the final epoch. This shows the model is excellent enough to forecast potato conditions from a picture of a leaf. Table 2. Accuracy and loss of training and validation data for 15 epoch - CNN algorithm. Epochs

Training Accuracy

Validation Accuracy

Training Loss

Validation Loss

1

0.5219

0.6145

0.8893

0.81509

2

0.7366

0.7916

0.6199

0.4456

234 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

(Table ) cont.....

Epochs

Training Accuracy

Validation Accuracy

Training Loss

Validation Loss

3

0.8599

0.8385

0.3531

0.341

4

0.8923

0.9218

0.2624

0.1929

5

0.9108

0.9375

0.2257

0.1889

6

0.9427

0.9479

0.145

0.1159

7

0.9548

0.9531

0.1157

0.1201

8

0.9513

0.8958

0.1268

0.2288

9

0.9577

0.9583

0.108

0.0988

10

0.97627

0.9843

0.0698

0.0605

11

0.9629

0.9531

0.0923

0.093

12

0.9768

0.9687

0.0663

0.0903

13

0.9861

0.9739

0.0405

0.0404

14

0.9878

0.9843

0.0308

0.0375

15

0.9913

0.9903

0.0224

0.0302

Fig. (8) displays the training and validation accuracy of every epoch. Magenta indicates training data accuracy, whereas black indicates validation data accuracy in the graph. Analyzing the figure reveals that over the first four epochs, training and validation accuracy rises rapidly. At epoch 15, the accuracy of training as well as testing is greater than 99 percent. Fig. (9) depicts the training and validation loss for each period. The magenta color plot in the graph reflects the loss of training data, while the black color represents the correctness of the validation data, identical to Fig. (8). Fig. (10) shows the CNN-LSTM model's confusion matrix for leaf disease. The chart illustrates the amount of correctly predicted and wrongly predicted leaf conditions. The 126 numbers in the figure demonstrate that the total correctly predicted an early blight leaf. In the same way, the numbers 114 and 13 correspond to a correctly predicted late blight and a healthy leaf, respectively. The numbers 2 and 1 show that the model mistakenly estimated that the leaf has late blight, whereas it is an early blight and a healthy leaf. Finally, the total of accurately predicted leaf conditions is 253, whereas the total of incorrectly predicted leaf conditions is 3.

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 235

Fig. (8). Training and Validation accuracy of CNN for 15 epochs.

Fig. 9. Training and Validation loss of CNN for 15 epochs.

236 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

Fig. (10). Confusion Matrix of CNN.

The accuracy of both models is plotted in the graph, and is shown in Fig. (11). The CNN accuracy is 97.65%, and the CNN-LSTM accuracy is 98.82%. This plot is done for better understanding.

Fig. (11). Accuracy comparison.

6. NOVELTY This research uses deep learning algorithms such as CNN and CNN-LSTM to detect potato leaf diseases with high accuracy, precision, and low loss percentage when compared to other existing models. This makes the modal more dependable and authentic. CONCLUSION The data is successfully collected from a trustable website (Kaggle). Then the data

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 237

is pre-processed to make the data ready for model training. The two models are taken, namely CNN and CNN-LSTM. The same amount of data is passed through the training, validation, and testing phase. The two models are constructed and trained using the pre-processed image. Both models are finalized using the accuracy and loss values obtained from the training and validation phase. The finalized model is used to predict the test data. The confusion matrix is used to analyse the model prediction quality. The total number of correct leaf conditions using the CNN model is 250 and using the CNN-LSTM model is 253. The analysis is made to conclude the CNN-LSTM gives better accuracy for potato leaf disease prediction. CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

M.E.H. Chowdhury, A. Khandakar, S. Ahmed, F. Al-Khuzaei, J. Hamdalla, F. Haque, M.B.I. Reaz, A. Al Shafei, and N. Al-Emadi, "Design, construction and testing of iot based automated indoor vertical hydroponics farming test-bed in qatar", Sensors (Basel), vol. 20, no. 19, p. 5637, 2020. [http://dx.doi.org/10.3390/s20195637] [PMID: 33023097]

[2]

R.N. Strange, and P.R. Scott, "Plant disease: a threat to global food security", Annu. Rev. Phytopathol., vol. 43, no. 1, pp. 83-116, 2005. [http://dx.doi.org/10.1146/annurev.phyto.43.113004.133839] [PMID: 16078878]

[3]

E.C. Oerke, "Crop losses to pests", J. Agric. Sci., vol. 144, no. 1, pp. 31-43, 2006. [http://dx.doi.org/10.1017/S0021859605005708]

[4]

S. Sankaran, A. Mishra, R. Ehsani, and C. Davis, "A review of advanced techniques for detecting plant diseases", Comput. Electron. Agric., vol. 72, no. 1, pp. 1-13, 2010. [http://dx.doi.org/10.1016/j.compag.2010.02.007]

[5]

S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D. Stefanovic, "Deep neural networks-based recognition of plant diseases by leaf image classification", Comput. Intell. Neurosci., vol. 2016, pp. 111, 2016. [http://dx.doi.org/10.1155/2016/3289801] [PMID: 27418923]

[6]

K.P. Ferentinos, "Deep learning models for plant disease detection and diagnosis", Comput. Electron. Agric., vol. 145, pp. 311-318, 2018. [http://dx.doi.org/10.1016/j.compag.2018.01.009]

[7]

M.A. Ebrahimi, M.H. Khoshtaghaza, S. Minaei, and B. Jamshidi, "Vision-based pest detection based on SVM classification method", Comput. Electron. Agric., vol. 137, pp. 52-58, 2017. [http://dx.doi.org/10.1016/j.compag.2017.03.016]

238 AI and IoT-based Intelligent Health Care & Sanitation

Upadhyaya et al.

[8]

S.R. Dubey, and A.S. Jalal, "Adapted approach for fruit disease identification using image", Int. J. Comput. Vis. Image Process., vol. 2, no. 3, pp. 44-58, 2012. [http://dx.doi.org/10.4018/ijcvip.2012070104]

[9]

A-L. Chai, B-J. Li, Y-X. Shi, Z-X. Cen, H-Y. Huang, and J. Liu, "Recognition of tomato foliage disease based on computer vision technology", Yuan Yi Xue Bao, vol. 37, no. 9, pp. 1423-1430, 2010.

[10]

V. Singh, N. Sharma, and S. Singh, "A review of imaging techniques for plant disease detection", Artificial Intelligence in Agriculture, vol. 4, pp. 229-242, 2020. [http://dx.doi.org/10.1016/j.aiia.2020.10.002]

[11]

A. Johannes, A. Picon, A. Alvarez-Gila, J. Echazarra, S. Rodriguez-Vaamonde, A.D. Navajas, and A. Ortiz-Barredo, "Automatic plant disease diagnosis using mobile capture devices, applied on a wheat use case", Comput. Electron. Agric., vol. 138, pp. 200-209, 2017. [http://dx.doi.org/10.1016/j.compag.2017.04.013]

[12]

J. Chen, H. Yin, and D. Zhang, "A self-adaptive classification method for plant disease detection using GMDH-Logistic model", Sustainable Computing: Informatics and Systems, vol. 28, p. 100415, 2020. [http://dx.doi.org/10.1016/j.suscom.2020.100415]

[13]

Z.H. Zhou, Y. Jiang, Y.B. Yang, and S.F. Chen, "Lung cancer cell identification based on artificial neural network ensembles", Artif. Intell. Med., vol. 24, no. 1, pp. 25-36, 2002. [http://dx.doi.org/10.1016/S0933-3657(01)00094-X] [PMID: 11779683]

[14]

B.C. Karmokar, and M.S. Ullah, "Md. K. Siddiquee, and K. Md. R. Alam.: “Tea leaf diseases recognition using neural network ensemble”", Int. J. Comput. Appl., vol. 114, no. 17, pp. 27-30, 2015. [http://dx.doi.org/10.5120/20071-1993]

[15]

K. Majid, Y. Herdiyeni, and A. Rauf, 2013 I-PEDIA: “Mobile application for paddy disease identification using fuzzy entropy and probabilistic neural network”. Int. Conf. on Adv. Comp. Sci. & Info. Sys. 2013, pp. 403-6. [http://dx.doi.org/10.1109/ICACSIS.2013.6761609]

[16]

D. Majumdar, D.K. Kole, A. Chakraborty, and D.D. Majumder, “An integrated digital image analysis system for detection, recognition and diagnosis of disease in wheat leaves”, In Proc. of the Third Int. Symp. on W. in Comp. & Info. pp. 400-5 2015. [http://dx.doi.org/10.1145/2791405.2791474]

[17]

H. Al Hiary, S. Bani Ahmad, M. Reyalat, M. Braik, and Z. ALRahamneh, "Fast and accurate detection and classification of plant diseases", Int. J. Comput. Appl., vol. 17, no. 1, pp. 31-38, 2011. [http://dx.doi.org/10.5120/2183-2754]

[18]

D. Hughes, and M Salathé, “An open access repository of images on plant health to enable the development of mobile disease diagnostics”. 2015. arXiv preprint arXiv:1511.08060.

[19]

J. Lu, J. Hu, G. Zhao, F. Mei, and C. Zhang, "An in-field automatic wheat disease diagnosis system", Comput. Electron. Agric., vol. 142, pp. 369-379, 2017. [http://dx.doi.org/10.1016/j.compag.2017.09.012]

[20]

D.H. Hubel, and T.N. Wiesel, "Receptive fields and functional architecture of monkey striate cortex", J. Physiol., vol. 195, no. 1, pp. 215-243, 1968. [http://dx.doi.org/10.1113/jphysiol.1968.sp008455] [PMID: 4966457]

[21]

K. Fukushima, and S. Miyake, “A self-organizing neural network model for a mechanism of visual pattern recognition”, In Competition and Cooperation in Neural nets 1982, pp. 267–285. [http://dx.doi.org/10.1007/978-3-642-46466-9_18]

[22]

X. Qiu, M. Li, L. Dong, G. Deng, and L. Zhang, “Dual-band maritime imagery ship classification based on multilayer convolutional feature fusion”, J. Sens, vol. 2020, 1-16. [http://dx.doi.org/10.1155/2020/8891018]

[23]

D. G. Lee, Y. H. Shin, and D.-C. Lee, “Land cover classification using SegNet with slope, aspect, and

Plant Diseases

AI and IoT-based Intelligent Health Care & Sanitation 239

multidirectional shaded relief images derived from digital surface model”. J. Sens 2020, 1-21. [http://dx.doi.org/10.1155/2020/8825509] [24]

S. Hochreiter, "The vanishing gradient problem during learning recurrent neural nets and problem solutions", Int. J. Uncertain. Fuzziness Knowl. Based Syst., vol. 6, no. 2, pp. 107-116, 1998. [http://dx.doi.org/10.1142/S0218488598000094]

240

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 240-254

CHAPTER 16

Discernment of Paddy Crop Disease by Employing CNN and Transfer Learning Methods of Deep Learning Arvind Kumar Shukla1,*, Naveen Nagendrappa Malvade2, Girish Saunshi3, P. Rajasekar4 and S.V. Vijaya Karthik4 Department of Computer Applications, IFTM University Moradabad, IFTM University, Moradabad, U.P, India-244001 2 Department of Information Science and Engineering, SKSVM Agadi College of Engineering and Technology, Lakshmeshwar, Karnataka, India 582116 3 Department of Computer Science and Engineering, KLE Institute of Technology, KLEIT, Gokul, Karnataka, India-580031 4 Department of ECE, Kings College of Engineering, Punalkulam, Gandarvakottai Taluk, Pudukkottai, Tamil Nadu, India-613005 1

Abstract: Agriculture is the backbone of human civilization since it is a requirement of every living entity. Paddy agriculture is extremely important to humans, particularly in Asia. Farmers are currently facing a deficit in agricultural yield owing to a variety of factors, one of which is illness. The composition of paddy crop diseases is complicated, and their presentation in various species is highly similar, making classification challenging. These agricultural infections must be discovered and diagnosed as soon as feasible to avoid disease transmission. The disease significantly impacts crop productivity, and early detection of paddy infections is critical to avoiding these consequences. These issues arise as a result of a lack of awareness regarding health. Identifying the disease needs the best expertise or previous knowledge to regulate it. This is both difficult and costly. To address the aforementioned problem, a Deep Learning (DL) model was created utilizing a Convolutional Neural Network (CNN) and the transfer learning approach. The model is trained using an image of a paddy crop as input. By comparing metrics like accuracy and loss, the optimum technique is identified.

