EARLY IDENTIFICATION AND

CLASSIFICATION OF THYROID NODULE

IN MEDICAL ULTRASOUND IMAGES

Thesis submitted in fulfilment of the requirements for the Degree of

DOCTOR OF PHILOSOPHY

By

RAJSHREE SRIVASTAVA

Enrolment No. 196204

Department of Computer Science & Engineering and Information Technology

JAYPEE UNIVERSITY OF INFORMATION TECHNOLOGY

Waknaghat, Solan-173234, Himachal Pradesh, INDIA

January, 2024
@ Copyright JAYPEE UNIVERSITY OF INFORMATION TECHNOLOGY (Established
under H.P.Legislative Assembly Act No. 14 of 2002 and Approved by UGC under section
2(f)) WAKHNAGHAT, SOLAN, H.P. (INDIA)
January, 2024
ALL RIGHTS RESERVED
 SUPERVISOR’S CERTIFICATE

This is to certify that the work in the thesis entitled “Early identification and
classification of thyroid nodule in medical ultrasound images” submitted by
Rajshree Srivastava at Jaypee University of Information Technology,
Waknaghat, India, is a bonafide record of her original work carried out under
my supervision. This work has not been submitted elsewhere for any other degree
or diploma.

(Signature of Supervisor)

Dr. Pardeep Kumar

Associate Professor
Department of Computer Science & Engineering and Information
Technology,Jaypee University of Information Technology,

Waknaghat, Solan, India

Date:

i
 DECLARATION BY THE SCHOLAR

I hereby declare that the work reported in the Ph.D. thesis entitled “Early
identification and classification of thyroid nodule in medical ultrasound
images” submitted at Jaypee University of Information Technology,
Waknaghat, India, is an authentic record of my work carried out under the
supervision of Dr. Pardeep Kumar. I have not submitted this work elsewhere for
any other degree or diploma. I am fully responsible for the contents of my Ph.D.
thesis.

Rajshree Srivastava
Department of Computer Science & Engineering and Information
Technology, Jaypee University of Information Technology,
Waknaghat, Solan, India

Date:

ii
 ACKNOWLEDGEMENT

With the divine grace of “Almighty God”, the journey of Ph.D comes to end. I would like take
this opportunity to thanks each and every person who has helped me to complete this journey.

It is an immense pleasure to express my profound gratitude towards my supervisor Dr. Pardeep

Kumar, Associate Professor, Department of Computer Science & Engineering and Information
Technology who graciously gave me the opportunity to work under his guidance. I am thankful
for his patience, continuous support, optimistic approach and time to time guidance throughout
this course. His valuable motivation, help, suggestion, affirmative vision, magnificent
supervision and enormous confidence in my abilities made me face tough circumstances during
the progress of the research work. It is a great honour to work under his supervision.

I would like to express my gratitude to our Honorable Vice Chancellor Prof. (Dr.) Rajendra
Kumar Sharma, Dean (Academics and Research) Prof. (Dr.) Ashok Kumar Gupta and Head
Department of CSE/IT Prof. (Dr.) Vivek Kumar Sehgal to promote the research and facilitate
resources in the institution. I would also like to thank my doctoral committee members for
their valuable feedback and critical reviews during presentations and time to time help.

I am also grateful for the support received from JUIT, Waknaghat. In particular, I thank, all
staff members at the Department of Computer Science and Engineering and Information
Technology, JUIT, Waknaghat who has been extremely helpful at numerous occasions. I
would like to thank my parents for supporting me spiritually throughout my life. Last but not
the least, I thank my fellow Ph.D. friends for their consistent help and valuable discussion.

Date: (Rajshree Srivastava)

iii
 ABSTRACT

In medical imaging, machine learning (ML) and deep learning (DL) play a significant role to
predict symptoms of early diseases. DL is one of the growing trends in general data analysis
since 2013. It is an improvement of artificial neural networks (ANN) as it consists of many
hidden layers to permit high level of abstraction from data. In particular, convolutional neural
networks (CNNs) have proven to be a potential tool for computer vision tasks. Deep CNNs
have a capability to automatically learn raw data especially images.
The accurate assessment or identification of disease depends on image interpretation as well as
acquisition. Due to the improvement in last decade in image acquisition, devices acquire data
at high rate with increased resolution. However, the interpretation of images has recently begun
to benefit from computer technology. Mostly these are made by the radiologist, physicians and
senior doctors but limited due to its subjectivity and high skilled physicians/ doctors.
Computerized tools in the medical imaging field are the key enablers to improve diagnosis by
facilitating the findings.
Analysis of thyroid ultrasonography (USG) images via visual inspection and manual
examination for early identification and classification of thyroid nodule has always been
cumbersome. This manual examination of thyroid USG images in order to identify benign and
malignant thyroid nodule can be tedious and time-consuming. Various deep learning models
have emerged in medical field especially in thyroid nodule classification with the rapid
advancement in technology and increase in computational resources. Early identification of
such nodules can improve the effectiveness of clinical interventions and treatments. Therefore,
many researchers now advocate the use of computer diagnostic system (CDS) to objectively
and quantitatively analyze the USG images of thyroid nodules. This helps the radiologists to
solve the differences in interpretation of results.
Incidents of thyroid tumor have dramatically increased in recent years. However, early
ultrasound diagnosis can reduce morbidity and mortality. The work in clinical situations relies
heavily on the subjective experience of the sonographer. Thus, developing of a reliable thyroid
nodule identification model based on ultrasound images is needed to ensure the accuracy and
efficiency of thyroid nodule diagnosis. The key motive of this thesis is to develop efficient
models using ML and DL techniques for thyroid nodule identification and classification. The
first phase of this thesis proposes two hybrid models namely ANN-SVM and CNN-SVM. These
models work in four stages namely data collection, preprocessing, classification using
iv
hybridization of ML and DL classifiers and analysis of results. The second phase of this thesis
proposes an optimized CNN based model using segmentation and boundary detection
techniques. In this work, Alex-Net, visual geometry group (VGG-16), deep neural network
(DNN) and Res-Net-50 models are optimized using grid search optimization (GSO) technique.
The third phase of this thesis proposes a deep-generative adversial network (Deep-GAN) model
to improve the accuracy of the VGG and Alex-Net models. VGG-GAN and Alex-GAN. The
proposed model will be useful to physicians, researchers, practitioner and small hospitals
having limited resources can benefit from using this model to assist with thyroid ultrasound
examination diagnosis.

v
 TABLE OF CONTENTS
CONTENT PAGE NO

Supervisor’s Certificate i

Declaration by the scholar ii

Acknowledgement iii

Abstract iv-v

Table of contents vi-ix

List of tables x-xi

List of figures xii-xiv

List Of Acronyms and Abbreviations xv-xviii

Chapter 1: Introduction 1-5

1.1 Thyroid Nodule Datasets 3

1.1.1 Thyroid Digital Imaging Database (TDID) 3

1.1.2 Collected Dataset 3

1.2 Performance Parameters 4

1.3 Research Objectives 4

1.4 Organization of Thesis 5

Chapter 2: Literature review 6-25

2.1 Simple Techniques 6

2.2 Improved Techniques 15

2.3 Hybrid Techniques 21

2.4 Inferences Drawn from Literature Review 25

2.5 Summary 25

vi
Chapter 3: ANN-SVM hybrid model for the 26-34
identification and classification of thyroid nodules

3.1 Proposed Methodology 26

3.1.1Data Collection 27

3.1.2 Pre-processing 27

3.1.3 Feature-extraction 27

3.1.4 Classification 28

3.1.4.1 SVM 28

3.1.4.2 KNN 29

3.1.4.3DT 29

3.1.4.4ANN 29

3.1.4.5RF 30

3.1.4.6 Hybridization 30

3.1.5 Result analysis 31

3.2 Summary 34

Chapter 4: CNN-SVM Hybrid model for the 35-45

identification and classification of thyroid nodule

4.1 Proposed Methodology 35

4.1.1 Data Collection 36

3.1.2 Pre-Processing 36

4.1.3 Classification Using Hybridization of CNN and SVM 37

classifiers

4.1.3.1 CNN 37

4.1.3.2 SVM 40

vii
4.1.3.3 Hybridization of CNN-SVM Model 41
4.1.4 Result Analysis 42

4.2 Summary 45

Chapter 5: An Optimized Convolutional Neural Network 47-60

Based Model for Identification and Classification of
Thyroid Nodule
5.1 Proposed Methodology 47

5.1.1 Data Collection 48

5.1.2 Pre-Processing 48

5.1.3Morphological Operation, Segmentation and 48

Boundary Detection
5.1.3.1 Morphological Operation 48

5.1.3.2 Segmentation 49

5.1.3.3 Boundary Detection 50

5.1.4 Optimization and Classification 50

5.1.4.1 Res-Net-50 51
5.1.4.2 Alex-Net 51

5.1.4.3 VGG-16 52

5.1.4.4 DNN 53

5.1.4.5 Optimization Using GSO Technique 53

5.1.5 Result Analysis 55

5.2 Summary 60

Chapter 6: An improved Deep-GAN model for the 61-71

identification and classification of thyroid nodule

6.1 Proposed Methodology 61

viii
6.1.1 Data Collection and Pre-processing 62
6.1.2 Data Augmentation 62

6.1.3 Classification Using Deep-GAN 63

6.1.4 Result analysis 65

6.2 Summary 71

Chapter 7: Conclusion and future work 72-83

7.1 Major Research Contribution 72

7.2 Future Work 73

References 74
List of publications 88

ix
 LIST OF TABLES

Table Number Caption Page number

1.1 TI-RADS Score 2

2.1 Summarization of literature review using 11

simple techniques

2.2 Summarization of literature review using 18

improved techniques

2.3 Summarization of literature review using 23

hybrid techniques

3.1 Performance comparison of ML classifiers 32

on Dset-1

3.2 Performance comparison of ML classifiers 32

on Dset-2

3.3 Classification results of the best two 33

classifiers and the proposed ANN-SVM
hybrid model on Dset-1 and Dset-2

3.4 Comparison of the ANN-SVM hybrid model 33

with existing models

4.1 Parameters set for training CNN-SVM 43

hybrid model on Dset-1 and Dset-2

4.2 The comparison of the CNN-SVM model on 43

Dset-1 and Dset-2 with existing models used
for classifying TND

5.1 Parameter setting for GSO-CNN model for 56

Dset -1 and Dset-2

5.2 Performance comparison of GSO-CNN 56

model with Alex-Net, Res-Net-50, VGG-16
and DNN on Dset -1 and Dset-2

x
5.3 Results of GSO technique obtained on 57
learning rate and dropout factor parameters
on GSO-CNN model, Alex-Net, Res-Net-50,
VGG-16 and DNN

5.4 Comparison of the proposed GSO- 58

CNN model and existing models on
Dset -1 and Dset-2

6.1 Parameter setting of Deep-GAN model on 66

Dset -1 and Dset-2

6.2 The grid search results over different values 66

of hyperparameters set for Alex-GAN and
VGG-GAN models

6.3 Comparative performance comparison of 66

VGG-16, Alex-Net, VGG-GAN and Alex-
GAN models

6.4 Performance comparison of the Deep- GAN 67

model (i.e., Alex-GAN and VGG-GAN)
with existing models for classification of
TND

xi
 LIST OF FIGURES
Figure Caption Page
Number Number

1.1 Thyroid gland 1

2.1 The assortment of literature review 6

2.2 Framework of simple techniques 7

2.3 Working of improved techniques 15

2.4 Working of hybrid techniques 21

3.1 Framework of ANN-SVM hybrid model 26

3.2 Optimal Hyperplane in SVM 29

3.3 Architecture of ANN 30

3.4 The ANN-SVM hyrbid model 31

4.1 Systematic flow of CNN-SVM hybrid model 35

4.2 Sample of benign and malignant TND 36

4.3 Workflow of convolutional neural network 37

4.4(a-e) The illustration of a convolutional operation with 3X3 kernel 38

size

4.5(a-b) Illustration of max-pooling operation 39

4.6 The flattening operation performed in fully connected layer 40

4.7 Proposed CNN-SVM hybrid model 41

4.8 Performance comparison of the accuracy parameter with the 44

existing and CNN-SVM model on Dset-1 and Dset-2

4.9 Performance comparison of specificity, f-measure and 44

sensitivity parameters with the existing and CNN-SVM model
on Dset-1 and Dset-2

xii
4.10 The epoch-accuracy graph for Dset-1 45

4.11 The epoch-accuracy graph for Dset-2 45

5.1 Systematic flow of the proposed GSO-CNN model 48

5.2 The morphological operation 49

5.3 Structure of Res-Net-50 51

5.4 Structure of Alex-Net 52

5.5 Structure of VGG-16 52

5.6 Architecture of DNN 53

5.7 Proposed GSO-CNN model 54

5.8 Comparative analysis of the accuracy parameter with models 58

on Dset -1 and Dset-2
5.9 Comparative analysis of models based on sensitivity, 59
specificity and f-measure on Dset -1 and Dset-2

5.10 Accuracy epoch graph for Dset -1 59

5.11 Accuracy epoch graph for Dset -2 60

6.1 The systematic flow of Deep-GAN model 62

6.2 Architecture of GAN 63

6.3 Proposed Alex-GAN model 64

6.4 Proposed VGG-GAN model 64

6.5 Generator network 68

6.6 Discriminator network 68

6.7(a-b) Training of Alex-Net architecture 68

6.8a Generator network 69

6.8b Discriminator network 69

xiii
6.8(a-d) Training for VGG-16 architecture 69

6.9 The accuracy-epoch graph of Alex-GAN model on Dset-1 70

6.10 The accuracy-epoch graph of Alex-GAN model on Dset-2 70

6.11 The accuracy-epoch graph of VGG-GAN model on Dset-1 70

6.12 The accuracy-epoch graph of VGG-GAN model on Dset-2 71

xiv
 LIST OF ACRONYMS & ABBREVIATIONS

AI Artificial Intelligence

ACWE Active Contour Without Edges

AC Active Contour

ANN Artificial Neural Network

AFUD-SVM Artificial Fish Coupled with Uniform Design – Support

Vector Machine

Ab-LSTM Attention-based Bi-Directional LSTM

ATA American Thyroid association

BTND Benign Thyroid Nodule

CT Computed Tomography

CMR Closest-matching CAD system

CAD Computer-aided Diagnostic System

CNN Convolutional Neural Network

CPU Central Processing Unit

DCNN Deep Convolutional Neural Network

Deep-GAN Deep- Generative Adversarial Network

D Discriminator

DL Deep Learning

DT Decision Tree

DA Data Augmentation

DNN Deep Neural Network

DRLSE Distance-Regularized-Level-Set-Evolution

EGFS Ensemble Gain Ratio Feature Selection

EHR Electronic Health Record

xv
ELM Extreme Learning Machine

FP False Positive

FN False Negative

FE Feature Engineering

FD-SMOTE Farthest-Distance-Based-Synthetic-Minority-
Oversampling Technique

FFNN Feed Forward Neural Network

GAN Generative Adversial Network

GSO Grid Search Optimization

GB Giga bytes

G Generator

GM Generative Models

GPU Graphical Processing Unit

GLCM Grey Level Co-Occurrence Matrix

GC-ACWE Graph Cut- Active Contour without edges

GLRLM Gray Level Run Length Matrix

GAN Generative Adversarial Network

GLCM Grey Level Co-occurrence Matrix

hyp Hyperparameter

IOT Internet of Things

IDMIE Instituto de Diagnostico Medico

LDA Linear Discriminant Analysis

LR Logistic Regression

LRN Local Response Normalization

MTN Malignant Thyroid Nodules

MHz Megahertz

xvi
ML Machine Learning

MLP Multi-Level Perceptron

MKSVM Multi-Kernel-Based Support Vector Machine

MTND Malignant Thyroid Nodules

MPR Maximum Retention of Inter-Pixel Relations

MCDLM Multi-Task Cascade DL

NABH National Accreditation Board for Hospitals

NN Neural Network

NLP Natural Language Processing

P Probability

RF Random Forest

RAM Random Access Memory

Res-GAN Resnet-Generative Adversial Neural Network

Re-Rx Recursive Rule Extraction Algorithm

RIFO Ranked Improved F-Score ordering

RTE Real-Time Elastography

SMOTE Synthetic-Minority-Oversampling Technique

SSD Solid State Drives

SVM Support Vector Machine

SVM-L SVM Linear

SVM-Q SVM Quadratic

SFTA Segmentation-Based Fractal Texture Analysis

SIFT Scale Invariant Feature Transform

SE Squeeze And Excitation

SURF Speeded Up Robust Features

xvii
TSH Thyroid-stimulating hormone

TNDs Thyroid nodules

T4 Thyroxine

T3 Triiodothyronine

TDID Thyroid Digital Imaging Database

TI-RADS Thyroid Imaging Reporting and Data system

TN True Negative

TP True Positive

TA Texture Analysis

USG Ultrasonography

U.S.A. United States of America

UCI University of California, Irvine

VLAD Vector of locally Aggregated Descriptor

VGG-16 Visual Geometry Group-16

xviii
 CHAPTER 1

INTRODUCTION

Disease prediction is important for medical institution as well as for the patients to make
possible medical care decisions. Sometimes, incorrect decision might lead to serious
consequences. Hence, early identification and classification of disease is important. According
to the latest report of American Thyroid Association (ATA), more than 12 percent of the U.S.
population will develop a thyroid condition during their lifetime [1]. Thyroid dysfunction is the
most prevalent endocrine disorder worldwide having an epidemiology of 4.6% in the United
Kingdom, 11% in India and 2% in the U.S among the overall population [2]. Thyroid gland is
a small butterfly shaped endocrine gland located in neck region below Adam’s apple [3]. It
produces three types of hormones namely thyroxine(T4), calcitonin and triiodothyronine (T3)
to maintain the metabolism of the body. The function of these hormones depends on the right
amount of supply of iodine in diet [4]. The ductless gland located at the bottom of our brain
sends signals to secret hormones called as Thyroid Stimulating Hormone (TSH). It informs the
thyroid gland to secret the hormones as per the requirement [5]. Thyroid Nodules (TNDs) are
solid or fluid filled lumps formed within the thyroid region/gland. Figure 1.1 shows an image
of thyroid gland. There are various imaging processes like Ultrasonography (USG) and
Computed Tomography (CT) used for the identification of diseases [6]. USG is highly
recommended for the identification of TNDs as it is cost friendly, radiation free, real-time and
painless.

Figure 1.1: Thyroid gland [7]

1
When there is an excess production of TSH in the body, it is said to be hyperthyroidism. Some
of the common hypothyroid signs noticed are abundance weight, dry skin, fatigue of body
quality and tiredness [8]. Hypothyroid is a condition when there is less production of TSH in
the body. Some of common sign noticed signs are tremors, weight loss and loose bowels.
Thyroid problems are common among people as its symptoms are mild. There are twocategories
of TNDs namely benign (BTND) and malignant (MTND) [9]. MTND are consideredas one of
the top 10 MTND and account for 1.1% of all the malignant tumors [10]. There are certain
distinctive features of TNDs which can be found in TI-RADS system [11]. Table 1.1 shows the
TI-RADS score.

