Covid-19 Chest X-Ray Images Detection and Classification using

Transfer Learning.

A THESIS

SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR DEGREE OF

MASTER OF TECHNOLOGY
IN
INFORMATION TECHNOLOGY

By

Ram Choudhary
MIT2020069
UNDER THE SUPERVISION OF
Prof. Shekhar Verma
IIIT-ALLAHABAD
INDIAN INSTITUTE OF INFORMATION TECHNOLOGY, ALLAHABAD
(A UNIVERSITY ESTABLISHED UNDER SEC.3 OF UGC ACT, 1956 VIDE NOTIFICATION NO.
F.9-4/99-U.3 DATED 04.08.2000 OF THE GOVT. OF INDIA)

A CENTRE OF EXCELLENCE IN INFORMATION TECHNOLOGY ESTABLISHED BY GOVT. OF

INDIA

JULY 2022
 CANDIDATE DECLARATION

I, Ram Choudhary, Enrolment No.MIT2020069, do hereby declare that this thesis

titled, “Covid-19 Chest X-Ray Images Detection and Classification using Trans-
fer Learning” submitted in the qualified perfection for the degree of Master of
Technology in Information Technology at Indian Institute of Information Tech-
nology, Allahabad is a record of bonafide work done by me under the able guidance
of Prof. Shekhar Verma and due acknowledgements have been made to all the other
material used. This report work was done in full compliance with the requirements
and constraints of the prescribed curriculum

Place : Allahabad
Date: / /

Ram Choudhary(MIT2020069)
 CERTIFICATE FROM SUPERVISOR

I do hereby recommend that the thesis prepared under my supervision by Ram

Choudhary (MIT2020069) entitled ”Covid-19 Chest X-Ray Images Detection and
Classification using Transfer Learning” be accepted in partial fulfillment for the
degree of M.Tech in Information technology for examination.

Date: / /
Place: Allahabad

Prof. Shekhar Verma

Thesis Supervisor

Countersigned by:

..................
Dean(Academics)
 CERTIFICATE OF APPROVAL

The foregoing thesis is hereby approved as a creditable study in Information Tech-

nology and its allied areas. It is carried out and presented in a satisfactory manner to
warrant its acceptance as a prerequisite to the degree for which it has been submitted.
It is understood that by this approval the undersigned do not necessarily endorse or
approve any statement made, opinion expressed or conclusion drawn therein but the
thesis only for the purpose for which it is submitted.

Committee Members for Evaluation of the Thesis for Final Examination:

..................................

..................................

..................................

..................................
 Acknowledgements

It is my honour and privilege to get an opportunity to study at such a great in-

stitute. “Indian Institute Information Technology –Allahabad”, where each day
was a chance to learn and grow personally and academically and professionally. The
wealth of knowledge and experience, it has given us is deep-rooted foundations in
our various fields of study.

I would like to express my sincere gratitude to all those who provided me with
the resources and guidance to complete my thesis. I am profoundly grateful to Prof.
Shekhar Verma for his expert guidance and continuous encouragement throughout
to see that this project rights its target since its commencement to its completion.
His instrumental contribution in the proposed work has made it conceivable.

I would also like to thank Mr. Adarsh Prasad Behera, Research Scholar in the
department of Information Technology, at IIIT Allahabad for his overwhelming sup-
port and for helping me out whenever and wherever I approached him. My friends,
classmates who helped me to widen intellectual scope and who supported me in
ways too great to forget. Special thanks to my family for their love, encouragement,
interesting discussions and continuous support throughout the course.

Finally at the end I would like to thank to all faculty members who inspired me to
finish my work.
Ram Choudhary
MTech I.T
 ABSTRACT

Coronavirus family is the source of the ailment known as COVID-19. the CXRs of
COVID-19 infected individuals are abnormal therefore radiography of chest is one
of the prominent detection method for COVID-19 diagnosis. However, an expert
with extensive knowledge is needed to review the CXRs (Chest-X-Ray). Therefore,
applying deep learning techniques to identify anomalies in X-ray pictures is fre-
quently suggested as a prepared option to aid in the diagnosis of the illness. Many
studies on COVID-19 CXR classification have been published, however the major-
ity of them used a limited sample of COVID-19 CXR pictures, which led to dataset
which is unbalanced and impacted the performance and the analysis of the models.
In this we suggest a number of image pre-processing methods to improve COVID-19
CXR pictures in order to provide a sizable and varied dataset to improve the effec-
tiveness of deep learning algorithms(DLA) in identifying the illness from CXR. We
also provide cutting-edge, reliable and efficient modified deep learning models to
detect COVID-19 from a huge collection of CXR pictures, based on DenseNet201,
VGG16, VGG19 and Inception-V3, we also make use of the transfer learning so that
we can train pre-trained models effectively those who are trained on more a million
images. Based on the previous work findings, we have improved dataset which is
large as well as balanced and contains better COVID-19 CXR pictures that is cur-
rently accessible.To identify more accurately COVID-19 using CXRs, we have also
added lung opacity as a new class to imporve precision in classification of COVID-
19.

This study concentrates on exploring Deep Neural Network Models on several

facets. We compare these models based on model architecture, pre-training meth-
ods, accuracy with CNN models, distribution of real, discriminator-loss function and
generator-loss function. On the binary class classification, our model scored 96.1%,
and on the multi-class classification, it scored 87.18%.Also based on these findings,
we found that accuracy and the performance more likely depends on dataset, number
of classes, number of images per class. Based on these factors we achieved optimal
results with binary class and multi class classification. The ability to more efficiently
identifying COVID-19 using CXRs makes this study of great interest to medical pro-
fessionals. On the benchmark, COVID-19 CXR image dataset consist of 4 classes,
Covid, Normal, Pneumonia and Lung Opacity.
 In this work, We have measured accuracy of our model and test it with other
models based on various parameters. we have proposed a method that combines
Deep Neural Network and Transfer Learning and analysed accuracy score claims of
different models with a range of epochs.

The main contributions made in this thesis are:

• We have applied transfer learning technique on different DNN models to en-

hance accuracy over smaller dataset of images

• We have performed experiments on COVID-19 CXR images that consist of 4

classes of normal, covid, pneumonia and lung opacity for comparative analysis
with different models.

• We have proposed methods for evaluating classification accuracy deep neural

network models combine with transfer learning.

Indian Institute of Information Technology, Allahabad 7

Contents

1 INTRODUCTION 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Deep Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 CNN for image classification . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 VGG-16 . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 VGG-19 . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.3 DenseNet . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.4 Inception-V3 . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 LITERATURE REVIEW 14

3 PROBLEM FORMULATION 18

4 PROPOSED METHODOLOGY 19
4.1 Synopsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

5 Training Process & Models 21

5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.1.1 Python 3.7 . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.1.2 Sklearn . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.1.3 Pandas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.1.4 Numpy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.1.5 Google Collaboratory . . . . . . . . . . . . . . . . . . . . . 22
5.2 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Image Pre-Processing and Augmentation . . . . . . . . . . . . . . 23
5.4 Pre-Trained Deep Learning Models . . . . . . . . . . . . . . . . . 25
5.5 Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

6 Result and Analysis 28

6.1 Binary class classification Evaluation . . . . . . . . . . . . . . . . . 28
 6.1.1 VGG-16 Model Evaluation . . . . . . . . . . . . . . . . . . 28
6.1.2 VGG-19 Model Evaluation . . . . . . . . . . . . . . . . . . 29
6.1.3 DenseNet Model Evaluation . . . . . . . . . . . . . . . . . 31
6.1.4 InceptionV3 Model Evaluation . . . . . . . . . . . . . . . . 32
6.2 Multi class classification Evaluation . . . . . . . . . . . . . . . . . 34
6.2.1 VGG-16 Model Evaluation . . . . . . . . . . . . . . . . . . 34
6.2.2 VGG-19 Model Evaluation . . . . . . . . . . . . . . . . . . 36
6.2.3 DenseNet201 Model Evaluation . . . . . . . . . . . . . . . 38
6.2.4 InceptionV3 Model Evaluation . . . . . . . . . . . . . . . . 40

7 Conclusion 44

References 44
List of Figures

1.1 CNN for image classification . . . . . . . . . . . . . . . . . . . . . 3

1.2 VGG-16 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 VGG-19 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 DenseNet Architecture . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 DenseNet Architecture . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Inception-V3 Architecture [46] . . . . . . . . . . . . . . . . . . . . 12
1.7 Inception-V3 Architecture . . . . . . . . . . . . . . . . . . . . . . 12

4.1 base working architecture [45] . . . . . . . . . . . . . . . . . . . . 20

5.1 Normal CXR image samples . . . . . . . . . . . . . . . . . . . . . 23

5.2 Covid CXR image samples . . . . . . . . . . . . . . . . . . . . . . 23
5.3 Pneumonia CXR image samples . . . . . . . . . . . . . . . . . . . 24
5.4 Lung Opacity CXR image samples . . . . . . . . . . . . . . . . . . 24
5.5 The proposed head architecture of pre-trained models . . . . . . . . 27