Keywords: Agriculture, Image Augmentation, Paddy, Preprocessing, Transfer Learning. Corresponding author Arvind Kumar Shukla: Department of Computer Applications, IFTM University Moradabad, IFTM University, Moradabad, U.P, India-244001; E-mail: [email protected]

*

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 241

1. INTRODUCTION Agriculture is an occupation that is more than just an occupation. This is because, unlike other occupations, agriculture provides food to people working in various other sectors. Agriculture not only provides food; it provides occupation to workers who are not directly linked to agriculture. The need for agriculture and agriculture-related professions will keep increasing with the increase in population. In India, agriculture is said to be the main occupation for the majority of the population. The establishment of sedentary human civilization in agriculture was a crucial breakthrough, allowing humans to live in cities thanks to surpluses of food generated by tamed species. But in recent times, agriculture has been considered an occupation with a bane as it requires a lot of manpower, water, etc. Though farmers try their best to survive the need for water and human power, they are then faced with another problem. The problem is plant diseases. Many types of plant diseases can affect the growth of the plant and also the yield produced. These plant diseases are common in food material crops like paddy, wheat, millets, etc. Many pesticides and chemicals that can prevent plant diseases are available on the market. But the problem arises when the farmer is unaware of the problem. This ignorance can lead to heavy loss for the farmer and eventually to stop farming. It is necessary to adopt some preventative actions to limit the possibility of farmers switching to other occupations. This research aims the development of a DL model which can be used to predict the type of plant disease a particular plant is affected by. To test and train the model, the paddy crop is used. The paddy crop is often affected by diseases like bacterial leaf blight, brown spots, and leaf smut. To detect the above-mentioned diseases, a DL model is constructed. This model is constructed twice using two different algorithms to find the most efficient algorithm for this process. The DL algorithms include CNNs and transfer learning. The transfer learning is implemented in the DL algorithm using the VGG-10 module. The images of paddy collected from the Kaggle are then used to train both models. Once the models are trained enough, they are tested for higher accuracy and lower loss. In the end, the accuracy and the loss levels of both the models are compared, and it is found that the transfer learning algorithm is the best algorithm for disease detection in plants, with an accuracy percentage of 98.75. 2. LITERATURE SURVEY The literature survey taken for researching paddy disease classification using the paddy leaf images is detailed in this section. 1. Agriculturalists are confronted with the difficulties of diagnosing rice crop diseases and are unable to identify an effective pesticide or insecticide to manage

242 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

the disease outbreak. The aforementioned problem can be resolved by developing a machine learning model based on the CNN (CNN) algorithm that detects images of Rice Blast, Bacterial Blight, and Healthy paddy leaf images, providing a solution in the form of insecticides or pesticides to control the Rice Blast and Bacterial Blight outbreaks. It has been observed that the system is more resilient, user-friendly, and produces a faster and more cost-effective result than the present technique [1]. 2. A combination of evolutionary algorithms, such as GA, and machine learning approaches, such as AdaBoost and Bagging Classifier, has been tried to detect paddy leaf diseases. The images of paddy leaves were processed to remove undesirable distortions and improve the image qualities. It includes image scaling, brightness adjustment, filtering, lighting corrections, noise reduction, geometric transformations, and greyscale transformations, among other things, as part of the whole process. A feature extraction approach is then applied to the processed photos after they have been pre-processed to extract the needed feature dataset from the images. The accuracy of the classification system is improved by applying cascaded classifiers with adjustments to the problem statement [2]. 3. Several image-processing approaches were tested for their effectiveness in diagnosing rice plant sickness. According to the current study, the scientists looked into image processing, disease diagnosis and feature extraction approaches, among other things. In addition to exploring the potential challenges and opportunities in the field of paddy leaf disease detection, this research investigated the issue of automated rice plant disease detection [3]. 4. Employs images taken directly from the farm field to obtain photographs of healthy paddy plants as well as images of paddy plants plagued with pests and illnesses for use in our photography. RGB pictures are converted to HSV before being utilized in pre-processing, where the hue component masking technique is employed to remove the background from the image. A clustering method is used to separate the damaged sections from the healthy parts during the segmentation process. The proposed DNN JOA approach is used to classify diseases, with the JOA selecting the best weights for each classification. A feedback loop is established in our system, which determines its stability. The results of analyzing and comparing experimental findings using ANN, DAE, and DNN offer several relevant and helpful metrics [4]. 5. Neutrosophic logic, which is derived from the fuzzy set, was proposed as a strategy for segmenting data for ROI evaluation. Three types of membership functions were used to segment the data for this study, and they were as follows: To determine whether or not a plant leaf was diseased, segmented regions were

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 243

used to make the determination. The random forest technique outperformed the others in the demonstration, which was done with a range of classifiers to demonstrate the concept. It was discovered that there were 400 leaf photographs in the collection, 200 of which were diseased and 200 of which were not [5]. 3. METHODOLOGIES A dataset containing 120 images of the paddy crop is collected from Kaggle. Out of the 120 images, forty images contain the healthy paddy leaf, forty images with the leaf affected by early leaf blight, and the rest of the images contains the leaves which are affected by late stages of leaf blight. The total images are divided into two parts in the ratio of 8:2. 80 percent of the entire dataset is used to train the model, and the rest 20 percent will be used for further testing. The images are then preprocessed using some techniques to maintain uniformity among the images. The images then undergo certain steps to train and test the model. The steps involved in the prediction of crop disease are explained in the below block diagram. As shown in Fig. (1), the images obtained from Kaggle are then processed using image augmentation. In this step, the images undergo processes like rescaling, shearing, zooming, and horizontal flip to ensure all the images have the same dimensions. Two DL models are constructed using two algorithms, namely the CNN and the transfer learning algorithm. The processed images are then used to train both models. Once the models are trained, the accuracy and the loss percentage of both models are analyzed.

Fig. (1). Workflow of plant disease prediction.

244 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

4. DISEASE AND PREPROCESSING Rice bacterial blight, commonly known as bacterial blight shown in Fig. (2), is a lethal bacterial disease that is one of the most devastating illnesses of cultivated rice. Bacterial blight is caused by the bacteria Xanthomonas oryzae. It causes seedlings to wilt, as well as the yellowing and drying of leaves.

Fig. (2). Bacterial Leaf Blight.

Brown spots, commonly known as Helminthosporium leaf spot, is one of Louisiana's most common rice illnesses, and it is depicted in Fig. (3). The fungus Cochliobolus miyabeanus causes brown spots. It has an impact on the crop's leaves and stalks. To control primary infection at the seedling stage, seeds should be soaked in hot water for 10 to 12 minutes before planting.

Fig. (3). Brown spots.

Leaf smut is defined by the accumulation of fungal spores in sootlike masses known as sori, which grow inside blisters in bulbs, stems, floral parts, leaves and seeds, and it is exposed in Fig. (4). The fungus kills the plant's flowering

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 245

components, preventing it from producing seeds. The fungus Ustilago maydis causes this plant disease.

Fig. (4). Leaf Smut.

The images are then preprocessed using some image augmentation techniques. Image augmentation is a technique for altering existing photos to obtain more data for model training. It's a convenient way to expand the training set without needing to take new photographs when building CNNs. The augmented images then undergo the below processes to ensure the uniform dimensions and other such characteristics. 4.1. Rescaling The process of rescaling or resampling is used to make a fresh version of an image of various sizes. These processes are not, as it turns out, seamless. This is a difficult process that necessitates a balance of sharpness, efficiency, speed, and smoothness. 4.2. Image Shearing Shearing is a transformation that alters an object's shape. Object layers slide back and forth. Shear can occur in either one or two directions, namely X and Y. The object's shape will be altered. 4.3. Zooming Zooming is a procedure in which additional pixels are added to a picture to increase its size, with the main task being the interpolation of the new pixels from the old pixels. Pixel replication or interpolation is used to zoom an image. 4.4. Horizontal Flip The spinning of an image around a horizontal or vertical axis is referred to as “flipping.” Horizontal flipping is used in this study. The vertical axis will be used

246 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

to flip the horizontal flip. For example, in a horizontal flip, the pixel at coordinate (x, y) in the new image will be at coordinate (width - x - 1, y). The processed images are then analyzed to detect the presence of plant diseases. 5. DL METHODS A DL model is a computer-generated model which is used to classify images, texts, and even sounds. This DL model can be developed using various DL algorithms. They include algorithms like CNNs, radial basis function networks, recurrent neural networks, transfer learning, etc. [6, 7]. In this research, two algorithms, namely the CNNs and the transfer learning algorithm, are used to develop the DL model. 5.1. CNN A CNN, or CNN, is a DL algorithm that can accept an image as input, attribute significance to different objects or aspects in the image, then distinguish one from the other. One of the main advantages of the CNN algorithm is that it does not require much preprocessing compared to other classification algorithms that use DL. The CNN algorithm's architecture is based on the arrangement of the Visual Cortex and is similar to the related to particular connective theories of neurons in the human brain. The CNN algorithm can successfully extract the Temporal and Spatial relationships in a picture by using appropriate filters [8]. The CNN algorithm is also capable of performing image classification in multiple convolutional layers, which helps in the reduction of the need for many DL models to perform a single task. The CNN algorithm is said to be one of the best algorithms which can be used in the extraction of high-level characteristics like edges from an image [9]. Fig. (5) explains the architecture of the CNN algorithm. This architecture is thoroughly trained according to the requirement, i.e., the prediction of plant diseases of spaddy. There is a pooling layer, an activation layer, a fully connected layer, and a convolutional layer in the above architecture, all of which are interconnected so that CNNs may process and interpret input to classify images. After that, the images are evaluated to see whether there are any signs of plant illness [10, 11]. 5.2. Transfer Learning The transfer learning algorithm is implemented using the VGG-19 architecture. The VGG 19 architecture, also known as the Oxford Net architecture, is a DL architecture used to classify images using a DL algorithm. CNNing systems

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 247

employ this architecture for image classification. The VGG-16 model has 16 layers, thirteen of which are convolutional and three of which are completely linked. The topmost and bottommost layers are the only two layers that can be adjusted to meet the user's needs [12]. VGG-16 has several advantages, including the fact that it is a pre-trained model that does not require a lot of training and testing. VGG is a type of network that may be categorized and tracked. The VGG-16 network is made up of 41 layers. There are 16 layers with learning weights, 13 of which are coevolutionary and three of which are linked [13]. The model that is created can be used to detect any type of crop disease in the image that will be submitted [14]. Fig. (6) contains the architecture of the VGG-19 module after the required changes are images in the modifiable parameters.