Table 1.1: TI-RADS Score

TI-RADS Type Category
Score
1 Benign
2 Not Suspicious BENIGN
3 Mildly Suspicious
4 Moderately suspicious
5 Highly suspicious MALIGNANT

In last few decades, success of Machine Learning (ML) and Deep Learning (DL) techniques in
image identification/diagnosis tasks leads to solve challenges in diagnostic imaging. ML
techniques have a great potential to deal problems in the field of medicine, drug discovery and
decision making. Medical images are mostly analyzed by the radiologists who are highly
experienced and have knowledge in the respective domain. A delayed or erroneous diagnosis
causes harm to the patients [12]. In ML, learning new features are well described by the patterns
in images. The sparse representation (i.e., feature extraction/selection) in ML techniques finds
informative patterns from a shallow architecture but they lack in its representational power [13].
DL techniques use the feature engineering (FE) step into a learning step [14]. It works by
discovering informative representation in a self-taught manner with some pre-processing steps
without extracting features manually [15]. Thus, the manual feature engineering has been
shifted to computers thereby allowing the non-experts of ML to use DL techniques efficiently
especially in the medical diagnosis filed. Some of the popular reasons lead to success of DL
techniques are: (1) development of advance algorithms [16], (2) availability of large amount of
data [17], (3) easy derivation of hierarchical features from low-level to high-level [18], (4)
improvement in the traditional Artificial Neural Networks (ANN) i.e., increase in multiple layer
[19] and (5) advancement in high-tech processing units [20].
2
 1.1 Thyroid Nodule Datasets
This section focuses on the datasets used for the identification and classification of TNDs in
this thesis. The dataset plays significant role while conducting any experiment. It contains
information that helps the model to have a high level of understanding. In this thesis, two
datasets are considered:

1.1.1 Thyroid Digital Imaging Database (TDID)

This dataset is collected from Instituto de Diagnostico Medico (IDIME) ultrasound department,
Colombia. The dataset has 295 thyroid USG images, having 107 BTND and 188 MTND images
[21]. The images were captured with Toshiba Nemio 30, MX Ultrasound devices and Toshiba
Nemio with 12MHz convex and linear transductors.

1.1.2 Collected Dataset

This dataset is collected from Kriti Scanning Center, Prayagraj, Uttar Pradesh (U.P.), India duly
approved by National Accreditation Board for Hospitals (NABH) and Healthcare Providers
[22]. The collected sample size was 654 images, out of which 428 and 226 were BTND and
MTND images. The duration of dataset collection was from July 2020 to March 2021. The
images were captured with SIEMENS Healthier, Mindray-Resona 7 and Voluson E- 10
Ultrasound devices with 3 to 11MHz convex and linear transductors. In both cases, Thyroid
Imaging Reporting and Data system (TI-RADS) scores is used to pre-classify the TNDs images.

The criteria for the inclusion and exclusion of thyroid USG images are as follows:
Inclusion-
• Images having clear boundary of the nodules.
• Images having one or multiple nodules.
• Patients having age greater than 16 years.
• Patients with the confirmed thyroid nodules problem.
Exclusion –
• Excluded the images which were zoomed in or zoomed out.
• Patients having age less than 16 years.
• Patients without confirmed thyroid nodules problem.

3
1.2 Performance Parameters

There are four different parameters considered for evaluation of proposed models in this thesis:

• Accuracy: It is defined as the ratio of correctly classified TNDs to the total number of
examples and computed using equation 1.1:
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = ∗ 100 (1.1)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

• Specificity: It is defined as the percentage (%) of people who test negative for a specific
disease among a group of people who do not have the disease and calculated using equation
1.2:
𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = ∗ 100 (1.2)
𝑇𝑁+𝐹𝑃

• Sensitivity: It is defined as the percentage (%) of people who test positive for a disease
that have disease and calculated using equation 1.3:
𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = ∗ 100 (1.3)
𝑇𝑃+𝐹𝑃

• F-measure: It measures the test accuracy of the model and calculated using equation 1.4:
2𝑇𝑃
𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = ∗ 100 (1.4)
2𝑇𝑃+𝐹𝑁+𝐹𝑃

1.3 Research Objectives

The objectives of this thesis are highlighted below:

1. To design a hybrid model for the identification and classification of TNDs.

a. To propose a hybrid model using hybridization of ML classifiers.
b. To propose a hybrid model using hybridization of ML and DL classifiers.
2. To propose an optimized DL model the identification and classification of TNDs
3. To design an improved DL model for the identification and classification of TNDs.

4
1.4 Organization of Thesis

The thesis consists of seven chapters as per the details below:

Chapter 1: This chapter provides the introduction about thyroid nodule problem along with
its types, datasets used, and performance parameters for the identification and classification
of TNDs.
Chapter 2: This chapter discusses the literature review of different ML and DL techniques
for identification and classification of TNDs.
Chapter 3: This chapter discusses the first part of objective one of the thesis i.e. design of a
hybrid model using hybridization of ML classifiers for the identification and classification
of TNDs.
Chapter 4: This chapter discusses the second part of objective one of the thesis i.e. design
of a hybrid model using hybridization of ML and DL classifier for the identification and
classification of TNDs.
Chapter 5: This chapter elaborates the second objective of the thesis i.e. design of an
optimized DL model for the identification and classification of TNDs.
Chapter 6: This chapter focuses on the third objective of the thesis i.e. design of an improved
DL model using Generative Adversial Network (GAN) for the identification and
classification of TNDs.
Chapter 7: This chapter concludes the thesis and throws light on future perspectives.

5
 CHAPTER 2

LITERATURE REVIEW

This chapter describes literature review related to thyroid disorder. The method of
categorization of work is based on techniques related to identification, classification and feature
selection for thyroid identification and detection. For this study, the selected articles are
assorted into three categories namely simple, improved and hybrid techniques. Figure 2.1
shows the assortment of literature review.

Figure 2.1: The assortment of literature review

2.1 Simple Techniques
Simple techniques are mainly ML techniques. It is a subset of Artificial Intelligence (AI) that
builds a mathematical model based on sample examples to predict without itself being
explicitly involved [23]. Simple techniques are widely used in medical diagnosis to detect
patterns of diseases from Electronic Health Record (EHR) and inform clinicians of any acute
problem [24]. It works in four phases namely (1) data collection, (2) data processing
(preparation of data, model selection), (3) train the model and (4) analysis of results. Figure
2.2 shows the framework of simple techniques.

6
 Figure 2.2: Framework of simple techniques [23]

Zhu et al. proposed ML-based CAD system [25]. Sonographic features are extracted from the
USG images and ANN is used for discriminating USG images. The model has achieved
specificity of 81.80%, accuracy of 83.10% and sensitivity of 83.80%.

Naryan et al. presented a multi-organ segmentation algorithm that is successful to perform

volumetric analysis and US-guided interventions [26]. The model was evaluated on two
datasets having 34 and 18 USG images. The model is yet to be evaluated on a large dataset.

Song et al. proposed a Grey Level Co-Occurrence Matrix (GLCM) method for feature
extraction and different ML classifiers for classification [27]. The logistic regression has
achieved a good result from the rest of the classifiers on 155 USG images.

Ardakani et al. evaluated the CAD system with Texture Analysis (TA), Linear Discrimant
Analysis, Principal Component Analysis, (PCA+LDA) for the classification of TNDs [28].

Nugroho et al. used bilateral filters for the pre-processing step results in smoothen images and
also preserves the boundaries of the object followed by Active Contour Without Edges
(ACWE) for the segmentation of the thyroid nodules [29]. The result shows not only
improvement in the accuracy of the segmentation of active contours (AC) but also makes
localization of the TNDs clear.

Song et al. proposed a GLCM based texture feature model using different ML algorithms for
classification [30]. The model was evaluated on 147 USG images. The best accuracy result is
achieved by Support Vector Machine (SVM) with 76.6%, Random Forest (RF) with 74.1%,
ANN with 72.8% and Decision Tree (DT) with 67.1% accuracy.

Pan et al. proposed Principal Component Analysis (PCA) ensemble method for feature

7
reduction and found that it performs better than Bagging and AdaBoost [31].

Alrubaidi et al. proposed an interactive segmentation method evaluated on 27 USG images of

public datasets [32].

Vanithamani et al. proposed CAD system using GLCM for feature extraction, segmentation
using fuzzy c- means clustering and classification using SVM on TDID dataset [33].

Nugroho et al. proposed a texture feature extraction using Gray Level Run Length Matrix
(GLRLM), histogram and GLCM. Multi-Level Perceptron (MLP) classifier was considered for
classification [34]. Model is evaluated on small public dataset.

Mustafa et al. proposed the Farthest-Distance--Based-Synthetic-Minority-Oversampling

Technique (FD-SMOTE) system and evaluated it on University of California, Irivine (UCI)
repository thyroid dataset [35].

Pasha et al. proposed Ensemble Gain Ratio Feature Selection (EGFS) model [36]. The model
ensemble RF & gain ratio algorithm to find the most efficient features followed by its alignment
with LR, K-Nearest Neighbor (KNN) and Naive Bayes.

Jothi et al. proposed a novel SVM linear+ SVM Quadratic+ Closest-matching CAD system i.e.
(SVM-L+SVM-Q+ CMR) [37]. The proposed CAD system has achieved a good result by
combining the advantages of SVM and CMR classifiers.

Jiang et al. proposed an intelligent model using ML-based techniques [38]. All ML-based
classifiers have achieved 84% accuracy except DT classifier.

Shankar et al. proposed a Multi-Kernel-Based SVM (MKSVM) system for thyroid disease
classification using the UCI and private datasets [39].

Ma et al. proposed a ML-based model (center clustering, deep neural network, SVM, KNN and
LR) for thyroid nodule classification [40]. A total of 508 USG images are used in the
experiment. The highest accuracy achieved is 87%.

8
Zheng et al. proposed the serum Raman spectroscopy technique combined with Artificial Fish
Coupled with Uniform Design (AFUD) –SVM discriminant model capable for the detection of
thyroid dysfunction [41].

Dandan et al. proposed CAD system using Wavelet Multi-Sub-Bands Co-Occurrence Matrix
(WMCM) and other feature extraction techniques [42]. The model is evaluated on 180 USG
images.

Poudel et al. proposed an autoregressive feature for thyroid ultrasound images without
modeling the images [43].

Prochazka et al. proposed a 2-threshold binary decomposition CAD system that uses direction
independent feature set [44]. Authors addressed the problem of direction-independent features
problem. Model was evaluated on 40 USG images and used histogram and segmentation for
classification. The results show that SVM gives good results in comparison with RF.

Wu et al. proposed texture analysis system using Local Binary Pattern (LBP) and Histogram
of Oriented Gradients (HOG) feature extraction techniques [45]. The model was evaluated on
4,574 USG images.

Colakoglu et al. proposed ML-based model to differentiate between BTND and MTND via
extracting 306 features from 235 USG images [46]. Among ML-based algorithms, the RF
classifier gives a good result with an accuracy of 86.8%.

Zhang et al. proposed ML-based thyroid nodule detection using sonographic features and Real-
Time Elastography (RTE) [47]. Nine different ML algorithms like RF, LR, SVM, Neural Net,
Adaptive Boosting, Naive Bayes and Convolutional Neural Network (CNN) were considered
but RF shows the best result with an accuracy of 84.1 % in comparison with other classifiers.

Kesarkar et al. proposed an ANN based CAD system [48]. The ACWE based segmentation
method and MLP for the classification of TNDs images.

Zhao et al. compared various ML classifiers for thyroid nodule detection using sonographic
features [49]. The model shows that the application of ML-based classifiers like DT, SVM,

9
KNN, RF, LR, MLP, Xgboost, Gradient Boosting Tree and Naive Bayes on sonographic
features has achieved the best result. The model was evaluated on 849 USG images.
Table 2.1 shows summarization of literature review using simple techniques.

10
Table 2.1: Summarization of literature review using simple techniques

Article Proposed Model Feature extraction/ Classification Dataset Findings Limitations

[Ref.] Detection
techniques
Zhu et al. CAD System Sonographic features ANN Public dataset (i)It has achieved 83.10 % (i)The result can further be
(2013) accuracy, 83.80% improved by using hybridization
sensitivity and 81.80% of two or more classifiers.
specificity
Narayan et Speckle Patch ---- Department of (i)It has achieved 92.1% (i)The model is yet to be
al. (2015) Similarity model Radiology, Sardjito accuracy. evaluated on large dataset.
Hospital, Yogyakart
Song et al. Texture based GLCM SVM, RF, DT, Shandong Provincial (i)LR has achieved 84% (i)Other different texture feature
(2015) thyroid detection LR, ANN Qianfoshan Hospital accuracy, 78.5% specificity extraction can be used to achieve
model and 78.5% sensitivity. better result.
Nugroho et Active Contour Department of (i)It is insensitive to noise. (i)Manual parameter tuning and
al. (2015) Bilateral Filtering Radiology, Sardjito initial contour is required.
model Hospital, Yogyakart
Song et al. ML based CAD GLCM SVM, RF, DT, Public dataset (i)SVM achieved the best (i)The model is evaluated on 147
(2015) System ANN result among rest of the thyroid USG images.
classifier.
Ardakani et CAD System Statistical features ------ Radiology Department (i) No operator dependency. (i)Model can further be
al. (2015) Imam Khomeini (ii)It takes additional cost improved by using other type of
hospital, Urmia, Iran and time loss. feature extraction method.
(ii) It has achieved accuracy
97.14% specificity 100%
and sensitivity 94.45%.
Vanithamani CAD System GLCM Fuzzy C- TDID (i) It has achieved accuracy (i)The size of the dataset is very
et al. (2016) means 86.6% specificity 85.7% less.
clustering, SVM and sensitivity 87.5%.
(ii) It achieves good results
in less numbers of features
extracted.
Nugroho et CAD System GLRM, GLCM, MLP ------ (i)It can be used as 2nd (i)The size of the dataset is very
al. (2016) Histogram opinion by clinicians. less.
(ii)Model has achieved
accuracy 89.74%
specificity 91.67% and

11
 sensitivity 88.89%.
Koundal et Automated ---- Spatial Public dataset (i) It is insensitive to nodule (i)It has strong resistance to
al. (2016) delineation of neutrosophic echogenicity. noise.
thyroid nodules distance (ii) It has achieved 95.45%
regularized level sensitivity.
set
Pasha et al. EGFS model EGFS LR, Naive UCI repository and local (i) It has achieved 96.45% (i) It has strong resistance to
(2017) Bayes, KNN dataset accuracy. noise.
(ii) The model has achieved (ii) Complex data pre-processing
good result with a few is used.
numbers of features.
Mustafa et FD-SMOTE Minimum Genetic UCI repository (i) The model works good (i)Work can be extended on
al. (2017) Technique Description Length algorithm on less dataset. rough theory problem.
Method, PCA (ii) It has achieved 84.12%
accuracy.
Jothi et al. SVM-L+SVM- PSO SVM Public dataset (i) The proposed CAD (i)Size of dataset is less.
(2017) Q+CMR CAD system differentiate PT C
System and NT H&E- stained
thyroid histo-pathology
images.
(ii) It has achieved accuracy
99.54% sensitivity 100%
and specificity 98.57%.
Jiang et al. ML based CAD Sonographic features SVM, KNN, LR, Public dataset (i)The model has achieved (i) Other feature extraction
(2017) system DT, RF 84% accuracy. techniques can be explored.
(ii) Size of the dataset is less.
Zheng et al. AFUD System Wavelet Multi-sub- SVM First Affiliated Hospital (i) It has achieved 82.74% (i)Size of the dataset used is less.
(2018) bands Co-occurrence accuracy.
matrix (ii) The proposed technique
can be used to develop a
portable device for
detecting TNDs function.
Dandan et al. ML based CAD WMCM SVM Public dataset (i)he model has achieved (i) Work can be extended on
(2018) system 87% accuracy. other different types of feature
extraction methods.
(ii) Size of the dataset is less.
Shankar et MKSVM method Auto-regressive SVM Public and UCI (i)It has achieved accuracy (i)High computation time.
al. (2019) features 97.49%, specificity 94.5%

12
 and sensitivity 99.05%.
(ii)The model has achieved
good result with a few
numbers of features.
Ma et al. ML based CAD Sonographic features KNN, LR, SVM Public dataset (i)The model has achieved (i) Work can be extended on
(2020) System 87% accuracy. other different types of feature
extraction methods.
(ii) Size of the dataset is less.
Poudel et al. Autoregressive Histogram features, SVM, RF SurgicEye GmbH, (i) The proposed approach (i)Bi-Spectral model can be used
(2019) Feature based SFTA University Clinic of outperforms from DSC and in future work.
CAD model Magdeburg, Germany had similar SE and SP.
(ii) It has achieved 90%
accuracy.
Prochazka et CAD System HOG, LBP SVM, RF Our clinic (i) It uses direction- (i)Work can be extended on
al. (2019) independent features. large dataset.
(ii) It has achieve-d 94.64%
accuracy.
Wu et al. CAD System LBP, GTF, HOG, Second Affiliated (i) Can be used as 2nd (i)High computation time.
(2019) FOS, GLCM, HWF, Hospital of Fujan opinion for diagnosis of
Autoregressive Medical University TNDs.
model, GRLM (ii) It has achieved 88.58%
accuracy.
Colakoglu et ML based CAD Ultrasonography, ML classifiers --- (i)It has achieved accuracy (i)Work can be extended on
al. (2019) system shear wavelet texture 86.8%, sensitivity 85.2% large dataset.
and specificity 87.9%.
Zhang et al. ML based CAD ----- RF, LR, SVM, First Affiliated Hospital (i)The model has achieved (i)Different other feature
(2019) system Neural net, 84.6% accuracy, 84.2% selection techniques can be used
adaptive specificity and 88.1% for better result.
boosting, naive sensitivity.
bayes and CNN
Kesarkar et ANN model for Sonographic features MLP ----- (i)It has achieved 93.84% (i)Size of the dataset is less.
al. (2021) thyroid nodule accuracy, 97.82%
diagnosis sensitivity and 93.75%
specificity.
Zhao et al. ML based EGFS DT, SVM, KNN, ------- (i)It has achieved 88.8% (i)Work can be extended to other
(2021) techniques for RF, LR, MLP, accuracy, 92.9% specificity feature extraction methods.
classification of Xgboost, and 81.7% sensitivity.
thyroid USG gradient

13
images boosting tree and
Naive baye

14
2.2 Improved Techniques
Improved techniques are the DL techniques based on ANN with representation learning. It
uses various layers to progressively extract a higher level of features from the user input [50].
DL techniques applications include audio recognition, medical image analysis, speech
recognition, bioinformatics and Natural Language Processing (NLP) [51]. The term “deep”
means large number of features to be extracted from raw image as input. E.g., lower layers and
higher layers identify edges and letters [52]. DL reduces the cost of feature engineering and
retrieves a lot of information that helps the model to achieve better results [53]. One of the
major differences between ML and DL is that ML cannot extract features automatically; it
requires labelled parameters whereas in the case of DL, it extracts features automatically
without human intervention [54]. Figure 2.3 shows the working of the improved techniques. It
also works in three steps: (1) input, (2) feature extraction and classification and (3)
prediction.