6.1 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 28

6.2 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 28
6.3 Predicted Outcomes by VGG-16 . . . . . . . . . . . . . . . . . . . 29
6.4 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 30
6.5 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 30
6.6 Predicted Outcomes by VGG-19 . . . . . . . . . . . . . . . . . . . 30
6.7 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 31
6.8 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 31
6.9 Predicted Outcomes by DenseNet . . . . . . . . . . . . . . . . . . 32
6.10 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 33
6.11 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 33
6.12 Predicted Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.13 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 34
6.14 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 34
6.15 Accuracy vs Val Accuracy with 30 epochs . . . . . . . . . . . . . . 35
6.16 Loss vs Val-loss with 30 epochs . . . . . . . . . . . . . . . . . . . 35
6.17 predicted outcomes by VGG-16 . . . . . . . . . . . . . . . . . . . 35
6.18 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 36
6.19 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 36
6.20 Accuracy vs Val Accuracy with 30 epochs . . . . . . . . . . . . . . 36
6.21 Loss vs Val-loss with 30 epochs . . . . . . . . . . . . . . . . . . . 36
6.22 predicted outcomes by VGG-19 . . . . . . . . . . . . . . . . . . . 37
6.23 Accuracy vs Validation Accuracy with 15 epochs . . . . . . . . . . 38
6.24 Loss vs Val loss with 15 epochs . . . . . . . . . . . . . . . . . . . 38
6.25 Accuracy vs Val Accuracy with 30 epochs . . . . . . . . . . . . . . 38
6.26 Loss vs Val-loss with 30 epochs . . . . . . . . . . . . . . . . . . . 38
6.27 predicted outcomes by DenseNet . . . . . . . . . . . . . . . . . . 39
6.28 Accuracy vs Val Accuracy with 15 epochs . . . . . . . . . . . . . . 40
6.29 Loss vs Val-loss with 15 epochs . . . . . . . . . . . . . . . . . . . 40
6.30 Accuracy vs Val Accuracy with 30 epochs . . . . . . . . . . . . . . 40
6.31 Loss vs Val-loss with 30 epochs . . . . . . . . . . . . . . . . . . . 40
6.32 predicted outcomes by InceptionV3 . . . . . . . . . . . . . . . . . 41
 CHAPTER 1. INTRODUCTION

Chapter 1

INTRODUCTION

1.1 Overview

Coronavirus disease is a severe and dangerous illness that first appeared in Dec
2019 and has since spread around the globe [1]. The Worldometer [2] is a site that
was built by a group of developers and researchers with the purpose of providing
information on numerical data from across the globe. According to the informa-
tion provided on the website, the COVID-19 pandemic has resulted in a total of
183,194,939 cases and 3,966,106 fatalities by the end of June 2021. Fever, cough-
ing, shortness of breath, and exhaustion are some of the symptoms that are associated
with COVID-19 [3]. One of the acute symptoms is the sensation of having discom-
fort in the chest and having difficulty breathing. An abnormal CXRs may be seen in
about half of the people who have COVID-19 [4]. When it comes to the early detec-
tion and treatment of patient who may be suffering with COVID-19, chest imaging
plays a very significant role. Imaging of the chest performed using X-rays is used in
order to conduct a thorough examination of the patient’s chest [5]. In addition, the
advancements that have been made in deep learning applications over the last several
years have been of assistance in properly recognising COVID-19 from the CXR. DL
is a subfield of ML that attempts to model the way in which humans acquire certain
categories of knowledge. We are using it to evaluate and recognise patterns within
images, such as those seen in radiography jobs. DLA have shown some encouraging
outcomes in the process of extracting information from X-rays and medical pictures
[7]. As a result, we emphasise the use of DL models in the process of recognising
COVID-19 instances based on CXR pictures. By using an automated COVID-19
predictor based on a CXR, medical professionals will be able to recognise the illness
more quickly and take appropriate treatment. This study suggests a novel approach
for transfer learning that makes use of deep learning along with transfer learning by
combining head and final layer to suit four pre-trained models. These models are

Indian Institute of Information Technology, Allahabad 1

 CHAPTER 1. INTRODUCTION

DenseNet201, VGG-16, VGG-19 and Inception-V3. In order to conduct a compre-

hensive analysis of the performance, we compute the accuracy, precision, recall, and
specificity scores, as well as the F1-score, multi-class classification. Deep learning
models, when applied to chest X-ray pictures, provide accurate findings when used
to make COVID-19 predictions, according to the framework that has been suggested.

Additionally, we assess the performance of our models on dataset which is bal-

anced that we created from pictures of 4 kinds of classes, normal, pneumonia,
COVID-19 and Lung Opacity chest X-rays. Prior studies show two main flaws, an
unbalanced dataset and a dearth of COVID-19 images both addressed by this dataset.
This dataset is compiled from 11 sources and contains a total of approx 14800 pic-
tures, including 3700 photos of normal chest X-rays, 3700 images of pneumonia,
2000 images of COVID-19 before the augmentation of images and 3700 images af-
ter the augmentation images and 3700 lung opacity images. The following are the
main contributions.

• To identify and know COVID-19 from CXR pictures, we suggest four new
modified DL models by the use of transfer learning based on Dense-Net, VGG16,
VGG19 and Inception-V3.

• We provide the COVID-Chest-X-ray balanced dataset, which was compiled

from eleven accessible datasets. By boosting the COVID-19 pictures from 2000
to 3700 photos, we also make use of several data augmentation methods to
create this balanced dataset. This gives a dataset of 14800 photos overall (3700
normal, 3700 pneumonia, 3700 COVID-19 and 3700 Lung Opacity).

1.2 Deep Neural Networks

Recently, deep neural networks have taken over as the go-to method for resolv-
ing several classification and computer vision issues. The capability includes im-
porting a pre-trained network and performing inference on it, however building a
neural network is beyond the scope. Deep neural networks, often known as DNNs,
have been very successful in the field of classification and computer vision. How-
ever, their extreme performance comes at the significant computational cost, which
greatly hinders their applications to many devices, such as mobile phones, detection
devices and Internet of Things (IoT) devices. Therefore, approaches and techniques
that decrease the computation cost while maintaining the high accuracy of DNN
are in high demand to allow a variety of edge AI applications. These methods and

Indian Institute of Information Technology, Allahabad 2

 CHAPTER 1. INTRODUCTION

strategies may be found in the literature. The motto of this chapter is to offer an
overview of effective deep learning approaches. In this, we begin by discussing
common model compression approaches such as factorization, pruning, and quan-
tized. After that, we outline strategies for the building of compact models, such as
efficient convolution layers and representational efficient CNN architectures.

1.3 CNN for image classification

Image classification has been primary topics in the computer vision domain. Con-
volutional neural networks (CNN) is a specific architecture of ANNs popularly used
in visual imagery and image recognition. The most crucial phase in detection of
COVID-19 virus using XR(X-ray) images is the extraction of features. The method
of extracting the information that is being distinct from pattern of lungs is known
as feature extraction. CNN is being used in extraction of optimal features from the
images in below Figure 1.1.

Figure 1.1: CNN for image classification

In traditional CNN architecture, an image goes through a number of convolution,

fully connected and pooling layers, to generate final output. Deep Convolutional
Networks(CNNs) usually worked better for image classification.CNN is combina-
tion of input layer, output layer, and numerous hidden layers in between. The hid-
den layers include convolution, pooling, Rectified Linear Unit, dropout Layer, and
normalisation. and the output is a prediction of the image. Taking the input as a

Indian Institute of Information Technology, Allahabad 3

 CHAPTER 1. INTRODUCTION

CXR image, the first layer is conv layer which is the basic layer of CNN. This layer
consists of some filters. Here, the input CXR image is made to pass through this ker-
nels/filter. The function map is obtained where the convolution operation performs
the multiplication of the sets of weights with the inputs. The multiplication of array
of input image and 2-dimensional collection of weights constitutes a filter. The filter
is considerably smaller that the size of input. This filter multiplies the input from
different angles and points. This layer covers the entire image so this filter is used to
identify some specific types of features .CNN presented a categorization model that
had been trained on before to work out experiments that had been practised. The
model believes that it has identified the chest XR image based on the label that was
present up to the training stage. The following stage is the evaluation, often known
as the categorization phase. The suggested model is going to be tested once more
utilising a dataset of chest x-ray pictures, but this time it won’t be labelled before-
hand. During the phase of training that came before, models will be expected to
categorise themselves based on the knowledge they have gained. At this level, the
categorization model will be measured or evaluated to see whether it is good or less
good than previous stages.

Stochastic gradient descent (SGD), which is used for optimization, is used in a se-
quential manner within the framework of the suggested classification model, which
has a rate of 0.01. The researchers will purposefully alter some parameters through-
out the classification process. Some of these factors include the epoch and the dataset
ratio. However, there are other factors that are constant. Other settings are define
with the same value for this classification. These settings include an architecture
with 2 convolutional layers, 3-layer pooling, 2 fully connected layers, there are 32
filters, a 3x3 filter/kernel that gathers size of 2x2 and finally soft-max as the classifi-
cation layer.

1.3.1 VGG-16

A ConvNet is also a naming convention for a CNN, which is a kind of ANN. An

input layer, many hidden layers and an output layer make up a CNN. One of the
top CV models to till date is the CNN variant also known as VGG-16. This models
analysed with the networks of model using an architecture with 3*3 convolutional
filters, which demonstrated a conventional advancement over the setup that have
state of the art convection. The depth was increased to 16–19 weight layers, yielding
around 138 trainable parameters.

Indian Institute of Information Technology, Allahabad 4

 CHAPTER 1. INTRODUCTION

We trained the VGG-16 model using data from X-rays of normal, covid,pneumonia
and lung opacity, as you can see in below Figure 1.2 our VGG-16 consists of thir-
teen conv layers and three fully connected layers, with the 3*3 kernels are combined
with convolutional and the pooling layers with 2*2 parameters. Blocks 1–5 of the
VGG-16 convolutional and pooling layers each comprise several conv layers along
with kernels and pool layers. Each of the 2-convolutional layers in block-1 uses 16
(filter)kernels to extract features and the pooling layer use to reduces the size of the
images. The design of succeeding blocks is identical, with the exception that blocks
1 and 2 uses 2 conv layers, whereas blocks 3–5 employ these 3 conv layers, each
with a different (filter)kernel number, to enhance the network and boost overall ac-
curacy of the model. Three completely interconnected layers combine and output
characteristics into two classes in the end.