Fig. (5). The architecture of the CNN algorithm.

Fig. (6). The VGG-19 architecture.

248 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

It can be deduced from the above Fig. that each convolutional block is used to train the model. 6. RESULTS AND DISCUSSION A dataset with 120 images of different types of plant diseases of paddy and images of healthy stalks is collected from Kaggle. The images are then preprocessed using image augmentation. This image augmentation technique includes processes like image rescaling, image shearing, image zooming, and horizontal flipping. A DL model which is capable of identifying the presence of plant diseases is designed using two different DL algorithms. These DL algorithms include CNNs and transfer learning algorithms. The models developed are then trained again and again until it reaches the maximum accuracy value and the minimum accuracy. While training, the accuracy, the loss, and the time taken for each epoch for the CNN algorithm are tabulated in Table 1. Table 1. Accuracy and loss for the CNN algorithm. Training Accuracy

Training Loss

Time

0.297619045

1.0003986

5s

0.34523809

0.9965899

3s

0.416666657

0.980482

4s

0.464285702

0.980004

3s

0.547619045

0.9876586

4s

0.702380955

0.9005926

4s

0.738095224

0.833176

3s

0.690476179

0.7586933

4s

0.785714269

0.6538414

4s

0.75

0.705868

3s

0.797619045

0.6016364

4s

0.84523809

0.5292866

4s

0.809523821

0.4944005

4s

0.821428597

0.4439789

4s

0.84523809

0.4193441

3s

0.892857134

0.3640049

4s

0.833333313

0.3798128

3s

0.916666687

0.3735335

4s

0.84523809

0.3444705

4s

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 249

(Table 1) cont.....

Training Accuracy

Training Loss

Time

0.869047642

0.3982665

4s

0.90476191

0.2886881

3s

0.892857134

0.2899225

4s

0.916666687

0.2495825

4s

0.90476191

0.2772922

4s

0.892857134

0.2752845

4s

0.952380955

0.1925959

4s

0.948571403

0.1836118

3s

0.956666687

0.1579013

4s

0.962380955

0.163199

3s

0.972380955

0.1359124

4s

From Table 1, it can be seen that the accuracy of the CNN algorithm is 29% in the first epoch, and the accuracy keeps increasing with an increase in the number of epochs. By the end of the thirtieth epoch, the accuracy of the CNN algorithm is increased to 97%, which is great for a disease detection algorithm. It can also be seen that the loss of the algorithm at the first epoch is almost equal to 100%, and it keeps on decreasing with the increase in epochs. In the end, the loss of the CNN algorithm is 13%. The data in the above table is pictorially represented as a graph in Fig. (7).

Fig. (7). Accuracy and loss graph for CNN algorithm.

The accuracy and loss details of the CNN algorithm followed by the accuracy and loss that occurred during every epoch of training the transfer learning algorithm is tabulated in Table 2.

250 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

Table 2. Accuracy and loss of the transfer learning model. Training Accuracy

Training Loss

Time

0.161904776

0.97782

78s

0.476190478

0.95112

66s

0.583333313

0.756215

65s

0.761904776

0.6358432

73s

0.797619045

0.6048943

64s

0.857142866

0.3616805

64s

0.880952358

0.3947337

72s

0.880952358

0.315268

58s

0.892857134

0.2491756

66s

0.928571403

0.2340974

60s

0.940476179

0.1649515

70s

0.964285731

0.1695967

63s

0.976190448

0.1026523

64s

0.952380955

0.1284684

64s

0.964285731

0.1385558

63s

0.988095224

0.0814996

60s

0.988095224

0.0804616

67s

0.988095224

0.0904971

67s

0.976190448

0.0883934

72s

0.988095224

0.0920053

66s

1

0.0597332

64s

1

0.0520293

65s

1

0.0492516

65s

1

0.0536174

64s

1

0.0543899

73s

1

0.049395

73s

0.988095224

0.0623318

73s

1

0.0363308

65s

1

0.0440592

73s

0.988095224

0.0429268

71s

From Table 2, it can be seen that the accuracy of the transfer learning algorithm is 16% in the first epoch, and the accuracy keeps increasing with an increase in the

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 251

number of epochs. By the end of the thirtieth epoch, the accuracy of the transfer learning algorithm is increased to 98%, which is higher than that of the CNN algorithm. It should also be noted that the accuracy of the transfer learning model even reached the 100% mark eight times during the training. It can also be seen that the loss of the algorithm at the first epoch is almost equal to 97%, and it keeps on decreasing with the increase in epochs. In the end, the loss of the transfer algorithm is 4%. This loss value is so small that it can even be neglected. But one of the disadvantages of the transfer learning model is the time taken to complete every epoch. But when the accuracy of the model is almost ideal, it can be concluded that the time taken to complete the process is one of the least concerns. The data depicted in the above table is pictorially represented as a graph, as shown in Fig. (8).

Fig. (8). Accuracy and loss graph for Transfer Learning algorithm.

The values obtained from both algorithms are compared for accuracy and loss, and Table 3 compares the overall accuracy and loss value of both algorithms.

252 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

Table 3. Comparison of two algorithms. DL Model

Test Accuracy

Test Loss

VGG19

98.75%

2.16%

CNN

95.42%

6.25%

From the above table, it can be concluded that the transfer learning model implemented using the VGG-19 model is more efficient in terms of both accuracy and loss value. 7. IDENTIFICATION To develop a DL model capable of identifying the presence of plant diseases, two distinct DL algorithms are applied by researchers. DL techniques include CNNs and transfer learning methods. The study is carried out to identify the best model. When the accuracy and loss of the model are investigated, it is determined that the transfer learning model is the most efficient algorithm for identifying the existence of paddy plant disease. CONCLUSION A dataset with 120 images of different types of plant diseases of paddy and images of healthy stalks is collected from Kaggle. The images are then preprocessed using image augmentation. This image augmentation technique includes processes like image rescaling, image shearing, image zooming, and horizontal flipping. A DL model which is capable of identifying the presence of plant diseases is designed using two different DL algorithms. These DL algorithms include CNNs and transfer learning algorithms. The models developed are then trained again and again until it reaches the maximum accuracy value and the minimum accuracy. When the accuracy of the model and the loss are analyzed, it is found that the transfer learning model is the efficient algorithm when employed to detect the presence of plant disease from paddy. When this DL model is used to detect the presence of crop disease, it would help the farmers to find the real pesticide or chemicals to get rid of the disease and to increase the yield of crops.

Crop Disease

AI and IoT-based Intelligent Health Care & Sanitation 253

CONSENT FOR PUBLICATON Declared None. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared None. REFERENCES [1]

B.S Prajwalgowda, M A Nisarga, M Rachana, S Shashank, and B.S. Sahana Raj, “Paddy Crop Disease Detection using Machine Learning”, Int. J. Eng. Res. Technol. (IJERT), NCCDS, vol. 8, no. 13, 2020.

[2]

N.S. Patil, "Identification of Paddy Leaf Diseases using Evolutionary and Machine Learning Methods", Turk. J. Comp. Math.Edu (TURCOMAT), vol. 12, no. 2, pp. 1672-1686, 2021. [http://dx.doi.org/10.17762/turcomat.v12i2.1503]

[3]

P.K. Sethy, N.K. Barpanda, A.K. Rath, and S.K. Behera, "Image Processing Techniques for Diagnosing Rice Plant Disease: A Survey", Procedia Comput. Sci., vol. 167, pp. 516-530, 2020. [http://dx.doi.org/10.1016/j.procs.2020.03.308]

[4]

S. Ramesh, and D. Vydeki, "Recognition and classification of paddy leaf diseases using Optimized Deep Neural network with Jaya algorithm", Inf. Process. Agric., vol. 7, no. 2, pp. 249-260, 2020. [http://dx.doi.org/10.1016/j.inpa.2019.09.002]

[5]

G. Dhingra, V. Kumar, and H.D. Joshi, "A novel computer vision based neutrosophic approach for leaf disease identification and classification", Measurement, vol. 135, pp. 782-794, 2019. [http://dx.doi.org/10.1016/j.measurement.2018.12.027]

[6]

S.P. Yadav, "Vision-based Detection, Tracking, and Classification of Vehicles", IEIE Transactions on Smart Processing & Computing, vol. 9, no. 6, pp. 427-434, 2020. [http://dx.doi.org/10.5573/IEIESPC.2020.9.6.427]

[7]

Yadav, Satya & Mahato, Dharmendra Prasad & Linh, Nguyen. “Distributed Artificial Intelligence: A Modern Approach”. Taylor & Francis (2020). [http://dx.doi.org/10.1201/9781003038467]

[8]

R. Chauhan, K.K. Ghanshala, and R.C. Joshi, "CNN (CNN) for Image Detection and Recognition", In: 2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC), 2018, pp. 278-282. [http://dx.doi.org/10.1109/ICSCCC.2018.8703316]

[9]

S. Albawi, T.A. Mohammed, and S. Al-Zawi, "Understanding of a CNN", In: 2017 International Conference on Engineering and Technology (ICET), 2017, pp. 1-6. [http://dx.doi.org/10.1109/ICEngTechnol.2017.8308186]

[10]

U.R. Acharya, S.L. Oh, Y. Hagiwara, J.H. Tan, M. Adam, A. Gertych, and S.T. Ru, "A deep CNN model to classify heartbeats", Comput. Biol. Med., vol. 89, 2017. [http://dx.doi.org/10.1016/j.compbiomed.2017.08.022] [PMID: 28869899]

[11]

H. Shin, "Deep CNNs for Computer-Aided Detection: CNN Architectures, Dataset Characteristics, and Transfer Learning", IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1285-1298, 2016. [http://dx.doi.org/10.1109/TMI.2016.2528162] [PMID: 26886976]

[12]

S.J. Pan, and Q. Yang, "A Survey on Transfer Learning", IEEE Trans. Knowl. Data Eng., vol. 22, no.

254 AI and IoT-based Intelligent Health Care & Sanitation

Shukla et al.

10, pp. 1345-1359, 2010. [http://dx.doi.org/10.1109/TKDE.2009.191] [13]

Y. Zhu, Y. Chen, Z. Lu, S.J. Pan, G-R. Xue, Y. Yu, and Q. Yang, "Heterogeneous Transfer Learning for Image Classification", In: Twenty-Fifth AAAI Conference on Artificial Intelligence, 2011.

