Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
ScienceDirect
AvailableScienceDirect
Procedia online at Science
Computer www.sciencedirect.com
00 (2022) 000–000
Procedia Computer Science 00 (2022) 000–000 www.elsevier.com/locate/procedia
ScienceDirect www.elsevier.com/locate/procedia
Procedia Computer Science 218 (2023) 357–366

International Conference on Machine Learning and Data Engineering

International Conference on Machine Learning and Data Engineering
A Deep Learning based model for the Detection of Pneumonia from
A Deep Learning based model for the Detection of Pneumonia from
Chest X-Ray Images using VGG-16 and Neural Networks
Chest X-Ray Images using VGG-16 and Neural Networks
Shagun Sharma and Kalpna Guleria*
Shagun Sharma and Kalpna Guleria*
Chitkara University Institute of Engineering & Technology, Chitkara University, Rajpura, 140401, Punjab, India
Chitkara University Institute of Engineering & Technology, Chitkara University, Rajpura, 140401, Punjab, India
Abstract
Abstract
Pneumonia is a viral infection which affects a significant proportion of individuals, especially in developing and penurious
countries where contamination, overcrowded, and unsanitary living conditions are widespread, along with the lack of healthcare
Pneumonia is aPneumonia
infrastructures. viral infection which
produces affects aeffusion,
pericardial significant proportion
a disease whereinof individuals,
fluids fill theespecially in developing
chest and create inhaling and penurious
problems. It is
countries
a difficultwhere
step to contamination,
recognize theovercrowded, and unsanitary
presence of pneumonia livinginconditions
quickly are widespread,
order to receive treatmentalong with and
services the lack of healthcare
improve survival
infrastructures. Pneumonia produces pericardial effusion, a disease wherein fluids fill the chest and create
chances. Deep learning, is a field of artificial intelligence which is used in the successful development of prediction models.inhaling problems. It is
aThere
difficult step to recognize the presence of pneumonia quickly in order to receive treatment services
are various ways of detecting pneumonia such as CT-scan, pulse oximetry, and many more among which the most common and improve survival
chances. Deep tomography.
way is X-ray learning, is aOn field
theofother
artificial
hand,intelligence
examiningwhich is used (CXR)
chest X-rays in the successful
is a toughdevelopment of prediction
process susceptible models.
to subjective
There are various
variability. In thisways
work,ofadetecting pneumoniamodel
deep learning(DL) such as CT-scan,
using VGG16pulse oximetry,
is utilized forand many more
detecting among which
and classifying the mostusing
pneumonia common
two
way isimage
CXR X-ray tomography.
datasets. On thewith
The VGG16 other hand,
Neural examining
Networks (NN)chest X-rays
provides (CXR) isvalue
an accuracy a tough process recall
of 92.15%, susceptible to subjective
as 0.9308, precision
variability.
as In this
0.9428, and work, a deepforlearning(DL)
F1-Score0.937 modelFurthermore,
the first dataset. using VGG16 theisexperiment
utilized forusing
detecting
NN and
withclassifying
VGG16 has pneumonia using two
been performed on
another CXR dataset containing 6,436 images of pneumonia, normal and covid-19. The results for the second datasetprecision
CXR image datasets. The VGG16 with Neural Networks (NN) provides an accuracy value of 92.15%, recall as 0.9308, provide
as 0.9428,recall,
accuracy, and F1-Score0.937
precision, andforF1-score
the firstasdataset.
95.4%,Furthermore,
0.954, 0.954, theand
experiment using NN with
0.954, respectively. The VGG16
researchhas been performed
outcome on
exhibits that
another
VGG16 CXR with dataset
NN providescontaining
better6,436 images ofthan
performance pneumonia,
VGG16 normal and covid-19.
with Support The results
Vector Machine for theVGG16
(SVM), second with
dataset provide
K-Nearest
accuracy,
Neighbor recall,
(KNN),precision,
VGG16 and withF1-score
RandomasForest 95.4%, 0.954,
(RF), and0.954,
VGG16 andwith
0.954, respectively.
Naïve Bayes (NB)The research
for both outcome
datasets. exhibits
Further, that
the
VGG16
proposedwithworkNN resultsprovides
exhibitbetter
improvedperformance
performance thanresults
VGG16 with datasets
for both Support 1Vector
and 2 in Machine (SVM),
comparison VGG16models.
to existing with K-Nearest
Neighbor (KNN), VGG16 with Random Forest (RF), and VGG16 with Naïve Bayes (NB) for both datasets. Further, the
proposed work results exhibit improved performance results for both datasets 1 and 2 in comparison to existing models.
© 2023The
© 2023 TheAuthors.
Authors. Published
Published by by ELSEVIER
Elsevier B.V. B.V.
This
This
© is
2023 anThe
is an open
open access
access
Authors. article
article under
under
Published the BY-NC-ND
CC BY-NC-ND
theELSEVIER
by CC license
B.V. license (https://creativecommons.org/licenses/by-nc-nd/4.0)
(https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review
Peer-review under
under responsibility
responsibility ofof the
the scientific
scientific committee
committee of of
the the International
International Conference
Conference on Machine
on Machine Learning
Learning
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) andEngineering
and Data Data
Engineering under responsibility of the scientific committee of the International Conference on Machine Learning and Data
Peer-review
Keywords: Deep Learning; VGG16; CNN; Pneumonia; Neural Networks; X-ray;
Engineering
Keywords: Deep Learning; VGG16; CNN; Pneumonia; Neural Networks; X-ray;

* Corresponding author. Tel.: +91-8826648937;

* Corresponding author. Tel.: +91-8826648937;
E-mail address: [email protected]
E-mail address: [email protected]
1877-0509 © 2023 The Authors. Published by ELSEVIER B.V.
1877-0509 © 2023 The Authors. Published by ELSEVIER B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and Data Engineering
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and Data Engineering

1877-0509 © 2023 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and
Data Engineering
10.1016/j.procs.2023.01.018
358 Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366
2 / Procedia Computer Science 00 (2019) 000–000