Indian Institute of Information Technology, Allahabad 5

 CHAPTER 1. INTRODUCTION

Figure 1.2: VGG-16 Architecture

Indian Institute of Information Technology, Allahabad 6

 CHAPTER 1. INTRODUCTION

1.3.2 VGG-19

The ”VGG-19 Neural Network” is referred to as a particular kind of neural net-

work that has been specifically trained over more huge number of images of different
classes from huge ”ImageNet” data. However, the ”VGG-19 Neural Network” dif-
fers from the ”VGG-16 Neural Network” in that it has 19 layers as opposed to 16
layers.

To expand the depth, VGG-19 just employs a 3 3 convnet set up as seen above.
Max-pool layers are used as a handler to reduce the dimension of feature frame.
There are two FCN layers, each of which has 4096 neurons. To lower the amount
of false positives, all sorts of lesions were taken into consideration while evalu-
ating the VGG, which is trained based on individual lesions. Convolution layers
apply the convolution algorithm on pictures at the pixel level, enabling the results to
flow through to the next layer. The 3x3-dimensional filters utilised in the convolu-
tion layer have been trained for feature extraction. Rectified Linear Unit stands for
(ReLU) layer and pool layer are then added to each stacked convolution layer. The
most well-known non-linear activation function that only accepts positive input is
ReLU at the moment.

ReLU is highly successful identify the optimal convergence behaviour that elimi-
nates the gradient problem when compared to the sigmoid function. After the ReLU
activation function, a down-sampling max-pooling layer is utilised. The filter with
a 2 2 dimension is often regarded as having the same step size. Every sub-region
will get the most value from the output. The activation function has to be created
for dense layer. In order to provide the dropout layer zero value, it was abandoned
during random layer activation. During the training phase, neurons are removed at
random to alleviate the problem of overfitting. This dropout is used when being
trained. below in Figure 1.3 is the architecture of VGG19.

Indian Institute of Information Technology, Allahabad 7

 CHAPTER 1. INTRODUCTION

Figure 1.3: VGG-19 Architecture

Indian Institute of Information Technology, Allahabad 8

 CHAPTER 1. INTRODUCTION

1.3.3 DenseNet

Deep learning is best for layer by layer feature extraction from the image data, ab-
straction of the features, and recognition, detection function implementation. Deep
learning was utilised by Krizhevsky et al. to win the ImageNet CV competition.
CNN, which reflect DL techniques, are becoming a common study area in pattern
detection or recognition. Due to its new architecture which is somehow different
form ResNet where it has dense connectivity in extra, DenseNet has outperformed
many other deep learning models in terms of results. Its interconnection among ar-
bitrary layers, skip connection mode, improved feature transfer to feed forward, and
feature will get reuse between network layers all contribute to a more condensed
network with fewer redundant features. Similar layer depth improves network con-
vergence performance, mitigates network degradation and gradient disappearance
issues brought on by convolutional network deepening, and considerably reduces
the total parameters of the netwok and computing efficiency.

Figure 1.4: DenseNet Architecture

Indian Institute of Information Technology, Allahabad 9

 CHAPTER 1. INTRODUCTION

Figure 1.5: DenseNet Architecture

structure using DenseNet. Figure 1.4 (A) Dense block, some of the transition
layer then conv layer, and lastly fully connected layer make up the foundation of
DenseNet. Figure 1.5 (B) Denseblock is a collection of densely linked dense units
with non-linear functions of mapping of Convolutional, BN and ReLU that are cre-
ated using a pre-activation technique to facilitate network training and improve gen-
eralisation performance. A large amount of features get generated through less
number of convolutional filters/kernels, and the final layer of DenseNet model is

Indian Institute of Information Technology, Allahabad 10

 CHAPTER 1. INTRODUCTION

relatively in scale because the input of each dense unit is combined with all of the
outputs of the preceding dense units, and newly features must also be carry forword
to next dense units. This method allows features of a dense block to get used effec-
tively. Figure 1.5 (C) The transition layer, which is the work as wall to each adjacent
dense blocks, is made up of a 1 - 1 conv layer and a 2 - 2 avg pool layer. It com-
presses the input from the dense block and all the feature information that have been
extracted so far, which reduces the size and as well as dimensionality of the feature
array and, in turn, reduces dense block parameters and guards against overfitting
the network. By including feature information in features network, the classification
prediction layer, which is the top layer with full connectivity, lessens the impact of
feature location on classification and classifies feature after weighing.

1.3.4 Inception-V3

Consideration is given to the CNN Inception-V3 model, and retraining of the net-
work is performed with our own acquired datasets. Figure 1.6 depicts the Inception-
V3 model that has been presented. This model is made up of the following com-
ponents: AvgPool, MaxPool, Convolution, Concat Layer, Fully Connected layer,
Dropout, and Softmax Function. The weights sharing and sparse connectivity offer
the characteristics of the convolution layer. This layer calculates the outcome of the
neuron that is related with present local areas from the layer that came before it. In
addition to this, it correlates to kernels at the same layer and shares the weights of
the neurons that are located underneath the separate feature maps. A fully connected
network layer is used in the portion of the classification that accepts the feature
outputs from Inception-V3 as well as the segmented feature that was specifically
produced by the user.

Indian Institute of Information Technology, Allahabad 11

 CHAPTER 1. INTRODUCTION

Figure 1.6: Inception-V3 Architecture [46]

Two primary components make up the CXR image classification model. These
components are the feature extraction process, which makes use of CNN, and the
classification process, which makes use of the softmax layer and the fully connected
layers. The input data is convolved using the convolution function, which then pro-
duces feature maps. This function accepts data as input.

Figure 1.7: Inception-V3 Architecture

Indian Institute of Information Technology, Allahabad 12

 CHAPTER 1. INTRODUCTION

Through the use of the pooling function, the feature resolution is lowered in order
to make the convolutionally generated features noise-resistant. The two forms of
pooling operations are avg-pooling and max-pooling, as shown in Figure 1.7. Here,
average pooling is a two-dimensional (2D) function with a size of (8, 8) that reduces
the computing burden while decreasing the variance of the data. It also permits
passing the output to the subsequent layer for the subsequent process. Additionally,
max pooling, a 2D pooling function, lowers computational cost and data variance.
When it comes to feature extraction, maximum pooling extracts significant edge
features whereas average pooling extracts smooth features.

Indian Institute of Information Technology, Allahabad 13

 CHAPTER 2. LITERATURE REVIEW

Chapter 2

LITERATURE REVIEW

Numerous studies on medical radiography analysis and AI pertaining to COVID-

19 and how do we identify it have been published since the COVID-19 pandemic first
surfaced. The available CXR image datasets to identify COVID-19 affected patient
through binary class classification on given dataset which has 2 classes COVID-19
or Normal and along with the classification on multi-class which are respectively
COVID-19, Normal, and pneumonia, along with the cutting-edge DL models used
to classify the Chest XR images, will be the two main areas of focus in this literature
review. Recent surveys on the released datasets and deep learning models for CXR
pictures have been published by a number of researchers [8, 9]. The best ML and
DL methods for utilising CXRs to help in diagnose of COVID 19 are listed by Alafif
et al. in their summary [9].

Different work and research has been done in the past for early detection of
COVID-19. Many algorithms have been tried to different datasets to check if a
better and perfect COVID-19 detection method could be found. Several research
and papers were analysed in order to reduce the research gap. The literature review
is divided into different sections which are important and influential for our research.

In paper [47] author uses CNN for classification and used CXR based images
that have 3 classes, covid, penumonia and lung opacity and they got test accuracy
of 82.6%. In [4] author used VGG16, DenseNet121, ResNet50, and ResNet152 on
CT-scan images to perform binary classification. These algorithms are more useful
for COVID-19 pneumonia detection rather than for any other lung issues. These al-
gorithms were applied on two different datasets to find several results. The datasets
used were huge Dataset of (OCT)Optical Coherence Tomography and CXR Images
and SARS-CoV-2 CT scan dataset. These are very small datasets which contain a
small number of OCT and CT scan images, hence provides an accuracy of 98-99%.

Indian Institute of Information Technology, Allahabad 14

 CHAPTER 2. LITERATURE REVIEW

The results of this study can help radiologists and virologists in finding more com-
plex lung diseases in different datasets. Apart from that, it also concluded that with
the help of these models, when applied on more public data sets can help radiologists
to find the COVID virus much faster in the lungs

With 196 photos (COVID-19 = 105, SARS = 11, normal = 80) used to train a
DeTraCResNet18-based binary model, Abbas et al[10] .’s analysis was able to detect
COVID with 95.12% accuracy with 97.91% sensitivity also having 91.87% speci-
ficity, and 93.36% precision. In order to detect COVID, Maguolo and Nanni [11]
gathered 338,270 pictures (145 images of COVID-19, 338 images of pneumonia),
using the AlexNet algorithm, and achieved a 99.97% AUC. ResNet-50 was pre-
trained on 455 chest X-rays by Hall et al. [12]. (CXR images which conatin 135
of COVID-19 and 320 of pneumonia). They attained 95% AUC and 89.2% accu-
racy. 3905 CXR pictures (450 images of COVID-19, 3455 mixed images other then
COVID) were gathered by the authors of [13], who then used pre-trained MobileNet-
v2 to classify them with an accuracy of 99.18%. With the use of a pre-trained
(CNN+HOG) + VGG19 model, Alam et al. [14] trained on 5090 pictures (1979
images of COVID-19, 3111 of non-covid), and their accuracy was 99.49%. In [15],
the authors gathered 6926 photos (2589 COVID-19, 4337 non-covid), classified the
images using a CNN, and attained the accuracy of 94.43%. In [16], the authors
trained a CNN model using TL on 610 pictures (305 COVID-19 and 305 normal),
and they attained a 97.4% accuracy rate. A deep learning model CoreDet with 99.1%
accuracy was proposed by authors in [17] using 900 pictures (500 COVID-19 and
400 normal). In [18], the authors gathered 3251 photos (371 images of COVID-
19, 2882 images of normal), classified them using the AlexNet model, and got an
accuracy of 99.16%.