[14]

Ling Shao, Fan Zhu, and Xuelong Li, "Transfer learning for visual categorization: a survey", IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 5, pp. 1019-1034, 2015. [http://dx.doi.org/10.1109/TNNLS.2014.2330900] [PMID: 25014970]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 255-268

255

CHAPTER 17

Deploying Deep Learning Model on the Google Cloud Platform For Disease Prediction C.R. Aditya1,*, Chandra Sekhar Kolli2, Korla Swaroopa3, S. Hemavathi4 and Santosh Karajgi5 Department of Computer Science & Engineering, Vidyavardhaka College of Engineering, Mysuru, Karnataka, India-570002 2 Aditya College of Engineering & Technology, Surampalem, East Godavari, India-570002 3 CSE, Aditya Engineering College, Kakinada, Andhra Pradesh, India-533437 4 CSIR - Central Electrochemical Research Institute (CECRI), CSIR-Madras Complex, Chennai, Tamil Nadu, India-600113 5 BLDEA's SSM College of Pharmacy and Research Centre Vijayapur 586103, Karnataka, India 1

Abstract: A brain tumor is defined by the proliferation of aberrant brain cells, some of which may progress to malignancy. A brain tumor is usually diagnosed via a magnetic resonance imaging (MRI) examination. These images demonstrate the recently observed aberrant brain tissue proliferation. Several academics have examined the use of machine learning and Deep Learning (DL) algorithms to diagnose brain tumors accurately A radiologist may also profit from these forecasts, which allow them to make more timely decisions. The VGG-16 pre-trained model is employed to detect the brain tumor in this study. Using the outcomes of training and validation, the model is completed by employing two critical metrics: accuracy and loss. Normal people confront numerous challenges in scheduling a doctor's appointment (financial support, work pressure, lack of time). There are various possibilities for bringing doctors to patients' homes, including teleconferencing and other technologies. This research creates a website that allows people to upload a medical image and have the website predict the ailment. The Google Cloud Platform (GCP) will be utilized to install the DL model due to its flexibility and compatibility. The customized brain tumor detection website is then constructed utilizing HTML code.

Keywords: Accuracy, Tumor detection, VGG-16, Loss, Kaggle, Google Cloud Search.

Corresponding author C.R. Aditya: Department of Computer Science & Engineering, Vidyavardhaka College of Engineering, Mysuru, Karnataka, India-570002; E-mail: [email protected] *

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

256 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

1. INTRODUCTION Many healthcare problems can be remedied quickly as a result of the rapid evolution of healthcare technology in the field of information and communication engineering. According to a study conducted by the National Cancer Institute Statistics and the World Health Organization, brain tumors kill 12,764 people in the United States each year. As it manages all other organs and aids in decisionmaking, the brain is the most important organ in the human body. It is the central nervous system's primary control center and is in charge of the body's everyday voluntary and involuntary functions. Our brains are undergoing uncontrollable tissue growth, which is replicating uncontrollably in the tumor. Brain cancer develops when abnormal cells proliferate within or near the brain's surface. A risk to the healthy brain occurs as a result of either piercing or squeezing and moving normal brain tissue as a result of the procedure. A developing brain tumor can potentially cause considerable injury to the brain due to the limited amount of space inside the skull. It is critical to understand brain tumors and their stages to properly prevent and treat the problem. Several imaging modalities, such as MRI and computed tomography (CT), can be used to assess brain tumors [1]. Brain tumors are one of the most severe disorders that can occur in both children and adults. Brain tumors are thought to account for 85 to 90 percent of all primary Central Nervous System malignancies. Appropriate therapy and planning, as well as accurate diagnoses and treatment recommendations, can all lead to better patient outcomes [2]. In 1969, Raymond V. Damadian invented the magnetic picture, which was the first of its kind. In 1977, the first MRI of the human body, as well as the most accurate technology, was developed. The inside anatomy of the brain, as well as the many other types of tissues in the human body, may be examined in fine detail using MRI. MRI images outperform X-ray and computer tomography images in terms of quality [3]. A brain tumor can be detected via MRI. Automatic data categorization using Artificial Intelligence (AI) is more accurate than manual data categorization. As a result, a system that employs DL Algorithms to detect and classify diseases would be extremely useful to clinicians all around the world. 2. LITERATURE SURVEY 1. The One-Pass Multi-Task Network (OM-Net) was created to address the issue of unequal data in medical brain volume. OM-Net learns discriminative features by using both task-specific and shared parameters for training. To improve performance in OM optimization, methods such as learning-based training and online training data transfer are used. Furthermore, the prediction results are exchanged between jobs via a CGA module [4].

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 257

2. Including the four MRI modalities in the architecture of brain tumor segmentation. As each modality has its own set of differentiating properties, the network will be able to discern between classes more efficiently. Working on a section of the brain image near the tumor tissue, it was proved that a CNN model can perform at least as well as human observers. Furthermore, a CNN model in cascade mode is used to extract both local and global features in two distinct ways, each with a different patch size [5]. 3. The study looked into using a single model to categorize brain MRI data on many classification issues, rather than using a new model for each classification assignment, as prior research had done. This can be accomplished with the assistance of a Convolutional Neural Network (CNN), a type of artificial intelligence. A CNN-based model can also be used to pinpoint the site of a brain tumor [6]. 4. CNN's and transfer learning are used in conjunction to examine MRI pictures of the human brain. Both methods have shown effective outcomes when compared to CNN; however, transfer learning requires a bigger investment of computational effort and memory [7]. 5. The researchers intend to build a computer-assisted brain tumor diagnosis system utilizing the CNN network in this study. The deep network model represents a network of links mathematically. CNN was trained using MRI datasets of brain tumors, and its performance was tested on 72 percent of previously unseen data. In the training phase, only 28% of the data was utilized. The algorithm accurately predicted the existence of a brain tumor using two datasets and achieved the greatest accuracy [8]. 6. A thorough hierarchical classification of brain tumors is established. Techniques for segmenting a brain tumor include thresholding, unsupervised and supervised machine learning, and DL. For diagnosing brain cancers, a comparison of machine learning techniques with DL-based algorithms is made [9]. 7. To identify and characterize brain tumors in seven different locations, a DL architecture was used. Scientists employed DL to classify gliomas, meningiomas, and pituitary cancers, among other things. The comparison of cropped and uncropped regions in a categorized brain image resulted in improved segmentation accuracy [10]. 3. METHODOLOGIES Fig. (1) depicts the step-by-step procedure for installing the brain tumor detection model in GCP. The data is initially obtained from a well-known website called

258 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

Kaggle. The raw data is then pre-processed to prepare it for the DL model. Image resizing, rescale, and augmentation are phases in the pre-processing process. Following data preparation, the DL model is built by adjusting the number of neurons in the output layer. Using the training and validation data, the model is then trained and validated. 80% of the data is used for training. The trained model is then evaluated using measures such as accuracy and loss. The website is then scheduled to be built. This necessitates the storage of the model in the cloud. The GCP is picked to deploy the finalized model.

Fig. (1). Flow diagram of proposed work.

3.1. Brain Tumor Data Brain tumors are classified into more than 120 different categories based on their location and the sort of cells that build them up. Each type has its own set of features. Certain tumors are normally noncancerous, whereas others, depending on the type of tumor, are usually cancerous (cancerous). The study's purpose is to identify four different types of cancer. The Kaggle website is being used to collect data about various cancers. Table 1 shows the number of samples gathered. There are samples for both training and testing in the table. It's described thoroughly here: The tumor that was used in the research was acquired from a patient. 1. Glioma is a common brain tumor type, and it can develop in the spinal cord. These malignancies are caused by glial cells, which are cells that cover and support neurons. 2. Meningioma is the most common type of brain tumor worldwide, accounting for more than one-third of all brain tumors. Meningiomas are normally

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 259

determined by three coats of meninges that cover and protect the brain underneath the base of the skull. Malignant tumors are far more common in women than in men. Non-cancerous, quiet tumors account for 85% of all meningiomas. Meningiomas are non-cancerous, however, relapses can occur after treatment has indeed been accomplished if not handled properly. 3. The most frequent form of pituitary gland tumor is an adenoma. This is a tumor that develops within the gland's tissues. Pituitary adenomas develop as the pituitary gland expands. They have a modest growth rate. Adenomas are responsible for about 10% of all brain tumors. They have the potential to induce eyesight and endocrinological issues. Adenomas are not malignant tumors that can be eliminated with surgery or treatment. Table 1. Tumor data. Tumor Type

Train

Test

Glioma Tumor

826

100

Meningioma tumor

822

115

Pituitary Tumor

827

105

No Tumor

395

74

3.2. Image Resizing Image resizing is an important preprocessing step in computer vision and is widely employed. Our machine learning algorithms can learn faster when given smaller training photos. Furthermore, the sizes of the photographs obtained may differ, complicating problems further because many DL model architectures demand images of uniform size. Before being fed into the VGG-16 neural network, all photos must be scaled to a certain size. The VGG-16 neural network only accepts 224*224 pixel inputs. It is easier to reduce the fixed size if it is larger than the variable size. As a result of the lessened shrinkage, there is less distortion of the image's features and patterns. 3.3. Image Rescaling All of the pixel values in the submitted image are within the permitted range of 0 to 255. Black and white are represented by the integers 0 and 255, respectively. There are red, green, and blue maps supplied to create a vivid image while maintaining each pixel inside the 0-255 color range. We'll choose 255 as an example because it's the highest possible pixel value. The rescale 1./255 function is used to shrink the range [0,255] to [0,1]. Some of the benefits are as follows:

260 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

1. The pixel ranges in high- and low-definition photographs are different. For all photos, the weights and learning rates are the same as before. As a general rule, high-range images cause more loss than low-range images; yet, when performing a backpropagation update, the total of the two images will be used. For visual comprehension, a contour is more significant than contrast, as long as the contour is kept to a bare minimum. If all of the images are scaled to the same range [0,1], they will contribute to the overall loss more equitably. 2. The ability to directly compare learning rates between two systems is made possible by the fact that they both perform scaling preprocessing on picture data sets. Ideally, a lower learning rate for high-pixel-range images should be used, and a higher learning rate for low-pixel-range images should be used to prevent these concerns. 3.4. Image Data Generator When given more data to work with, DL neural networks perform better. Data augmentation techniques are used to create additional training data from previously gathered training data [11]. Domain-specific approaches are used to develop training data examples, resulting in fresh and unique training data examples. Picture data augmentation is a common sort of data augmentation in machine learning. It entails altering photos in the training dataset so that the original image can be recreated. There are numerous approaches for data augmentation and data synthesis that are commonly used. Augmentation techniques such as zoom, shear, and flip are used in this piece. This technique can be used to focus on a specific area of an image. This method allows you to arbitrarily enlarge an image by zooming in or adding pixels around the image. The use of input data accounts for eighty percent of the zooming. These two forms of augmentations are known as vertical and horizontal flip augmentations. Shear refers to the technique of warping a picture along an axis to shift the angle at which the observer perceives it. Image augmentation is frequently used to assist computers in seeing objects from a variety of perspectives. The Shear Transformation is used to slant the shape of an item. The letters X-Shear and YShear represent shear transformations. On both sides of the equation, the same person changes the X and Y coordinates. According to the findings of this investigation, the shear has increased by 20%. Fig. (2) shows how data augmentation can help DL improve its output by explaining how it works.