Figure 2.3: Working of improved techniques [54]

Du and Sang et al. proposed an improved Distance-Regularized-Level-Set-Evolution (DRLSE)

method for the segmentation of 2D TNDs USG images [55]. It consists of three steps namely
image processing, extracting boundaries and image post-processing. It does not perform well
to automatically segment the TNDs USG images.

Archarya et al. proposed Gabor transform features from high-resolution TNDs images [56].
The model was evaluated on 240 USG images.

Peng et al. show the first order features selection in TNDs ultrasound image recognition as
image biomarkers [57].

Zhu et al. proposed an image augmentation CAD system using transfer learning technique as
Generative Adversarial Network (GAN) [58].

15
Song et al. proposed a multi-scale DL single-shot detection network method guided by nodule
with the help of a CNN [59].

Zhang et al. proposed a multi-scale single-shot detection network guided by nodule [60]. The
model was evaluated on 45 USG images.

Sundar et al. explored different DL techniques like CNN, VGG-16 model and Inception-V3
model for the diagnosis of TNDSs [61].

Wang et al. proposed a novel EM-based model to train Alex-Net for the diagnosis of the TNDs
[62]. Authors addressed the problem of labeled data and evaluation was done on public and
private datasets.

Chen et.al proposed 2- Level Attention-Based Bi-Directional LSTM (Ab-LSTM) for TNDs
classification [63].

Li et al. proposed a Deep Convolutional Neural Network (DCNN) model for the detection of
TNDs [64]. The result shows good improvement in sensitivity, accuracy and specificity in
comparison with a group of skilled radiologists.

Ko et al. proposed a DCNN for TNDs diagnosis [65]. The model was evaluated on 589 USG
images dully approved by the radiologist and used 3 layers of CNN for the classification of
TNDs.

Wang et al. proposed a DL-based CAD system using YOLOv2 NN [66]. The proposed method
achieved a real-time and synchronous diagnosis CAD system for TNDs diagnosis.

Liu et al. proposed clinical-knowledge-guided CNN model for the TNDs classification [67].

Aboudi et al. proposed a CAD system combining different attributes after optimization to
achieve good accuracy [68].

Guo et al. proposed an improved DL technique for TNDs identification [69]. A new CNN

16
model is proposed which integrates Squeeze and Excitation (SE) and Maximum Retention of
Inter-Pixel Relations Module (CNN-SE-MPR). The model was evaluated on 407 USG images.

Ajilisa et al. used various pre-trained CNN network [70]. K-means clustering is employed to
deal with imbalanced datasets.

Nguyen et al. used a fusion of spatial and Fourier based methods for feature extraction and also
used two CNN models together to improve the accuracy [71].

Vasile et al. proposed an intelligent diagnosis model using ensemble DL methods [72]. Model
was designed with 5 CNN layers and then ensemble with pre-trained VGG-19 model. The
model was evaluated on 230 USG images.

Zhu et al. proposed an automatic DCNN CAD system [73]. The model was evaluated on 719
USG images.

Yang et al. proposed a Multi-Task Cascade DL (MCDLM) model [74]. It works in three steps
namely pre-processing, U-Net model for segmentation of nodules and semi-supervised method
for feature extraction and classification. The model has achieved good result on small dataset.
Table 2.2 shows summarization of literature review using improved techniques.

17
Table 2.2: Summarization of literature review using improved techniques
Article Proposed Feature Classification Dataset Findings Limitations
[Ref.] Model extraction/
Detection
techniques
Du et al. CAD System -- DLRSE Wuhan Tongji (i) It is insensitive to noise and (i)It requires manual initial contour.
(2015) Hospital, China nodule echogenicity.
(ii) Achieved 85% sensitivity.
Acharya CAD System Gabor MLP, SVM, Chiang Mai (i)It has achieved 94.3%. (i)Work can be extended to non-linear feature
et al. Transform KNN University Hospital, extraction algorithms.
(2016) Thailand
Peng et CAD System FOS SVM Ruian People (i)It has achieved specificity (i)Work can also extend on large thyroid USG
al. Hospital, Zhejiang, 93.3%, accuracy 88% and dataset.
(2017) China sensitivity 82.1%.
Zhu et. al DL based --- CNN TDID (i) It can be used as a 2nd opinion (i)Handcrafted and DL model can be used for
(2017) CAD system by the radiologist. better result.
(ii) It has achieved 93.75%
accuracy.
Song et. MC-CNN ---- VGG-16 TDID (i) It has achieved accuracy (i)Sample size is less
al (2019) Framework 92.1%, sensitivity 94.1% and
specificity 96.2%.
(ii) Embedded spatial pyramid
module into traditional CNNs.

Zhang et FAFL Model Multi- --- Local hospital (i) New feature association model (i) Improvement can be done by fusion of two
al. channel is proposed. or more classifiers.
(2018) feature (ii) It has achieved 83% accuracy. (ii) The dataset size is very small.
association
Sundar DL CAD VGG-16 and - Public and private (i) Achieved good result on a (i)The dataset size is very small.
et al. System Inception- smaller number of datasets.
(2018) V3 model (ii) It has achieved accuracy 93%
with Inception-v3 + CNN and
with VGG-16 with 89%.

Wang et Semi- VGG-16 ---- TDID (i)The model solves multi- (i)The dataset size is very small.
al. supervised instance problems and also
(2018) method improves the classification
accuracy.

18
 (ii)It has achieved accuracy
80.91%, sensitivity 81.82% and
specificity 80%.
Chen et Ab-LSTM -- CNN Ruijin Hospital, (i) Model perform good. (i)The dataset size is very small.
al. Memory Shanghai (ii) It has achieved 86.18%
(2018) method accuracy.
Li et al. DCNN ---- CNN Traditional Chinese (i) It can be used as 2nd opinion (i)The dataset size is very small.
(2019) Weihai Municipal for diagnosis.
Hospital, Shandong, (ii) It has achieved sensitivity
Tianjin Cancer 93.4% and specificity 86.1%.
Hospital and Western
Medicine Hospital,
Jilin
Ko et al. DCNN --- CNN ----- (i)The model has achieved 86% (i)Sample size is less
(2019) accuracy.
Wang et DL- CAD ---- CNN Tertiary class A (i) Can be used as 2nd opinion for (i)Work can be extended on large database.
al. System hospital the diagnosis.
(2019) (ii) It has achieved sensitivity
90.5% and accuracy 90.31%.
Liu et al. DL- CAD --- CNN Cancer Hospital ofthe (i) The pixel annotation used in (i)Work can be extended on large database.
(2019) System Chinese Academy of this work is coarse.
Medical Sciences (ii) It has achieved accuracy on
dataset-I 97.5% and
accuracy on dataset-II 97.1%.
Aboudi CAD system Statistical, CNN CIM Laboratory of (i) Fusion of (intensity + textural (i)Work can be extended on large database.
et al. GLCM the National features + shape) is used.
(2020) University of (ii) It has achieved 96.13%
Columbia accuracy.
Guo et CNN-SE- --- CNN Co-operate hospitals (i) With the help of MPR module, (i)Work can be extended on big database.
al. MPR System loss of inter pixel relation is
(2020) solved.
(ii) It has achieved accuracy
90.17%, specificity 92.35% and
sensitivity 86.99%.
Ajilisa et CAD System ---- Alexnet, DNN, TDID (i) DNN performs best. (i) Work can be experiment on large database.
al. Inception v-3, (ii) Failed to address the problem of noise.
(2020) Googlenet,
VGG-19,

19
 Xception,
Resnet-10,
VGG-16,
Resnet-50,
Nguyen DL based FFT CNN TDID (i) Fusion of Res-Net50 and (i)Work can be extended on large database.
et al. CAD system Inception Model is considered for
(2021) classification.
(ii) It has achieved 92.05%
accuracy.
Vasile et Ensemble ---- CNN Public dataset (i)Ensemble CNN and VGG-19. (i)Work can be extended on large database.
al. model (ii)It has achieved sensitivity
(2021) 95.75%, accuracy 97.35% and
specificity 98.43%.
(iii)It can be used as 2nd opinion
by the physician.
Zhu et al. DL based --- CNN Ethics committee (i) It has achieved 86.5% avg. (i)Hand-crafted feature extraction methods
(2021) model Hospital of Shanghai accuracy, 87.1% specificity and can also be used to improve accuracy.
86.6% sensitivity.
(ii) It can be used as 2nd opinion
by the physician.
Yang et DL CAD Sonographic CNN Partner hospitals (i)Used augmentation technique. (i)Size of the dataset is less.
al. system features (ii)It has achieved 90.01%
(2021) accuracy, 92.15% specificity and
87.47% sensitivity.

20
2.3 Hybrid Techniques
Hybrid techniques are the combination of two or more classifiers to improve performance of
model [75]. It works in three steps namely, (1) data pre-processing, (2) model selection and
(3) prediction. Figure 2.4 shows the working of hybrid techniques.

Figure 2.4: Working of hybrid techniques [75]

Liu. et al. proposed a combination of hand-crafted feature and DL feature-based extraction

technique for the thyroid image classification i.e., VGG-F model with Scale Invariant Feature
Transform (SIFT) + Vector of Locally Aggregated Descriptor (VLAD) and GLCM [76].

Xia et al. proposed a hybrid Extreme learning machines (ELM) based approach for TNDs
diagnosis [77]. The model was evaluated on 230 USG images.

Nguyen et al. proposed a model based on fusion of spatial and frequency domain for feature
extraction from USG images and achieved an accuracy of 90.88% [78].

Qin et al. proposed (conventional ultrasound+ elasticity imaging) model for the classification
of TNDs [79].

Xie et al. proposed a (DL +handcrafted features) model and achieved an accuracy of 85% [80].
The model was evaluated on 623 USG images.

Sun et al. proposed a (DL-based technique + statistical) features together for classification of
TNDs [81]. The model was evaluated on 490 USG images.

Hang et al. suggested a Res-GAN model [82]. In this work, authors have compared ResNet-
18 with Res-GAN model and Speeded Up Robust Features (SURF) with deep features along

21
with RF as classifier. The proposed Resnet-Generative Adversial Neural Network (Res-GAN)
model has achieved an accuracy of 95%.

Srivastava R., et al. proposed a hybrid model using ANN and SVM for the classification of
TNDs [83]. The model shows that when two classifiers hybrid together, the accuracy is
increased. Table 2.3 shows summarization of literature review using hybrid techniques.

22
Table 2.3: Summarization of literature review using hybrid techniques
Article [Ref.] Proposed Model Feature extraction/ Classification Dataset Findings Limitations
Detection
techniques
Liu et al. (2017) Hybrid method SIFT, HOG, VLAD, CNN Cancer Hospital of (i) The feature used (i)There is no
LBP the Chinese shows that both hand- comparison with the
Academy crafted features and DL other models.
features contribute to the
thyroid image
classification.
(ii) It has achieved
specificity 94.1%,
accuracy 92.9% and
sensitivity 90.8%.
Xia et al. (2017) Hybrid method Relief method SVM Wenzhou Central (i) It can be used as 2nd (i)Small dataset.
Hospital opinion for TNDs
diagnosis.
(ii) It has achieved
accuracy 87.72%,
sensitivity 78.89% and
specificity 94.55%.

Nguyen et al. (2019) DL-CAD System --- CNN TDID (i) Used the combination (i)Work can be
of spatial and frequency extended on other
domain problem. different type of
(ii) It has achieved feature extraction
90.88% accuracy. method.

Qin et al. (2019) Hybrid method ----- VGG-16 Huiying Medical (i)Use of end-to-end (i)The model is yet
Technology implementation leads to to be evaluated on
(Beijing) Co.Ltd good efficiency result. large dataset.
(ii)It has achieved
accuracy 94.70%,
specificity 92.7% and
sensitivity 97.96%.

23
Xie et al. (2020) Hybrid deep LBP CNN Shanghai Tenth (i) Can be used as 2nd (i)Different other
learning based People’s hospital opinion for the diagnosis. feature methods can
feature method (ii) It has achieved 85% be explored.
accuracy.
Sun et al. (2020) Hybrid method Statistical features CNN First Affiliated (i) A voting system is (i) Different other
Hospital of Nanjing used. feature methods can
Medical University (ii) The model has be explored.
achieved 86.5% (ii)The model is yet
accuracy. to be evaluated on
large dataset.
Hang et al. (2021) Res-GAN model Statistical features ResNet-18, Random TDID (i) The model has (i) Different other
Forest, AdaBoost performed well in less feature methods can
sample size. be explored.
(ii) The Res-GAN model (ii)Different other
performs well in DL techniques can
comparison with ResNet- be explored for the
identification of
18.
TNDs.
(iii) In this work, fusion
of SURF + deep
features+ RF is used for
classifier.
(iv) Accuracy: 95%
Srivastava, R et al. Hybrid model GLCM ANN, SVM, DT, TDID, collected (i) The model has (i) Different other
(2021) KNN, RF performed well in less feature methods can
sample size. be explored.
(ii) The issue of noise in (ii) DL models can
images is addressed. be explored.
(iii) Model shows that
accuracy of the model is
increased in comparison
with standalone classifier
when it is hybrid.
(iv) The model is
evaluated on 2 datasets.

24
2.4 Inferences Drawn from Literature Review
The literature review shows significant research carried out for the identification and
classification of TNDs. However, there are many gaps which are partially unexplored and needs
immediate attention. The following are the inferences drawn from the detailed review:

1. Most of the existing study relies on either ML or DL classifier for the identification and
classification of TNDs. Thus, there is a scope to build hybrid model.

2. This review sheds light on the possible use of segmentation, clustering, boundary and
edge detection techniques to enhance the performance of improved techniques.

3. It is inferred from that improved techniques model require continuous training and
adjustment of hyper parameters. It is still an unresolved issue as it requires complete
domain knowledge.

4. It is evident that less attention has been paid by the researchers for noise removal in
thyroid USG images.

5. This review sheds light on the importance of TNDs disorder. Such thyroid disorder has
also attracted a large number of health care experts subject to availability of USG images
with advancement in medical imaging. Hence, such images are a good inspiration for
the healthcare analysts to identify and detect TNDs disorder.

2.5 Summary

This chapter described the literature review conducted for TNDs identification and
classification. The inferences drawn are inked from the detailed literature review carried out in
this chapter. The coming chapters in the thesis will present effective and efficient models for
the identification and classification of TNDs.

25
 CHAPTER 3

ANN-SVM HYBRID MODEL FOR THE IDENTIFICATION

AND CLASSIFICATION OF THYROID NODULE

ML and DL techniques have been widely used in recent years in medical imaging. USG is one
of the suggested diagnostic methods for accurately differentiating BTND and MTND images.
In this chapter, ANN-SVM hybrid model is proposed for the identification and classification of
TNDs USG images.

3.1 Proposed Methodology

It consists of five steps namely (1) data collection phase, (2) pre-processing, (3) feature
extraction phase, (4) classification and (5) result analysis. Figure 3.1 shows the framework of
ANN-SVM hybrid model.

Figure 3.1: Framework of ANN-SVM hybrid model

26
3.1.1 Data Collection
Two datasets are considered namely public TDID and collected dataset. The public dataset has
188 malignant and 107 benign images i.e., a total of 295 USG images whereas collected dataset
has 226 malignant and 428 benign i.e., a total of 654 USG images. The complete detailof dataset
has given in under subsection 1.1.1.

3.1.2 Pre-processing

It is an important phase for the development of any model. In this phase, image resizing, noise
removal using median filter and RGB to gray scale conversion are performed to maintain the
uniformity of the dataset. All these steps are computed using equation 3.5-3.7:

𝑟𝑒𝑠𝑖𝑧𝑒 = 𝑡𝑓. 𝑘𝑒𝑟𝑎𝑠. 𝑆𝑒𝑞𝑢𝑒𝑛𝑡𝑖𝑎𝑙([𝑙𝑎𝑦𝑒𝑟𝑠. 𝑅𝑒𝑠𝑖𝑧𝑖𝑛𝑔(𝐼𝑀𝐺 _𝑆𝐼𝑍𝐸, 𝐼𝑀𝐺 _𝑆𝐼𝑍𝐸)]) (3.5)

where IMG_SIZE: size of the image

𝑡𝑓. 𝑖𝑚𝑎𝑔𝑒. 𝑔𝑟𝑎𝑦𝑠𝑐𝑎𝑙𝑒_𝑡𝑜_𝑟𝑔𝑏(𝑖𝑚𝑎𝑔𝑒𝑠, 𝑛𝑎𝑚𝑒 = 𝑁𝑜𝑛𝑒) (3.6)

where images: the RGB tensor to convert the images, name: name of the operation

𝑦[𝑚, 𝑛] = 𝑚𝑒𝑑𝑖𝑎𝑛{𝑥|𝑖, 𝑗|, (𝑖, 𝑗)𝜖 𝑣} (3.7)

where ѡ is a neighborhood, [m, n] pixel is location in the image

The main reason for using median filter is that it removes the noise and also preserves the
edges of the images for the accurate prediction [84]. Benign images are labelled as “0” and
malignant images as “1”.

3.1.3 Feature extraction

GLCM texture analysis is used in this work to find underlying characteristics of textures and
represent them in unique manner to have accurate classification of an image [85]. It works by
calculating how often the pairs of pixels with a specific value and in a specified spatial
relationship occur in input image [86]. Let there be a square matrix of size “N x N”, where N is
considered as the number of levels specified and f (a, b) is the pixel value at the point (a, b) of
an image size N x N. A total of 4 features are extracted using GLCM technique. The description
is given as:

27
3.1.3.1 Contrast: It measures the local variance in GLCM and can be computed by using
equation 3.8:

𝑁−1
𝑁−1
𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡 = ∑ 𝑘2 ∑ (𝑓(𝑎, 𝑏)) (3.8)
𝑎,𝑏=0
𝑘=0

where k: is range [0 to (size (GLCM)-1))2]

3.1.3.2 Correlation: It computes the combined probability rate of the particular pixel pairs and
can be computed by using equation 3.9:

𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 = ∑𝑁−1 (𝑎 − 𝜇)2𝑓(𝑎, 𝑏) (3.9)

𝑎,𝑏=1

where µ is mean

3.1.3.3 Homogeneity: It measures the closeness of the distribution of the elements in the
GLCM to the GLCM diagonal using equation 3.10:

𝐻𝑜𝑚𝑜𝑔𝑒𝑛𝑒𝑖𝑡𝑦 = ∑𝑁−1 {𝑓(𝑎, 𝑏)}2 (3.10)

𝑎,𝑏=0

3.1.3.4 Energy: It computes the total number of squared elements in the GLCM and can be
computed by equation 3.11:

𝐸𝑛𝑒𝑟𝑔𝑦 = ∑𝑎,𝑏=0
𝑁−1 𝑓(𝑎, 𝑏)2 (3.11)

3.1.4 Classification

Four different types of ML classifiers are considered namely ANN, SVM, DT, KNN and RF.
These are discussed below:

3.1.4.1 SVM

It is one of the frequently used supervised ML classifiers, widely used in image processing
and pattern recognition [87]. In SVM, an optimal hyperplane decides the separation between
individual classes of patterns. It must be maximum so that the problem of underfitting and
overfitting can be reduced [88-89]. Figure 3.2 shows the optimal hyperplane of SVM,

28
C1(diamonds) and C2(circles). Here, SVM tries to fit a linear boundary known as a hyperplane
between two different classes and tries to maximize the distance between the boundary and its
nearest point of each class.