1. Introduction

Epidemics and chronic diseases have claimed the lives of countless individuals throughout historical time,
causing huge crises that took many years to resolve [1][2]. Pandemics and eruptions are two different ways to define
infectious diseases within communities over the years. Similarly, a pandemic is the emergence of more occurrences
of injury, illness, or other medical problems than anticipated in a specific location between many or a particular
group of people over a given period. The majority of the cases are ostensibly connected. The outbreak differs from a
pandemic in that it is more confined and less liable to create public concern. One of the pandemics in the past
includes pneumonia, which is a dangerous disease causing a large number of health issues.
In the United States alone, over 1 million individuals are hospitalized due to pneumonia each year, with nearly half a
million deaths [3]. Apart from this, it was also reported by WHO in 2019 that, pneumonia killed 14% of children
under the age of 5 worldwide. This motivates the expansion of a DL-based model for early pneumonia prediction so
that precious human lives can be saved by providing timely medication and reducing the increased mortality rate
caused due pneumonia. Chest X-ray (CXR) images is a highly effectual and accessible technique for detecting
pneumonia, and serve an essential application in epidemiological studies and healthcare. Pneumonia identification in
CXR images is a difficult and challenging task requiring professional radiologists' presence. It is a respiratory
disease related to the chest due to the spreading of bacteria, infection, and fungus via breathing air [4]. This disease
spreads to the chest, causing more breathing issues. Pneumonia is a leading cause of increased mortality in children
as young as five. The most effective way of pneumonia detection is an X-ray image, which is more cost-effective
and broad rather than other standard diagnostics procedures. Radiologists meticulously identify and analyze the
X-ray images for detecting pneumonia which is a time-consuming process, and might lead to conflict among
radiologists regarding disease prediction. Hence, this issue leads to the development of pneumonia detection
methods that are safe to use in the healthcare department for real-time pneumonia diagnosis without any errors.
In disease diagnostic systems, machine learning (ML), DL and statistical methods are extremely efficient tools
[5][6]. They may be used to solve highly sophisticated vision tasks in the healthcare imaging sector, such as lung
disease classification, lung segmentation, and many more. Recent DL advancements have helped achieve and
perhaps even significantly exceed human performance in many activities. DL can also be utilized to determine the
outcome of treatments, such as chevalier studies and cancer treatment. Labeled data and DL-based algorithms are
linked to encouraging results in thoracic illness categorization utilizing an X-ray image modality. Deep neural
network (DNN) models have traditionally been built and tested by human professionals in a continuous
trial-and-error technique that takes time, resources, and expertise [7].
To address this issue, an innovative model is presented which utilizes DNN architecture to efficiently execute
ideal classification. The proposed model is designed exclusively for the classification and prediction of pneumonia
by utilizing CXR radiographs. The technique works on the basis of neural network (NN) architecture, which uses
several neurons to concatenate, identify and extract significant features from a set of images. Despite the
state-of-the-art models, NN offers similarly focused network architectures of the train and test system, which has
also been their developing premise. NN model has motivated DL-based algorithm to bring a standard choice for
prediction and classifications of healthcare-based image dataset [7].
DL plays an important role while predicting the classification results [8]. Similarly, there are various models in DL
which are used in the healthcare department for disease prediction [9]. But the usage of the DL models is dependent
on the type of data. The used dataset in the paper is in the form of images, which shows the NN model is the suitable
model for the type of dataset taken for the experiment. The NN is a model for analyzing information with a uniform
grid, such as radiographs in the healthcare department, that is influenced by the organization of the human vision
network and created to acquire spatial stratification of characteristics, spanning from the lowest to the highest level
structures, autonomously and flexibly [9][10][11]. The convolution layer, pooling layer along with fully connected
layers are the three levels of hierarchy that create a NN architecture where the initial two layers extract information,
while the final layer is used to link the nodes completely and transfers those characteristics to the output layer and
shows the classification results. There are various architectures of NN models which can be used for classification
tasks. The architectures include Visual Geometry Group (VGG), ResNet (Residual Network), GoogleNet etc., for
predicting accurate results in various sectors [9]. This work presents the use of the VGG16 transfer learning model
for feature extraction due to its applicability with no restrictions for smaller strides and window size. It has the
ability to insert 16 deep layers and hence can perform better on larger datasets. The proposed work presents a DL
based model for quick and early pneumonia diagnosis by utilizing one of the NN architectures called, VGG16, in
which the number of layers has been kept to 16 to decrease the higher computation time issues.
Further, section II introduces the state-of-the-art model. Section III describes the datasets and methodology used
in the proposed method for pneumonia classification and prediction. Apart from this, the performance metrics along
 / Procedia Computer Science 00 (2019) 000–000 3
/ Procedia Computer Science 00 (2019) 000–000 3
Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366 359
/ Procedia Computer Science 00 (2019) 000–000 3
with the results have been mentioned in Section IV. Lastly, the conclusion and future scopes have been discussed in
with the results have been mentioned in Section IV. Lastly, the conclusion and future scopes have been discussed in
section
with theV.
section V.results have been mentioned in Section IV. Lastly, the conclusion and future scopes have been discussed in
section V.
2. Present State-of-the-art
2. Present State-of-the-art
2. Present State-of-the-art
This section describes works done using pre-trained and ensemble models for pneumonia prediction in CXR
This A
images. section
hybrid describes
ensemble works
model donehasusing pre-trained
been pre-trained
presented by andSirazitdinov
ensemble models et al. [5] forfor
pneumonia
predictingprediction
pneumonia. in CXR
This
This A
images. section
hybrid describes
ensemble works
model donehas using
been presented andSirazitdinov
by ensemble models et al. forfor
[5] pneumonia
predicting prediction
pneumonia. in CXR
This
model
images. was created
A created using
hybrid ensemble Adam optimizer,
model has been batch size
presented of 8, where
by where the
Sirazitdinov learning rate,
et al. rate, training-testing
[5] for predicting pneumonia. ratio, and input
This
model size
image was using Adam optimizer, batch size of 8, the learning training-testing
are concluded inratio, and input
model
image was have
size created
have
been
using
been
keptAdam
kept
as 0.0001,
as 0.0001,
75:25batch
optimizer,
75:25
and 512*512,
and size of 8, respectively.
512*512, where the learning
respectively.
The results
The rate, training-testing
results are concluded in
the recall,
ratio,
the and input
recall,
and
and
precision,
image along
sizealong with
have with F1-score
beenF1-score values
kept as 0.0001, as 0.284, 0.288, and 0.286, respectively.
precision, values as75:250.284,and 512*512,
0.288, respectively.
and 0.286, respectively.The results are concluded in the recall, and
precision,
Ahmad etalong withhave
al. [12] F1-score values
utilized as 0.284,
a deep 0.288, and
convolutional 0.286,
neural respectively. technique for feature extraction from
network(CNN)
Ahmad
the CXR etradiographs
al. [12] have utilized
dataset. The a deep
data convolutional
were categorized neural
into network(CNN)
AUC, technique the
which indicated for severity
feature extraction
of the patients' from
Ahmad
the CXR etradiographs
al. [12] have utilized
dataset. The a deep
data convolutional
were categorized neural
into network(CNN)
AUC, which technique the
indicated for severity
feature extraction
of the from
patients'
health
the CXRchallenges.
radiographsThe dataset.
conclusion Thehashas been
databeen
werementioned
categorized in the form
intoform
AUC, of AUCwhichas asindicated
0.98. Rajpurkar
the severity et al.of[3]thedesigned
patients'aa
health challenges.
121-layer CheXNetThe The conclusion
model for predicting mentioned
pneumonia in the
ininCXR of
data. The AUCresults 0.98. Rajpurkar
were Rajpurkar
calculated et et al.
asal. [3]
an[3] designed
F1-score anda
health challenges.
121-layer conclusion has been mentionedin the form
data. of AUC as 0.98. designed
contrasted CheXNet
121-layer to an average
CheXNet
model
model
formedian
real
for
predicting
predicting
pneumonia
F1-score
pneumonia
CXR
value ofin0.387,
CXR which
data.
The results
is substantially
The results
werelower
were
calculated
than that
calculated
as an
as of F1-score
an the CheXNet
F1-score
and
and
contrasted to an average real median F1-score value of 0.387, which is substantially
algorithm. to an average real median F1-score value of 0.387, which is substantially lower than that of the CheXNet lower than that of the CheXNet
contrasted
algorithm.
algorithm.
Zech et al. [13] have presented a CNN approach for pneumonia diagnosis and observed that the presented model
Zech et
might al. [13] have
exaggerate presented
real-world a CNN approach
pneumonia forecastingfor pneumonia
accuracy. The diagnosis and observed