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 261

Fig. (2). Data Augmentation importance.

3.5. Construct VGG-16 Transfer learning is a knowledge-sharing approach that can be used to reduce the amount of training data, the amount of time spent on it, and the amount of money spent on computing resources. Transfer learning is a technique that allows one to learn from a previously trained model and then apply that knowledge to another model. Transfer learning is beneficial in a wide variety of situations and applications. A convolutional neural network (CNN) architecture known as VGG16 [12] was used by VGG16 to compete in the 2014 ImageNet competition, and the architecture of the model is shown in Fig. (3). It has been acclaimed as one of the best vision model architectures ever designed and has received numerous accolades. Instead of a large number of hyper-parameters, the convolution layers of a 3x3 filter with a stride 1 and the padding and max pool layer of a 2x2 filter with a stride 2 were used to create the most distinctive feature of VGG16. The convolution and max pool layers are arranged in the same way throughout the architecture, which is consistent. Fully-Connected (FC) layers are used after the convolutional layer stacks. The first two of the FC layers each have 4096 channels, while the third is used for four tumor classifications, and so has four channels (one for each class). This network contains an estimated 138 million parameters, according to estimates. The presence of ReLU nonlinearity may be found in all of the buried layers.

262 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

Fig. (3). VGG-16 model.

3.6. Train VGG-16 The training data is used to train the model. 8 epochs are used for training. To reduce error, the weight and bias parameters are modified during the training phase. The model's accuracy is 0.69 in the first epoch, and 0.85 in the second epoch, and the accuracy level has moved from a lower to a higher number. The accuracy obtained from 3rd epochs is 0.87, 0.87, 0.91, 0.93, 0.93, and 0.94. The loss values are then computed. As the epoch rises, the loss value drops in the following order: 0.82, 0.39, 0.33, 0.31, 0.23, 0.17, 0.17, and 0.15. Fig. (4) depicts the training accuracy and loss. The green plot reflects accuracy, while the orange plot represents a loss.

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 263

Fig. (4). Performance evaluation of VGG-16 on training data.

3.7. Validate VGG-16 Validation data is used to validate the model. The procedures of training and validation are carried out concurrently. Validation and training are both performed during an eight-epoch period. The model's accuracy is 0.75 in the first epoch and 0.69 in the second epoch, suggesting that the accuracy level has risen. The accuracy from the third epochs was 0.74, 0.78, 0.79, 0.82, and 0.81. This is followed by the loss computation. The loss value drops with each epoch in the following range: 0.66, 0.65, 0.56, 0.50, 0.45, 0.33, 0.29, and 0.20. Fig. (5) depicts the accuracy and loss of validation. The color orange represents accuracy, whereas the color blue represents a loss. 3.8. Deploy in GCP Large-scale apps can be built with the assistance of GCP, a Google service. The article [13] discusses some of the advantages of using Google's App Engine to develop cloud-based applications, such as how to save time and money. Because graphics processing units (GPUs) are so widely available, this platform's popularity has skyrocketed. GCP, on the other side, is extremely user-friendly and demands very little from you. GCP also just announced the availability of a BETA version of their Continuous Evaluation Tool, which allows you to test and track the progress of your model immediately after deployment. DL flow can be

264 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

used to track quality indicators as the model is developed before putting it into production. The fig. (6) depicts the model's deployment in Google's cloud. When calculating the total number of requests, latencies and CPU use are combined with memory utilization and other parameters.

Fig. (5). Performance evaluation of VGG-16 on validation data.

Fig. (6). Deploy model in GCP.

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 265

4. Results and Discussion After assessing performance parameters such as accuracy and loss, the model is constructed and finalized. The model is running on GCP. Based on the submitted photograph, the website is designed to predict the tumor. Fig. (7) depicts the creation of the home page. The website is built so that everyone may readily access it. There are only two buttons on the webpage. One for uploading the brain scan and the other for determining whether the brain is affected by a tumor or not. This website is intended to distinguish between the three types of brain tumors: glioma, meningioma, and pituitary. The online tumor detection process takes only three stages.

Fig. (7). Step 1: Home page of the website.

1. The first step is to log in to the website. 2. The second step is to upload the brain scan. The webpage after uploading the image is shown in Fig. (8). If anything is uploaded wrongly it is also possible to remove the image and re-upload the correct one.

266 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

Fig. (8). Step 2: Upload MRI image.

3. The third step is to forecast brain health. Fig. (9) depicts the webpage after hitting the categorized button. To test the website's functionality, a sample image from the dataset is submitted and predicted on the web. The scanned image of a man with meningioma was submitted, and the web correctly predicted it. The probability score is also available on the internet. This allows the patient to get a clear picture. For example, if the website indicates that the tumor type has a probability of less than 60%, the individual can seek a second opinion.

Fig. (9). Step the type of tumor.

5. REAL-TIME IMPLEMENTATIONS The use of DL techniques in this study to detect brain tumors from MRI scan images and store the information in GCP makes this model unique and more

Disease Prediction

AI and IoT-based Intelligent Health Care & Sanitation 267

reliable than prior methods. The expense of storing the finalized DL model is eliminated by this GCP. And the web development to upload the scan and predict the disease makes it easier for people who are temporarily unable to attend the doctor. CONCLUSION Still, brain tumors are difficult to detect since their appearance, size, form, and structure vary from person to person. Although image processing may detect cancers in MRI, several advancements are still needed to correctly classify tumorrelated data. DL algorithms have made an important contribution to the solution of this challenge. In this study, the pre-trained model is used to solve the specified problem. This study collects about 3,000 picture samples. The sample is split into two parts: training data and validation data. The model is trained and validated concurrently, and it is evaluated using accuracy and loss. If the loss is within the tolerable range, the model is finished. The final DL model has been deployed in GCP and will aid in the development of the website. The created website helps those who are unable to visit a doctor and also increases the patient's survival rate. CONSENT FOR PUBLICATON Not applicable. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared none. REFERENCES [1]

S. Prakash Yadav, and S. Yadav, "Fusion of medical images in wavelet domain: a hybrid implementation", Comput. Model. Eng. Sci., vol. 122, no. 1, pp. 303-321, 2020. [http://dx.doi.org/10.32604/cmes.2020.08459]

[2]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images", Med Biol Eng Comput, vol. 58, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6]

[3]

R.A. Novellines, M. D. “Squire’s fundamentals of radiology”, Six Edition. UPR, 2004.

[4]

C. Zhou, C. Ding, X. Wang, Z. Lu, and D. Tao, "One-pass multi-task networks with cross-task guided attention for brain tumor segmentation", IEEE Trans. Image Process., vol. 29, pp. 4516-4529, 2020. [http://dx.doi.org/10.1109/TIP.2020.2973510] [PMID: 32086210]

[5]

R. Ranjbarzadeh, A. Bagherian Kasgari, S. Jafarzadeh Ghoushchi, S. Anari, M. Naseri, and M. Bendechache, "Brain tumor segmentation based on deep learning and an attention mechanism using

268 AI and IoT-based Intelligent Health Care & Sanitation

Aditya et al.

MRI multi-modalities brain images", Sci. Rep., vol. 11, no. 1, p. 10930, 2021. [http://dx.doi.org/10.1038/s41598-021-90428-8] [PMID: 34035406] [6]

B. Kokila, M.S. Devadharshini, A. Anitha, and S. Abisheak Sankar, "Brain Tumor Detection and Classification Using Deep Learning Techniques based on MRI Images", J. Phys. Conf. Ser., vol. 1916, no. 1, p. 012226, 2021. [http://dx.doi.org/10.1088/1742-6596/1916/1/012226]

[7]

G. Raut, A. Raut, J. Bhagade, J. Bhagade, and S. Gavhane, “DL Approach for Brain Tumor Detection and Segmentation”. 2020 International Conference on Convergence to Digital World - Quo Vadis (ICCDW), pp. 1-5, 2020. [http://dx.doi.org/10.1109/ICCDW45521.2020.9318681]

[8]

A. Naseer, T. Yasir, A. Azhar, T. Shakeel, and K. Zafar, "Computer-Aided Brain Tumor Diagnosis: Performance Evaluation of Deep Learner CNN Using Augmented Brain MRI", Int. J. Biomed. Imaging, vol. 2021, pp. 1-11, 2021. [http://dx.doi.org/10.1155/2021/5513500] [PMID: 34234822]

[9]

A. Tiwari, S. Srivastava, and M. Pant, "Brain tumor segmentation and classification from magnetic resonance images: Review of selected methods from 2014 to 2019", Pattern Recognit. Lett., vol. 131, pp. 244-260, 2020. [http://dx.doi.org/10.1016/j.patrec.2019.11.020]

[10]

K. Simonyan, and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition”. The 3rd International Conference on Learning Representations, 2015. [http://dx.doi.org/10.48550/arXiv.1409.1556]

[11]

A. Mikołajczyk, and M. Grochowski, Data augmentation for improving DL in image classification problem. In 2018 international interdisciplinary PhD workshop (IIPhDW). IEEE, 2018, pp. 117-122. [http://dx.doi.org/10.1109/IIPHDW.2018.8388338]

[12]

A.M. Alqudah, H. Alquraan, I. Abu Qasmieh, A. Alqudah, and W. AlSharu, "Brain tumor classification using DL technique—a comparison between cropped, uncropped, and segmented lesion images with different sizes", Int. J. Adv. Trends. Comp. Sci. Eng., vol. 8, no. 6, pp. 3684-3691, 2019. [http://dx.doi.org/10.30534/ijatcse/2019/155862019]

[13]

Isak Shabani, and Agni Dika, "The Benefits of Using Google Cloud Computing for Developing Distributed Applications", J. math. syst. sci., vol. 5, no. 4, 2015. [http://dx.doi.org/10.17265/2159-5291/2015.04.004]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 269-284

269

CHAPTER 18

Classification and Diagnosis of Alzheimer’s Disease using Magnetic Resonance Imaging K.R. Shobha1,*, Vaishali Gajendra Shende2, Anuradha Patil2, Jagadeesh Kumar Ega3 and Kaushalendra Kumar4 Department of Electronics and Telecommunication Engineering, Ramaiah Institute of Technology, MSR Nagar, MSRIT Post, Bangalore, Karnataka, India-560054 2 Department of ECE, Faculty of Engineering and Technology Exclusive for Women Sharn Basva University Kalburagi, Anuradha Patil, ECE Department, SB Campus, Vidya Nagar, Sharn Basva University Kalburagi, Karnataka, India-585103 3 Chaitanya (Deemed to be University), Hanamkonda, Telangana State, India 4 Department of Bio-Science, Galgotias University, Greater Noida. Uttar Pradesh, India-201310 1

Abstract: Different types of brain illnesses can affect many parts of the brain at the same time. Alzheimer's disease is a chronic illness characterized by brain cell deterioration, which results in memory loss. Amnesia and ambiguity are two of the most prevalent Alzheimer's disease symptoms, and both are caused by issues with cognitive reasoning. This paper proposes several feature extractions as well as Machine Learning (ML) algorithms for disease detection. The goal of this study is to detect Alzheimer's disease using magnetic resonance imaging (MRI) of the brain. The Alzheimer's disease dataset was obtained from the Kaggle website. Following that, the unprocessed MRI picture is subjected to several pre-processing procedures. Feature extraction is one of the most crucial stages in extracting important attributes from processed images. In this study, wavelet and texture-based methods are used to extract characteristics. Gray Level Co-occurrence Matrix (GLCM) is utilized for the texture approach, and HAAR is used for the wavelet method. The extracted data from both procedures are then fed into ML algorithms. The Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) are used in this investigation. The values of the confusion matrix are utilized to identify the best technique.