VGG16, VGG19, InceptionResNet V2, InceptionV3, and Xception were com-

pared by the authors of [19]. They gathered 327 pictures (125 COVID-19, 152
normal, and 50 images of pneumonia), and 84.1% accuracy, 87.7% sensitivity, and
AUC was 97.4%. 16,756 CXRs images were introduced by the authors in [20].
(358 images of COVID-19 cases, total of 8066 patient that were with no pneumonia,
and 5530 patients who was suffering from non-COVID-19 lung pneumonia). They
honed COVID-Net to attain 91.0% sensitivity and 92.4% accuracy. The authors in
[21] employed CNN to obtain 94.03% accuracy using 2971 pictures (285 images of
COVID-19, 1341 images of normal, and 1346 images). Parallel-dilated CNN was
used to train 2905 pictures (218 of COVID-19, 1340 of normal, and 1340 of pneu-

Indian Institute of Information Technology, Allahabad 15

 CHAPTER 2. LITERATURE REVIEW

monia), and Chowdhury et al. [22] attained an accuracy of 96.58%. A dataset of

2700 pictures (900 images of COVID, 900 normal image, and 900 images of pneu-
monia) was proposed by Murugan et al. [23], along with an E-DiCoNet model that
had an accuracy rate of 94.07%. In [24], the authors gathered total of 6100 photos
(225 photos of COVID-19, 1583 photos of normal, and 4292 photos of pneumo-
nia) and used CNN to classify the images, achieving 98.50% accuracy. CoreDet
was used by Hussain et al. [17] to classify total of 1300 pictures (500 photos of
COVID-19, 400 photos of normal, and 400 photos of pneumonia) with 94.2% ac-
curacy. Using 7331 pictures (371 COVID-19, 2882 normal, and 4078 pneumonia).
[18] they also suggested a multi-class categorization and trained AlexNet to reach
94.00% accuracy.

In[5] author used DL methods for COVID-19 image detection with the help of
lung ultrasound imaging which works as an alternative when there is no access to CT
scans or chest XR images. In [6] the author conducted a survey and applied different
deep learning models on many different data sets available publicly. These deep
learning models are applied separately on Chest XR images and CT scan images.

In [15], the author has used different CNN models to perform two binary clas-
sifications to check whether it gives a better result and accuracy or not when com-
pared with a multi-class classification method. Two binary classification method first
checked the COVID and NON-COVID images and then it further classified normal
and pneumonia using the NON-COVID images. It performed better than one multi-
class classification method and achieved 98.38% of accuracy.

For image data, deep learning techniques are very top-notch. These techniques
are used to create medical images. These are also used for classification and detec-
tion of these images as well. The CNN detects Pneumonia in CXR im- ages with
best accuracy. The results show above 95% accuracy. Apart from Pneumonia detec-
tion, this architecture also showed good results in other chest problems like bacterial
infections, viruses, and viral, etc

In [8] a latest classification method grounded on deep CNN is introduced for

COVID-19 detection from CXR. This model achieved 96% accuracy. In [9], author
again conducted a systematic review in which 37 different deep learning architec-
tures were applied to study how these deep learning architectures would fare when

Indian Institute of Information Technology, Allahabad 16

 CHAPTER 2. LITERATURE REVIEW

applied on CT scan and (X-ray)XR image data sets and how these radiology images
are the future for better detection on COVID-19 images.

In [10] author has applied federated learning model called electronic medical
record(EMR) chest X-ray AI model. Using vital signs, laboratory data and CXR
as inputs this model predicted the future oxygen requirements for COVID patient.
Data was collected from 20 different institutes. The data is not publicly available
and it was made sure that no data is shared on the centralized server as well as with
any other institute.

Deep learning algorithms, especially CNNs, have gotten a lot of attention. Al-
though CNN-based algorithms accomplish automated feature extraction and excel-
lent results, they often need more samples of training and computing com- plexity is
more compared to traditional methods.

The majority of the studies that were analysed for this literature review diagnosed
COVID-19 based on chest XR pictures and DL. This demonstrates the significance
of chest XR pictures in the diagnosis of virus like COVID-19 and assists medical
professionals in the expeditious detection of COVID from chest XR. However, we
discovered several drawbacks in prior efforts, such as unbalanced datasets and less
amount of COVID CXR photos to classify them. Both of these shortcomings have
a substantial influence on the performance and the outcome of these methods and
create a wrong impression on their level of accuracy. This study introduces a novel
pre-trained DL model that based on CNN using transfer learning, which we then ap-
ply to VGG-19, Inception-V3, VGG-16 and Dense-Net201and the results are quite
promising. In addition, we make use of image augmentation techniques in order to
build a dataset which is balanced that is capable of overcoming the constraint posed
by unbalanced datasets. Last but not least, in order to circumvent the issue of inade-
quately sized datasets, we compiled a massive one consisting of 14800 photos (3700
normal, 3700 pneumonia, 3700 COVID-19 and 3700 lung opacity) and used deep
learning to categorise them. We perform both binary and multi class classification

Indian Institute of Information Technology, Allahabad 17

 CHAPTER 3. PROBLEM FORMULATION

Chapter 3

PROBLEM FORMULATION

To end the spread of COVID, a significant number of people who may be infected
with the virus need to undergo testing so that they may be properly isolated and
treated. Pathogenic research facility testing represents the suggested highest quality
level, despite the fact that it is laborious and frequently yields significant false neg-
ative results. We need analytical tactics that are both swift and accurate in order to
combat the illness. In light of the radiographic changes that COVID-19 causes in XR
images, we planned to develop a DL methods with transfer learning that would be
able to extract characteristics of COVID CXRs. This would enable us to provide a
clinical analysis in advance to saving valuable time that could be better spent on dis-
ease control. A method of classification based on machine learning is utilised in this
study in order to categorise the X-ray pictures of the chest. It is possible to enhance
the accuracy of the current methodology and models, which is the most important as-
pect of this problem, by imposing transfer learning on that because less data in each
class, increasing the number of iterations, as well as the number of photos used for
training the models by augmenting images through augmentation. Accuracy is the
most important aspect of this problem. Google’s Inception V3 , VGG-16, VGG-19
and DenseNet are the CNN based architectures.

Indian Institute of Information Technology, Allahabad 18

 CHAPTER 4. PROPOSED METHODOLOGY

Chapter 4

PROPOSED METHODOLOGY

4.1 Synopsis

In this work, DL is suggested for the COVID-19 automated diagnostic.The CNN

is the DL model utilised in this investigation (CNN). A Multi-Layer Perceptron
(MLP) advancement called a (CNN) is made to analyse two-dimensional or pic-
ture input. The outcomes of chest x-ray image data will be categorised using this
model. The information compiled in a dataset from Kaggle, Github and from many
sources is split into four categories: the COVID-19-positive patient’s CXR picture,
the patient’s healthy CXR image, patient with Pneumonia and patient with Lung
Opacity. The model will be trained and evaluated using this dataset by looking for
the COVID classification that gain the highest accuracy. This COVID detection and
classification method has not been able to take the place of doctors or other health-
care professionals in determining whether a particular patient is positive or negative
for the COVID-19 virus. But in addition to saving time, the availability of this sys-
tem can also aid medical professionals in scoring and diagnosing patients based on
the interpretation of CXR that have been completed by patients.The goal of this
study is to help healthcare providers and facilities cope with Covid-19 affected pa-
tients, particularly with regard to the use of gained score that may be utilised as a
filter to detect decease like COVID-19.

We are going to present an automated covid patient prediction model in Figure 4.1
to categorise COVID-19 by making use of (CNN), with the help of CXR pictures.
This model is totally utilised because, according to certain studies that employed
models that were comparable to this one, it provided high accuracy performance.
Additionally, this model was built primarily to handle two-dimensional data or pic-
ture data, and it is suited for substantial data. In the next chapter, we will discuss a
dataset consisting of CXRs pictures. In order to categorise CXRs pictures, the sug-

Indian Institute of Information Technology, Allahabad 19

 CHAPTER 4. PROPOSED METHODOLOGY

gested model places a greater emphasis on locating the appropriate parameters for a
CNN architecture network with making use of any pre-trained models and making
use of extra methodologies like transfer learning.

In this, we have four different types of chest x-ray scans, and they are labelled
as (0) for Covid-19, (1) for Normal, (2) for Pneumonia and (3) for Lung Opacity.
Therefore, we need to categorise them in a way that yields effective outcomes with
efficient accuracy. In this article, we improve the effectiveness of convolutional neu-
ral networks by applying the notion of transfer learning. As a result of the fact that
these models need additional data in order to be trained, we have resorted to transfer
learning in order to get around this limitation. Specifically, we make use of pre-
trained CNN models and then train our primary data on a classification layer that we
have developed on our own. Image augmentation was another strategy that helped
us get around the data image issues. We came up with some pictures just for the
classes with augmentation.

Figure 4.1: base working architecture [45]

Indian Institute of Information Technology, Allahabad 20

 CHAPTER 5. TRAINING PROCESS & MODELS

Chapter 5

Training Process & Models

5.1 Experimental Setup

5.1.1 Python 3.7

The programming language utilised for appearing the recommended method is

Python. Python is now a mainstream coding language that changed into made via
way of means of Guido van Rossum in 1991. It observes use in a superb deal of im-
prove topics with many devices gaining knowledge of undertakings too. The Python
network have created several programs for helping software program engineers with
coping with unique capabilities in device gaining knowledge of. The more a part
of the famous device gaining knowledge of capabilities can be actualized using best
one line of code. Because of its straightforwardness with to this point realised li-
braries, python is typically applied for plenty regions of device gaining knowledge
of. Repositories particularly anaconda, mini conda, and so on are handy that incor-
porates the bulk of the libraries which might be essential. Hence, the set up and
coding system is extraordinarily easy for software program engineers.