Figure 3.2: Optimal hyperplane in SVM [89]

3.1.4.2 KNN
It is one of supervised classifier that can be used for classification and regression. In both the
conditions, the training part consists of k closest training examples while testing part computes
the distance between the samples and then gives output based on classification or regression
[90].

3.1.4.3 DT
Its structure is like a tree where the top most nodes is called as “root node”, leaf node as the
“class label”, branches as result on the test, decision node represent the test of an attribute. One
of the advantages of using DT classifier is its ability to give good classification result but it may
lead to the problem of overfitting [91].

3.1.4.4 ANN
It is inspired by the “nervous system” where nodes act as an artificial neuron and the directed
edges between neurons are defined by weights [92]. Figure 3.3 shows structure of ANN. In this
work, we have used three layered ANN.

29
 Figure 3.3: Architecture of ANN [92]

3.1.4.5 RF
It is a kind of ensemble learning technique involving a large number of individual DT. It works
by first selecting the random samples from a given dataset and then constructs DT for each
sample. Thereafter, it performs prediction based on maximum voting. RF does not suffer from
the problem of overfitting and is used for both regression problems and classification [93].

3.1.4.6 Hybridization

An ANN- based 3-layer Feed Forward Neural Network (FFNN) is used along with
hybridization of SVM classifier for the classification of TNDs. A simple ANN having one
hidden layer with 4 neurons in the input layer was created. The neurons present in MLP
network are trained consequently on backpropagation algorithm until an output value is
obtained at each of the output neurons. ANN was trained with the extracted features and the
output of the hidden layer was used to train the SVM classifier. Figure 3.4 shows the ANN-
SVM hybrid model.

30
 Figure 3.4 : The ANN-SVM hyrbid model
3.1.5 Result Analysis

This section summarizes the results of the ANN-SVM hybrid model for classification of TNDs.
The work is done on MATLAB 2016b software using intel core i5 8th generation computer.
The k-fold cross validation technique is used to validate the results of the ANN-SVM model.
In this work, 5 cross validation is considered i.e., D= [c1, c2,.……c5], where 4-fold is used for
training and 1-fold is used for testing. Once, the pre-processing phase is over and features are
extracted using GLCM method, various ML algorithms like RF, SVM, KNN, DT and ANN are
employed for the classification. For better representation, public dataset is named as: Dset-1
and collected dataset is named as: Dset-2. Figure 3.5 shows the parameter setting for ANN-
SVM hybrid model. The various parameters considered are batch size, optimizer, etc. Table
3.1 shows the standalone ML classifiers comparison on Dset-1. The ANN classifier has
achieved an accuracy of 79.28%, specificity of82%, F-measure of 80.9% and sensitivity of
77.16% and SVM has achieved an accuracy of 78.4%, specificity of 80.4%, sensitivity of
76.3% and F-measure of 79%. Table 3.2 shows the performance of the standalone ML
classifiers on Dset-2. The ANN classifier has achieved an accuracy of 84.96%, specificity of
85.88%, F-measure of 85.4% and sensitivity of 81.56% andSVM has achieved an accuracy of
82.65%, specificity of 83.50%, sensitivity of 80.12% and F-measure of 84%. From the table 3.2
and table 3.3, it can be seen that ANN performs better in comparison with the rest of the other
classifiers i.e., DT, KNN, RF followed by SVM classifier. Table 3.3 show classification results
of the best two classifiers and the proposed hybrid model on Dset-1 and Dset-2. It can be
analyzed from the table that when ANN and SVM classifiers are hybrid together then the
performance of the model is increased.
31
Thus, it can be said that the proposed ANN-SVM performs better in comparison with the
standalone ANN and SVM classifier. Table 3.4 shows comparison of the ANN-SVM hybrid
model with existing models.An improvement of 2% to 4% is reported with ANN-SVM hybrid
model in comparison with reported literature. The ANN-SVM hybrid model has achieved an
accuracy of 84.12%,specificity of 82.95%, F-measure of 85.23% and sensitivity of 85.14% on
Dset-1 and an accuracy of 90%, specificity of 87.5%, F-measure of 91.67% and sensitivity of
91.66% on Dset-2.

Figure 3.5: Parameter setting of ANN-SVM hybrid model [83]

Table 3.1: Performance comparison of ML classifiers on Dset-1

Models Accuracy Sensitivity F- Specificity
(%) (%) measure (%)
(%)
DT 74.5 73.26 75 79
SVM 78.4 76.3 79 80.4
KNN 71.4 68.69 72 74.59
RF 74.5 72.60 76 77.99
ANN 79.28 77.16 80.9 82

Table 3.2: Performance comparison of ML classifiers on Dset-2

Models Accuracy Sensitivity F- Specificity
(%) (%) measure (%)
(%)
DT 75.90 73.55 77.1 80.68
SVM 82.65 80.12 84 83.50
KNN 70.1 68.11 71.6 75.55
RF 80.60 77.79 81.11 81.41
ANN 84.96 81.56 85.4 85.88

32
Table 3.3: Classification results of the best two classifiers and the proposed ANN-SVM hybrid
model on Dset-1 and Dset-2
Datasets Classifiers/ Accuracy Specificity F-measure Sensitivity
Models (%) (%) (%) (%)
ANN 79.28 77.16 80.9 82
Dset-1 SVM 78.4 76.3 79 80.4
ANN+SVM 84.12 82.95 85.23 85.14
(Hybrid model)
ANN 84.96 81.56 85.4 85.88
Dset-2 SVM 82.65 80.12 84 83.50
ANN+SVM 90 87.5 91.67 91.66
(Hybrid model)

Table 3.4: Comparison of the ANN-SVM hybrid model with existing models
Model Feature Classifier Accuracy Specificity Sensitivity F-measure
Extraction (%) (%) (%) (%)
Technique
(Zhu et. al Sonographic ANN 83.10 81.80 83.80 -
(2013)) features

(Song et. al GLCM SVM, ANN, RF, 84 78.5 78.9 78.4

(2015)) LR, logistic

(Jiang et. al Sonographic SVM, KNN, LR, 84 - - -

(2017)) features DT, RF

(Dandan et. al WMCM SVM 87 - - -

(2018))

(Zhang et. al Sonographic RF, logistic 84.6 84.2 88.1 -

(2019)) features regression, SVM,
Neural net, adaptive
boosting, naïve
bayes, CNN
(Colakoglu et. Texture RF 86.8 87.95 85.2 -
al (2019)) features

(Ma et. al Sonographic KNN, LR, SVM 87 - - -

(2020)) features

Proposed GLCM Hybrid of ANN- Dset-1: Dset-1: Dset-1: Dset-1:

model SVM 84.12; 82.95; 85.14; 85.23;
Dset-2: Dset-2: Dset-2: Dset-2:
90 87.5 91.66 91.67

33
3.2 Summary
This chapter presents a hybrid model using hybridization of two ML classifiers for
identification and classification of TNDs. The proposed ANN-SVM model is compared with
standalone classifiers as well with hybrid model. From the experiments, it can be inferred that
the performance of the proposed model is competitive when the two ML classifiers are hybrid
together.

34
 CHAPTER 4

CNN-SVM HYBRID MODEL FOR THE IDENTIFICATION

AND CLASSIFICATION OF THYROID NODULE

CNN and its variants are one of the most famous and commonly employed algorithms in the
field of medical imaging [94]. Some of the popular reasons include automatic identification of
relevant features from the images without any human supervision. Its structure is inspired from
neurons processing in human brains [95]. This chapter gives an elaborative explanation on the
hybridization of CNN with SVM for the identification and classification of TNDs. The
efficiency of the CNN-SVM hybrid model is tested on collected and TDID datasets.

4.1 Proposed Methodology

The working of CNN-SVM model for the identification and classification of TNDs is discussed
in this section. The model works in four stages namely: (1) data collection, (2) pre-processing,
(3) classification and (4) result analysis. Figure 4.1 shows the systematic flow of the CNN-
SVM model.

Figure 4.1: Systematic flow of CNN-SVM hybrid model

35
4.1.1 Data Collection Phase

In this work, public TDID dataset and collected dataset are considered. The public dataset has
188 malignant and 107 benign images i.e., a total of 295 thyroid USG images whereas collected
dataset has 226 malignant and 428 benign i.e., a total of 654 thyroid USG images. Figure 4.2
shows the sample of benign and collected dataset.

Figure 4.2: Sample of benign and malignant TNDs

4.1.2 Pre-processing

It is defined as a way of converting raw data into a desired form [96-97]. This derived
information will be fed into the training model for the successful diagnosis [98]. The various
steps involve in pre-processing are image resizing, noise removal using gaussian blur function,
RGB to grayscale conversion and data augmentation respectively. Initially the size of image
was 560 X 360 pixels, for the uniformity it’s being resized to 256 X 256 pixels. In this work,
four types of data augmentation techniques are used namely like rescale, zoom, shear and
horizontal flip respectively. All these steps are computed using equation [4.12-4.15]:
𝑟𝑒𝑠𝑖𝑧𝑒 = 𝑡𝑓. 𝑘𝑒𝑟𝑎𝑠. 𝑆𝑒𝑞𝑢𝑒𝑛𝑡𝑖𝑎𝑙([𝑙𝑎𝑦𝑒𝑟𝑠. 𝑅𝑒𝑠𝑖𝑧𝑖𝑛𝑔(𝐼𝑀𝐺 _𝑆𝐼𝑍𝐸, 𝐼𝑀𝐺 _𝑆𝐼𝑍𝐸)]) (4.12)

where IMG_SIZE: size of the image

𝑡𝑓. 𝑖𝑚𝑎𝑔𝑒. 𝑔𝑟𝑎𝑦𝑠𝑐𝑎𝑙𝑒_𝑡𝑜_𝑟𝑔𝑏(𝑖𝑚𝑎𝑔𝑒𝑠, 𝑛𝑎𝑚𝑒 = 𝑁𝑜𝑛𝑒) (4.13)

where images: the RGB tensor to convert the images, name: name of the operation

𝑔𝑎𝑢𝑠𝑠𝑖𝑎𝑛 = 𝑛𝑝. 𝑟𝑎𝑛𝑑𝑜𝑚. 𝑛𝑜𝑟𝑚𝑎𝑙(𝑚𝑒𝑎𝑛, 𝑠𝑖𝑔𝑚𝑎, 𝑖𝑚𝑎𝑔𝑒. 𝑠ℎ𝑎𝑝𝑒) (4.14)

where sigma: standard deviation

36
𝑡𝑟𝑎𝑖𝑛_𝑑𝑎𝑡𝑎𝑔𝑒𝑛 = 𝐼𝑚𝑎𝑔𝑒𝐷𝑎𝑡𝑎𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑜𝑟(𝑁𝑎𝑚𝑒, 𝑉𝑎𝑙𝑢𝑒) (4.15)

where Value: values assigned to techniques, Name: name of the techniques

4.1.3 Classification using Hybridization of CNN and SVM Classifiers

4.1.3.1 CNN
The CNN architecture consists of four parts namely: (1): input layer, (2): convolutional layer,
(3): max-pooling layer, (4): fully connected layer and (5): output layer [99]. Figure 4.3 shows
the workflow of CNN.

Figure 4.3: Workflow of convolutional neural network [100]

1. Input Layer
It specifies a complete description of an image (Nh x Nw x Nc) were h: height, w: width and c:
channel size of an image [101].

2. Convolutional Layer
In this layer, 2-D kernels are learned during the initial phase. The coefficients(coeff.) of all
these kernels encodes the important information from the input image [102]. There are some
hyper-parameters like number of kernels, padding, size of kernels and stride to improve the
performance [103]. Convolution works by finding the sum of the dot(.) products between the
input image and the kernel [104]. The output of the convolution becomes the input of the next
layer. It works as the spatial filtering [105]. Here, the size of kernel (k) is defined by equation
16:
𝑊𝑥𝐻 (4.16)
where W: width and W=2a+1, a and b: positive integer, H: height of the image pixel and
H=2b+1

37
During the entire convolutional operation, the size of the image and the kernel are same.
Padding is one of the techniques of the convolution process to introduce zeros (0s) around the
border of the image to maintain the aspect ratio of the resultant image after convolution [106].
This can be computed using equation 4.17:
𝑁+2𝜌−𝑤 𝑁+2𝜌−𝑤
[𝑁, 𝑁, 𝑍] ∗ [𝑊, 𝑊, 𝑁 ] = ([ + 1] , [ + 1] , 𝑁 ) (4.17)
𝐶 𝑠 𝑠 𝑓

where [N, N, Z]: Size of an image, [W, W, N_C]: Size of a kernel, p: no. of padding, s: stride
length, N: filter depth, Z: image depth, Nf: filters of the kernel

Figure 4.4(a-e) explains the illustration of a convolutional operation with 3X3 kernel size. In
this case, the input to the image is 5X5 matrix with stride 1. After the 9th convolution operation,
3X3 feature map is obtained. This shows the size of the input image is reduced extracting some
useful global and local features.

Figure 4.4(a-e): The illustration of a convolutional operation with 3X3 kernel size [106]

3. Max-pooling Layer
The output of the convolutional layer becomes the input of the max-pooling layer [107]. It
38
extracts the most essential features from each feature map, thereby giving the best feature
output. The major benefits of using this layer are as follows: (1) it prevents overfitting, (2)
feature invariance and (3) dimensionality reduction of the features [108]. It can be computed
using equation 4.18 and 4.19:
𝑦𝑖 = max 𝑅𝑥 𝑅{𝑦𝑖𝑟𝑥𝑟}𝑓(𝑟, 𝑟) (4.18)
𝑓(𝑟, 𝑟) = 𝜀. 𝑦𝑘−1 𝑥 𝑣𝑘 + 𝑒𝑘 (419)
𝑖 𝑖,𝑗 𝑗

where max RxR: max-pooling in RxR, 𝑦𝑖𝑟𝑥𝑟: i-th output of rxr window, ϵ: trainable variant

Figure 4.5(a-b) explains the illustration of max-pooling operation. The complete operation
works from left(L) to right(R), the first picture is the input, next is the max-pooling and so on.

Figure 4.5(a-b): Illustration of max-pooling operation [107]

4. Fully Connected Layer

This layer combines all the features extracted from the input images learned by the previous
layers to classify the images [109]. Here flattening function is performed to convert a 2D matrix
to the 1D array which helps in reducing dimension and computation complexity [110]. In this,
each node of neuron has some weight and bias associated with it to help the input image to gain
information [111]. Figure 4.6 shows the flattening operation performed in fully connected layer.

39
 Figure 4.6: The flattening operation performed in fully connected layer [109]

5. Output Layer
In this layer we have the classified result which is mostly followed by softmax function (i.e.,
the normalized exponential function) which gives the output mapped into the range of [0,1]
[112]. It can be computed using equation 4.20:
𝑒𝑧𝑖
𝜎(𝑧 ) = (4.20)
𝑖 𝑛
𝛴𝑖=1 𝑒 𝑧𝑖

where n: maximum no. of classes, zi : input

4.1.3.2 SVM
It is a class of ML algorithms works by finding (n-1) dimensional hyper-plane. It finds a plane
by maximizing the dist. between the data points available or given in the classes with the help
of support vectors [113]. SVM employs the use of kernels like sigmoid, radial basis function
(rbf), linear, polynomial and non-linear which takes the data as the input and gives output
accordingly [114-115].
Let there be a sample dataset input: {(𝑥1, 𝑦1 ), (𝑥2, 𝑦2), … … (𝑥𝑛, 𝑦𝑛)}, 𝑥𝑖 𝜖 𝑅𝑑, target output: 𝑖 =1,2,
… … 𝑁, 𝑦 ∈ {∓1}[116]. The optimal classification hyperplane is defined as given in equation 4.21:
𝑣𝑇𝑥𝑖 + 𝑏 = 0 (4.21)
where b: bias, 𝑣: weight vector
Then b and 𝑣 can be determined using equation 4.22:
𝑦𝑖(𝑤𝑇𝑥𝑖 + 𝑏) ≥ 1 − 𝑥𝜉𝑖 (422)

40
where ξi is the slack variable(var.).
According to Lagrange multiplier method, the solution to optimal classification hyperplane is
converted into optimization using equation 4.23:
𝑁
𝑁 𝑁
1
𝑄(𝑎) = ∑ 𝑎𝑖 − ∑ ∑ 𝑎𝑖 𝑎𝑗 𝑦𝑖 𝑦𝑗 𝐾 (𝑥𝑖 , 𝑥𝑗 ) (4.23)
2
𝑖=1 𝑗=1
𝑖=1

where {𝑎𝑖}𝑁𝑖=1 is lagrange multiplier, K(xi,xj) is kernel function

4.1.3.3 Hybridization of CNN-SVM Model

This model works by removing the last output layer of CNN with SVM model. It can be said
that the output of the last layer will become the input to SVM for better classification of TNDs.
The CNN model applies different convolutional (conv.) and subsampling technique in
convolutional layer [117]. These functions speed up the extraction of important data from the
input given to the model. At the end, SVM trains the feature vectors obtained from fully
connected layer and take decision in an effective manner. Figure 4.7 shows the CNN-SVM
model.

Figure 4.7: Proposed CNN-SVM hybrid model

4.1 Algorithm: Proposed CNN-SVM hybrid model

Algorithm 4.1: Proposed algorithm for CNN-SVM hybrid model
Input: Imd→ Image, AugImd →Augmented Images
Output: Type of TNDs
BEGIN
1. Upload the image dataset on the google drive account
2. for i→row(Imd)
for j→column(Imd)
3. B= tf.keras.Sequential ([Resize(IMG_SIZE,IMG_SIZE)])
4. B= tf.image.rgb_to_grayscale(images,name=None)
5. denoise← Gaussian_blur
6. Data Augmentation←imageDataGenerator (Name, Value)
7. Initialize the parameters of CNN
41
 - dropoutfactor
- maxepoch
- batchsize
- learning rate
- activation function
- optimizer
8. Remove the last layer of CNN
9. Add SVM classifier
10. Set training and testing ratio
Train model← Train (Net, Training data)
Classification result← Train (Net, Training data, test data)
11. if (Classification result = true)
Classify type of TNDs
END
END

4.1.4 Result Analysis

The proposed model is performed on Google Collaboratory, intel core i5 8th Generation, 16GB
RAM, 1TB SSD, 8 GB NVIDIA GPU. The libraries used in this work are matplotlib, keras,
sklearn, sys, imageio, PIL and tensorflow. For better representation, public TDID and collected
datasets are re-named as Dset-1 and Dset-2. Further for comparing the proposed CNN-SVM
model, those research (Rs.) papers are considered that have used the public TDID dataset. Data
augmentation technique is used to increase the sample size of the datasets after applying the
various pre-processing steps. The initial public and collected sample size were 295 and 654
USG images. After data augmentation, initial and collected sample size are 1180 and 2616 USG
images. For both the datasets, 8:2 ratios are set for CNN-SVM model training and testing. Table
4.1 shows the various parameters set for the CNN-SVM hybrid model on Dset-1 & Dset-2.
Table 4.2 shows the comparison of the CNN-SVM hybrid model on Dset-1 and Dset-2 with the
existing models used for classifying TNDs. It can be analyzed that the CNN-SVM hybrid model
has performed well on both the datasets. The model has achieved an accuracy of 94.57%, F-
measure of 95.64%, sensitivity of 96.70% and specificity of 91.89% on Dset-1 and an accuracy
of 96%, F-measure of 98.33%, sensitivity of 97.80% and specificity of 93.93% on Dset-2.
Figure 4.8 shows the performance comparison of the accuracy parameter with existing models
and proposed model on Dset-1 and 2. It can be analyzed from the figure that there is an
improvement of 3% to 5% in the performance of the proposed model. Figure 4.9 shows
performance comparison of specificity, F-measure and sensitivity parameters with existing and
proposed model on Dset-1 and Dest-2. The proposed model has achieved sensitivity of 96.70%,

42
F-measure as 95.64% and specificity of 91.89% on Dset-1 and specificity of 93.93%, F-measure
of 98.33% and sensitivity of 97.80% on Dset-2. An improvement of 2% to 3% is reported by
proposed model in figure 4.9. Figure 4.10 shows the epoch-accuracy graph for Dset-1. Figure
4.11 shows the epoch-accuracy graph for Dset-2.