Keywords: Alzheimer, Confusion Matrix Values, Feature Extraction, HAAR, Magnetic Resonance Imaging. Corresponding author K.R. Shobha: Department of Electronics and Telecommunication Engineering, Ramaiah Institute of Technology, MSR Nagar, MSRIT Post, Bangalore, Karnataka, India-560054; E-mail: [email protected] *

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque (Eds.) All rights reserved-© 2023 Bentham Science Publishers

270 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

1. INTRODUCTION Alzheimer's disease is a neurological illness that causes dementia in the elderly. It gradually decreases the person's cognitive function, resulting in declining memory and psychotic symptoms, and eventually death. Alzheimer's disease affects roughly 24 million people worldwide, and the number is anticipated to climb every 20 years until at least 2040. Alzheimer's disease affects women by a factor of 1.2 to 1.5% higher than men [1]. This disease affects roughly 10% - 15% of people over the age of 65, and 47% of people over the age of 80. People with this disease, according to Alois Alzheimer, have a damaged brain with neurofibrillary tangles and neuritis senile plaques [2]. Alzheimer's disease is the sixth greatest cause of death and is a devastating neurological disorder. Though signs of this disease can be found in the brain as early as middle life, symptoms do not develop until after the age of 65. Although the primary cause of this ailment is assumed to be old age, it is not the only one. In addition to age, risk factors include one or more apolipoprotein gene E4 alleles, a family history of Alzheimer's, low educational and vocational attainment, and moderate to severe concussions. The most important aspect of detecting Alzheimer's disease in humans is still assessing cognitive deficits, behavioral problems, functional impairment, and a variety of psychosocial therapies [3]. According to current research, reasonably benign Alzheimer's disease is rarely detected, and moderately severe Alzheimer's disease is under-recognized by diagnostic practice, even when the most thorough examination is undertaken. Numerous studies have suggested that neuroimaging, namely high-resolution MRI, could be utilized to identify or diagnose Alzheimer's disease. Structural MRI imaging provides additional information on diseased tissue degeneration or other abnormal biomarkers that can be detected early in the disease with great sensitivity. This implies that a computer-aided diagnostic tool for detecting Alzheimer's disease before lasting neuronal damage occurs is extremely desirable [4]. Although early detection of Alzheimer's disease is critical, some researchers have used ML approaches to assess data and diagnose the disease. According to the researcher [5] ML is the process that occurs between 'supervised' and 'unsupervised' learning. A supervised learning strategy was used, with a collection of MRI images and the accompanying prognosis training set fed into an ML model, in this case, an SVM. The SVM “learned” to identify between those with and without Alzheimer's disease [6]. According to the article [7], longitudinal processing of sequential MRIs is required to create and study the course of the disorder over time for a more definitive diagnosis. To differentiate between the three groups, the actual procedure uses anatomical abnormalities of the brain, longitudinal variations in MRI, and a built-in classifier.

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 271

The author proposed a multilayer classification strategy using ML algorithms such as NB, SVM Algorithm, and K-NN to differentiate between various themes [8]. The MM-SDPN method, which is based on MRI and PET imaging data, was offered as a new algorithm for diagnosing Alzheimer's disease in the study [9]. The journal is structured as follows: Section 2 discusses the materials and processes for identifying Alzheimer's disease. Section 3 discusses the preprocessing procedures employed in this investigation. Section 4 will go through how to extract features with GLCM and HAAR. Section 5 discusses classification algorithms such as SVM and LDA. Section 6 investigates the combined performance of numerous feature extraction and classification approaches. Finally, section 7 suggests integrating feature extraction with several classifier methods to diagnose Alzheimer's disease. 2. METHODS AND MATERIALS The study uses brain MRI images from the Kaggle database, which contains over 6000 images. Fig. (1) given below, is a sample MRI image of the brain.

Fig. (1). a) Normal MRI b) Alzheimer MRI.

To train and test these MRI images, ML algorithms are used, with 80% of the image data used for training and 20% of the image data used for testing. In this scenario, the whole image data collection is divided into 8:2 ratios for training and testing operations (i.e., 4800:1200). In this study, preprocessing techniques such as picture scaling and smoothing using the Gaussian method are employed to further examine and improve these images. Fig. (2) depicts a flow diagram of the

272 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

texture and wavelet-based feature extraction, as well as the types of classifiers used in this study.

Fig. (2). Flow diagram of Alzheimer’s disease detection model.

3. PREPROCESSING TECHNIQUES Medical images, in general, are massive data stacks that are difficult to store, analyze, and calculate. As a result, picture preprocessing techniques are used to reduce the data's complexity so that the high-dimensional data can be reduced to an optimal dimensional resolution [10]. The obtained picture data is preprocessed in this study utilizing image scaling and smoothing using the Gaussian approach.

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 273

3.1. Image Resizing Image resizing is an important feature in medical image analysis. The picture magnification or reduction is performed by a 2-dimensional interpolation function that resamples a collection of discrete signal samples using a specified methodology. The best image scaling approach should preserve the image's quality characteristics. This means that the final image should be free of flaws such as blurring, edge discontinuities, or checkerboard effects. Furthermore, when utilized in real-time imaging applications, the techniques must meet specific quantitative constraints [11]. 3.2. Smoothing using the Gaussian Method The smoothing of the image data is done using the Gaussian filter. The standard deviation (i.e., the spread parameter) of the Gaussian function determines the degree of image smoothing [12]. The Gaussian filter is commonly used to smooth out the sample image and remove noise. Convolution is used to achieve Gaussian smoothing in most cases [13]. The Gaussian operator in 2 dimensions is given in the below equation. (1) 4. FEATURE EXTRACTION Feature extraction is a crucial step in image processing since it reduces the size of the raw picture dataset to make it more manageable. The original raw image dataset is huge, and processing these photos will require a significant amount of computing power. The method of turning variables in a raw image database into features while reducing the amount of data and precisely representing the original data is known as feature extraction [14]. To extract critical data from brain MRI images, texture-based and wavelet-based extraction algorithms are used. This study employs the HAAR and GLCM techniques, which are detailed in the section below. 4.1. GLCM The GLCM- a texture-based feature extraction technique that determines the cooccurrence matrix of each image data in the obtained set of data by estimating how frequently a pixel of a given intensity I would seem concerning another pixel 'j' at a given distance 'd' and orientation θ [15]. GLCM describes the 2nd order of

274 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

statistical information regarding the gray-level between neighboring pixels in a picture. In general, GLCM extracts 11 different texture features, a few of which are explained below through equations: 1. Homogeneity: (2) 2. Inertia: (3) 3. Energy: (4) 4. Entropy: (5) 5. Shade: (6) 6. Prominence: (7) 7. Variance: (8) 8. Correlation:

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 275

(9)

4.2. HAAR HAAR is a wavelet-based feature extraction method that generates an accurate description, captures edges, records multi-resolution data, and is quick to process. As the HAAR basis generates an orthogonal basis, the transform provides a nonredundant representation of an input image. As a result, the HAAR wavelet should be used to extract features [16]. The HAAR wavelet, which has no overlapping regions, only reflects the difference between two neighboring pixels. As it only uses two functions, scale and wavelet [17], it also computes the average and variance of a pair. 5. CLASSIFICATION TECHNIQUES For classifying the MRI into Alzheimer's and Normal two different supervised ML algorithms are used, such as SVM and LDA. The mathematics behind both algorithms is detailed in this section. 5.1. SVM The SVM is a well-known and successful ML method for dealing with large-scale data categorization problems. SVM is widely used in neuroimaging investigations due to its ease of use and effectiveness in solving a wide range of classification problems [18]. An SVM decision function is a more precise formulation of a “hyper-plane” that divides (i.e., “classifies”) data into many classes based on particular patterns concerning those observations known as features. An SVM decision function can be trained by finding a repeatable hyperplane that optimizes the margin between the support vectors in the two classes to which it is applied. In rare cases, when using a kernel approach, the support vectors may need to be translated into a higher-dimensional input space. In other words, this extra step transforms a set of features that are not linearly distinct into a set of linearly distinct traits. SVM has the best generalization capacity, and it not only produces good classification results but also gives a lot of room to update data categories. The final classifier equation is as follows: (10)

276 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

5.2. LDA Regarding image data categorization, a technique known as LDA is extensively used. LDA is a simple strategy to consider when the within-class frequencies are out of whack and their performance has been evaluated using randomly generated sample data. Unlike other classifiers, LDA does not change the position of the starting data set when it is transferred to different spaces; instead, it improves class separation and produces a decision area between specified classes. This categorization technique also aids in a better understanding of feature data distribution. LDA allows you to modify and categorize your dataset using one of two methods: class-dependent transformation or class-independent transformation [19]. The class classifier can be learned through decision analysis. Assume as the Y class density in the H=k class [20]. The following equation given below represents the application of the Bayes theorem:

(11)

Here, Mean If a multinomial Gaussian model is used to represent each assumed density class, then:

(12)

Here, Mean Covariance When the covariance is unique across all classes, it becomes necessary to use LDA to solve the problem. As a result, the discriminant function is considered to

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 277

be (13) Parameter estimation,

(14)

(15)

(16)

LDA generates predictions by calculating the likelihood that a given set of inputs will fall into each of the classes. The resulting class has the best chance of succeeding, and a prognosis is generated. 6. RESULT COMPARISON The Kaggle database was used to obtain 6000 brain MRI images for this study. In an 8:2 ratio (i.e., 80% and 20%), the data is divided into two parts: one for training and one for testing the classifier model. Performance measures are used to evaluate the feature extraction technique and the resulting classification model. Accuracy, Specificity, Sensitivity, Precision, False Negative and False Positive Rates, and Error Rates were the performance metrics employed in this investigation. The components of the confusion matrix are utilized to determine the performance measures. The Confusion matrix has four elements: TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative). These four characteristics are crucial in computing performance measures. The mathematical equation for calculating the metrics is provided below.

278 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

7 Accuracy: Measures total correct predictions from the total sample of test image data.

(17)

1. Specificity: Measures total correct normal predictions from the total sample of actual normal image data. (18) 2. Sensitivity: Measures total correct Alzheimer predictions from the total sample of actual Alzheimer image data. (19) 3. FNR: Measures total wrong normal predictions from the total sample of actual Alzheimer image data. (20) 4. FPR: Measures total wrong Alzheimer predictions from the total sample of actual normal image data. (21) 5. Precision: Measures total correct Alzheimer predictions from the total sample of predicted Alzheimer image data.