5.1.2 Sklearn

Scikit-study is a python module that’s applied to execute more than one device
mastering capabilities. This is primarily based totally on NumPy, SciPy and Mat-
plotlib modules. Scikit-study offers clear in conjunction with powerful capabilities
for more than one supervised and unsupervised mastering algorithms. The additives
utilised from Scikit-study are random wooded area classifier, confusion matrix, class
document and accuracy rating from metrics for plotting and getting the results

Indian Institute of Information Technology, Allahabad 21

 CHAPTER 5. TRAINING PROCESS & MODELS

5.1.3 Pandas

This has been the majorly utilised module in facts analysis. This offers enor-
mously more suitable execution code carried out in C or Python. The more than
one method through which the facts will be tested with Pandas is Series and Data
Frames.

5.1.4 Numpy

NumPy is the maximum large module important for handling the bulk of the log-
ical processing inside python. It is usually applied for dealing with arrays and its
computations NumPy resources a multi-dimensional array item having extraordinar-
ily advanced execution along with capabilities critical for handling this item. These
facts systems are extraordinarily beneficial for systematic computation of excessive
dimensional records.

5.1.5 Google Collaboratory

This is unfastened cloud software provisioned through Google which toughen un-
fastened GPU. This bolsters Python three.6 which has been best in executing pro-
found deep mastering modules particularly Keras, Tensor Flow, OpenCV and Py-
Torch. This is installed on Jupyter Notebook subsequently smooth for operating

5.2 Dataset Description

We employ chest X-ray scans from normal people, those with pneumonia, peo-
ple with COVID-19 and as well Lung Opacity for this investigation. We establish
Dataset by collecting eleven publicly accessible sub-databases. This dataset con-
tains a approx 2000 COVID-19 pictures before image augmentation, 3700 images
of pneumonia, 3700 images of normal chest x-rays and 3700 of lung opacity. We
combine and modify eleven distinct sets of publicly available data, including those
from the dataset with name as ChestX-ray8 [25], the CXR Images of pneumonia
dataset [26], the BIMCV-COVID-19 CXR dataset [27], the COVID-19 CXRs Col-
lection [28], the COVID -19 CXR Dataset [29], the Actual-Med COVID-19 CXR
Dataset Initiative [30], the SIRM CXR COVID Dataset [32], COVID-CXNet [34],
MOMA- Dataset [35], and COVID-19 Image Repository [33]. These eleven datasets
were selected and they are freely accessible to academics and they are openly pub-
lished.

Indian Institute of Information Technology, Allahabad 22

 CHAPTER 5. TRAINING PROCESS & MODELS

Figure 5.1: Normal CXR image samples

Figure 5.2: Covid CXR image samples

5.3 Image Pre-Processing and Augmentation

We make use of image augmentation in order to enhance the performance of DL

models when applied to datasets with less CXR images and to build a dataset that
is more balanced. Following the augmentation of the data, the number of COVID
images was raised to 3700 from the previous 2000. In order to produce the pictures,
we make use of three different image enhancement techniques as follows distor-
tion,rotation and flipping image. After completing the rotation process for picture
augmentation, which involves rotating the photos either in the direction of clockwise

Indian Institute of Information Technology, Allahabad 23

 CHAPTER 5. TRAINING PROCESS & MODELS

Figure 5.3: Pneumonia CXR image samples

Figure 5.4: Lung Opacity CXR image samples

or counterclockwise by a max of 10 deg, we then randomly make damage the image.

In the final step, we invert both the horizontal and the vertical orientation of the pic-
tures with a probability of half. Also provides an illustration of the dataset, including
examples of normal, pneumonia, and COVID-19 pictures. Before we begin training
on the photos, we do certain preprocessing processes on them to get them ready for
categorization. We start by converting the photographs to grayscale, then we resize

Indian Institute of Information Technology, Allahabad 24

 CHAPTER 5. TRAINING PROCESS & MODELS

them to 224 pixels by 224 pixels before adding them to the feature array. In order to
train, test and validate the models, we separate the photographs into four categories:
Normal, COVID-19, pneumonia(lung infection) and Lung Opacity images, and we
label them with the numbers 0, 1, 2 and 3. In the final step, we apply a one-hot
encoding method to each of the labels.

5.4 Pre-Trained Deep Learning Models

For our trials, we used the VGG16, VGG19, DenseNet201, and Inception-V3
deep learning classifiers. TensorFlow and Keras libraries contain all the models.
These models serve as our foundational models, and we give each one a fresh, un-
trained brain. A convolutional neural network (CNN) design called VGG16[36] has
a pool layer that is repeated three times, a pool layer, and 3 conv filter layers (3
3). Then, 1 pool layer and 3 conv kernel layers (3 3) were performed twice. The
head of the architecture, which includes three completely linked layers and SoftMax
output, is the last component. The convolutional neural network (CNN) architecture
VGG19[36] consists of 1 pool layer repeated 3 times, 2 conv filter/kernel layers (3
3), and 2 convolution kernel layers. Then, one pooling layer and four convolution
filter layers (3 3) were repeated twice. The head of the architecture, which includes
three completely linked layers and softmax output, is the last component. A highly
linked convolutional network is DenseNet201[37]. DenseNet201’s architecture is
divided into four sections. The first one has a layer of 7 7 convolution, a layer of 2 2
max-pool,then a dense block of 1*1 conv, and a layer of 3*3 conv layers repeated 6
times.Second section consists of a thick block of 1*1 conv layer and 3*3 conv layer
comes again and again 12 times, followed by a layer of 1*1 conv layer and a layer
of 2*2 max-pool. The second portion’s layers are comes again and again 48 times
in the third part. The same layers from sections 2 and 3 are repeated 32 times in the
fourth segment. Classification layers, often known as the head layer, are made up of
1000 completely linked layers with SoftMax and a 7 by 7 global average pooling.
Convolutional neural network (CNN) architecture Inception-V3.

5.5 Transfer Learning

TL is now one of the most effective methods utilised in the field of DL. It is
possible for us to train our model using smaller datasets in a shorter amount of
time by receiving knowledge from models that have been pre-trained using larger

Indian Institute of Information Technology, Allahabad 25

 CHAPTER 5. TRAINING PROCESS & MODELS

datasets and passing that information to our model. Because there are typically only
limited datasets accessible for medical data, such as photographs, this scenario arises
rather frequently. We can train DL models on relatively limited datasets without
risking of over-fitting if we make use of transfer learning. Some of the last layers
of the existed pre-trained network get removed, which is a necessary step in order
to accommodate the new categorization challenge. After that, we remove it and
replace it with our own newly defined layers that correspond to the new classes that
our problem presents. In addition to this, we change the network as fully connected
layer’s dimension to 64 and the dropout layer to 0.5. The average pooling size is
now 4 by 4. The topmost layer is made up of two different class heads, which are the
standard chest X-ray pictures as well as the COVID CXR images that have a binary
cross-entropy loss function. In addition, we apply a loss function that is categorical
cross-entropy loss function on the pictures of normal chest X-rays, pneumonia, lung
opacity and COVID-19 to produce a third and final layer of classification. This layer
is composed of four class heads. Below Figure 5.5 provides an in-depth look at the
head that was added to the existed pre-trained models.

Indian Institute of Information Technology, Allahabad 26

 CHAPTER 5. TRAINING PROCESS & MODELS

Figure 5.5: The proposed head architecture of pre-trained models

Indian Institute of Information Technology, Allahabad 27

 CHAPTER 6. RESULT AND ANALYSIS

Chapter 6

Result and Analysis

6.1 Binary class classification Evaluation

6.1.1 VGG-16 Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.1 and 6.2, we see the validation accuracy is 94.3% and train accuracy is 97%.
also see the loss vs validation loss as shown in the figure 6.2. With VGG-16 model,
we plot data with mentioned dataset where from whole dataset we took 80% as train
data and 20% test data , and from train data we took 80% as train validation data and
20% test validation data. we use transfer learning for our model with this, our model
was able to predict right 1786 data out 1900 test data and it predicated 114 wrong
data as our test data is of 1900 images of binary classes and we get the test accuracy
of 94%. Below is the Figure 6.3 random output that our model predicted between
normal and covid.

Figure 6.1: Accuracy vs Val Accuracy with 15 Figure 6.2: Loss vs Val-loss with 15 epochs
epochs

Indian Institute of Information Technology, Allahabad 28

 CHAPTER 6. RESULT AND ANALYSIS

Figure 6.3: Predicted Outcomes by VGG-16

6.1.2 VGG-19 Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.4 and 6.5, we see the validation accuracy is 93.2% and train accuracy is 95.7%.
also see the loss vs validation loss as shown in the figure 6.5. With VGG-19 model
above we plot data with mention dataset where from whole dataset we took 80% as
train data and 20% test data , and from train data we took 80% as train validation
data and 20% test validation data. we use transfer learning for our model with this,
our model was able to predict right 1783 data out 1900 test data and it predicated 117
wrong data as our test data is of 1900 images of binary classes and we get the test
accuracy of 93.83%.Below is the Figure 6.6 random output that our model predicted
between normal, covid.