experiment resultedthat in thean presented
AUC of 0.931. model
Zech et
might al. [13] have
exaggerate presented
real-world a CNN approach
pneumonia forecastingfor pneumonia
accuracy. diagnosis
The and observed
experiment resulted that
in the
an presented
AUC of model
0.931.
Rahimzadeh
might exaggerate et al. [14] createdpneumonia
real-world a Deep CNN frameworkaccuracy.
forecasting that predictsTheresults in the form
experiment resultedof efficiency
in an AUC by integrating
of 0.931.
Rahimzadeh
the ResNet50V2 et al. [14] created
and created
XceptionNeta Deep CNN
models. framework
The resultsthatthat
were predicts
compared results in
to otherthe form of efficiency by integrating
Rahimzadeh
the ResNet50V2 et al. [14]
and XceptionNeta Deep CNN framework
models. The results were predicts
compared results in the existing
tohad
other existing
models with
form of efficiency
models withby the
the
highest
integrating
highest
accuracy
the of
ResNet50V2 91.4%. A
and CNN algorithm
XceptionNet for
models. the identification
The results were of pneumonia
compared tohad
other been developed
existing models by Ieracitano
with et al.
the highest
accuracy
[15]. of 91.4%. A CNN algorithm for the identification of pneumonia been developed by Ieracitano et al.
accuracy
[15]. of 91.4%. A CNN algorithm for the identification of pneumonia had been developed by Ieracitano et al.
[15].
This model performs feature extraction from X-ray data of the chest and fuzzy images. Furthermore, the proposed
This model
model has been performs feature extraction
also compared from X-ray data and
to pre-existing of the chest and fuzzy images. Furthermore, the proposed
This model
model has beenperforms
also feature extraction
compared to from algorithms
pre-existing X-ray data and
algorithms of theresulted
chest and
resulted
an accuracy
an fuzzy images.
accuracy
rate of
rate of
up to 81%. the
Furthermore,
up to 81%.
Zhang
Zhang
et al.
proposed
et al.
[16]
model have evaluated
has evaluated a variety
been also acompared of AI-based
to AI-based
pre-existing algorithms
algorithms to detect pneumonia
and resulted and
an accuracy concluded
rate of up that Inf-Net
toInf-Net
81%. Zhang algorithms
et al.
[16] have
mighthave
perform variety
much abetter of
afterofdoing algorithms
a thorough analysisto detect pneumonia
of the existing and
algorithms.concluded
Kundu that that
et al.Inf-Net
[17] have algorithms
used an
[16]
might perform evaluated
much variety
better AI-based
after doing algorithms
a thorough analysisto detect pneumonia
of the existing and concluded
algorithms. Kundu etpneumonia,
al. [17] have algorithms
used an
ensemble
might perform of three
muchCNN betterframeworks, namely
after doingnamely
a thoroughResNet, DenseNet,
analysis and
of the existing GoogleNet
algorithms.to diagnose
Kundu etpneumonia,
al. [17] have withusedgood an
ensemble
results when of three CNN
compared frameworks, ResNet, DenseNet, and GoogleNet to diagnose with good
ensemble
results when of three CNN to
compared to
previous design
frameworks,
previous designnamelytechniques.
techniques.
Two datasets
ResNet, DenseNet,
Two datasets
were used to
and GoogleNet to perform pneumonia,
were used todiagnose
the research in which
perform the researchwith good
in which
results were
results when shown
compared in the form
to previous of accuracy as 87.02%, and 98.8% for Radiological Society the of North in America
results were shown in the form of design
accuracy techniques.
as 87.02%, Two anddatasets
98.8% were used to perform
for Radiological Society research
of North America which
(RSNA)
results and Kermany’s datasets,
form respectively.
(RSNA)were shown in the
and Kermany’s datasets, of accuracy as 87.02%, and 98.8% for Radiological Society of North America
respectively.
(RSNA) and
Yaseliani Kermany’s
et al. [18] havedatasets,
presented respectively.
an ensemble hybrid DL system, consisting of an SVM, radial basis function, and
Yaselianiregression,
logistic et al. [18]inhave which presented
three an ensemble
different hybrid DLprocesses
classification system, consisting of an SVM,
are used, namely a fullyradial basis function,
connected layer forand the
Yaselianiregression,
logistic et al. [18]inhave which presented
three an ensemble
different hybrid DLprocesses
classification system, consisting
are used, of an SVM,
namely a the
fullyradial basis function,
connected layer forandthe
categorization
logistic regression,of images.
in which In the
threesecond step,
different the weights
classification have
processesbeen set through
are used, which
namely a the features
fullyfeatures
connected from the images
categorization of images.
are extracted. Lastly, In the second
the computer-aided step,
modelthehasweights
been have
utilizedbeenforset through
classifying which
CXR the images. fromlayer
Mabrouk the for the
images
et al. [19]
categorization
are extracted. of images.
Lastly, the In the second step,
computer-aided modelthehasweights
been have
utilizedbeenfor set through which
classifying CXR features
images. from the
Mabrouk et images
al. [19]
have expanded
are extracted. a DL
Lastly, model by combining DenseNet169, MobileNetV2, and Vision Transformer models for the
have expanded
prediction of a DLthe
pneumonia modelcomputer-aided
using byCXRcombining model
images. The
has been utilized
DenseNet169,
three models
for classifying
MobileNetV2,
were employedand Vision
to
CXR Transformer
extract
images. Mabrouk
the features models
from
et al.
for[19]
images, the
have expanded a DL model
prediction usingbyCXR combining DenseNet169, MobileNetV2, and Vision Transformer models for the
and finally,ofthepneumonia
prediction extracted features images.
haveimages.
been usedThe three
Thefor
models
the models
experimentalwere employed
were process.
to extract the features from images,
and finally,ofthepneumonia using CXR
extracted features have been used three
for the experimental employed to extract the features from images,
process.
and finally, the extracted features have been used for the experimental process.
3. Materials and Methods
3. Materials and Methods
3. Materials and Methods
This section introduces the method used for pneumonia prediction in CXR images along with the dataset used for
This section introduces the method used for pneumonia prediction in CXR images along with the dataset used for
the
Thisexperiment. In this the
section introduces work, the VGG16
method used forwhich
pneumoniais a pre-trained
prediction feature
in CXRextractor
images along is used with forthe feature
dataset extraction.
used for
the experiment. In this work, the VGG16 which is a pre-trained feature extractor is used for feature extraction.
VGG16
the is an NN-based
experiment. In this architecture
work, the that gained
VGG16 which interest
is a after the 2014
pre-trained featureILSVR (Imagenet)
extractor is used challenge
for [1]. NN
feature comes
extraction.
VGG16 is an NN-based architecture that gained interest after the 2014 ILSVR (Imagenet) challenge [1]. NN comes
under
VGG16oneoneis anof the best image categorization model architectures. It is remarkable in that it contains 3x3 filter
under of NN-based
the best image architecture that gained
categorization interest
model after the 2014
architectures. It isILSVR (Imagenet)
remarkable in that challenge
it contains [1]. NN 3x3comes
filter
convolution
under one ofand and uses
theuses the
best the imagesame max pooling and padding layer of a 2x2 filter rather than a huge number of
convolution samecategorization
max poolingmodel architectures.
and padding layer of It is remarkable
a 2x2 in that
filter rather than it acontains
huge number 3x3 filter of
hyper-parameters.
convolution The convolutional
and uses and max pooling layers are structured infilter
a similar manner across the system.
hyper-parameters. The the same maxand
convolutional pooling and padding
max pooling layers layer of a 2x2in
are structured rathermanner
a similar than aacross huge the numbersystem. of
Finally, for output,
hyper-parameters. there are two fully connected softmax layers [20][21]. The number 16 in VGG16 is attributed to
Finally, for output,The thereconvolutional
are two fullyand max pooling
connected softmax layers
layersare[20][21].
structured in number
The a similar16manner
in VGG16 across the system.
is attributed to
the idea
Finally, that it has 16 layers of varying weights with around 138 million parameters.
the idea for
thatoutput,
it has 16 there
layersareoftwo fully weights
varying connected withsoftmax
aroundlayers [20][21].
138 million The number 16 in VGG16 is attributed to
parameters.
the idea that it has 16 layers of varying weights with around 138 million parameters.
3.1 Datasets Description
3.1 Datasets Description
3.1 Datasets Description
The first dataset is gathered from Kaggle [22], which is split into two primary directories: train folder, and test
The first dataset is gathered from Kaggle [22], which is split into two primary directories: train folder, and test
folder; both
The both directories
first dataset comprise
is gathered fromtwo subdirectories, including one containing pneumonia X-ray radiographs and
folder; directories comprise twoKaggle [22], which
subdirectories, is split one
including into containing
two primary directories:
pneumonia trainradiographs
X-ray folder, and and test
another
folder; containing
both CXR
directories radiographs
comprise two for normal
subdirectories, lungs. A
including number
one of 5,856
containing anteroposterior
pneumonia X-ray CXR images
radiographs were
and
another containing CXR radiographs for normal lungs. A number of 5,856 anteroposterior CXR images were
efficiently selected from
another containing historical
CXRhistorical
radiographs pediatric patients having
for normal an age between 1 to 5 years [20]. There are two labels
efficiently selected from pediatric patientslungs.
having A an number of 5,856
age between 1 toanteroposterior
5 years [20]. There CXR are images were
two labels
efficiently selected from historical pediatric patients having an age between 1 to 5 years [20]. There are two labels
360 Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366
4 / Procedia Computer Science 00 (2019) 000–000