Table 4.1: Parameters set for training CNN-SVM hybrid model for Dset-1 and Dset-2
Parameter Batch Kernel Activation Stride Pooling Activation Optimizer Epoch Regularizer
Size Size layer l2
Values 32 3 Relu 2 Max- Linear Adam 100 0.01
pooling
layer

Table 4.2: The comparison of the CNN-SVM model on Dset-1 and Dset-2 with existing models
used for classifying TNDs
Models Accuracy F- Sensitivity Specificity
(%) measure (%) (%)
(%)
(Wang et al. 80.91 - 81.82 80
(2018))
(Nguyen et al. 90.88 - - -
(2019))
(Song et al. 92.1 - 94.1 96.2
(2019))
(Ajilisa et al. 89.93 89.43 92.76 -
(2020))
(Nguyen et al. 92.05 - - -
(2021))
(Proposed 94.57 95.64 96.70 91.89
CNN-SVM
hybrid
model on
Dset-1)
(Proposed 96 98.33 97.80 93.93
CNN-SVM
hybrid
model on
Dset-2)

43
 100
96
94.57
95 92.1 92.05
90.88 89.93
90
Percentage

85
80.91
80

75

70
Wang et al. Song et al. Nguyen et Ajilisa et al. Nguyen et Proposed Proposed
(2018) (2019) al. (2019) (2020) al. (2021) CNN-SVM CNN-SVM
hybrid hybrid
model on model on
dataset-1 dataset-2
Models

Figure 4.8: Performance comparison of the accuracy parameter with the existing and CNN-
SVM model on Dset-1 and Dset- 2

100 97.8 98.33

96.2 96.7 95.64
94.1 93.93
95 92.76 91.89
89.43
90
Percentage

85 81.82
80
80
75
70
65
60
Wang et al. (2018) Song et al. (2019) Ajilisa et al. (2020) Proposed CNN- Proposed CNN-
SVM hybrid model SVM hybrid model
on dataset-1 on dataset-2
Models

Sensitivity (%) Specificity (%) F-measure (%)

Figure 4.9: Performance comparison of specificity, f-measure and sensitivity parameters with
existing and CNN-SVM model on Dset-1 and Dset-2

44
 Epoch-accuracy graph for dataset-1
100

97

94

91
Accuracy(%)

88

85

82

79

76

73
10 20 30 40 50 60 70 80 90 100
Epoch

Figure 4.10: The epoch-accuracy graph for Dset-1

Epoch-accuracy graph for dataset-2

98
96
94
92
Accuracy(%)

90
88
86
84
82
80
78
10 20 30 40 50 60 70 80 90 100
Epoch

Figure 4.11: The epoch-accuracy graph for Dset-2

4.2 Summary
An effective CNN-SVM hybrid model is proposed for classification of BTND and MTND USG
images in this chapter. The work has been evaluated on 1180 and 2616 public TDID and
collected TNDs image datasets on Google Colaboratory. The proposed model is found to
perform well even when the sample size of the dataset is less. It has shown an improvement of

45
3% to 5% on Dset-1 and 4% to 6% on Dset-2 in comparison with existing models reported in
literature for the identification and classification of TNDs.

46
 CHAPTER 5

AN OPTIMIZED CONVOLUTIONAL NEURAL NETWORK

MODEL FOR THE IDENTIFICATION AND CLASSIFICATION
OF THYROID NODULE

In DL, hyperparameter (hyp) optimization is problem of choosing a set of optimal parameters

for a learning algorithm. It chooses the best parameter to control the learning process [118]. In
this chapter, Grid Search Optimization (GSO) based CNN model is proposed for the
identification and classification of TNDs. Several DL models like Alex-Net, Deep Neural
Network (DNN), Res-Net-50 and Visual Geometry Group (VGG-16) are explored along with
segmentation and boundary detection techniques. The proposed approach is tested on public
TDID and collected datasets using GSO technique. The result shows that proposed GSO-CNN
model outperforms from reported literature and improves the accuracy of the model.

5.1 Proposed Methodology

This section covers various phases involved for the working of the GSO-CNN model. The model
works in five phases: (1) data collection, (2) pre-processing, (3) morphological, segmentation
and boundary detection, (4) optimization and classification and (5) result analysis. Figure 5.1
shows the systematic flow of the proposed GSO-CNN model.

47
 Figure 5.1: Systematic flow of the proposed GSO-CNN model

5.1.1 Data Collection

Public TDID and collected datasets are considered for the evaluation of the model as discussed
in section 4.1.1

5.1.2 Pre-processing
The various pre-processing steps involved in this work are image resizing, noise removal, RGB
to grayscale conversion as discussed in section 3.2.2. The various data augmentation techniques
involved in this work are RandRotation, RandXTranslation, RandYTranslation,
RandXReflection, RandYReflection and Padding.

5.1.3 Morphological Operation, Segmentation and Boundary Detection

5.1.3.1 Morphological Operation
Morphological operation is applied on the image to extract out the exact region of TNDs with
using basic operation [119]. Figure 5.2 shows the morphological operation.

48
 Figure 5.2: The morphological operation [120]

In this work, dilation and erosion operation are considered. The structure used in this work is
“disk”. The dilation and erosion are computed using equation 5.24 and 5.25:
𝐼(𝑑𝑖𝑙𝑎𝑡𝑒)𝑆 = {𝐼 + 𝑆; 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑝𝑖𝑥𝑒𝑙𝑠 𝑖𝑛 𝐼 ∈ 𝑆} (5.24)
𝐼(𝑒𝑟𝑜𝑑𝑒𝑑)𝑆 = {𝐼(𝑒𝑟𝑟𝑜𝑟𝑒𝑑)𝑆}(𝑒𝑟𝑜𝑑𝑒𝑑)𝑆 (5.25)

The morphological gradient (G) function is computed using equation 5.26:

𝐺 = {𝐼(𝑑𝑖𝑙𝑎𝑡𝑒)𝑆 − {𝐼(𝑒𝑟𝑜𝑑𝑒𝑑)𝑆} (5.26)

5.1.3.2 Segmentation
It is defined as the process of partitioning image into image segments formerly known as image
objects [121]. It is mostly used to change the representation into more meaningful form for easy
analysis [122]. Active Contour (AC) segmentation technique was first coined by “Kass” [123].
It is basically the curve in the image space whose deformation is based on energy minimization.
Let x and y be the position of co-ordinates of an 2D image I(x,y), and curve be:

49
𝑉(𝑠) = (𝑥(𝑠), 𝑦(𝑠)) (5.27)
where s є [0,1]
The energy function (E) is defined using equation 5.28 and 5.29:
1
𝐸𝑡𝑜𝑡𝑎𝑙 = ∫ 𝐸(𝑉(𝑠))𝑑𝑠 (5.28)
0
1
𝐸𝑡𝑜𝑡𝑎𝑙 = ∫ [𝐸𝑖𝑛𝑡 (𝑉(𝑠)) + 𝐸𝑒𝑥𝑙 (𝑉(𝑠))] 𝑑𝑠 (5.29)
0

where Eint: internal & Eext: external energy

To control the deformation and displacement, Eint and Eext are converted into internal and
external forces.
The Eint and Eext energy component can be defined as:
1 1
𝐸𝑖𝑛𝑡(𝑉 (𝑠)) = [ 𝛼(𝑠 )|𝑉𝑠 (𝑠)|2 + 𝛽(𝑠)|𝑉𝑠𝑠(𝑠)|2] (5.30)
2 2

where α: weight factor, β(s): control factor

1 1
𝐸𝑖𝑛𝑡(𝑉 (𝑠)) = [ 𝛼(𝑠 )|𝑉𝑠 (𝑠)|2 + 𝛽(𝑠)|𝑉𝑠𝑠(𝑠)|2] (5.31)
2 2

where γ: weight factor, G_σ (x,y): 2D gaussian function, *:convolution operator, ∇: gradient
operator, σ: standard deviation
At equation 5.32, the Eint and Eext attain the minimum value. Hence goal is achieved.
𝐹𝑖𝑛𝑡+ 𝐹𝑒𝑥𝑡 = 0 (5.32)

5.1.3.3 Boundary Detection

It works by finding the boundaries between the light and dark pixels in an input image [124]. It
is computed using equation 5.33:
𝐵(𝐴) = 𝐴 − (𝐴 ⊝ 𝐵) (5.33)
where A: Matrix of an image M x N, B(A): Boundary detection, ⊝: Erosion function, B:
Structuring element

5.1.4 Optimization and Classification

Different variants of the CNN models are considered such as Res-Net-50, Alex-Net, VGG-16
and DNN for classification of TNDs. The working of CNN has been explained in section
4.1.3.1.
50
5.1.4.1 Res-Net-50
Res-Net-50 model is a residual deep learning network having 50 layers developed by He, Zhang
[125]. Res-Net-50 attempts to address the problem of vanishing gradient problem that occur
during back-propagation of CNN [126]. The network allows training of networks by
constructing through modules called as residual model. Figure 5.3 shows the structure of Res-
Net-50.

Figure 5.3: Structure of Res-Net-50[127]

5.1.4.2 Alex-Net
It was coined in 2012 and employed for generalized image classification for the first time [128].
In the Alex-Net, ReLU function is introduced to solve the problem of non-linearity and speed
up of the network [129]. Alex-Net consists of five convolutional(conv.) layer, three fully
connected layers and one output layer. There is a Local Response Normalization (LRN) layer
that follows the first as well as the second convolutional layer [130]. Figure 5.4 shows the
structure of Alex-Net.

51
 Figure 5.4: Structure of Alex-Net [131]

5.1.4.3 VGG-16
The idea was coined from the “University of Oxford” and trained on ImageNet dataset [132].
A traditional VGG-16 consists of multiple “3X3 kernel-sized filters” that help the model in
learning complex features by increasing depth of the network [133]. Figure 5.5 shows the
architecture of VGG-16.

Figure 5.5: Structure of VGG-16 [134]

52
5.1.4.4 DNN
It is the improved versions of ANN with multiple layers [135]. The DNN is highly preferred
over traditional ML algorithms like SVM, KNN and DT. These include extracting high level
features thereby learning more complex patterns [136-137]. Due to the mentioned reasons,
DNN is highly preferred image processing, computer vision, machine translation and Natural
Language Processing (NLP) [138]. Figure 5.6 shows the structure of DNN.

Figure 5.6: Structure of DNN [139]

5.1.4.5 Optimizing Using GSO Technique

Optimization of drop-out and learning rate factor using GSO technique is discussed in this
section. Let the hyperparameters be hyp1,……..hypn of the DL model and α1,…..αn be domains.
The model is trained with hyp (DLtrain) on the training of TNDs images. The main aim behind
hyp is to find the best parameter hyp1*, hyp2*. Figure 5.7 shows the proposed GSO-CNN model.

53
 Figure 5.7: Proposed GSO-CNN model

5.1 Algorithm for optimization of CNN model using GSO technique

Algorithm 5.1: Optimization of CNN model using GSO technique

Input: hyp1*
Output: hyp1*, hyp2*
create deeplmodel ()
ml.add(hyp1*, values_of_drop-out _actor, optimization_algo)
gridsearch= GriddSearchCV(par grid=par, grid,est=ml)
result=gridsearchd.fit( )
Print (“best_drop-out_factor, best_learnin_rate”);
Return hyp1 *, hyp2 *

5.2 Proposed GSO-CNN Algorithm for the identification and classification of TNDs

Algorithm: Classification using GSO-CNN model

Input: TNDs USG images
Output: Type of TNDs
BEGIN
1. Upload sample images of TNDs.
2. Use pre-processing step for the uniformity of the TNDs images.
3. Use active contour segmentation technique.
4. Apply B(A) to find boundary of the TNDs.
5. Set the training and testing ratio of the model.
6. Initialize the parameters of CNN
-dropoutfactor
- maxepoch
- batchsize
- learning rate
- activation function
- optimizer
7. Use gd_technique to find the best hyp1 * and hyp2 *.
54
 8. Train the model
Classify type of TNDs
END

5.1.5 Result Analysis

This section focuses on the various results obtained for the GSO-CNN model. In this work,
datasets are renamed as follows: Dset-1: public TDID and Dset-2: collected dataset. Initially,
the sample size of public and collected datasets were 295 and 654. After considering six
different augmentation techniques (RandRotation, RandXTranslation, RandYTranslation,
RandXReflection, RandYReflection and Padding), the new sample size of the datasets are 1770
and 3924 USG images. 8:2 ratio is set for GSO-CNN model for training and testing purpose.
Table 5.1 shows the various parameters setting for GSO-CNN model for Dset -1 and Dset-2.
Table 5.2 shows the comparison of the proposed GSO-CNN model with Alex-Net, Res-Net-
50, VGG-16 and DNN based on with (CASE-1) and without segmentation and boundary
detection techniques (CASE-2) on Dset -1 and Dset-2. From the table, it can be observed that
the results obtained in CASE-2 are better than CASE-1. The proposed GSO-CNN, VGG-16,
Alex-Net, and Res-Net-50 models have shown an improvement of 1% to 2% in CASE-2 with
respect to performance parameters. Table 5.3 shows results of GSO obtained on dropout factor
and learning rate parameters on proposed GSO-CNN, Alex-Net, Res-Net-50, VGG-16 and
DNN models. The best results are obtained on 0.01 learning rate and 0.2 dropout factor. The
best results obtained by the models on Dset -1 are 95.30%, 91.01%, 82.84%, 86.01%, 86.66%
for GSO-CNN, DNN, Alex-Net, Res-Net-50 and VGG-16 models respectively. The models
achieved 96.02%, 92.30%, 83.08%, 91.76% and 87.16% for GSO-CNN, DNN, Alex-Net, Res-
Net-50 and VGG-16 models respectively on Dset-2. Table 5.4 shows the comparison of the
GSO-CNN model and existing models on Dset -1 and Dset-2. From the table, it can be
analyzed that the proposed GSO-CNN model has achieved better results in comparison with
the existing models. The proposed GSO-CNN model achieved 95.30%, 96.66%, 97.20%,
94.87% on Dset -1 and 96.02%, 96.70%, 98.34%, 95% on Dset -2 on accuracy, sensitivity, F-
measure and specificity parameters. Figure 5.8 shows the comparative analysis of the accuracy
parameter with the models on Dset -1 and Dset-2. The results obtained on accuracy parameter
of the GSO-CNN model have shown an increment of 3% to 4% on Dset -1 and Dest-2. The
model has achieved an accuracy of 95.30% on Dset -1 and 96.02% on Dset -2. Figure 5.9
shows the comparative analysis of sensitivity, specificity and F-measure parameters with
55
existing models on Dset -1 and Dset-2. The model has achieved sensitivity of 96.66% and
96.70% on Dset -1 and Dset-2 & F-measure of 97.20% and 98.34% on Dset -1 and Dset-2.
The GSO-CNN model has shown rise of 4 to 5% on sensitivity and 8% to 9% on F-measure
parameters. However, the model has not achieved better results on specificity parameter. The
model has achieved specificity of 94.87% on Dset -1 and 95% on Dset -2 which is lower than
the results obtained by Song et al. (2019) with 96.2% and higher with Wang et al. (2018) model
with 80%. Figure 5.10 and figure 5. 11 show the accuracy-epoch graph for Dset -1 and Dset-
2 respectively.