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 279

(22)

6. F-Measure: Measure harmonic mean of recall and precision (23) 7. Error Rate: Measures total wrong predictions from the total sample of test image data. (24) 8. Balance Accuracy: Measures the average of specificity and sensitivity. The average depends on the total unique classes presented in the output. (25) These performance criteria help to determine the most effective feature extraction and classification model for Alzheimer's disease detection. Table 1 shows the output of classifiers such as SVM and LDA when the HAAR feature extraction technique is utilized. Table 1 demonstrates that the level of accuracy for SVM is exceptionally high, at 93%. The accuracy level is then moderate, equal to 90% for LDA. The precision of SVM and LDA is 95% and 91%, respectively. SVM has a greater specificity of roughly 95% when compared to LDA, which has a specificity of 92%. The sensitivity of SVM is 90%, whereas the sensitivity of LDA is 87%. SVM has an error rate of 0.06% compared to LDA's error rate of around 0.09%, making it a perfect classifier.

280 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

Table 1. Analysis of Alzheimer’s disease MRI images using HAAR feature extraction with various Classifier. Model/ Metrics

SVM

LDA

Accuracy

93.16667

90.08333

Specificity

95.8194

92.1797

Sensitivity

90.53156

87.97997

FNR

9.468439

12.02003

FPR

4.180602

7.8203

Precision

95.61404

91.81185

F-Measure

93.00341

89.85507

Error Rate

0.068333

0.099167

Balance Accuracy

93.17548

90.07983

The performance study of Alzheimer's disease identification using HAAR with multiple classifiers is depicted in graphical form in Fig. (3) for better comprehension.

Fig. (3). Graphical analysis of Brain MRI image for a combination of HAAR feature extraction with the different classifiers.

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 281

Table 2 demonstrates the output of classifiers such as SVM and LDA when the GLCM feature extraction technique is utilized. The accuracy level is very high, as shown in Table 2, with SVM accuracy equal to 96%. Then there's LDA, which has a 95% accuracy level. LDA and SVM, on the other hand, have precisions of up to 98.2% and 98.1%, respectively. LDA has a higher specificity than SVM, which has a specificity of 98.1%. SVM's sensitivity is 90% of the total, while LDA's sensitivity is 92%. SVM has an error rate of 0.03% when compared to LDA using HAAR feature extraction, whereas LDA has an error rate of around 0.04%, making SVM the best classifier. Table 2. Analysis of Brain MRI images using GLCM feature extraction with the different classifiers. Model/ Metrics

SVM

LDA

Accuracy

96.08333

95.33333

Specificity

98.12287

98.33333

Sensitivity

94.13681

92.33333

FNR

5.863192

7.666667

FPR

1.877133

1.666667

Precision

98.13243

98.22695

F-Measure

96.0931

95.189

Error Rate

0.039167

0.046667

Balance Accuracy

96.12984

95.33333

Fig. (4) shows a graphical representation of the performance analysis of Alzheimer's disease detection using GLCM with various classifiers for better comprehension. 7. DISCUSSION The performance measure aids in determining the most successful feature extraction and classification model combination for Alzheimer's disease diagnosis. (Fig. 4) represent an investigation of Alzheimer's disease detection performance using HAAR and GLCM with several classifiers. In this study, wavelet and texture-based feature selection techniques are used in conjunction with classical classifiers such as SVM and LDA to detect Alzheimer's disease from MRI images, yielding high accuracy output with a low error rate, making this unique combination of feature selection and classifiers an ideal and most efficient ML modal. According to the results of the analyses, the GLCM feature extraction of the SVM classifier is the best for Alzheimer's disease classification.

282 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

Fig. (4). Graphical analysis of Brain MRI image for a combination of GLCM feature extraction with various Classifiers.

CONCLUSION The number of people affected by Alzheimer's disease is steadily increasing. It is critical to identify the illness early to cure it; otherwise, the person may experience neurological and psychological problems. The Kaggle database is used in this work to collect and process the Brain MRI image dataset. The raw images are pre-processed to make them suitable for classification. Pre-processed data is used in feature extraction to identify the image's relevant information. Finally, the retrieved feature data is fed into a variety of Alzheimer's disease detection classifier models. The model and feature extraction tactics are evaluated using performance metrics. The SVM classifier with the GLCM feature extraction has a high accuracy of 96% and a precision of 98.1%, according to the data analysis. Following that, the LDA with GLCM achieves a second high level of accuracy of

Magnetic Resonance Imaging

AI and IoT-based Intelligent Health Care & Sanitation 283

95%. The SVA classifier achieves the third high accuracy level by utilizing HAAR feature extraction algorithms. This study lends support to the selection of the best feature extraction and ML classifier model for detecting Alzheimer's disease in brain MRI images. CONSENT FOR PUBLICATON Not applicable. CONFLICT OF INTEREST The author declares no conflict of interest, financial or otherwise. ACKNOWLEDGEMENT Declared none. REFERENCES [1]

J.L. Cummings, and G. Cole, "Alzheimer Disease", JAMA, vol. 287, no. 18, pp. 2335-2338, 2002. [http://dx.doi.org/10.1001/jama.287.18.2335] [PMID: 11988038]

[2]

M.A. Smith, "Alzheimer Disease", Int. Rev. Neurobiol., vol. 42, pp. 1-54, 1998. [http://dx.doi.org/10.1016/S0074-7742(08)60607-8] [PMID: 9476170]

[3]

L.G. Apostolova, "Alzheimer Disease", Continuum (Minneap. Minn.), vol. 22, no. 2, Dementia, pp. 419-434, 2016. [http://dx.doi.org/10.1212/CON.0000000000000307] [PMID: 27042902]

[4]

Y. Yuan, H. Feng, L. Wang, Z. Li, Y. Shi, L. Zhao, Z. Feng, and H. Zhu, "Potential of endophytic fungi isolated from cotton roots for biological control against verticillium wilt disease", PLoS One, vol. 12, no. 1, p. e0170557, 2017. [http://dx.doi.org/10.1371/journal.pone.0170557] [PMID: 28107448]

[5]

V.N. Vapnik, “Introduction: Four periods in the research of the learning problem”. In The nature of statistical learning theory Springer, New York, pp. 1-15, 2000. [http://dx.doi.org/10.1007/978-1-4757-3264-1_1]

[6]

L.E. Collij, F. Heeman, J.P.A. Kuijer, R. Ossenkoppele, M.R. Benedictus, C. Möller, S.C.J. Verfaillie, E.J. Sanz-Arigita, B.N.M. van Berckel, W.M. van der Flier, P. Scheltens, F. Barkhof, and A.M. Wink, "Application of ML to arterial spin labeling in mild cognitive impairment and Alzheimer disease", Radiology, vol. 281, no. 3, pp. 865-875, 2016. [http://dx.doi.org/10.1148/radiol.2016152703] [PMID: 27383395]

[7]

J. Neelaveni, and M.S.G. Devasana, "Alzheimer Disease Prediction using ML Algorithms," 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS), pp. 101104, 2020. [http://dx.doi.org/10.1109/ICACCS48705.2020.9074248]

[8]

K.R. Kruthika, Rajeswari, and H.D. Maheshappa, "Multistage classifier-based approach for Alzheimer’s disease prediction and retrieval", Informatics in Medicine Unlocked, vol. 14, pp. 34-42, 2019. [http://dx.doi.org/10.1016/j.imu.2018.12.003]

[9]

J. Shi, X. Zheng, Y. Li, Q. Zhang, and S. Ying, "Multimodal neuroimaging feature learning with multimodal stacked deep polynomial networks for diagnosis of Alzheimer’s disease", IEEE J. Biomed.

284 AI and IoT-based Intelligent Health Care & Sanitation

Shobha et al.

Health Inform., vol. 22, no. 1, pp. 173-183, 2018. [http://dx.doi.org/10.1109/JBHI.2017.2655720] [PMID: 28113353] [10]

I.E. Kaya, A.Ç. Pehlivanlı, E.G. Sekizkardeş, and T. Ibrikci, "PCA based clustering for brain tumor segmentation of T1w MRI images", Comput. Methods Programs Biomed., vol. 140, pp. 19-28, 2017. [http://dx.doi.org/10.1016/j.cmpb.2016.11.011] [PMID: 28254075]

[11]

A. Amanatiadis, and I. Andreadis, "Performance evaluation techniques for image scaling algorithms", In: 2008 IEEE International Workshop on Imaging Systems and Techniques IEEE., 2008, pp. 114-118. [http://dx.doi.org/10.1109/IST.2008.4659952]

[12]

H.C. Lin, L.L. Wang, and S.N. Yang, "Automatic determination of the spread parameter in Gaussian smoothing", Pattern Recognit. Lett., vol. 17, no. 12, pp. 1247-1252, 1996. [http://dx.doi.org/10.1016/0167-8655(96)00082-7]

[13]

D. Nandan, J. Kanungo, and A. Mahajan, “An error-efficient Gaussian filter for image processing by using the expanded operand decomposition logarithm multiplication”. J. Ambient. Intell. Humaniz. Comput. pp.1-8. 2018. [http://dx.doi.org/10.1007/s12652-018-0933-x]

[14]

S.P. Yadav, and S. Yadav, "Image fusion using hybrid methods in multimodality medical images", Med. Biol. Eng. Comput., vol. 58, no. 4, pp. 669-687, 2020. [http://dx.doi.org/10.1007/s11517-020-02136-6] [PMID: 31993885]

[15]

D. Singh, and K. Kaur, "Classification of abnormalities in brain MRI images using GLCM, PCA and SVM", Int. J. Eng. Adv. Technol., vol. 1, no. 6, pp. 243-248, 2012. [IJEAT].

[16]

X. Wen, H. Yuan, C. Yang, C. Song, B. Duan, and H. Zhao, "Improved Haar wavelet feature extraction approaches for vehicle detection", In: 2007 IEEE Intelligent Transportation Systems Conference IEEE., 2007, pp. 1050-1053. [http://dx.doi.org/10.1109/ITSC.2007.4357743]

[17]

P.S. Sanjekar, and P.S. Dhabe, “Fingerprint verification using haar wavelet”. In 2010 2nd International Conference on Computer Engineering and Technology, vol. 3, pp. V3-361. 2010. [http://dx.doi.org/10.1109/ICCET.2010.5485896]

[18]

D.A. Pisner, and D.M. Schnyer, Support vector machine. ML. Academic Press, 2020, pp. 101-121. [http://dx.doi.org/10.1016/B978-0-12-815739-8.00006-7]

[19]

S. Balakrishnama, and A. Ganapathiraju, Linear discriminant analysis-a brief tutorial, 1998.

[20]

H. Rajaguru, and S.K. Prabhakar, “Time frequency analysis (dB2 and dB4) for Epilepsy classification with LDA classifier”. In 2017 2nd international conference on communication and electronics systems (ICCES), pp. 708-711. 2017. [http://dx.doi.org/10.1109/CESYS.2017.8321172]

AI and IoT-based Intelligent Health Care & Sanitation, 2023, 285-290

285

SUBJECT INDEX A Algorithm-ANN algorithm 41 Alzheimer’s disease 160, 269, 270, 281, 282 Applications 55, 146, 179, 208 analog-to-digital synchronization 179 computer-aided diagnostic 146 computer vision 208 wearable electronics 55 Artificial 16, 20, 21, 23, 24, 25, 27, 33, 38, 39, 145, 224 intelligence technology 145 neural network (ANN) 16, 20, 21, 23, 24, 25, 27, 33, 38, 39, 224 Autonomous wheat disease 225