being used to split all the folders: pneumonia and normal, which were applied to all the other pneumonia X-ray
images. The initial data classifications were amended and merged; thereafter, the entire image dataset was
reorganized into 70% for training purposes, and 30% for testing. Hence, the measure of information is assigned to
the complete dataset. A number of 5,216 X-ray images were allocated to the training, and 640 radiographs to the
testing data to test the system. The dataset is a publicly accessible dataset: containing JPEG images only.

Fig. 1 Proposed deep learning model for pneumonia prediction

Originally, the anterior-posterior CXR dataset was collected from retroactive groups of children patients under
the age of 1-5 at Guangzhou Medical Center for Women and Children [23].
The second dataset has also been collected from Kaggle [24], which is split into two directories, namely training
and testing. Each of these directories has three subfolders: pneumonia, covid-19, and normal. The dataset contains
6,432 radiographs in JPEG format, out of which training is 80% data and testing is 20% data. This dataset is
originally collected from several sources such as GitHub and kaggle. The train and test folders have subfolders
containing 3418 pneumonia, 1266 normal, and 460 covid-19 images and the train folder contains 855 pneumonia,
317 normal, and 116 covid-19 images. The first dataset and second dataset are used for binary and multi-class
classification, respectively.
 Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366 361
/ Procedia Computer Science 00 (2019) 000–000 5

3.2 Methodology

The proposed DL model is categorized into several parts: data collection, preprocessing, feature extraction [25],
training, testing, classification, and pneumonia prediction, as shown in Fig. 1. The data preprocessing is done for
balancing and normalizing the data, this technique is used to set the data in a normalized form between a range of
[0-255]. Thereafter, the input data is fed to the VGG16 model for feature extraction. This step is capable of
extracting features from the images and feeding them to the classification for further prediction process. The VGG16
has 16 layers, including input, convolution, pooling, dense, and output, which are responsible for an in-depth and
better feature extraction process. The proposed model has been implemented in the Orange 3.31.1 simulator. It uses
the Adam optimizer, 0.0001(learning rate), and the ReLu activation function is utilized. Afterward, the extracted
features have been fed to various classifiers, namely NN, SVM, KNN, random forest (RF), and naïve Bayes (NB),
for analyzing the prediction results, whether the data sample has pneumonia or not. The training of whole model is
carried out on the training dataset having 70% out of the whole data and 30 % for the test data set 1. Whereas the
training and test ratio of dataset 2 has been taken as 80:20. Finally, the results of both the datasets are compared to
validate the model.

4. Performance Evaluation

The results of the proposed model are estimated using various metrics, including accuracy, precision, F1-score,
and recall, as shown in Table 1. These metrics require a brief definition of four terms, namely, "false positive", "true
positive”, “false negative”, and “true negative”. "false positive (FP)" refers to the samples which belong to negative
classes but are predicted to belong to positive classes. "true positive (TP)" refers to samples that are classified as
positive and also belong to the positive class. “false negative (FN)” samples are those that belong to the positive
classes but are anticipated to belong to the negative classes. “true negative (TN)” refers to the samples which belong
to the negative classes, and are predicted correctly. The metrics used for the proposed models’ prediction are defined
below:

𝐴𝐴𝑐𝑐𝑐𝑐𝑢𝑢𝑟𝑟𝑎𝑎𝑐𝑐𝑦𝑦 = (𝑇𝑇𝑃𝑃 + 𝑇𝑇𝑁𝑁)/(𝑇𝑇𝑃𝑃 + 𝑇𝑇𝑁𝑁 + 𝐹𝐹𝑃𝑃 + 𝐹𝐹𝑁𝑁) (1)

𝑅𝑅𝑒𝑒𝑐𝑐𝑎𝑎𝑙𝑙𝑙𝑙 = 𝑇𝑇𝑃𝑃/(𝑇𝑇𝑃𝑃 + 𝐹𝐹𝑁𝑁) (2)
𝑃𝑃𝑟𝑟𝑒𝑒𝑐𝑐𝑖𝑖𝑠𝑠𝑖𝑖𝑜𝑜𝑛𝑛 = 𝑇𝑇𝑃𝑃/(𝑇𝑇𝑃𝑃 + 𝐹𝐹𝑃𝑃) (3)
𝐹𝐹1 − 𝑠𝑠𝑐𝑐𝑜𝑜𝑟𝑟𝑒𝑒 = 2/((1/𝑃𝑃𝑟𝑟𝑒𝑒𝑐𝑐𝑖𝑖𝑠𝑠𝑖𝑖𝑜𝑜𝑛𝑛) + (1/𝑅𝑅𝑒𝑒𝑐𝑐𝑎𝑎𝑙𝑙𝑙𝑙)) (4)

The accuracy rate is a metric that indicates how many right predictions the algorithm has made. Nevertheless, if
the data is unbalanced, the higher accuracy of the technique does not guarantee its capability to discriminate distinct
categories effectively and efficiently. In healthcare, image categorization is a technique that can be applied to every
class in the dataset. The "recall" and "precision" values provide information about the model's efficacy in such
instances. The term "precision" refers to the rate of positive label accuracy. It provides the percentage of accurately
predicted values to the total number of predictions classified using the model. In contrast, "recall" refers to the value
%age of true positives anticipated accurately by the model. The "F1-score" balances "precision" with "recall," taking
into account both FNs and FPs. Extreme values of "recall" and "precision," each of which is at the loss of the other,
are assessed. To acquire an accurate determination of a normal patient along with diseased patients, it is worthwhile
to examine evaluation measurement metrics instead of only the accuracy metric in diagnostic image identification.