Indian Institute of Information Technology, Allahabad 29

 CHAPTER 6. RESULT AND ANALYSIS

Figure 6.4: Accuracy vs Val Accuracy with 15 Figure 6.5: Loss vs Val-loss with 15 epochs
epochs

Figure 6.6: Predicted Outcomes by VGG-19

Indian Institute of Information Technology, Allahabad 30

 CHAPTER 6. RESULT AND ANALYSIS

6.1.3 DenseNet Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in fig-

ure 6.7 and figure 6.8, we see the validation accuracy is 96.8% and train accuracy is
97.8%. also see the loss vs validation loss as shown in the figure 6.8.With DenseNet
model above we plot data with mention dataset where from whole dataset we took
80% as train data and 20% test data , and from train data we took 80% as train val-
idation data and 20% test validation data. we use transfer learning for our model
with this, our model was able to predict right 1826 data out 1900 test data and it
predicated 74 wrong data as our test data is of 1900 images of binary classes and we
get the test accuracy of 96.1%.Below is the Figure 6.9 random output that our model
predicted between normal, covid.

Figure 6.7: Accuracy vs Val Accuracy with 15 Figure 6.8: Loss vs Val-loss with 15 epochs
epochs

Indian Institute of Information Technology, Allahabad 31

 CHAPTER 6. RESULT AND ANALYSIS

Figure 6.9: Predicted Outcomes by DenseNet

6.1.4 InceptionV3 Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in fig-

ure 6.10 and figure 6.11, we see the validation accuracy is 94.5% and train accuracy
is 97.3%. also see the loss vs validation loss as shown in the figure 6.11.With Incep-
tionV3 model above we plot data with mention dataset where from whole dataset
we took 80% as train data and 20% test data , and from train data we took 80% as
train validation data and 20% test validation data. we use transfer learning for our
model with this, our model was able to predict right 1797 data out 1900 test data and
it predicated 103 wrong data as our test data is of 1900 images of binary classes and
we get the test accuracy of 94.57%.Below is the Figure 6.12 random output that our
model predicted between normal, covid.

Indian Institute of Information Technology, Allahabad 32

 CHAPTER 6. RESULT AND ANALYSIS

Figure 6.10: Accuracy vs Val Accuracy with 15 Figure 6.11: Loss vs Val-loss with 15 epochs
epochs

Figure 6.12: Predicted Outcomes

On the training and validation data, we train four pre-trained models with binary
class classification, where label 0 is for normal and label 1 for COVID-19. Accuracy
and loss function on the training data and validation data for the all models are men-
tioned above. The DenseNet201 model exhibits highest The training set’s accuracy

Indian Institute of Information Technology, Allahabad 33

 CHAPTER 6. RESULT AND ANALYSIS

Model Epochs Train Accuracy Val-Accuracy Test-Accuracy

VGG-16 15 97 94.3 94
VGG-19 15 95.7 93.2 93.83
DenseNet 15 97.8 96.8 96.1
InceptionV3 15 97.5 94.5 94.57
Table 6.1: Binary class classification evaluation with different models

Figure 6.13: Accuracy vs Val Accuracy with 15 Figure 6.14: Loss vs Val-loss with 15 epochs
epochs

is 97.8% among all and the validation data correctness is 96.8% and it also set high-
est test accuracy of 96.1% with binary class classification on covid and non-covid
data. This finding suggests that the model acquire sufficient knowledge throughout
its learning process to accurately predict the outcomes. This outcome shows that the
model can correctly categorise the validation data using the knowledge it has gained
from the training set and also predicts right on unseen data.

6.2 Multi class classification Evaluation

6.2.1 VGG-16 Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.13, we see the validation accuracy is 85% and train accuracy is 89%. also see
the loss vs validation loss as shown in the figure 6.14.With VGG-16 model above we
plot data with mention dataset where from whole dataset we took 80% train data and
20% test data , and from train data we took 80% as train validation data and 20% test
validation data. we use transfer learning for our model with this, our model was able
to predict right 2551 data out 2926 test data and it predicated 375 wrong data as our
test data is of 2926 images of all classes and we get the test accuracy of 87.18%. we
also train our model with 30 epochs where we found relatively same train-accuracy

Indian Institute of Information Technology, Allahabad 34

 CHAPTER 6. RESULT AND ANALYSIS

vs val-accuracy graph and also relatively same loss vs val-loss graph which shown
in below figures 6.15 and figures 6.16.

Figure 6.15: Accuracy vs Val Accuracy with 30 Figure 6.16: Loss vs Val-loss with 30 epochs
epochs

Below in figure 6.17 is the random output that our model predicted between nor-
mal, covid, pneumonia and lung opacity.

Figure 6.17: predicted outcomes by VGG-16

Indian Institute of Information Technology, Allahabad 35

 CHAPTER 6. RESULT AND ANALYSIS

6.2.2 VGG-19 Model Evaluation

Figure 6.18: Accuracy vs Val Accuracy with 15 Figure 6.19: Loss vs Val-loss with 15 epochs
epochs

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.18, we see the validation accuracy is 82% and train accuracy is 84%. also see
the loss vs validation loss as shown in the figure 6.19.With VGG-19 model above
we plot data with mention dataset where from whole dataset we took 80% as train
data and 20% test data , and from train data we took 80% as train validation data and
20% test validation data. we use transfer learning for our model with this, our model
was able to predict right 2450 data out 2926 test data and it predicated 476 wrong
data as our test data is of 2926 images of all classes and we get the test accuracy
of 84.45%. we also train our model with 30 epochs where we found train accuracy
as 85% and validation accuracy as 83% in below graph and relatively same loss vs
val-loss graph which is shown in below figure 6.20 and figure 6.21.

Figure 6.20: Accuracy vs Val Accuracy with 30 Figure 6.21: Loss vs Val-loss with 30 epochs
epochs

Indian Institute of Information Technology, Allahabad 36

 CHAPTER 6. RESULT AND ANALYSIS

Below in figure 6.22 is the random output that our model predicted between nor-
mal, covid, pneumonia and lung opacity.

Figure 6.22: predicted outcomes by VGG-19

Indian Institute of Information Technology, Allahabad 37

 CHAPTER 6. RESULT AND ANALYSIS

6.2.3 DenseNet201 Model Evaluation

Figure 6.23: Accuracy vs Validation Accuracy Figure 6.24: Loss vs Val loss with 15 epochs
with 15 epochs

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.23, we see the validation accuracy is 87% and train accuracy is 95%. also see
the loss vs validation loss as shown in the figure 6.24.With DenseNet model above
we plot data with mention dataset where from whole dataset we took 80% as train
data and 20% as test data , and from train data we took 80% as train validation data
and 20% test validation data. we use transfer learning for our model with this, our
model was able to predict right 2548 data out 2926 test data and it predicated 378
wrong data as our test data is of 2926 images of all classes and we get the test accu-
racy of 87.08%. we also train our model with 30 epochs where we found relatively
same train-accuracy vs val-accuracy graph and also relatively same loss vs val-oss
graph which shown in below Figure 6.25 and Figure 6.26.

Figure 6.25: Accuracy vs Val Accuracy with 30 Figure 6.26: Loss vs Val-loss with 30 epochs
epochs

Indian Institute of Information Technology, Allahabad 38

 CHAPTER 6. RESULT AND ANALYSIS

Below in Figure 6.27 is the random output that our model predicted.

Figure 6.27: predicted outcomes by DenseNet

Indian Institute of Information Technology, Allahabad 39

 CHAPTER 6. RESULT AND ANALYSIS

6.2.4 InceptionV3 Model Evaluation

Observing the graph of train-accuracy vs val-accuracy and loss vs val-loss in Fig-

ure 6.28, we see the validation accuracy is 83% and train accuracy is 90%. also
see the loss vs validation loss as shown in the figure 6.29.With InceptionV3 model
above we plot data with mention dataset where from whole dataset we took 80% as
train data and 20% as test data , and from train data we took 80% as train validation
data and 20% test validation data. we use transfer learning for our model with this,
our model was able to predict right 2488 data out 2926 test data and it predicated
438 wrong data as our test data is of 438 images of all classes and we get the test
accuracy of 85%. we also train our model with 30 epochs where we found train
accuracy as 95% and validation accuracy as 83% in below graph and relatively same
loss vs val-loss graph which shown in below Figure 6.30 and Figure 6.31.

Figure 6.28: Accuracy vs Val Accuracy with 15 Figure 6.29: Loss vs Val-loss with 15 epochs
epochs

Figure 6.30: Accuracy vs Val Accuracy with 30 Figure 6.31: Loss vs Val-loss with 30 epochs
epochs

Indian Institute of Information Technology, Allahabad 40

 CHAPTER 6. RESULT AND ANALYSIS

Below in Figure 6.32 is the random output that our model predicted.

Figure 6.32: predicted outcomes by InceptionV3

Model Epochs Train Accuracy Val-Accuracy Test-Accuracy

VGG-16 15 89.3 85 87.18
VGG-19 15 84 82.5 84.45
DenseNet 15 95.2 87 87.08
InceptionV3 15 90.4 83 85
Table 6.2: Multi class classification evaluation with diff models

Indian Institute of Information Technology, Allahabad 41

 CHAPTER 6. RESULT AND ANALYSIS

Model/Method Data-Samples Test-Accuracy

CNN [38] 1124 85
Decision Tree [39] 324 80
CNN [47] 5000 82.6
DenseNet [37] 327 84.1
KNN, Regression [40] 279 82
AI Framework [41] 53 80
Neural Network [42] 5644 82
GNB classifier — 83
RestNet [43] 746 92.15
InceptionRestNetV2 [43] 746 91.35
DarkCovidNet [44] 125 86

Table 6.3: Multi class classification comparison between previous work and our work

We recommended dataset which is balanced with an adequate no. of COVID-19

instances as well as pictures associated with chest x-rays. Even though they achieved
a high accuracy result, a meaningful classification cannot be generated from a dataset
that contains imbalance, and this is especially true if the total COVID-19 pictures in
the dataset is low. However, with a relatively low number of photos in comparison
to our dataset, there is a possibility that the results will be inaccurate. It is also a
cause for concern that our proposed techniques based on transfer learning produced
classification prediction for the multi-class classification and achieved an tain accu-
racy of 95.2% with DenseNet model and get highest test accuracy of 87.18% with
VGG-16 model, whereas other techniques used datasets that were unbalanced, and
also the COVID-19 CXRs are 219 and 225, respectively. This raises a number of
issues. This suggests that the COVID-19 photos are insufficient for the algorithm to
make an accurate determination on their classification.