B Back propagation neural network (BPNN) 224 Blood 21, 38 oxygenated 21 pressure, diastolic 38 Brain 160, 270 damaged 270 disorder 160 MRI images 271, 273, 277, 280, 281, 282, 283 Breast mammograms 85

C Carcinoma 84, 114, 115, 116 cystic 84 infiltrating ductal 84 infiltrating lobular 84 mammograms 84 Cardiac arrest 19, 23 Chest X-ray images 130, 132, 136, 140, 141 Chlamydia pneumoniae 130 Cloud 20, 21, 23, 27, 28, 31, 35, 50, 51, 52, 54, 57, 58, 62, 63

applications 35 CNNing 100, 102, 246 algorithms 100 systems 102, 246 CNNs 112, 248, 252 and transfer learning algorithms 112, 248, 252 and transfer learning methods 252 Completed local binary pattern (CLBP) 224 Computed tomography (CT) 256 Computer 148, 257 assisted brain tumor diagnosis system 257 generated categorization system 148 Computing resources 261 Conditions, neurological 4 Convolutional neural network (CNN) 4, 98, 100, 102, 103, 110, 111, 145, 146, 222, 228, 229, 230, 232, 235, 237, 240, 242, 246, 257 long short term memory 222 Corona diagnosis 130 Crops 222, 223, 225, 241, 244, 252 food material 241 CT images 116, 117 Cutting-edge classification technique 88 CXR images 129

D Damage 17, 98, 159, 160, 161, 270 brain cell 160 neuronal 270 Database, mini-MIAS mammography 85 Deep learning 16, 23, 31, 145, 146, 148, 225, 227, 228, 237 algorithms 16, 23, 145, 148, 227, 237 methods 31, 146, 225 network 225, 228 piping system 227 Dementia 171, 173 detecting 173

Shashank Awasthi, Mahaveer Singh Naruka, Satya Prakash Yadav, & Victor Hugo C. de Albuquerque Eds.) All rights reserved-© 2023 Bentham Science Publishers

286 AI and IoT-based Intelligent Health Care & Sanitation

forecast 171 Detecting lung infections 130 Detection 52, 99, 102, 145, 146, 147, 208, 209, 210, 211, 222, 224 developing stress 52 Diabetes mellitus 36 Diseases 18, 191, 193, 204, 224, 225, 230, 240 cardiac 191, 193, 204 potato 225, 230 transmission 18, 240 wheat 224 Disorders 145, 160, 181, 270 mental 160 neurological 270

E Electrocardiogram 19 Electrocardiography 179 Electroencephalography Electroencephalography (EEG) 179 Electromyography 180 Extraction techniques 159, 161, 162, 163, 171, 174

F Fourier analysis 87 Fundamental data analysis technique 89 Fungal spores 244

G Gaussian 271, 272, 273 approach 272 filter 273 function 273 method 271, 273 RBF Kernel function 88 Geometrical transformations 102 Geometric transformations 242 Gini index 89 Glaucoma 100, 101, 111, 112

Awasthi et al.

infection 100 prediction 101, 111, 112 therapy 100 Glaucomatous 111 GLCM 115, 163, 174, 281 extraction technique 174 feature extraction technique 281 FE method 115 techniques 163, 273 Glioma tumor 259 GLRM 123, 163, 164, 171 method 163, 164, 171 technique 123, 164 Google service 263 Gray-level run-length matrix method 164, 171 Growth 3, 36, 84, 114, 178, 241 aberrant cell 114 abnormal 84 excessive 3 malignant 36 rapid economic 178

H HAAR 85, 87, 93, 279 feature extraction technique 279 FE method 93 FE techniques 85 methods 87 technique 87 transform 87 Healthcare 17, 18, 177, 256 services 17 systems 18, 177 technology 256 Health 36, 52 monitoring technologies 52 problems, cardiovascular-related 36 Heart 17, 23, 35, 69, 191, 192, 193, 194, 198, 199, 200, 201, 204 disease 17, 23, 191, 192, 193, 194, 198, 199, 200, 201, 204 infection 35 transfer resistance 69

Subject Index

AI and IoT-based Intelligent Health Care & Sanitation 287

transmission 69 Hereditary disease 99, 160 Homogeneity 119, 163, 274 Hypothermia 58

IoT-based 52, 53, 54, 56 health monitoring system 52 patient health surveillance system 53 wearable sensors 54, 56

I

K

Illnesses 44, 46, 176, 178, 188, 193, 197, 201, 204, 223, 225, 240, 242, 244, 246, 270 cardiovascular 193, 204 forecast heart 192 neurological 270 plant 246 Image(s) 9, 85, 87, 132, 161, 162, 163, 165, 179, 224, 228, 245, 248, 252

Kernel function 84, 88, 121, 166 Kidney disease 33 KNN 122 classification technique 122 technique 122

augmentation techniques 132, 245, 248, 252 biopsy 85 compression technique 87 converting vertically-oriented 228 detector 179 flipping 133 magnetic resonance 162, 163, 165 magnetic resonant 161 mammogram breast 87 mammography 85 preprocessing techniques 9 processing techniques 224 ImageNet 130, 133, 146 classification 130, 146 database 133 Integral 65, 79 square error (ISE) 65, 79 time square error (ITSE) 65, 79 Intelligent health care monitoring system 182 Internet 35, 177, 178 of medical things (IoMT) 35 of things for healthcare applications 177 of things in healthcare 178 IoT 16, 52, 53, 63 enabled personal health surveillance system 53 sensors 52 technology 16, 50, 53, 63

L Leaf disease detection 225 Learning techniques 36, 181, 257 machine 36, 257 Lesions, lung carcinoma 116 Lethal bacterial disease 244 Leukemia 8 affected blood cell 8 Long 182, 183, 184, 188, 2 22, 225, 229, 230 short term memory (LSTM) 182, 183, 184, 188, 222, 225, 229, 230 term memory 188 term reliance problem 184 Lung cancer 115, 116, 123, 124, 127 CT scan images 116, 124 detecting 116, 127 screening 115 detection 116, 123 diagnosis 116 prediction 116, 127

M Machine 18, 39, 99, 102, 161, 168, 181, 208, 259 interpretation 39 learning algorithms 18, 99, 102, 161, 168, 181, 208, 259 Machine learning 18, 99, 161, 242 approaches 99, 242

288 AI and IoT-based Intelligent Health Care & Sanitation

classification techniques 161 classifier algorithms 161 technologies 18 Malignancy, primary central nervous system 256 Malignant melanoma 145 Mammography 83 Mask R-CNN 207, 208, 209, 213, 219 algorithm 208, 209, 213, 219 technique 207 Mask 180, 207 regional convolutional neural network 207 RNN approach 180 Medical 18, 133, 207 databases 18 imaging 133, 207 Melanoma detection process 147, 148 Methods, oscillometric 179 Microelectronics 18 Mineral-bearing water 67 Mobile application 33, 34, 37, 45, 46, 178 Morphometry technique 161 MRI 256, 257 and computed tomography 256 datasets of brain tumors 257 MRI images 256, 270, 271, 281 outperform X-ray and computer tomography images 256 Multi-layer transfer training system 133

N Neural network 65, 148, 151 algorithm 148 method 151 predictive controller (NNPC) 65 Nicotine, smoking 115 NNPC technique 76 Noise reduction 242 Noncommunicable diseases 34

O Object 207, 209

Awasthi et al.

detection techniques 207 instance segmentation 209 Oxygen saturation 52, 179

P Paddy 240, 241, 242, 252 infections 240 leaf diseases 242 leaf images 241 plant disease 252 plants 242 Pain, respiratory 130 Pesticides 222, 241, 242 PID controller 65, 66, 78, 79, 80, 81 for drug preparation 81 traditional 78 Pneumonia 129, 130, 131, 132, 136, 140, 141, 142 affected X-ray 132 detecting 129, 131, 136 diagnose 129, 130, 131 viral 130 Potato 227 diseased 227 Potato leaf 222, 227 diseases 233, 237 Predicting mobile application 46 Pressure 57, 59, 60, 69 diastolic 60 systolic 59 Principal component analysis (PCA) 36, 163, 224 Process 65, 66, 67, 104, 131, 132, 149, 161, 192, 193, 241, 242, 243, 245, 246 decision-making 23 Prognosis 50, 90, 277 Python program 176, 178, 182, 188

R Radiotherapy 145 Random decision forests 168 Random forest 168, 181, 243

Subject Index

algorithm 168 calculation method 181 technique 243 Real-time healthcare systems 18 Recognition 63, 228 health problem 63 leaf disease 228 Recurrent neural network (RNNs) 4, 100, 102, 148, 183, 229, 246 ResNet systems 130 RF algorithm 161, 172, 173

S Security frameworks 177 Segmentation 116, 192, 209 efficient automatic 209 techniques 116, 192 Self-driving automobiles 211 Sensor(s) 17, 19, 20, 21, 23, 52, 53, 55, 56, 57, 58, 61, 63, 181, 185, 208 blood pressure 19 electrocardiogram 19, 52 eye blink 52 infrared 53 security 181 technologies 53 Skin cancer 83, 144, 145, 156 prediction 156 Softmax activation algorithms 133 Software 41, 55 generating application 41 Stress, mental 51 Support vector machine (SVM) 83, 84, 85, 88, 92, 93, 121, 123, 161, 165, 166, 171, 173, 193, 196, 203, 224, 275, 279, 281 SVM 161, 165 modeling 165 technique 161 training method 165 System, sensor 208

AI and IoT-based Intelligent Health Care & Sanitation 289

T Techniques 6, 38, 63, 83, 88, 89, 99, 101, 115, 116, 130, 131, 161, 165, 192, 210, 224, 260 behavior-monitoring 63 detecting 210 natural 38 neuroimaging 161 Technology 18, 34, 224 image fusion 34 imaging 224 wireless communication 18 Telemedicine 18 Therapies, photodynamic 145 Thoracic radiographs 132 Traditional machine-learning methods 4 Training accuracy and loss of CNN algorithm 108 Transferable learning technique 130 Transfer learning 98, 99, 100, 102, 104, 105, 106, 107, 109, 111, 112, 133, 141, 246, 248, 249, 250, 251 algorithm 98, 99, 100, 104, 105, 106, 107, 109, 111, 112, 246, 248, 249, 250, 251 method 102, 105, 109, 133, 141, 252 Tumor 3, 84, 114, 256, 258, 259, 265, 266 malignant 144, 259 meningioma 259 pituitary gland 259

V Validation accuracy 108, 136, 230, 231, 232, 234, 235 Validation accuracy and loss of transfer learning 107 data 230, 231, 234, 258, 263, 264, 267 loss 106, 107, 108, 109, 136, 230, 231, 232, 234, 235 Voxel’s method 161

290 AI and IoT-based Intelligent Health Care & Sanitation

W Watershed segmentation technique 116 Wireless medical sensor network (WMSN) 177 Sensors 51, 53, 179 transmission 51

Awasthi et al.