4.1 Results and Discussion

This subsection introduces and illustrates a detailed discussion of the obtained outcomes using the proposed model
for pneumonia prediction.

● NN classifier with VGG16: The performance of the proposed model has been found better for the prediction of
pneumonia, in which various performance metrics have been calculated with the least value of error. The model
has predicted a quite better performance with an overall accuracy of 92.15%, precision as 0.9428, and recall as
0.9308, whereas F1-score is 0.937 for dataset 1 as tabulated in Table 1. The model exhibits overall accuracy of
94.5%, precision as 0.954, recall as 0.954, and F1-score as 0.954 for dataset 2.
6362 / Procedia
Shagun SharmaComputer Science Computer
et al. / Procedia 00 (2019) Science
000–000218 (2023) 357–366

Table 1. Performance outcome of proposed work using two different datasets

Dataset 1 Dataset 2
Models
Accuracy Precision Recall F1-score AUC Accuracy Precision Recall F1-score AUC

NN with VGG16 92.15% 0.9428 0.9308 0.937 0.974 95.4% 0.954 0.954 0.954 0.988

SVM with 91.5% 0.916 0.916 0.914 0.968 94% 0.942 0.940 0.940 0.988
VGG16

KNN with 91% 0.91 0.91 0.91 0.959 91.9% 0.922 0.919 0.919 0.969
VGG16

RF with VGG16 87.2% 0.871 0.872 0.871 0.951 91.1% 0.912 0.911 0.911 0.971

NB with VGG16 84.8% 0.856 0.848 0.849 0.908 83.4% 0.872 0.834 0.839 0.834

The results are also shown in ROC graph as shown in Fig. 2 for first dataset and in Fig.3 for second dataset
applicable to representing the degree of separatability. ROC illustrates the depiction of the relationship between
recall and precision. It is used to show the evaluation of binary classification graphically. Whereas other methods
result in a single value for representing the performance. ROC is a well-known curve to identify a good classifier the
closer the ROC is to the top left of the graph better the model is.

Fig. 2 ROC of NN with VGG16 for first dataset Fig. 3 ROC of NN with VGG16 for second dataset

● SVM with VGG16: SVM is a model which is used for classification tasks and reduces over fitting issues in the
datasets [26]. In this work, the SVM with VGG16 model is compared with the classification outcomes of the
proposed NN model with VGG16 model. The outcome SVM with VGG16 has been identified in the form of
recall, precision, accuracy, and F1-score as 0.915, 0.916, 91.5%, 0.914, respectively, and ROC as shown in Fig. 4
for dataset 1. However, dataset 2 achieves accuracy, recall, precision, and F1-score of 94%, 0.94, 0.942 and 0.94,
respectively, using SVM with VGG16. The comparison has shown that the proposed NN with VGG16 performs
better than the SVM model with VGG16, as the accuracy of NN with VGG16 is quite higher for both the datasets
than SVM with VGG16 for pneumonia classification.
 / Procedia Computer Science 00 (2019) 000–000 7
Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366 363

Fig. 4 ROC of SVM with VGG16 for first dataset Fig. 5 ROC of KNN with VGG16 for first dataset

● KNN with VGG16: KNN is a supervised ML model utilized for classification tasks [27]. The proposed NN
model with VGG16 classification results are compared to the KNN model with VGG16 classification results. The
ROC is illustrated in Fig. 5 for dataset 1. The outcome has been identified as precision, recall, accuracy, and
F1-Score as 0.91, 0.91, 91%, and 0.910, respectively, for dataset 1. Dataset 2 achieves accuracy, recall, precision,
and F1-Score as 91.9%, 0.919, 0.922, and 0.919, respectively. As a result of the comparison, the proposed NN
model with VGG16 performs better than the KNN model with VGG16 for both the datasets.

Fig. 6 ROC of Naive Bayes with VGG16 for first dataset Fig. 7 ROC of Random Forest with VGG16 for first dataset

● NB with VGG16: The classification results of NB with VGG16 model are compared with the proposed NN
with VGG16 model The ROC is depicted in Fig. 6 for dataset 1. The outcomes of NB with VGG16 for dataset 1
are recorded in terms of recall, precision, accuracy, and F1-score as 0.848, 0.856, 84.8%, and 0.849, respectively
; however the dataset 2 achieves precision, recall, accuracy, and F1-score as 0.872, 0.834, 83.4%, 0.839
respectively. As a result of the comparison, the proposed NN model with VGG16 performs better than the NB
model with VGG16 for both datasets, because the accuracy for pneumonia classification is substantially higher.
● Random Forest (RF) with VGG16: In comparing classification results of the proposed NN with VGG16 model
and the results of RF with VGG16, it has been found that the proposed NN model with VGG16 performs better in
case of a recall, precision, accuracy, and F1-score for both datasets. The results of RF with VGG16 for dataset 1
are noted as recall, precision, accuracy, and F1-score as, 0.872, 0.871, 87.2%, 0.871, respectively; however,
dataset 2 achieves recall, precision, accuracy, and F1-score as 0.911, 0.912, 91.1%, 0.911 respectively.
Furthermore, the ROC curve of the RF with VGG16 for dataset 1 is also shown in Fig. 7.
364 Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366
8 / Procedia Computer Science 00 (2019) 000–000

4.2 Performance comparison of NN, SVM, NB, RF, and KNN using VGG16 Feature Extractor

This section illustrates the comparative analysis of the various models for both datasets with respect to various
performance measures:

Fig. 8 Accuracy comparison of several models with the proposed Fig. 9 Precision comparison of several models with the proposed
work (NN with VGG16) work (NN with VGG16)

Fig. 8 depicts the graph of accuracy comparison of several models. The proposed NN with VGG16 has been
determined as the outperforming pneumonia prediction model having the highest accuracy in both datasets as
compared to SVM with VGG16, KNN with VGG16, RF with VGG16, and NB with VGG16 model. As the accuracy
depicts the models' performance, NN with VGG16 can be seen as the best performing model among all other
models. Furthermore, for the dataset 2, the accuracy of NN with VGG16 has also been found highest among all
95.40%.

Fig. 9 presents a graph of the precision comparison of various models, in which the proposed technique NN with
VGG16 is determined to be the best model for predicting pneumonia, with the highest precision for both datasets.
The precision defines the number of positive predictions made by the model, NN with VGG16 can be considered the
best performing model among the others used in the study. The precision value for the second dataset has been
found very high as 0.954 using NN model with VGG16 which outperforms the other models.