Table 5.1: Parameter setting for GSO-CNN model for Dset -1 and Dset- 2
Parameters Value
Batch size 15
Activation function Softmax
Max-pooling 2X2
Drop-out factor 0.2
Max-epoch 10
Learning rate 0.01
Optimizer SGDM

Table 5.2: Performance comparison of the GSO-CNN model with Alex-Net, Res-Net-50,
VGG-16 and DNN on Dset -1 and Dset-2
Dset Techniques Models Accuracy Specificity F-measure Sensitivity
(%) (%) (%) (%)
GSO-CNN
Without segmentation and boundary

model 94.11 93.33 96.20 95.65

DNN 90.90 89.85 89 92.63
Dset -1

Alex-Net 82.98 80.98 83.96 83.96

detection techniques

Res-Net-50 86.48 84.52 87.25 87.25

VGG-16 85.71 83.90 86.40 86.40
CASE 1

GSO-CNN
model 95.33 95 97.20 96.66
Dset -2

DNN 91.81 90.78 92.14 92.63

Alex-Net 82.28 81.92 82.60 83.51
Res-Net-50 90.11 89.33 93.61 91.66
VGG-16 86.63 84.88 87.25 87.25

56
 GSO-CNN

With segmentation and boundary

model 95.30 94.87 97.20 96.66
Dset -1 DNN 91.01 90.27 93.54 92.55
Alex-Net 83.85 81.81 84.76 84.76

detection techniques
Res-Net-50 87.91 86.58 88.55 89
CASE 2 VGG-16 86.66 83.95 87.56 87.12
GSO-CNN
model 96.02 95 98.34 97.02
DNN 92.30 90.16 93.19 93.68
Dset -2

Alex-Net 83.08 82.97 85.98 84.76

Res-Net-50 91.76 90.66 92.14 92.63
VGG-16 87.16 86.04 87.68 88.11

Table 5.3: Results of GSO technique obtained on dropout factor and learning rate parameters
obtained on GSO-CNN model, Alex-Net, Res-Net-50, VGG-16 and DNN

Learning 0.01 0.04 0.07 0.09 0.1 0.05 0.08

meters
Para-

rate
Dropout 0.2 0.02 0.06 0.09 0.1 0.04 0.08
factor
(GSO-
CNN) 95.30 80.90 85.16 89 76.06 82.32 87.19
Model
DNN 91.01 69 76.88 82.76 62.3 71.1 81.1
Accuracy

Models
Dset-1

Alex-Net 82.84 59.99 68.3 77.9 58 64 73.6

Res-Net- 86.01 62.7 64 78.41 54 69.6 74.29
50
VGG-16 86.66 68.2 78.55 81.27 59 73.3 80.3
(GSO-
CNN) 96.02 84.96 88.86 92.70 77.39 86.40 90.81
Model
DNN 92.30 65.2 79.27 88.9 59.5 72.2 85.1
Accuracy

Models
Dset -2

Alex-Net 83.08 64.44 73.23 78.9 57.11 69.9 75

Res-Net- 91.76 76 86 91 72.6 81.49 88.1
50
VGG-16 87.16 64.7 75 81.9 60 71.34 79.81

57
Table 5.4: Comparison of the proposed GSO-CNN model and existing models on Dset -1 and
Dset-2
Models F-measure Accuracy Specificity Sensitivity
(%) (%) (%) (%)
(Wang et. al (2018)) - 80.91 80 81.82

(Nguyen et. al (2019)) - 90.88 - -

(Song et. al (2019)) - 92.1 96.2 94.1

(Ajilisa et. al (2020)) 89.43 89.93 - 92.76

(Nguyen et. al (2021)) - 92.05 - -

Proposed (GSO-CNN) 97.20 95.30 94.87 96.66

model on Dset -1

Proposed (GSO-CNN) 98.34 96.02 95 96.70

model on Dset -2

100

95.3 96.02
95
92.1 92.05
90.88
89.93
90
Accuracy(%)

85
80.91
80

75

70
Wang et. al Song et. al Nguyen et. Ajilisa et. al Nguyen et. Proposed Proposed
(2018) (2019) al (2019) (2020) al (2020) model on model on
dataset-1 dataset-2
Models

Figure 5.8: Comparative analysis of the accuracy parameter with models on Dset -1 and Dset-
2

58
 120

96.7 95 98.34
100 94.1 92.76 92.53
96.2 89.43 94.62 94.09
81.82
80 80
Percentage

60

40

20

0
Wang et. al (2018) Song et. al (2019) Ajilisa et. al (2020) Proposed model onProposed model on
dataset-1 dataset-2
Models

Sensitivity Specificity F-measure

Figure 5.9: Comparative analysis of models based on sensitivity, f-measure and specificity on
Dset -1 and Dset-2

100 95.3
86.43
90
81
78.12
80 73.82
69.23
Accuracy (%)

70 63.59
55.62
60

50 43.73

40
3 0.75
30
1 2 3 4 5 6 7 8 9 10
Epoch

Figure 5.10: Accuracy-epoch graph for Dset -1

59
 100 92.53 96.02
95
86.37
90
81
85
76.45
80
71.33
75
Accuracy(%) 65.69
70
65
55.67
60
55
45.67
50
45
40
0.33
353
30
1 2 3 4 5 6 7 8 9 10
Epoch

Figure 5.11: Accuracy-epoch graph for Dset -2

5.2 Summary

GSO-CNN model is proposed for the identification and classification of TNDs in this chapter.
The model is compared with Res-Net-50, DNN, VGG-16, Alex-Net and various reported
models. It is concluded from the results that the proposed GSO-CNN outperforms when
compared with and without segmentation and boundary detection techniques. The GSO
technique is used to optimize the drop-out factor and learning rate parameters to find the best
results.

60
 CHAPTER 6

AN IMPROVED DEEP-GAN MODEL FOR

IDENTIFICATION AND CLASSIFICATION OF
THYROID NODULE

Health care is conventionally regarded as essential determinant in promoting the general mental,
physical, and social well-being of people around the world [140]. A reliable system forTNDs is
necessary for early identification and classification of BTND and MTND to save effort, time
and human life [141-145]. In this chapter, Deep- Generative Adversarial Network (Deep- GAN)
based model is proposed for the early identification and classification of TNDs. GSO technique
is used for tunning of the Deep-GAN (i.e., Alex-GAN and VGG-GAN) model by optimizing
the optimizer, batch size and learning rate parameters.

6.1 Proposed Methodology

This section presents the proposed methodology adopted for Deep-GAN model. It works in four
phases namely: (1) data collection and pre-processing, (2) data augmentation, (3) classification
using Deep-GAN model and (4) result analysis. Figure 6.1 shows the systematic flow of the
proposed Deep-GAN model.

61
 Figure 6.1: The systematic flow of Deep-GAN model

6.1.1 Data Collection and Pre-processing

This phase has been discussed in section 4.1.1

6.1.2 Data Augmentation

Data Augmentation (DA) plays an important role in medical image analysis for synthetically
increasing the sample size by generating new data points from existing data. [146]. Here, GAN
based DA technique is used to increase the sample size.[147]. In 2014, Ian Goodfellow
introduced the adversial process to learn Generative Models (GM) [148]. The core idea of
GANs is inspired by a two-player zero-sum minimax game between the Discriminator (D) and
the Generator (G) [149]. Here, there are two players namely (G) and (D) who participates in
this game and tries to win by defeating the other and vice-versa. The D tries to determine
whether the sample belongs to fake or real distribution whereas G aims to deceive D by
generating fake sample distribution. The D tries to produce probability that the sample is real
sample. The value close to zero interpret that the sample is fake sample whereas the probability
62
(P) near to 0.5 indicates D is unable to differentiate fake or real sample [150]. Figure 6.2 shows
the architecture of GAN.

Figure 6.2 : Architecture of GAN[151]

6.1.3 Classification Using Deep-GAN

In this chapter, VGG-16 and Alex-Net model are considered for classification of TNDs. Some
of the popular reasons for considering Alex-Net and VGG-16 are: (i) ReLU function is used in
pre trained architectures as it has a proven record of increasing the training speed of the models
in literature [152], (ii) Such pertained architectures reduce the over fitting problem by using
drop-out factor, (iii) Alex-Net allows multi-GPU training, (iv) Alex-Net shows good accuracy
on different proposed models despite having 8 layers[153], (v) VGG-16 network introduced the
concept of grouping multiple convolutional layers with smaller kernel size to increases the
performance of the model [154], and (vi) VGG-16 is a deeper model having 16 layers with
increased capability to extract features from images[155]. Its architecture explanation is
discussed in chapter 5 under subsection 5.1.4.2 and 5.1.4.3. Figure 6.3 shows the proposed
Alex-GAN model and Figure 6.4 shows the proposed VGG-GAN model.

63
 Figure 6.3: Proposed Alex-GAN model

Figure 6.4: Proposed VGG-GAN model

6.1Algorithm for classification of TNDs using Deep-GAN

Algorithm 6.1: Classification of TNDs using Deep-GAN
Input: Image dataset
Output: Type of TNDs
BEGIN
1. Upload dataset.
2. Perform pre-processing steps.

64
 3. Use GAN for data augmentation
4. Divide train and test part of TNDs images
5. Initialize the parameters of Deep-GAN model
-dropoutfactor
- maxepoch
- batchsize
- learning rate
- activation function
- optimizer
6. Use GSO technique for optimization of Deep-GAN.
7. Train the model
Classify type of TNDs
END

6.1.4 Result Analysis

The section will focus on the results of the Deep-GAN model. GAN model is used to increase
the sample size of the datasets for building and evaluating the Deep-GAN model. The public
and collected datasets are divided into 8:2 ratio. The public dataset and collected datasets are
renamed as Dset-1 and Dset-2. Table 6.1 shows the parameter settings of Deep-GAN on Dset-
1 and Dset-2. Table 6.2 shows the grid search results over different values of hyperparameters
set for the proposed Alex-GAN and VGG-GAN models. The hyper tuning of the following
parameters like optimizer, batch size and learning rate is optimized using GSO technique. It can
be concluded from table 6.2 that the proposed models perform best on Adam optimizer with
learning rate of 0.001 and batch size of 35 on both datasets. Table 6.3 shows comparative
performance comparison of VGG-16, Alex-Net, VGG-GAN and Alex-GAN models.
Experiments are performed on Alex-Net and VGG-16 architectures CASE-1: without data
augmentation and CASE 2: with GAN as augmentation i.e., Alex-GAN and VGG-GAN. Table
6.3 shows the results of VGG-16, Alex-Net, VGG-GAN and Alex-GAN models. The result
obtained in CASE-1 shows that the obtained results were not good. Hence, CASE-2 is carried
out with GAN to achieve better results. In CASE-1, Alex-Net has achieved (96%, 94.91%,
96.70%, 96.65%) and VGG-16 has achieved (93.60%, 90.32%,95.18%, 96.73%) in terms of
accuracy, specificity, sensitivity and F-measure parameters. While in CASE-2, Alex-GAN
model has achieved (97.03%, 95.83%, 97.70%, 98.26%) and VGG-GAN achieved (94.28%,
96.73%,90.32%, 96.7%) on both datasets in terms of considered parameters. Table 6.4 shows
the performance comparison of the Deep- GAN model (i.e., Alex-GAN and VGG-GAN) with
reported models used for classification of TNDs. The Alex-GAN model has shown performance

65
 improvement of 2% to 4% with VGG-GAN and existing models. Figure 6.5 shows the generator
network and figure 6.6 shows the discriminator network. Figure 6.7(a-b) shows the training of
Alex-Net architecture whereas figure 6.8(a-d) shows the training of VGG-16 architecture.
Figure 6.9 and 6.10 show the accuracy epoch graph of Alex-GAN model on Dset-1 and Dset-
2. Figure 6.11 and 6.12 show the accuracy epoch graph of VGG-GAN model on Dset-1 and
Dset-2.

Table 6.1: Parameter setting of Deep-GAN model on Dset-1 and Dset-2

Parameter Epoch Optimizer Learning Batch size Stride Pooling layer Kernel size
rate
Values 80 Adam 0.001 35 2 Max-pooling 3

Table 6.2: The grid search results over different values of hyperparameters set for Alex-GAN
and VGG-GAN models
Results

Parameters Alex-GAN VGG-GAN

Dset-1 Dset-2 Dset-1 Dset-2
Sr. Epochs Batch Optimizer Learning Accuracy Accuracy Accuracy Accuracy
No size rate
1 80 55 SGD 0.1 79.64 75.9 70.1 72.9
2 80 55 Adam 0.1 80.1 77.7 74.89 78.23
3 80 45 SGD 0.01 84.7 83.55 85.27 81.46
4 80 45 Adam 0.01 87.89 88.6 86 87.34
5 80 35 SGD 0.001 93.6 94.89 91.88 92.12
6 80 35 Adam 0.001 96 97.03 93.60 94.28
7 80 25 SGD 0.0001 91.3 92.6 92 93.4
8 80 25 Adam 0.0001 92.5 93.22 90.43 92
9 80 15 SGD 0.00001 86.20 88.89 84.5 85.7
10 80 15 Adam 0.00001 87.45 91.12 88.99 82

Table 6.3: Comparative performance comparison of VGG-16, Alex-Net, VGG-GAN and

Alex-GAN models
Models Dsets Accuracy Sensitivity Specificity F-measure
(%) (%) (%) (%)
Alex-Net without Dset-1 92.16 93.68 91.42 95.18
data augmentation Dset-2 94.89 95.69 91.11 95.69
technique
(CASE-1)
VGG-16 without Dset-1 92.40 94.68 91.35 93.19
data augmentation
technique Dset-2 93.37 94.62 91.37 94.11
(CASE-1)
66
 Alex-GAN (CASE- Dset-1 96 96.70 94.91 95.65
2) Dset-2 97.03 97.70 95.83 98.26
VGG-GAN Dset-1 93.60 96.73 90.32 95.18
(CASE-2)
Dset-2 94.28 95.98 91.48 96.7

Table 6.4: Performance comparison of the Deep- GAN model (i.e., Alex-GAN and VGG-GAN) with
existing models for classification of TNDs
Models F-measure Accuracy Sensitivity Specificity
(%) (%) (%) (%)
(Wang et al. (2018)) - 80.91 81.82 80
(Nguyen et al. (2019)) - 90.88 - -
(Song et al. (2019)) - 92.1 94.1 96.2
(Ajilisa et al. (2020)) 89.43 89.93 92.76 -
(Hang et al. (2021)) - 95 - -
(Nguyen et al. (2021)) - 92.05 - -
Proposed model 95.65 96 96.70 94.91
(Alex-GAN) on
Dset-1
Proposed model 98.26 97.03 97.70 95.83
(Alex-GAN) on
Dset-2
Proposed model 95.18 93.60 96.73 90.32
(VGG-GAN) on
Dset-1
Proposed model 96.7 94.28 95.98 91.48
(VGG-GAN) on
Dset-2

67
Figure 6.5: Generator network Figure 6.6: Discriminator network

Figure 6.7a: part-1 Figure 6.7b:part-2

Figure 6.7(a-b): Training of Alex-Net architecture

68
Figure 6.8a: Generator network Figure 6.8b: Discriminator network

Figure 6.8c: part-1 Figure 6.8d: part-2

Figure 6.8(a-d): Training for VGG-16 architecture

69
 Accuracy-epoch graph for dataset-1
100
95

Accuracy(%) 90
85
80
75
70
10 20 30 40 50 60 70 80
Epoch

Figure 6.9: The accuracy-epoch graph of Alex-GAN model on Dset-1

Accuracy-epoch graph for dataset-2

100

95
Accuracy(%)

90

85

80

75
10 20 30 40 50 60 70 80
Epoch

Figure 6.10: The accuracy-epoch graph of Alex-GAN model on Dset-2

Accuracy-epoch graph for dataset-1

100
95
90
Accuracy(%)

85
80
75
70
65
10 20 30 40 50 60 70 80
Epoch

Figure 6.11: The accuracy-epoch graph of VGG-GAN model on Dset-1

70
 Accuracy-epoch graph for dataset-2
100
95
90
Accuracy(%)

85
80
75
70
65
60
10 20 30 40 50 60 70 80
Epoch

Figure 6.12: The accuracy-epoch graph of VGG-GAN model on Dset-2

6.2 Summary
The proposed Deep-GAN model has achieved promising results for the early identification and
classification of TNDs. GSO technique is used for tunning of the Deep-GAN (i.e. Alex-GAN
and VGG-GAN) model. The experimental results demonstrate that the proposed Alex-GAN
performs better in comparison with VGG-GAN and other DL models reported in literature with
an improvement of 2% to 4%. The developed models have been evaluated on TDID and
findings are generalized by evaluating such models on collected dataset. The proposed model
will help the radiologists for early identification and classification of TNDs in USG images.

71
 CHAPTER 7

CONCLUSION AND FUTURE WORK

This chapter concludes research work written in the thesis. The major research contributions and
future research directions are discussed as follows.

7.1 Major Research Contributions

The following are the major contributions of this research work.

▪ The proposed model has been evaluated on public TDID and collected dataset taken from
Kriti Scanning Center, Prayagraj, Uttar Pradesh, India duly approved by NABH and
Healthcare Providers. The duration of dataset collection was from July 2020 to March
2021.The goal behind using collected dataset is to generalize the results of the proposed
models for the identification and classification of benign and malignant TNDs.
▪ Two hybrid models are proposed: (1) hybridization of two ML classifiers i.e., ANN-SVM
hybrid model and (2) hybridization of DL and ML classifier i.e., CNN-SVM hybrid model.
The experimental results for these models show that hybridization of DL and ML classifier
outperforms in comparison with hybridization of two ML classifiers. The findings
demonstrate that the proposed CNN-SVM hybrid model performs better as compared to the
standalone classifier.
▪ A GSO-CNN model is proposed for the classification of TNDs in medical USG images. The
variants of CNN models like Res-Net-50, DNN, Alex-Net and VGG-16 are explored for the
classification of TNDs. The proposed model is optimized using GSO technique in terms of
hyperparameters like optimizer, batch size and learning rate. The proposed GSO-CNN model
experimental results are compared with Res-Net, DNN, Alex-Net, VGG-16 and other state-
of-the-art models. The results shows that the proposed GSO-CNN model is competitive in
comparison with the reported models in literature.
▪ Deep-GAN model is proposed to improve the accuracy for the identification and classification
of TNDs. GSO technique is used for tunning of the Deep-GAN (i.e. Alex-GANand VGG-
GAN) model by optimizing the optimizer, learning rate and batch size hyperparameters. The

experimental results demonstrate that the proposed Alex-GAN performs better in

72
 comparison with proposed VGG-GAN and other reported DL models in literature with an
improvement of 2% to 4 %.
▪ The proposed (i) CNN-SVM hybrid model, (ii) GSO-CNN model and (iii) Deep -GAN
models are evaluated in terms of accuracy, F-measure, specificity and sensitivity parameters
on the public and collected datasets. Based on the performance evaluation of these proposed
model among themselves, Deep-GAN model is ranked 1, GSO-CNN model as ranked 2 and
CNN-SVM model is ranked 3 for the identification and classification of TNDs.

The proposed models can be used by the clinicians, doctors, researchers and practitioners for the
diagnostic purpose. The proposed models will act as an early alarm for patients having TNDs.

7.2 Future Work

• Internet of Things (IOT) enabled healthcare application can be developed using DL
techniques.
• Different types of optimization techniques can be explored to improve accuracy of model.
• A mobile based application can be designed for better feasibility to the doctors.
• Other kinds of diseases related to lung, kidneys etc. can be considered to further verify the
universality of the proposed approaches.

73
 REFERENCES

[1] Das, D., Banerjee, A., Jena, A. B., Duttaroy, A. K., & Pathak, S., “Essentiality, relevance,
and efficacy of adjuvant/combinational therapy in the management of thyroid
dysfunctions”, Biomedicine & Pharmacotherapy, vol. 146, 2022.
[2] https://www.thyroid.org/professionals/ata-professional-guidelines
[3] Nilsson, Mikael, and Henrik Fagman, “Development of the thyroid gland”, Development,
vol. 144, pp. 2123-2140, 2017.
[4] Zhang, X., Malik, B., Young, C., Zhang, H., Larkin, D., Liao, X. H., & Arvan , P.,
“Maintaining the thyroid gland in mutant thyroglobulin-induced hypothyroidism requires
thyroid cell proliferation that must continue in adulthood”, Journal of Biological
Chemistry, 2022.
[5] Darras, Veerle M., "Thyroid gland", In Sturkie's Avian Physiology, pp. 795-812, Academic
Press, 2022.
[6] Sadeghi, D., Shoeibi, A., Ghassemi, N., Moridian, P., Khadem, A., Alizadehsani, R., ... &
Acharya, U. R, “An overview of artificial intelligence techniques for diagnosis of
Schizophrenia based on magnetic resonance imaging modalities: Methods, challenges, and
future works”, Computers in Biology and Medicine, 2022.
[7] Abbasian Ardakani, A., Reiazi, R., & Mohammadi, A., “A Clinical Decision Support
System Using Ultrasound Textures and Radiologic Features to Distinguish Metastasis
from Tumor‐Free Cervical Lymph Nodes in Patients with Papillary Thyroid Carcinoma”,
Journal of Ultrasound in Medicine, vol. 37(11), pp. 2527-2535, 2018.

[8] Hegedus, L., Bianco, A. C., Jonklaas, J., Pearce, S. H., Weetman, A. P., & Perros, P.,
“Primary hypothyroidism and quality of life", Nature Reviews Endocrinology, vol. 18(4),
pp. 230-242, 2022.