The method that we have suggested focuses on two elements that distinguish
themselves from the other recent efforts. First, we have observed that the research in
question only includes a limited number of COVID-19 photos and an uneven distri-
bution of datasets derived from the gathered datasets. This issue has an impact on the
results, most notably with the differentiation of COVID-19 from the other classes.
We tackled this issue head-on by utilising a well-rounded dataset that contained a
significant number of COVID-19 pictures. This dataset comprises COVID-19 CXR
pictures that have been augmented with data from eleven various sources. There
are a total of 3,700 normal photos, 3,700 pneumonia images, 3,700 COVID-19 im-
ages, and 3,700 lung opacity images in this dataset. Second, we present a technique
called transfer learning, which is derived from a variety of deep learning algorithm

Indian Institute of Information Technology, Allahabad 42

 CHAPTER 6. RESULT AND ANALYSIS

techniques and has shown some encouraging results. Pre-trained DL models like
VGG16, DenseNet201, VGG19, and InceptionV3 are trained, validated, and tested
by our team. We suggest a final layer, sometimes referred to as the head(top / main)
layer, for the pre-trained DL modelss that achieve excellent performance and suit
our data. We were able to attain an accuracy of 95.2% percent for the multi-class
categorization and test accuracy of 87.18%. Even though we are unable to compare
the published studies owing to the fact that each study employs a unique dataset and
set of algorithms, this issue persists because of the fast change that occurs daily in
the chest XR datasets that are available on online sources. This comparison is pro-
vided to highlight the prior research that has been done in this field, as well as the
improvements we have achieved in comparison to the prior work, and to clarify the
limits of the prior study in order to fix them. Last but not least, the results/outcomes
of this study are comparable to the results of the most recent research.

Indian Institute of Information Technology, Allahabad 43

 CHAPTER 7. CONCLUSION

Chapter 7

Conclusion
The detection of the COVID-19 case more quickly and the prevention of the dis-
ease’s spread on the chest are two benefits of employing CXRs to forecast COVID-
19. Within the scope of this investigation, we use transfer learning to train, validate,
and test four widely used DL algorithms. When it comes to classifying CXR pic-
tures from COVID-19, we put the pre-trained models Inception-V3, DenseNet201,
VGG16, and VGG19 through their paces. According to the findings, with binary
class classification The DenseNet201 model exhibits highest The training set’s ac-
curacy is 97.8% among all and the validation data correctness is 96.8% and it also
set highest test accuracy of 96.1% with binary classification on covid and non-covid
data. For multi class classification the DenseNet model gain the highest training
accuracy from the four models is 95.2% accuracy on the train set and among all
VGG-16 gain the highest test accuracy of 87.18%. This was determined by compar-
ing the models’ performance. As a direct consequence of this, our overall test accu-
racy is 87.18% and the train accuracy of 95.2%. In addition, this research presents
the COVID-Chest-X-ray balanced dataset, which was compiled from eleven distinct
sources and had a total of 3700 normal pictures, 3700 pneumonia images, 3700
COVID-19 CXR images following the application of data augmentation and 3700
lung Opacity. In order to solve the problem of an imbalanced dataset, this dataset has
a drastically increase in the amount of COVID-19 photos compared to earlier stud-
ies. In light of our findings, the dataset that we have provided can assist researchers
in training ML and DL models with dataset which is balanced that contains a signif-
icant number of COVID-19 photos. In addition, the collected data can aid medical
professionals in recognising COVID-19 from the CXR at an initial stage, which will
allow them to make judgments more quickly. For the future, we plan to expand
the proposed data set to include additional CXR pictures that come from a wider
variety of diseases as well as increase the number of images in the dataset. If any
open-source data is already accessible, this will be possible.

Indian Institute of Information Technology, Allahabad 44

 REFERENCES

References

[1] Paules, C.I.; Marston, H.D.; Fauci, A.S, [Coronavirus Infections—More Than
Just the Common Cold ], JAMA 2020, 323, 707.

[2] Coronavirus Cases. Available online: [ https://www.worldometers.info/coronavirus/].

[3] Bell, D.J. [COVID-19: Radiology Reference Article. Available online:

https://radiopaedia.org/articles/COVID-19-4 ].

[4] Rousan, L.A.; Elobeid, E.; Karrar, M.; Khader, Y. [Chest X-ray Findings and
Temporal Lung Changes in Patients with COVID-19 Pneumonia],BMC Pulm.
Med. 2020, 20, 1–9.

[5] Wong, H.; Lam, H.; Fong, A.H.; Leung, S.T.; Chin, T.W.; Lo, C.; Lui, M.M.;
Lee, J.; Chiu, K.W.; Chung, T.W.; et al, [Frequency and Distribution of Chest
Radiographic Findings in Patients Positive for COVID-19. Radiology], 2020,
296, E72–E78.

[6] McBee, M.P.; Awan, O.A.; Colucci, A.T.; Ghobadi, C.W.; Kadom, N.;
Kansagra, A.P.; Trid, A.S.; Auffermann, W.F. [Deep Learning in Radiol-
ogy],Acad. Radiol. 2018, 25, 1472–1480.

[7] Kim, M.; Yan, C.; Yang, D.; Wang, Q.; Ma, J.; Wu, G. [Deep Learning in
Biomedical Image Analysis. In Biomedical Information Technology],Academic
Press: Cambridge, MA, USA, 2020; pp. 239–263.

[8] slam, M.M.; Karray, F.; Alhajj, R.; Zeng, J [A Review on Deep Learning Tech-
niques for the Diagnosis of Novel Coronavirus (COVID-19)] IEEE Access 2021,
9, 30551–30572.

[9] Alafif, T.; Tehame, A.M.; Bajaba, S.; Barnawi, A.; Zia, S[Machine and Deep
Learning towards COVID-19 Diagnosis and Treatment: Survey, Challenges, and
Future Directions.],Int. J. Environ. Res. Public Health 2021, 18, 1117.

Indian Institute of Information Technology, Allahabad 45

 REFERENCES

[10] Abbas, A.; Abdelsamea, M.M.; Medhat Gaber, M[Classification of COVID-

19 in Chest X-ray Images Using DeTraC Deep Convolutional Neural Network],
Appl. Intell. 2021, 51, 854–864.

[11] Maguolo, G.; Nanni, L[A Critic Evaluation of Methods for COVID-19 Auto-
matic Detection from X-ray Images], Inform. Fusion 2021, 76, 1–7.

[12] Hall, L.; Goldgof, D.; Paul, R.; Goldgof, G.M[Finding COVID-19 from
Chest X-rays Using Deep Learning on a Small Dataset], arXiv 2020,
arXiv:2004.02060.

[13] Apostolopoulos, I.D.; Aznaouridis, S.I.; Tzani, M.A. [Extracting Possibly Rep-
resentative COVID-19 Biomarkers from X-ray Images with Deep Learning Ap-
proach and Image Data Related to Pulmonary Diseases], J. Med. Biol. Eng.
2020, 40, 462–469.

[14] Alam, N.A.; Ahsan, M.; Based, M.A.; Haider, J.; Kowalski, M. [COVID-19
Detection from Chest X-ray Images Using Feature Fusion and Deep Learning].
Sensors 2021, 21, 1480.

[15] Duran-Lopez, L.; Dominguez-Morales, J.P.; Corral-Jaime, J.; Vicente- Diaz,

S.; Linares-Barranco, A. COVID-XNet: [A custom deep learning system to di-
agnose and locate COVID-19 in chest X-ray images.] Appl. Sci. 2020, 10, 5683.

[16] Mahmud, T.; Rahman, M.A.; Fattah, S.A. [CovXNet: A multidilation convolu-
tional neural network for automatic COVID-19 and other pneumonia detection
from chest X-ray images with transferable multi-receptive feature optimization.]
Comput. Biol. Med. 2020, 122, 103869

[17] Hussain, E.; Hasan, M.; Rahman, M.A.; Lee, I.; Tamanna, T.; Parvez,
M.Z.[CoroDet: A deep learning based classification for COVID-19 detection
using chest X-ray images.] Chaos Solitons Fractals 2021, 142, 110495.

[18] Ibrahim, A.U.; Ozsoz, M.; Serte, S.; Al-Turjman, F.; Yakoi, P.S.[ Pneumonia
Classification Using Deep Learning from Chest X-ray Images During COVID-
19.] Cogn. Comput. 2021, 1–13.

[19] Moutounet-Cartan, P.G.B. [Deep convolutional neural networks to diagnose

COVID-19 and other pneumonia diseases from posteroanterior chest x-rays.]
arXiv 2020, arXiv:2005.00845.

Indian Institute of Information Technology, Allahabad 46

 REFERENCES

[20] Wang, L.; Lin, Z.Q.; Wong, A. [COVID-Net: A Tailored Deep Convolutional
Neural Network Design for Detection of COVID-19 Cases from Chest X-ray
Images.] Sci. Rep. 2020, 10, 1–12.

[21] Ahammed, K.; Satu, M.S.; Abedin, M.Z.; Rahaman, M.A.; Islam, S.M.S.
[Early Detection of Coronavirus Cases Using Chest X-ray Images Employing
Machine Learning and Deep Learning Approaches.]medRxiv 2020.