Fig. 10 Recall comparison of several models with the proposed Fig. 11 F1-score comparison of several models with the proposed work
work (NN with VGG16) (NN with VGG16)

Fig. 10 illustrates the graph of recall comparison of numerous models, in which the proposed technique NN with
VGG16 has been determined as the outperforming pneumonia prediction model with the highest recall for both
datasets. As recall defines how many positive predictions are truly right out of a total number of positive predictions,
NN with VGG16 can be considered the highest performing model among all other models used in the study. The
recall of the second dataset using NN with VGG16 is the highest among all 0.954 compared to other models.
 // Procedia
Procedia Computer
Computer Science
Science 00
00 (2019)
(2019) 000–000 99
000–000
/ Procedia Computer Science 00 (2019)
Shagun Sharma et al. / Procedia Computer 000–000
Science 218 (2023) 357–366 3659
/ Procedia Computer Science 00 (2019) 000–000 9
F1-score
F1-score isis aa performance
performance metric
metric which
which isis used
used to
to identify
identify the
the performance
performance of of aa classifier.
classifier. The
The range
range of
of the
the F1-score
F1-score
F1-score
is between is a
0-1.performance
The value metric
of the which is
F1-score used
nearto 1identify
shows the
the performance
better of
performancea classifier.
of the The range
model. of
Fig.11thedepicts
F1-score
an
is between
F1-score is 0-1. The value metric
a performance of thewhich
F1-score nearto1identify
is used shows the
the performance
better performance of the The
of a classifier. model.
rangeFig.11
of thedepicts an
F1-score
is between
F1-score 0-1.
comparison The value
of of the
various F1-score
models, in near
which 1 shows
the the
proposed better
NN performance
with VGG16 of the
has model.
been Fig.11
identified depicts
as the an
best
F1-score
is betweencomparison
0-1. The value of various
of themodels,
F1-score in near
which the proposed
1 shows NN performance
the better with VGG16ofhas thebeen identified
model. Fig.11 as the best
depicts an
F1-score
model comparison of various models, in which the proposed NN with VGG16 has been identified as the best
model for
F1-score predicting
predicting pneumonia,
forcomparison pneumonia, with
with the
of various models, the inhighest
whichF1-score
highest F1-score value
value NN
the proposed for
for both
withthe
both the datasets,
haswhen
datasets,
VGG16 beencompared
when identifiedto
compared toasthetheSVM
the SVM
best
modelVGG16,
with for predicting
KNN pneumonia, with the highest F1-score value for both the datasets, when comparedthe to harmonic
the SVM
with
modelVGG16, KNN with
for predicting with VGG16,
VGG16,with
pneumonia, RF with
RF the VGG16,
withhighest
VGG16, and
and NB
F1-score with
NBvalue
withforVGG16.
VGG16.
both theThe
The F1-score
F1-score
datasets, whendetermines
determines
comparedthe to harmonic
the SVM
with
mean VGG16,
of KNN with VGG16, RF with VGG16, and NB with VGG16. The F1-score determines the harmonic
mean of precision
precision
with VGG16, KNN and recall;
andwith
recall; hence,
hence,RF
VGG16, NN with
NNwith VGG16
withVGG16,
VGG16 and can be
be considered
can NB considered
with VGG16. the
the best-performing
best-performing model
model among
The F1-score determines among the
the other
the harmonic other
mean
models ofused
precision
in the and recall;
study. Thehence,
NN NN VGG16
with with VGG16 has can be considered
achieved an F1-score thevalue
best-performing
of 0.954 formodel
datasetamong
2, which the is
other
the
modelsofused
mean in theand
precision study. Thehence,
recall; NN withNN VGG16
with VGG16 has achieved an F1-score
can be considered thevalue of 0.954 formodel
best-performing datasetamong
2, which the is the
other
models
highest used inall.
among the study. The NN with VGG16 has achieved an F1-score value of 0.954 for dataset 2, which is the
highest among all.
models used in the study. The NN with VGG16 has achieved an F1-score value of 0.954 for dataset 2, which is the
highest among all.
highest among all.
4.3
4.3 Performance
Performance outcome
outcome comparison
comparison of of the
the proposed
proposed model
model with
with the
the existing
existing models
models
4.3 Performance outcome comparison of the proposed model with the existing models
4.3 Performance outcome comparison of the proposed model with the existing models
In
In this
this section
section of of the
the paper,
paper, the
the results
results of
of NN
NN with
with the
the VGG16
VGG16 modelmodel are
are compared
compared with with the
the results
results of
of the
the models
models
In this section
presented in theofliterature
the paper, the results
review as shownof NNin with the
Table. 2. InVGG16
this model areitcompared
comparison, has been with the
found thatresults
the of the models
proposed model
presented in theofliterature
In this section the paper,review as shown
the results of NNin Table. 2. InVGG16
with the this comparison,
model areitcompared
has been found thatresults
with the the proposed model
of the models
presented
outperforms in the literature
the review asmodels
other shown in Table.
i.e. 2.ResNet50V2+XceptionNet
In this comparison, it has been[14], found that the proposed model
CovNNet[15] and
outperforms
presented the
in the literature other
review asmodels i.e. 2.ResNet50V2+XceptionNet
shown in Table. In this comparison, it has been[14], found thatCovNNet[15]
the proposed model and
outperforms the
GoogleNet+ResNet18+DenseNet[17] other models
in terms i.e.
of ResNet50V2+XceptionNet
accuracy. Whereas the precision [14],
and recall CovNNet[15]
have been found and
quite
GoogleNet+ResNet18+DenseNet[17]
outperforms the other in
models terms of
i.e. accuracy. Whereas the
ResNet50V2+XceptionNet precision and recall
[14], have been
CovNNet[15] found quite
and
GoogleNet+ResNet18+DenseNet[17]
higher in terms of accuracy. Whereas the precision and recall have been found quite
higher than
than RetinaNet+Mask
RetinaNet+Mask R-CNN
GoogleNet+ResNet18+DenseNet[17] R-CNN in presented
terms ofin
presented [5].
[5]. Finally,
inaccuracy.
Finally, the
the F1-score
Whereas F1-score is
is also
the precision found
alsoand
found higher
havethan
higher
recall beenin
than 121-Layer
infound
121-Layer
quite
higher than RetinaNet+Mask
ChexNet[3] and R-CNN presented in [5]. Finally, the F1-score is also foundbetter
higher than for
in 121-Layer
ChexNet[3]
higher and RetinaNet+Mask
RetinaNet+Mask
than RetinaNet+Mask R-CNNR-CNN[5].
R-CNN[5].
presented Conclusively,
Conclusively, the
in [5]. Finally, proposed
the the
proposed
F1-score model
model exhibits
exhibits
is also foundbetter results
results
higher than for datasets
datasets 11
in 121-Layer
ChexNet[3]
and 2 than and models
other RetinaNet+Mask
presented R-CNN[5].
in the Conclusively, the proposed model exhibits better results for datasets 1
literature.
and 2 than other
ChexNet[3] and models presented in
RetinaNet+Mask the literature.
R-CNN[5]. Conclusively, the proposed model exhibits better results for datasets 1
and 2 than other models presented in the literature.
and 2 than other models presented in the literature.
Table
Table 2.
2. Performance
Performance comparison
comparison of
of NN
NN with
with VGG16
VGG16 feature
feature extractor
extractor with
with the
the existing
existing models
models
Table 2. Performance comparison of NN with VGG16 feature extractor with the existing models
Table 2. Performance comparison of NN with VGG16 feature extractor with the existing models
Reference Model Dataset Accuracy Precision Recall F1-score
Reference Model Dataset Accuracy Precision Recall F1-score
Reference Model Dataset Accuracy Precision Recall F1-score
Reference
[14] Model
ResNet50V2+XceptionNet Dataset
Chest X-ray
X-ray (11,302
(11,302 images)
images) Accuracy
91.4% Precision Recall F1-score
[14] ResNet50V2+XceptionNet Chest 91.4% -- -- --
[14] ResNet50V2+XceptionNet Chest X-ray (11,302 images) 91.4% - - -
[14]
[15] ResNet50V2+XceptionNet
ConvNet Chest X-ray
Chest X-ray (11,302