[9] Demet, M. M., Ozmen, B., Deveci, A., Boyvada, S., Adiguzel, H., & Aydemir, O,
“Depression and anxiety in hyperthyroidism”, Archives of medical research, vol. 33(6),
pp. 552-556, 2002.

[10] Kwong, N., Medici, M., Angell, T. E., Liu, X., Marqusee, E., Cibas, E. S., ... & Alexander,
E. K., “The influence of patient age on TND formation, multinodularity, and thyroid cancer
risk”, The Journal of Clinical Endocrinology & Metabolism, vol. 100(12), pp. 4434-4440,
2015.
74
[11] Richman, D. M., Benson, C. B., Doubilet, P. M., Wassner, A. J., Asch, E., Cherella, C. E.,
& Frates, M. C, “Assessment of American college of radiology thyroid imaging reporting
and data system (TI-RADS) for pediatric thyroid nodules”, Radiology, vol. 294(2), pp.
415-420.
[12] Ker, J., Wang, L., Rao, J., & Lim, T, “Deep learning applications in medical image
analysis”, IEEE Access, vol. 6, pp. 9375-9389, 2017.
[13] Yap, P. T., Zhang, Y., & Shen, D., “Multi-Tissue Decomposition of Diffusion MRI Signals
via $\ell _ {0} $ Sparse-Group Estimation”, IEEE Transactions on Image Processing, vol.
25(9), pp. 4340-4353, 2016.
[14] Schmidhuber, J, “Deep learning in neural networks: An overview”, Neural networks, vol.
61, pp. 85-117, 2015.
[15] Bengio, Y., “Learning deep architectures for AI”, Now Publishers Inc., 2009.
[16] Hinton, G. E., & Salakhutdinov, R. R, “Reducing the dimensionality of data with neural
networks”, Science, vol. 313(5786), pp. 504-507, 2006.
[17] Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., Manzagol, P. A., & Bottou, L, “Stacked
denoising autoencoders: Learning useful representations in a deep network with a local
denoising criterion”, Journal of machine learning research, 2010.
[18] Nair, V., & Hinton, G. E, “Rectified linear units improve restricted boltzmann machines”,
In Icml, 2010.
[19] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R., “Dropout:
a simple way to prevent neural networks from overfitting”, The journal of machine
learning research, vol. 15(1), pp. 1929-1958, 2014.
[20] Ioffe, S., & Szegedy, C, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift”, In International conference on machine learning, pp.
448-456, 2015.
[21] Pedraza, L., Vargas, C., Narváez, F., Duran, O., Munoz, E., & Romero, E., “An open
access thyroid ultrasound image database”, In 10th International Symposium on Medical
Information Processing and Analysis, vol. 9287, pp.188-193, SPIE, 2015.
[22] Quality Council of India, Accessed: 07- Sep. 2022, [Online], Available:
https://www.nabh.co/frmViewCGHSRecommend.aspx?Type=Diagnostic%20Centre&cit
yID=94
[23] Guvenir, H. A., Acar, B., Demiroz, G., & Cekin, A, “A supervised machine learning
algorithm for arrhythmia analysis”, In Computers in Cardiology, pp. 433-436, IEEE, 1997.
75
[24] Gupta, V., Sachdeva, S., & Dohare, N, “Deep similarity learning for disease prediction”,
Trends in Deep Learning Methodologies, pp.183-206, 2021.
[25] Zhu, L. C., Ye, Y. L., Luo, W. H., Su, M., Wei, H. P., Zhang, X. B., & Zou, C. L, “A
model to discriminate malignant from benign thyroid nodules using artificial neural
network”, PLoS One, vol. 8(12), 2013.
[26] Narayan, N. S., Marziliano, P., Kanagalingam, J., & Hobbs, C. G., “Speckle patch
similarity for echogenicity-based multiorgan segmentation in ultrasound images of the
thyroid gland”, IEEE journal of biomedical and health informatics, vol. 21(1), pp.172-
183, 2015.
[27] Song, G., Xue, F., & Zhang, C., “A model using texture features to differentiate the nature
of thyroid nodules on sonography”, Journal of Ultrasound in Medicine, vol. 34(10), pp.
1753-1760, 2015.
[28] Ardakani, A. A., Gharbali, A., & Mohammadi, A., “Application of texture analysis method
for classification of benign and malignant thyroid nodules in ultrasound images”, Iranian
journal of cancer prevention, vol. 8(2), 2015.
[29] Nugroho, H. A., Nugroho, A., & Choridah, L., “Thyroid nodule segmentation using active
contour bilateral filtering on ultrasound images”, In 2015 International Conference on
Quality in Research (QiR), pp. 43-46, 2015.
[30] Song, G., Xue, F., & Zhang, C, “A model using texture features to differentiate the nature
of thyroid nodules on sonography”, Journal of Ultrasound in Medicine, vol. 34(10), pp.
1753-1760, 2015.
[31] Pan, Q., Zhang, Y., Zuo, M., Xiang, L., & Chen, D., “Improved ensemble classification
method of thyroid disease based on random forest”, In 2016 8th International Conference
on Information Technology in Medicine and Education (ITME), pp. 567-571, 2016.
[32] Alrubaidi, W. M., Peng, B., Yang, Y., & Chen, Q., “An interactive segmentation algorithm
for thyroid nodules in ultrasound images”, In International Conference on Intelligent
Computing, pp. 107-115, Springer, Cham, 2016.
[33] Vanithamani, R., & Dhivya, R., “Thyroid Nodule Classification in Medical Ultrasound
Images”, In International Conference on Soft Computing and Pattern Recognition, pp.
509-514, Springer, Cham, 2016.
[34] Nugroho, H. A., Rahmawaty, M., Triyani, Y., & Ardiyanto, I., “Texture analysis for
classification of thyroid ultrasound images”, In 2016 International Electronics Symposium
(IES), pp. 476-480, 2016.
76
[35] Mustafa, N., Memon, R. A., Li, J. P., & Omer, M. Z., “A classification model for
imbalanced medical data based on PCA and farther distance based synthetic minority
oversampling technique”, Int. J. Adv. Comput. Sci. Appl, vol.8(1), pp.61-67, 2017.
[36] Pasha, S. J., & Mohamed, E. S., “Ensemble gain ratio feature selection (EGFS) model with
machine learning and data mining algorithms for disease risk prediction”, In 2020
International Conference on Inventive Computation Technologies (ICICT), pp. 590-596,
2020.
[37] Jothi, J. A. A., & Rajam, V. M. A., “Automatic classification of thyroid histopathology
images using multi-classifier system”, Multimedia Tools and Applications, vol. 76(18), pp.
18711-18730, 2017.
[38] Jiang, Y., Deng, Z., Chen, J., Wu, H., Choi, K. S., & Wang, S, “Intelligent diagnostic
methods for thyroid nodules”, Journal of Medical Imaging and Health Informatics, vol.
7(8), pp. 1772-1779, 2017.
[39] Shankar, K., Lakshmanaprabu, S. K., Gupta, D., Maseleno, A., & De Albuquerque, V. H.
C., “Optimal feature-based multi-kernel SVM approach for thyroid disease classification”,
The Journal of Supercomputing, vol. 76(2), pp. 1128-1143, 2020.
[40] Ma, X., Xi, B., Zhang, Y., Zhu, L., Sui, X., Tian, G., & Yang, J., “A Machine Learning-
based Diagnosis of Thyroid Cancer Using Thyroid Nodules Ultrasound Images.”, Current
Bioinformatics, vol.15(4), pp. 349-358, 2020.
[41] Zheng, X., Lv, G., Du, G., Zhai, Z., Mo, J., & Lv, X., “Rapid and low-cost detection of
thyroid dysfunction using Raman spectroscopy and an improved support vector machine”,
IEEE Photonics Journal, vol. 10(6), pp. 1-12, 2018.
[42] Dandan, L., Yakui, Z., Linyao, D., Xianli, Z., & Yi, S., “Texture analysis and classification
of diffuse thyroid diseases based on ultrasound images”, In 2018 IEEE International
Instrumentation and Measurement Technology Conference (I2MTC), pp. 1-6, 2018.
[43] Poudel, P., Illanes, A., Ataide, E. J., Esmaeili, N., Balakrishnan, S., & Friebe, M., “Thyroid
ultrasound texture classification using autoregressive features in conjunction with machine
learning approaches”, IEEE Access, vol. 7, pp. 79354-79365, 2019.
[44] Prochazka, A., Gulati, S., Holinka, S., & Smutek, D., “Classification of thyroid nodules in
ultrasound images using direction-independent features extracted by two-threshold binary
decomposition”, Technology in cancer research & treatment, vol. 18, 2019.
[45] Wu, Y., & Liu, P., “A Classification Algorithm of Ultrasonic Thyroid Standard Planes
Using LBP and HOG Features”, In 2019 IEEE 13th International Conference on Anti-
77
 counterfeiting, Security, and Identification (ASID), pp. 103-107, 2019.
[46] Colakoglu, B., Alis, D., & Yergin, M., “Diagnostic value of machine learning-based
quantitative texture analysis in differentiating benign and malignant thyroid nodules”,
Journal of oncology, 2019.
[47] Zhang, B., Tian, J., Pei, S., Chen, Y., He, X., Dong, Y., & Zhang, S., “Machine learning–
assisted system for thyroid nodule diagnosis”, Thyroid, vol. 29(6), pp. 858-867, 2019.
[48] Kesarkar, X. A., & Kulhalli, K. V., “Thyroid Nodule Detection using Artificial Neural
Network”, In 2021 International Conference on Artificial Intelligence and Smart Systems
(ICAIS), pp. 11-15, 2021.
[49] Zhao, C. K., Ren, T. T., Yin, Y. F., Shi, H., Wang, H. X., Zhou, B. Y., ... & Xu, H. X., “A
comparative analysis of two machine learning-based diagnostic patterns with thyroid
imaging reporting and data system for thyroid nodules: diagnostic performance and
unnecessary biopsy rate”, Thyroid, vol. 31(3), pp. 470-481, 2021.
[50] Bengio, Y, “Learning deep architectures for AI”, Foundations and trends in Machine
Learning, vol. 2(1), pp. 1-127, 2009.
[51] Rusk, N, “Deep learning”, Nature Methods, vol. 13(1), pp. 35-35, 2016.
[52] Lauzon, F. Q, “An introduction to deep learning” In 2012 11th International Conference
on Information Science, Signal Processing and their Applications (ISSPA), pp. 1438-1439,
2012.
[53] Voulodimos, A., Doulamis, N., Doulamis, A., & Protopapadakis, E, “Deep learning for
computer vision: A brief review”, Computational intelligence and neuroscience, 2018.
[54] Ronao, C. A., & Cho, S. B, “Human activity recognition with smartphone sensors using
deep learning neural networks”, Expert systems with applications, vol. 59, pp. 235-244,
2016.
[55] Du, W., & Sang, N., “An effective method for ultrasound thyroid nodules segmentation”,
In 2015 International Symposium on Bioelectronics and Bioinformatics (ISBB), pp. 207-
210, 2015.
[56] Acharya, U. R., Chowriappa, P., Fujita, H., Bhat, S., Dua, S., Koh, J. E., ... & Ng, K. H,
“Thyroid lesion classification in 242 patient population using Gabor transform features
from high resolution ultrasound images”, Knowledge-Based Systems, vol. 107, pp. 235-
245, 2016.
[57] Peng, W., Liu, C., Xia, S., Shao, D., Chen, Y., Liu, R., & Zhang, Z., “Thyroid nodule
recognition in computed tomography using first order statistics”, Biomedical engineering
78
 online, vol.16(1), pp. 1-14, 2017.
[58] Zhu, Y., Fu, Z., & Fei, J., “An image augmentation method using convolutional network
for thyroid nodule classification by transfer learning”, In 2017 3rd IEEE International
Conference on Computer and Communications (ICCC), pp. 1819-1823, 2017.
[59] Song, W., Li, S., Liu, J., Qin, H., Zhang, B., Zhang, S., & Hao, A., “Multitask cascade
convolution neural networks for automatic thyroid nodule detection and recognition”,
IEEE journal of biomedical and health informatics, vol. 23(3), pp. 1215-1224, 2018.
[60] Zhang, R., Liu, Q., Cui, H., Wang, X., Song, S., Huang, G., & Feng, D., “Thyroid
classification via new multi-channel feature association and learning from multi-modality
MRI images”, In 2018 IEEE 15th international symposium on biomedical imaging (ISBI
2018), pp. 277-280, 2018.
[61] Sundar, K. S., Rajamani, K. T., & Sai, S. S. S., “Exploring image classification of thyroid
ultrasound images using deep learning”, In International Conference on ISMAC in
Computational Vision and Bio-Engineering, pp. 1635-1641, Springer, Cham, 2018.
[62] Wang, J., Li, S., Song, W., Qin, H., Zhang, B., & Hao, A., “Learning from weakly-labeled
clinical data for automatic thyroid nodule classification in ultrasound images”, In 2018
25th IEEE International Conference on Image Processing (ICIP), pp. 3114-3118, 2018.
[63] Chen, D., Zhang, J., & Li, W., “Thyroid Nodule Classification Using Two Levels
Attention-Based Bi-Directional LSTM with Ultrasound Reports”, In 2018 9th
International Conference on Information Technology in Medicine and Education (ITME),
pp. 309-312, 2018.
[64] Li, Z., Qin, J., Zhang, X., & Wan, Y., “A Hybrid Intelligent Framework for Thyroid
Diagnosis”, In Cyberspace Data and Intelligence, and Cyber-Living, Syndrome, and
Health, pp. 441-451, Springer, Singapore, 2019.
[65] Ko, S. Y., Lee, J. H., Yoon, J. H., Na, H., Hong, E., Han, K., ... & Kwak, J. Y., “Deep
convolutional neural network for the diagnosis of thyroid nodules on ultrasound”, Head &
neck, vol. 41(4), pp. 885-891, 2019.
[66] Wang, L., Yang, S., Yang, S., Zhao, C., Tian, G., Gao, Y., ... & Lu, Y., “Automatic thyroid
nodule recognition and diagnosis in ultrasound imaging with the YOLOv2 neural
network”, World journal of surgical oncology, vol. 17(1), pp. 1-9, 2019.
[67] Liu, T., Guo, Q., Lian, C., Ren, X., Liang, S., Yu, J., ... & Shen, D., “Automated detection
and classification of thyroid nodules in ultrasound images using clinical-knowledge-
guided convolutional neural networks”, Medical image analysis, vol. 58, 2019.
79
[68] Aboudi, N., Guetari, R., & Khlifa, N., “Multi-objectives optimization of features selection
for the classification of thyroid nodules in ultrasound images”, IET Image Processing,
2020.
[69] Guo, X., Zhao, H., & Tang, Z., “An Improved Deep Learning Approach for Thyroid
Nodule Diagnosis”, In 2020 IEEE 17th International Symposium on Biomedical Imaging
(ISBI), pp. 296-299, 2020.
[70] Ajilisa, O. A., Jagathyraj, V. P., & Sabu, M. K, “Computer-Aided Diagnosis of Thyroid
Nodule from Ultrasound Images Using Transfer Learning from Deep Convolutional
Neural Network Models”, In 2020 Advanced Computing and Communication
Technologies for High Performance Applications (ACCTHPA), pp. 237-241, 2020.
[71] Nguyen, D. T., Kang, J. K., Pham, T. D., Batchuluun, G., & Park, K. R., “Ultrasound
image-based diagnosis of malignant thyroid nodule using artificial intelligence”, Sensors,
vol. 20(7), 2020.
[72] Vasile, C. M., Udriștoiu, A. L., Ghenea, A. E., Popescu, M., Gheonea, C., Niculescu, C.
E., ... & Alexandru, D. O., “Intelligent Diagnosis of Thyroid Ultrasound Imaging Using an
Ensemble of Deep Learning Methods”, Medicina, vol. 57(4), 2021.
[73] Zhu, Y. C., AlZoubi, A., Jassim, S., Jiang, Q., Zhang, Y., Wang, Y. B., ... & Hongbo, D.
U., “A generic deep learning framework to classify thyroid and breast lesions in ultrasound
images”, Ultrasonics, vol. 110, 2021.
[74] Yang, W., Dong, Y., Du, Q., Qiang, Y., Wu, K., Zhao, J., ... & Zia, M. B., “Integrate
domain knowledge in training multi-task cascade deep learning model for benign–
malignant thyroid nodule classification on ultrasound images”, Engineering Applications
of Artificial Intelligence, vol. 98, 2021.
[75] Dabowsa, N. I. A., Amaitik, N. M., Maatuk, A. M., & Aljawarneh, S. A., “A hybrid
intelligent system for skin disease diagnosis”, In 2017 International Conference on
Engineering and Technology (ICET), pp. 1-6, 2017.
[76] Liu, T., Xie, S., Zhang, Y., Yu, J., Niu, L., & Sun, W., “Feature selection and thyroid
nodule classification using transfer learning”, In 2017 IEEE 14th International Symposium
on Biomedical Imaging (ISBI 2017), pp. 1096-1099, 2017.
[77] Xia, J., Chen, H., Li, Q., Zhou, M., Chen, L., Cai, Z., ... & Zhou, H., “Ultrasound-based
differentiation of malignant and benign thyroid Nodules: An extreme learning machine
approach”, Computer methods and programs in biomedicine, vol. 147, pp. 37-49, 2017.
[78] Nguyen, D. T., Pham, T. D., Batchuluun, G., Yoon, H. S., & Park, K. R., “Artificial
80
 intelligence-based thyroid nodule classification using information from spatial and
frequency domains”, Journal of clinical medicine, vol. 8(11),2019.
[79] Qin, P., Wu, K., Hu, Y., Zeng, J., & Chai, X., “Diagnosis of benign and malignant thyroid
nodules using combined conventional ultrasound and ultrasound elasticity imaging”, IEEE
journal of biomedical and health informatics, vol. 24(4), pp. 1028-1036, 2019.
[80] Xie, J., Guo, L., Zhao, C., Li, X., Luo, Y., & Jianwei, L., “A Hybrid Deep Learning and
Handcrafted Features based Approach for Thyroid Nodule Classification in Ultrasound
Images”, In Journal of Physics: Conference Series, vol. 1693, 2020.
[81] Sun, H., Yu, F., & Xu, H., “Discriminating the Nature of Thyroid Nodules Using the
Hybrid Method”, Mathematical Problems in Engineering, 2020.
[82] Hang, Y, “Thyroid nodule classification in ultrasound images by fusion of conventional
features and res-GAN deep features”, Journal of Healthcare Engineering, 2021.
[83] Srivastava, R., Kumar, P., “A hybrid model for the identification and classification of
thyroid nodules in medical ultrasound images”, International Journal of Modelling,
Identification and Control (IJMIC), 2021.
[84] Kwak, J. Y., Han, K. H., Yoon, J. H., Moon, H. J., Son, E. J., Park, S. H., & Kim, E. K,
“Thyroid imaging reporting and data system for US features of nodules: a step in
establishing better stratification of cancer risk.”, Radiology, vol. 260(3), pp. 892-899,
2011.
[85] Karanja, E. M., Masupe, S., & Jeffrey, M. G, “Analysis of internet of things malware using
image texture features and machine learning techniques”, Internet of Things, vol. 9, pp.
100-153, 2020.
[86] Garg, M., & Dhiman, G, “A novel content-based image retrieval approach for
classification using GLCM features and texture fused LBP variants”, Neural Computing
and Applications, pp. 1-18, 2020.
[87] Chapelle, O., Haffner, P., & Vapnik, V. N, “Support vector machines for histogram-based
image classification”, IEEE transactions on Neural Networks, vol. 10(5), pp. 1055-1064,
1999.
[88] Keramidas, E. G., Maroulis, D., & Iakovidis, D. K, “ΤND: a thyroid nodule detection
system for analysis of ultrasound images and videos”, Journal of medical systems, vol.
36(3), pp. 1271-1281, 2012.
[89] Zhang, B., Tian, J., Pei, S., Chen, Y., He, X., Dong, Y., & Zhang, S, “Machine learning–
assisted system for thyroid nodule diagnosis”, Thyroid, vol. 29(6), pp. 858-867, 2019.
81
[90] Jothi, A. A. J., & Rajam, M. A, “Automatic classification of thyroid histopathology images
using multi-classifier system”, Multimedia Tools and Applications, vol. 76(18), pp. 18711-
18730, 2017.
[91] Zhou, H., Zhang, J., Zhou, Y., Guo, X., & Ma, Y, “A feature selection algorithm of
decision tree based on feature weight”, Expert Systems with Applications, vol.164, 113842,
2021.
[92] Abiodun, O. I., Jantan, A., Omolara, A. E., Dada, K. V., Mohamed, N. A., & Arshad, H,
“State-of-the-art in artificial neural network applications: A survey”, Heliyon, vol. 4(11),
e00938, 2018.
[93] A. Liaw, and M. Wiener, “Classification and regression by random Forest”, R news, vol.
2, pp. 18-22, 2002.
[94] Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B., ... & Chen, T, “Recent
advances in convolutional neural networks”, Pattern recognition, vol. 77, pp. 354-377,
2018.
[95] Zhou, D. X, “Theory of deep convolutional neural networks: Down sampling”, Neural
Networks, vol. 124, pp. 319-327, 2020.
[96] Srivastava, R., Kumar, P, “A contemporary review on soft computing techniques for
thyroid disorder identification and detection”, International Journal of Computer
Applications in Technology (IJCAT), 2021.
[97] Kang, M., & Tian, J, “Machine Learning: Data Pre‐processing. Prognostics and Health
Management of Electronics: Fundamentals”, Machine Learning, and the Internet of
Things, pp. 111-130, 2018.
[98] Masoudi, S., Harmon, S. A., Mehralivand, S., Walker, S. M., Raviprakash, H., Bagci, U.,
... & Turkbey, B, “Quick guide on radiology image pre-processing for deep learning
applications in prostate cancer research”, Journal of Medical Imaging, vol. 8(1), 010901,
2021.
[99] Qi, P. F., Meng, Q. H., & Zeng, M, “A CNN-based simplified data processing method for
electronic noses”, In 2017 ISOCS/IEEE International Symposium on Olfaction and
Electronic Nose (ISOEN), pp. 1-3, IEEE, 2017.
[100] Pitaloka, D. A., Wulandari, A., Basaruddin, T., & Liliana, D. Y, “Enhancing CNN with
preprocessing stage in automatic emotion recognition”, Procedia computer science,
vol.116, pp. 523-529, 2017.
[101] Vimal Kurup, R., Sowmya, V., & Soman, K. P, “Effect of data pre-processing on brain
82
 tumor classification using capsulenet”, In International Conference on Intelligent
Computing and Communication Technologies, pp. 110-119, 2019.
[102] Haryanto, T., Wasito, I., & Suhartanto, H, “Convolutional Neural Network (CNN) for
gland images classification”, In 2017 11th International Conference on Information &
Communication Technology and System (ICTS), pp. 55-60, IEEE, 2017.
[103] Rajan, S., Chenniappan, P., Devaraj, S., & Madian, N, “Novel deep learning model for
facial expression recognition based on maximum boosted CNN and LSTM”, IET Image
Processing, vol. 14(7), pp. 1373-1381, 2020.
[104] Soomro, T. A., Afifi, A. J., Shah, A. A., Soomro, S., Baloch, G. A., Zheng, L., & Gao, J,
“Impact of image enhancement technique on CNN model for retinal blood vessels
segmentation”, IEEE Access, vol. 7, pp. 158183-158197, 2019.
[105] Garg, G., Kumar, D., Sonker, Y., & Garg, R, “A Hybrid MLP-SVM Model for
Classification using Spatial-Spectral Features on Hyper-Spectral Images”, arXiv preprint
arXiv:2101.00214, 2021.
[106] Guo, T., Dong, J., Li, H., & Gao, Y, “Simple convolutional neural network on image
classification”, In 2017 IEEE 2nd International Conference on Big Data Analysis
(ICBDA), pp. 721-724, IEEE, 2017.
[107] Han, D., Liu, Q., & Fan, W, “A new image classification method using CNN transfer
learning and web data augmentation”, Expert Systems with Applications, vol. 95, pp. 43-
56, 2018.
[108] Lei, X., Pan, H., & Huang, X, “A dilated CNN model for image classification”, IEEE
Access, vol. 7, pp. 124087-124095, 2019.
[109] Li, Q., Cai, W., Wang, X., Zhou, Y., Feng, D. D., & Chen, M, “Medical image
classification with convolutional neural network”,In 2014 13th international conference
on control automation robotics & vision (ICARCV), pp. 844-848, IEEE, 2014.
[110] Sultana, F., Sufian, A., & Dutta, P, “Advancements in image classification using
convolutional neural network”, In 2018 Fourth International Conference on Research in
Computational Intelligence and Communication Networks (ICRCICN), pp. 122-129,
IEEE, 2018.
[111] Jmour, N., Zayen, S., & Abdelkrim, A, “Convolutional neural networks for image
classification”, In 2018 international conference on advanced systems and electric
technologies (IC_ASET), pp. 397-402, 2018.
[112] Li, W., Chen, C., Zhang, M., Li, H., & Du, Q, “Data augmentation for hyperspectral image
83
 classification with deep CNN”, IEEE Geoscience and Remote Sensing Letters, vol. 16(4),
pp. 593-597, 2018.
[113] Alkhaleefah, M., & Wu, C. C, “A hybrid CNN and RBF-based SVM approach for breast
cancer classification in mammograms” In 2018 IEEE International Conference on
Systems, Man, and Cybernetics (SMC), pp. 894-899, 2018.
[114] Sun, X., Liu, L., Li, C., Yin, J., Zhao, J., & Si, W, “Classification for remote sensing data
with cimproved CNN-SVM method”, IEEE Access, vol. 7, pp. 164507-164516, 2019.
[115] Rosasco, L., De Vito, E., Caponnetto, A., Piana, M., & Verri, A, “Are loss functions all
the same?”, Neural computation, vol. 16(5), pp. 1063-1076, 2004.
[116] Ilhan, U., & Ilhan, A, “Brain tumor segmentation based on a new threshold approach”,
Procedia computer science, vol.120, pp. 580-587, 2017.
[117] Yu, S., Jia, S., & Xu, C, “Convolutional neural networks for hyperspectral image
classification”, Neurocomputing, vol. 219, pp. 88-98, 2017.
[118] Priyadarshini, I., & Cotton, C, “A novel LSTM–CNN–grid search-based deep neural
network for sentiment analysis”, The Journal of Supercomputing, vol. 77(12), pp. 13911-
13932, 2021.
[119] Louverdis, G., Vardavoulia, M. I., Andreadis, I., & Tsalides, P, “A new approach to
morphological color image processing”, Pattern recognition, vol. 35(8), pp. 1733-1741,
2002.
[120] Merzban, M. H., & Elbayoumi, M, “Efficient solution of Otsu multilevel image
thresholding: A comparative study”, Expert Systems with Applications, vol. 116, pp.299-
309, 2019.
[121] Araujo, R. L., de Araujo, F. H., & Silva, R. R, “Automatic segmentation of melanoma skin
cancer using transfer learning and fine-tuning”, Multimedia Systems, vol. 28(4), pp. 1239-
1250, 2022.
[122] Khan, M. Q., Hussain, A., Rehman, S. U., Khan, U., Maqsood, M., Mehmood, K., & Khan,
M. A, “Classification of melanoma and nevus in digital images for diagnosis of skin
cancer”, IEEE Access, vol.7, pp. 90132-90144, 2019.
[123] Kass, M., Witkin, A., & Terzopoulos, D, “Snakes: Active contour models”, International
journal of computer vision, vol.1(4), pp. 321-331, 1988.
[124] Bovik, A. C, “Basic binary image processing. In The Essential Guide to Image
Processing”, Academic Press, pp. 69-96, 2009.
[125] He, K., Zhang, X., Ren, S., & Sun, J, “Deep residual learning for image recognition”, In
84
 Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-
778, 2016.
[126] Fulton, L. V., Dolezel, D., Harrop, J., Yan, Y., & Fulton, C. P, “Classification of
Alzheimer’s disease with and without imagery using gradient boosted machines and Res-
Net-50”, Brain sciences, vol.9(9), pp. 1-16, 2019.
[127] Yao, X., Wang, X., Wang, S. H., & Zhang, Y. D, “A comprehensive survey on
convolutional neural network in medical image analysis”, Multimedia Tools and
Applications, pp. 1-45, 2020.
[128] Krizhevsky, A., Sutskever, I., & Hinton, G. E, “Imagenet classification with deep
convolutional neural networks”, Communications of the ACM, vol. 60(6), pp. 84-90, 2017.
[129] Simonyan, K., & Zisserman, A, “Very deep convolutional networks for large-scale image
recognition”, arXiv preprint arXiv:1409.1556, 2014.
[130] Liu, Y., Yin, B., Yu, J., & Wang, Z, “Image classification based on convolutional neural
networks with cross-level strategy”, Multimedia Tools and Applications, vol.76(8), pp.
11065-11079, 2017.
[131] Srivastava, S., Kumar, P., Chaudhry, V., & Singh, A, “Detection of ovarian cyst in
ultrasound images using fine-tuned VGG-16 deep learning network”, SN Computer
Science, vol.1(2), pp.1-8, 2020.
[132] Hosny, K. M., Kassem, M. A., & Fouad, M. M, “Classification of skin lesions into seven
classes using transfer learning with AlexNet”, Journal of digital imaging, vol. 33(5), pp.
1325-1334, 2020.
[133] Islam, S., Khan, S. I. A., Abedin, M. M., Habibullah, K. M., & Das, A. K, “Bird species
classification from an image using VGG-16 network”, In Proceedings of the 2019 7th
international conference on computer and communications management, pp. 38-42, 2019.
[134] Liu, X. Y., Fang, Y., Yang, L., Li, Z., & Walid, A, “High-performance tensor
decompositions for compressing and accelerating deep neural networks”, In Tensors for
Data Processing, Academic Press, pp. 293-340, 2022.
[135] Rezaee, M., Zhang, Y., Mishra, R., Tong, F., & Tong, H, “Using a vgg-16 network for
individual tree species detection with an object-based approach”, In 2018 10th IAPR
Workshop on Pattern Recognition in Remote Sensing (PRRS) pp. 1-7, 2018.
[136] Montavon, G., Samek, W., & Müller, K. R, “Methods for interpreting and understanding
deep neural networks”, Digital signal processing, vol. 73, pp. 1-15, 2018.
[137] Mahmood, A., Bennamoun, M., An, S., Sohel, F., Boussaid, F., Hovey, R., ... & Fisher, R.
85
 B, “Deep learning for coral classification. In Handbook of neural computation, Academic
Press, pp. 383-401, 2017.
[138] Bingham, E., Kaski, S., Laaksonen, J., & Lampinen, J, “Advances in independent
component analysis and learning machines”, Academic Press, 2015.
[139] Shams, S., Platania, R., Zhang, J., Kim, J., Lee, K., & Park, S. J, “Deep generative breast
cancer screening and diagnosis”, In International Conference on Medical Image
Computing and Computer-Assisted Intervention, pp. 859-867, Springer, 2018.
[140] Lipkind, H. S., Vazquez-Benitez, G., DeSilva, M., Vesco, K. K., Ackerman-Banks, C.,
Zhu, J., ... & Kharbanda, E. O, “Receipt of COVID-19 vaccine during pregnancy and
preterm or small-for-gestational-age at birth—Eight Integrated Health Care
Organizations”, Morbidity and Mortality Weekly Report, United States, December 15,
2020–July 22, 2021, vol. 71(1), 26, 2022.
[141] Mahajan, H. B., Rashid, A. S., Junnarkar, A. A., Uke, N., Deshpande, S. D., Futane, P. R.,
... & Alhayani, B, "Integration of Healthcare 4.0 and blockchain into secure cloud-based
electronic health records systems", Applied Nanoscience, vol.13(3), pp.2329-2342, 2023.
[142] Chengoden, R., Victor, N., Huynh-The, T., Yenduri, G., Jhaveri, R. H., Alazab, M., ... &
Gadekallu, T. R, "Metaverse for healthcare: A survey on potential applications, challenges
and future directions", IEEE Access, 2023.
[143] Dicuonzo, G., Donofrio, F., Fusco, A., & Shini, M, "Healthcare system: Moving forward
with artificial intelligence", Technovation, vol. 120, 102510, 2023.
[144] Santana, I. R., Mason, A., Gutacker, N., Kasteridis, P., Santos, R., & Rice, N, "Need,
demand, supply in health care: working definitions, and their implications for defining
access", Health Economics, Policy and Law, vol. 18(1), pp. 1-13, 2023.
[145] Ali, O., Abdelbaki, W., Shrestha, A., Elbasi, E., Alryalat, M. A. A., & Dwivedi, Y. K, "A
systematic literature review of artificial intelligence in the healthcare sector: Benefits,
challenges, methodologies, and functionalities", Journal of Innovation & Knowledge, vol.
8(1), 100333, 2023.
[146] Rumelhart, D. E., Hinton, G. E., & Williams, R. J, “Learning representations by back-
propagating errors”, Nature, vol. 323(6088), pp. 533-536, 1986.
[147] Wang, D., Lu, Z., Xu, Y., Wang, Z., Santella, A., & Bao, Z, “Cellular structure image
classification with small targeted training samples”, IEEE Access, vol.7, pp. 148967-
148974, 2019.
[148] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... &
86
 Bengio, Y, “Generative adversarial nets”, Advances in neural information processing
systems, vol. 27, 2014.
[149] Aggarwal, A., Mittal, M., & Battineni, G, “Generative adversarial network: An overview
of theory and applications” International Journal of Information Management Data
Insights, vol.1(1), 100004, 2021.
[150] Alqahtani, H., Kavakli-Thorne, M., & Kumar, G, “Applications of generative adversarial
networks (gans): An updated review”, Archives of Computational Methods in Engineering,
vol.28(2), pp. 525-552, 2021.
[151] https://developers.google.com/machine-learning/gan/gan_structure
[152] Akbar, S. B., Thanupillai, K., & Sundararaj, S, “Combining the advantages of Alex-Net
convolutional deep neural network optimized with anopheles search algorithm-based
feature extraction and random forest classifier for COVID‐19 classification”, Concurrency
and Computation: Practice and Experience, pp. 1-15, 2022.