[22] Chowdhury, N.K.; Rahman, M.M.; Kabir, M.A. [PDCOVIDNet: A parallel-

dilated convolutional neural network architecture for detecting COVID-19 from
chest X-ray images.] Health Inf. Sci. Syst. 2020, 8, 1–14.

[23] Murugan, R.; Goel, T. [E-DiCoNet: Extreme Learning Machine Based Classi-
fier for Diagnosis of COVID-19 Using Deep Convolutional Network.] J. Ambi-
ent. Intell. Humaniz. Comput. 2021, 12, 8887–8898.

[24] Sekeroglu, B.; Ozsahin, I. [Detection of COVID-19 from Chest X-ray Images
Using Convolutional Neural Networks.] SLAS Technol. Transl. Life Sci. Innov.
2020, 25, 553–565.

[25] Wang, X.; Peng, Y.; Lu, L.; Lu, Z.; Bagheri, M.; Summers, R.M. [ChestX-ray8:
Hospital-Scale Chest X-ray Database and Benchmarks on Weakly-Supervised
Classification and Localization of Common Thorax Diseases.] In Proceedings
of the 2017 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), Honolulu, HI, USA, 21–26 July 2017.

[26] Kermany, D.S.; Goldbaum, M.; Cai, W.; Valentim, C.C.; Liang, H.; Baxter,
S.L.; Zhang, K. [Identifying medical diagnoses and treatable diseases by image-
based deep learning. ] Cell 2018, 172, 1122–1131.

[27] BIMCV. Available online: [https://bimcv.cipf.es/bimcv-projects/bimcv-

covid19/1590858128006-9e640421-6711]

[28] Cohen, J.P.; Morrison, P.; Dao, L.; Roth, K.; Duong, T.Q.; Ghassemi, M.
[COVID-19 image data collection: Prospective predictions are the future.]
IarXiv 2020, arXiv:2006.11988.

[29] Wang, X.; Peng, Y.; Lu, L.; Lu, Z.; Bagheri, M.; Summers, R.M. [ChestX-ray8:
Hospital-Scale Chest X-ray Database and Benchmarks on Weakly-Supervised
Classification and Localization of Common Thorax Diseases.] In Proceedings
of the 2017 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), Honolulu, HI, USA, 21–26 July 2017.

Indian Institute of Information Technology, Allahabad 47

 REFERENCES

[30] Zargari Khuzani, A., Heidari, M. Shariati, S.A. [COVID-Classifier: an auto-

mated machine learning model to assist in the diagnosis of COVID-19 infection
in chest X-ray images.]

[31] Redazione. COVID-19 DATABASE [Available online:

https://www.sirm.org/category/senza-categoria/COVID-19/]

[32] Twitter COVID-19 CXR Dataset. [Available online:

http://twitter.com/ChestImaging/]

[33] Winther, H.B.; Laser, H.; Gerbel, S.; Maschke, S.K.; Hinrichs, J.B.; Vogel-
Claussen, J.; Meyer, B.C. [COVID-19 Image Repository.] Figshare Dataset
2020.

[34] Armiro. [Available online: https://github.com/armiro/COVID-CXNet]

[35] Shams, M.; Elzeki, O.; Abd Elfattah, M.; Hassanien, A. [Chest X-ray images
with three classes: COVID-19, Normal, and Pneumonia. ] Mendeley Data 2020,
V3.

[36] Simonyan, K.; Zisserman, A. [Very deep convolutional networks for large-
scale image recognition.] arXiv 2014, arXiv:1409.1556.

[37] Huang, G.; Liu, Z.; Van Der Maaten, L.; Weinberger, K.Q. [Densely connected
convolutional networks.] In Proceedings of the IEEE conference on Computer
Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017;

[38] A. Waheed, M. Goyal, D. Gupta, A. Khanna, F. Al-Turjman and P. R. Pin-

heiro. [CovidGAN: Data Augmentation Using Auxiliary Classifier GAN for Im-
proved Covid-19 Detection.] n IEEE Access, vol. 8, pp. 91916-91923, 2020, doi:
10.1109/ACCESS.2020.2994762;

[39] Yoo SH, Geng H, Chiu TL, Yu SK, Cho DC, Heo J, Choi MS, Choi IH, Cung
Van C, Nhung NV, Min BJ and Lee H (2020) [Deep Learning-Based Decision-
Tree Classifier for COVID-19 Diagnosis From Chest X-ray Imaging.] Front.
Med. 7:427. doi: 10.3389/fmed.2020.00427;

[40] Brinati, D., Campagner, A., Ferrari, D. et al. [Detection of COVID-19 Infection
from Routine Blood Exams with Machine Learning: A Feasibility Study.] J Med
Syst 44, 135 (2020);

Indian Institute of Information Technology, Allahabad 48

 REFERENCES

[41] Jiang X., Coffee M., Bari A., Wang J., Jiang X., Huang J.J., et al. [Towards
an artificial intelligence framework for data-driven prediction of coronavirus
clinical severity.] Comput. Mater. Continua, 63 (1) (2020), pp. 537-551;

[42] Schwab P., Schütte A.D., Dietz B., Bauer S. [Clinical predictive models for
COVID-19: Systematic study.] J. Med. Internet Res., 22 (10) (2020), Arti-
cle e21439.

[43] Jiang X., Coffee M., Bari A., Wang J., Jiang X., Huang J.J., et al. [Towards
an artificial intelligence framework for data-driven prediction of coronavirus
clinical severity.] Comput. Mater. Continua, 63 (1) (2020), pp. 537-551;

[44] Ozturk T, Talo M, Yildirim EA, Baloglu UB, Yildirim O, Acharya UR (2020)
[Automated detection of COVID-19 cases using deep neural networks with X-ray
images.] Comput Biol Med 121:103792;

[45] Aijaz Ahmad Reshi, Furqan Rustam, Arif Mehmood, Abdulaziz Alhossan,
Ziyad Alrabiah, Ajaz Ahmad, Hessa Alsuwailem, Gyu Sang Choi, [An Effi-
cient CNN Model for COVID-19 Disease Detection Based on X-Ray Image
Classification], Complexity, vol. 2021, Article ID 6621607, 12 pages, 2021.
https://doi.org/10.1155/2021/6621607

[46] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens,

Zbigniew Wojna, [Rethinking the Inception Architecture for Computer Vi-
sion],arXiv:1512.00567.

[47] F. W. Wibowo and Wihayati, [Classification of Lung Opacity, COVID-19, and

Pneumonia from Chest Radiography Images Based on Convolutional Neural
Networks,] 2021 4th International Seminar on Research of Information Tech-
nology and Intelligent Systems (ISRITI), 2021, pp. 173-177, doi: 10.1109/IS-
RITI54043.2021.9702841.

Indian Institute of Information Technology, Allahabad 49

thesis
ORIGINALITY REPORT

7 %
SIMILARITY INDEX
3%
INTERNET SOURCES
4%
PUBLICATIONS
2%
STUDENT PAPERS

PRIMARY SOURCES

1
Mohammad Khalid Imam Rahmani, Fahmina
Taranum, Reshma Nikhat, Md. Rashid Farooqi,
1%
Mohammed Arshad Khan. "Automatic Real-
Time Medical Mask Detection Using Deep
Learning to Fight COVID-19", Computer
Systems Science and Engineering, 2022
Publication

2
Submitted to National Institute of Technology,
Agartala
1%
Student Paper

3
www.mdpi.com
Internet Source 1%
4
Submitted to Monash University
Student Paper 1%
5
Submitted to University of Dundee
Student Paper <1 %
6
WWW.MDPI.COM
Internet Source <1 %
7
clok.uclan.ac.uk
Internet Source <1 %
 <1 %
8
"Intelligent Computing Theories and
Application", Springer Science and Business
Media LLC, 2017
Publication

9
Tao Zhou, XinYu Ye, HuiLing Lu, Xiaomin
Zheng, Shi Qiu, YunCan Liu. "Dense
<1 %
Convolutional Network and Its Application in
Medical Image Analysis", BioMed Research
International, 2022
Publication

10
link.springer.com
Internet Source <1 %
11
mdpi-res.com
Internet Source <1 %
12
Dulani Meedeniya, Hashara Kumarasinghe,
Shammi Kolonne, Chamodi Fernando, Isabel
<1 %
De la Torre Díez, Gonçalo Marques. "Chest X-
ray analysis empowered with deep learning: A
systematic review", Applied Soft Computing,
2022
Publication

13
G. H. G. S. A. D. Dhanapala, S. Sotheeswaran.
"Transfer Learning Techniques with SVM For
<1 %
Covid-19 Disease Prediction Based On Chest
X-Ray Images", 2022 2nd International
Conference on Advanced Research in
Computing (ICARC), 2022
Publication
14
Submitted to Universiti Teknikal Malaysia
Melaka
<1 %
Student Paper

15
bura.brunel.ac.uk
Internet Source <1 %
16
Muzammil Khan, Muhammad Taqi Mehran,
Zeeshan Ul Haq, Zahid Ullah, Salman Raza
<1 %
Naqvi, Mehreen Ihsan, Haider Abbass.
"Applications of artificial intelligence in COVID-
19 pandemic: A comprehensive review",
Expert Systems with Applications, 2021
Publication

17
www.researchgate.net
Internet Source <1 %
18
Dimitrios Kollias, Mihalis A. Nicolaou, Irene
Kotsia, Guoying Zhao, Stefanos Zafeiriou.
<1 %
"Recognition of Affect in the Wild Using Deep
Neural Networks", 2017 IEEE Conference on
Computer Vision and Pattern Recognition
Workshops (CVPRW), 2017
Publication

19
ebin.pub
Internet Source <1 %
20
"Image Analysis for Moving Organ, Breast,
and Thoracic Images", Springer Science and
<1 %
Business Media LLC, 2018
Publication