(26,684 images)
X-ray (26,684 images) 91.4%
81% - - -
[15] ConvNet Chest images) 81% -- -- --
[15] ConvNet Chest X-ray (26,684 images) 81% - - -
[15]
[17] ConvNet
GoogleNet+ResNet+DensNet121 Chest
RSNAX-ray (26,684 images) 81%
86.85% - - -
[17] GoogleNet+ResNet+DensNet121 RSNA 86.85% -- -- --
[17] GoogleNet+ResNet+DensNet121 RSNA 86.85% - - -
[17]
[5] GoogleNet+ResNet+DensNet121
RetinaNet+Mask R-CNN
R-CNN RSNA
Chest X-ray
X-ray (26,684
(26,684 images)
images) 86.85% -0.288 -0.284 -0.286
[5] RetinaNet+Mask Chest -- 0.288 0.284 0.286
[5] RetinaNet+Mask R-CNN Chest X-ray (26,684 images) - 0.288 0.284 0.286
[5]
[3] RetinaNet+Mask R-CNN
121-layer ChexNet
ChexNet Chest
FrontalX-ray
view(26,684 images) images)
X-ray (1,00,000
(1,00,000 - 0.288 0.284 0.286
[3] 121-layer Frontal view X-ray images) -- -- -- 0.387
0.387
[3] 121-layer ChexNet Frontal view X-ray (1,00,000 images) - - - 0.387
[3] NN with
Proposed 121-layer
VGG16ChexNet
(Dataset 1)
1) Frontal view(5,856
Chest X-ray
X-ray X-ray (1,00,000
images) images) -92.15% -0.9428 -0.9308 0.387
0.937
Proposed NN with VGG16 (Dataset Chest (5,856 images) 92.15% 0.9428 0.9308 0.937
Proposed NN with VGG16 (Dataset 1) Chest X-ray (5,856 images) 92.15% 0.9428 0.9308 0.937
Proposed
Proposed NN with VGG16 (Dataset 1) Chest X-ray (5,856 images) 92.15% 0.9428 0.9308 0.937
Proposed NN
NN with
with VGG16
VGG16 (Dataset
(Dataset 2)
2) Chest
Chest X-ray
X-ray (6,436
(6,436 images)
images) 95.4%
95.4% 0.954
0.954 0.954
0.954 0.954
0.954
Proposed NN with VGG16 (Dataset 2) Chest X-ray (6,436 images) 95.4% 0.954 0.954 0.954
Proposed NN with VGG16 (Dataset 2) Chest X-ray (6,436 images) 95.4% 0.954 0.954 0.954
5.
5. Conclusion
Conclusion and and Future
Future ScopeScope
5. Conclusion and Future Scope
5. Conclusion and Future Scope
Early
Early identification
identification of of pneumonia
pneumonia is is essential
essential for
for selecting
selecting the
the best
best alternative
alternative treatments
treatments and and providing
providing the
the timely
timely
Early identification
medication to save of pneumonia
precious lives. is essential
X-ray for
images selecting
are among the best
the mostalternative
common treatments
methods and
for providing
detecting the timely
pneumonia;
medication to save of
Early identification precious lives.isX-ray
pneumonia images
essential are among
for selecting thethe
bestmost commontreatments
alternative methods for anddetecting
providingpneumonia;
the timely
medication
unfortunately, to save
they precious lives. X-ray
are susceptible
susceptible to crossimages
cross are among
variability, and the
the most common is
identification methods
dependentfor detecting
on the pneumonia;
the ability
ability of the
the
unfortunately,
medication they
to save are
precious to
lives. X-ray variability,
images and
are among the identification
most common is dependent
methods on
for detecting of
pneumonia;
unfortunately,
clinician to they initial
detect are susceptible
pneumonia tosymptoms.
cross variability,
In this and thea identification
article, DL-based is dependent
model is presentedon the
to ability of
support the
health
clinician to detect
unfortunately, they initial pneumoniatosymptoms.
are susceptible In this article,
cross variability, and thea identification
DL-based model is presented
is dependent to support
on the health
ability of the
clinician to detect
professionals, and initial
it pneumonia
leverages the symptoms.classification
DL-based In this article,of aCXR
DL-based model into
radiographs is presented
2 to support
categories: normal health
and
professionals,
clinician and initial
to detect it leverages
pneumoniathe DL-based
symptoms.classification
In this article,of aCXRDL-basedradiographs
model into 2 categories:
is presented normalhealth
to support and
professionals,
pneumonia. The Theand it leverages
results of the DL-based classification of CXR radiographs into 2 categories: normal and
pneumonia.
professionals, results
and of the
the research
it leverages research exhibit
exhibit that
the DL-based that NN
NN with
with VGG16
classification VGG16
of CXR model
model provides better
better2 performance
provides into
radiographs performance for
for datasets
categories: normal datasets
and
pneumonia.
11pneumonia.
and The results of the research exhibit that NN with VGG16 model provides better performance for datasets
and 22 than
thanThe SVM
SVM with
withofVGG16
results VGG16 model,
model,
the research KNN
KNNthat
exhibit with
withNNVGG16
VGG16 model,
model,model
with VGG16 Random
Random Forest with
Forestbetter
provides VGG16
VGG16 model,
with performance for and
model, and NB
NB
datasets
1with
andVGG16
2 than model.
SVM with The VGG16
NN with model,
VGG16 KNN
model with VGG16
provides an model,
accuracy Random
value Forest
of 92.15%,with VGG16
precision model,
as and
0.9428, NB
recall
1with
andVGG16
2 than model.
SVM with The VGG16
NN withmodel,
VGG16KNN model provides
with VGG16 an model,
accuracy value ofForest
Random 92.15%,withprecision
VGG16as 0.9428,
model, andrecall
NB
with
as VGG16 model. The NN with VGG16 model provides an accuracy value of 92.15%, precision as 0.9428, recall
as 0.9308,
with0.9308,
VGG16 and F1-score
andmodel.
F1-scoreThe 0.937
0.937
NN with for dataset
for VGG16 1;
1; however
dataset model however the
the an
provides dataset
dataset 22 achieves
accuracy achieves accuracy
value ofaccuracy value
92.15%, value of 95.4%,
of
precision 95.4%, precision
precision
as 0.9428, recall
as 0.9308,
0.954, recalland F1-score
0.954, 0.937 for
and F1-score
F1-score dataset
0.954 which 1; ishowever
highest the
amongdataset It22 is
all. It isachieves
concluded accuracy
that value of 95.4%, precision
0.954,
as recall
0.9308, and0.954, and
F1-score 0.937 for 0.954 which
dataset 1; ishowever
highest among
the all.
dataset concluded
achieves that the
accuracy proposed
thevalue
proposed model NN
NN with
modelprecision
of 95.4%, with
0.954, recall
VGG16 exhibits0.954, andperformance
better F1-score 0.954 thanwhich
other isclassification
highest among all. It mentioned
models is concluded in that
the the proposed
literature reviewmodel NN with
section. The
VGG16 exhibits better performance than other classification models mentioned in the
0.954, recall 0.954, and F1-score 0.954 which is highest among all. It is concluded that the proposed model NN with literature review section. The
VGG16
highest exhibits
accuracy better
results performance
achieved than
with other
proposed classification
NN with models
VGG16 for mentioned
pneumonia in the literature
prediction arereview
92.15% section.
and The
95.4%
highest accuracy
VGG16 resultsperformance
exhibits better achieved with proposed
than NN with VGG16
other classification models formentioned
pneumonia in prediction
the literaturearereview
92.15%section.
and 95.4%The
highest
for accuracy results achieved with proposed NN with VGG16 for pneumonia prediction are 92.15% andVGG16
95.4%
for dataset
dataset
highest 11 and
accuracy dataset2
andresults
dataset2 respectively.
respectively.
achieved The
The research
with proposed research
NN withoutcome
outcome
VGG16 exhibits
exhibits that
that the
for pneumonia the proposed
proposed
prediction model
model NN
NN with
are 92.15% with
andVGG16
95.4%
for dataset
exhibits the 1 and dataset2 respectively. The research outcome exhibits that the proposed model NN normal
with VGG16
exhibits
for the1highest
dataset highest accuracy
accuracy
and dataset2 of
of 95.4%
95.4% for
respectively. for
The dataset
dataset 22 as
research compared
as outcome to
to other
comparedexhibits otherthat existing
existing models
models to
the proposed diagnose
tomodel
diagnose
NN normal lungs,
lungs,
with VGG16
exhibits the highest accuracy of 95.4% for dataset 2 as compared to other existing models to diagnose normal lungs,
exhibits the highest accuracy of 95.4% for dataset 2 as compared to other existing models to diagnose normal lungs,
366 Shagun Sharma et al. / Procedia Computer Science 218 (2023) 357–366
10 / Procedia Computer Science 00 (2019) 000–000