[153] Rasheed, A., Ali, N., Zafar, B., Shabbir, A., Sajid, M., & Mahmood, M. T, “Handwritten
Urdu characters and digits recognition using transfer learning and augmentation with
AlexNet”, IEEE Access, vol. 10, pp. 102629-102645, 2022.

[154] Hammad, I., & El-Sankary, K, “Impact of approximate multipliers on VGG deep learning
network”, IEEE Access, vol. 6, pp. 60438-60444, 2018.
[155] Jibhakate, P., Gole, S., Yeskar, P., Rangwani, N., Vyas, A., & Dhote, K, “Early Glaucoma
Detection Using Machine Learning Algorithms of VGG-16 and Resnet-50”, In 2022 IEEE
Region 10 Symposium (TENSYMP), pp. 1-5, 2022.

87
LIST OF PUBLICATIONS

Journal(s)

Published

1. Srivastava, R., & Kumar, P., “Deep-GAN: an improved model for thyroid nodule
identification and classification”, Neural Computing and Application, Springer.
Index: SCIE, Scopus, ACM digital library, Impact factor: 6

2. Srivastava, R., & Kumar, P., “Optimizing CNN Based Model for Thyroid Nodule
Classification Using Data Augmentation, Segmentation and Boundary Detection
Techniques”, Multimedia Tools and Applications, Springer.
Index: SCIE, Scopus, ACM digital library, Impact factor:3.6

3. Srivastava, R., & Kumar, P., “A hybrid model for the identification and classification of
benign and malignant thyroid nodules on medical USG image”, International Journal of
Modelling, Identification and Control, Inderscience., 2022.
Index: ESCI, Scopus, ACM digital library
DOI: 10.1504/IJMIC.2022.10051644
4. Srivastava, R., & Kumar, P., “A contemporary review on soft computing techniquesfor
thyroid identification and detection”, International Journal of Computer Applications in
Technology (IJCAT), Inderscience, 2022.
Index: ESCI, Scopus, ACM digital library
DOI: 10.1504/IJCAT.2022.10051647
5. Srivastava, R., & Kumar, P., “GSO-CNN based model for the identification and
classification of thyroid nodule in medical USG images”, Network Modeling Analysis in
Health Informatics and Bioinformatics (NHIB), Springer, 2022.
Index: ESCI, Scopus, ACM digital library
DOI: 10.1007/s13721-022-00388-w
6. Srivastava, R., & Kumar, P., “A CNN-SVM hybrid model for the classification of thyroid
nodules in medical ultrasound images”, International Journal of Grid and Utility
Computing (IJGUC), Inderscience, 2022.
Index: ESCI, Scopus, ACM digital library

88
 DOI: 10.1504/IJGUC.2022.10052488

7. Srivastava, R., & Kumar, P., “Performance comparison of various machine learning
algorithms for the identification and classification of thyroid nodules”, International
Journal of Grid and Utility Computing (IJGUC), Inderscience.
[Index: ESCI, Scopus, ACM digital library]
Communicated

1. Srivastava, R., & Kumar, P., “Alex-GAN based synthetic ultrasound images
augmentation model for increasing the performance of thyroid nodules classification”,
Multimedia Tools and Applications.
[Submitted] [Index: SCIE, Scopus, ACM digital library, Impact factor:3.6]

Conference(s)
1. Srivastava, R., & Kumar, P, “BL_SMOTE ensemble method for prediction of thyroid
disease on imbalanced classification problem”, In Proceedings of Second International
Conference on Computing, Communications, and Cyber-Security, pp. 731-741, Springer,
Singapore.
[Index: Scopus, DBLP, Google Scholar, EI-Compendex, SCImago]
DOI: https://doi.org/10.1007/978-981-16-0733-2_52
2. Srivastava, R., & Kumar, P, “A novel model for novel model for the identification and
classification of thyroid nodules using deep neural network”, 4th International Conference
on Machine Intelligence and Signal Processing (MISP2022), NIT Raipur, Springer, 12-
14 March 2022.
[Index: Scopus, DBLP, Google Scholar, EI-Compendex, SCImago]

89