covid-19, and pneumonia. In the future work, to achieve higher performance outcomes of the DL model the dataset
can be increased with augmentation and the number of hidden layers may be increased for the deep convolutions in
the model.

References

[1] El Asnaoui K, Chawki Y, Idri A (2021) Automated methods for detection and classification pneumonia based on x-ray images using deep
learning. In: Artificial intelligence and blockchain for future cyber security applications, Springer, Cham, pp 257-284.
[2] Sarangi PK, Guleria K, Prasad D, Verma DK (2021) Stock movement prediction using neuro genetic hybrid approach and impact on growth
trend due to COVID-19. International Journal of Networking and Virtual Organisations. 25(3-4):333-52.
[3] Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H, Duan T, Ding D, Bagul A, Langlotz C, Shpanskaya K, Lungren MP (2017) Chexnet:
Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225.
[4] Singh S, Tripathi BK(2022) Pneumonia classification using quaternion deep learning. Multimedia Tools and Applications. 81(2):1743-64.
[5] Sirazitdinov I, Kholiavchenko M, Mustafaev T, Yixuan Y, Kuleev R, Ibragimov B (2019) Deep neural network ensemble for pneumonia
localization from a large-scale chest x-ray database. Computers & electrical engineering. 78:388-99.
[6] Sharma, A., Guleria, K., & Goyal, N. (2021) Prediction of Diabetes Disease Using Machine Learning Model. In: International Conference on
Communication, Computing and Electronics Systems, Springer, Singapore, pp 683-692.
[7] Stephen O, Sain M, Maduh UJ, Jeong DU (2019) An efficient deep learning approach to pneumonia classification in healthcare. Journal of
healthcare engineering..
[8] Saravagi D, Agrawal S, Saravagi M, Rahman MH (2022) Diagnosis of Lumbar Spondylolisthesis Using a Pruned CNN Model. Computational
and Mathematical Methods in Medicine.
[9] Hasan MD, Ahmed S, Abdullah ZM, Monirujjaman Khan M, Anand D, Singh A, AlZain M, Masud M (2021) Deep learning approaches for
detecting pneumonia in COVID-19 patients by analyzing chest X-ray images. Mathematical Problems in Engineering.
[10]Gayathri JL, Abraham B, Sujarani MS, Nair MS (2022) A computer-aided diagnosis system for the classification of COVID-19 and
non-COVID-19 pneumonia on chest X-ray images by integrating CNN with sparse autoencoder and feed forward neural network. Computers
in Biology and Medicine. 141:105134.
[11] Sharma S, Khanra P, Ramkumar KR (2021) Performance Analysis of Biomass Energy using Machine and Deep Learning Approaches. In:
Journal of Physics: Conference Series. 2089(1): 012003.
[12] Ahmad J, Saudagar AK, Malik KM, Ahmad W, Khan MB, Hasanat MH, AlTameem A, AlKhathami M, Sajjad M (2022) Disease
Progression Detection via Deep Sequence Learning of Successive Radiographic Scans. International journal of environmental research and
public health. 19(1):480.
[13] Zech JR, Badgeley MA, Liu M, Costa AB, Titano JJ, Oermann EK (2018) Variable generalization performance of a deep learning model to
detect pneumonia in chest radiographs: a cross-sectional study. PLoS medicine. 15(11):e1002683.
[14] Rahimzadeh, M., and Attar, A. (2020) A modified deep convolutional neural network for detecting COVID-19 and pneumonia from chest
X-ray images based on the concatenation of Xception and ResNet50V2. Informatics in medicine unlocked. 19:100360.
[15] Ieracitano, C., Mammone, N., Versaci, M., Varone, G., Ali, A. R., Armentano, A., and Morabito, F. C. (2022) A Fuzzy-enhanced Deep
Learning Approach for Early Detection of Covid-19 Pneumonia from Portable Chest X-Ray Images. Neurocomputing. 202-215.
[16] Zhang, F. (2021) Application of machine learning in CT images and X-rays of COVID-19 pneumonia. Medicine, 100(36).
[17] Kundu R, Das R, Geem Z.W, Han G.T, and Sarkar R, (2021) Pneumonia detection in chest X-ray images using an ensemble of deep learning
models. PloS one. 16(9): e0256630.
[18] M. Yaseliani, AZ. Hamadani, AI. Maghsoodi, and A. Mosavi, (2022) Pneumonia Detection Proposing a Hybrid Deep Convolutional Neural
Network Based on Two Parallel Visual Geometry Group Architectures and Machine Learning Classifiers, IEEE Access.
[19] A. Mabrouk, RP. Díaz Redondo, A. Dahou, M. Abd Elaziz, M. Kayed, (2022) Pneumonia Detection on Chest X-ray Images Using Ensemble
of Deep Convolutional Neural Networks, Applied Sciences, 12(13):6448.
[20] Hammoudi K, Benhabiles H, Melkemi M, Dornaika F, Arganda-Carreras I, Collard D, and Scherpereel A, (2021) Deep learning on chest
X-ray images to detect and evaluate pneumonia cases at the era of COVID-19. Journal of medical systems. 45(7):1-10.
[21] Toğaçar M, Ergen B, Cömert Z, and Özyurt F, (2020) A deep feature learning model for pneumonia detection applying a combination of
mRMR feature selection and machine learning models. Irbm. 41(4): 212-222.
[22] Chest X-Ray images (Kaggle). https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia
[23] Alqudah A.M, Qazan S, and Masad I.S, (2021) Artificial intelligence framework for efficient detection and classification of pneumonia using
chest radiography images. Journal of Medical and Biological Engineering. 41(5): 599-609.
[24] Chest X-ray (Covid-19 & Pneumonia). https://www.kaggle.com/datasets/prashant268/chest-xray-covid19-pneumonia
[25] A. Kaur, K. Guleria, and N. Kumar Trivedi, (2021) Feature selection in machine learning: Methods and comparison, In: 2021 International
Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE), pp 789-795.
[26] S. Bhushan, M. Alshehri, N. Agarwal N, I. Keshta, J. Rajpurohit, A. Abugabah, (2022) A Novel Approach to Face Pattern Analysis,
Electronics, 11(3):444.
[27] S. Bhushan, M. Alshehri M, I. Keshta, AK. Chakraverti, J. Rajpurohit, A. Abugabah, (2022) An Experimental Analysis of Various Machine
Learning Algorithms for Hand Gesture Recognition, Electronics, 11(6):968.