1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-

2022)

Detection of COVID-19 using CNN from Chest X-ray

2022 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES) | 978-1-6654-8004-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/CISES54857.2022.9844330

Images
1st Pavipra Singh 2nd Shashank Sahu

MTech Scholar, Department of Computer Science Engineering Professor, Department of Computer Science Engineering
Ajay Kumar Garg Engineering College Ajay Kumar Garg Engineering College
Ghaziabad, India Ghaziabad, India
[email protected] [email protected]

Abstract - COVID-19 (novel coronavirus disease) is a deadly high cost, more time consumed to give result, the quality of
illness, has infected and killed a very large number of people the test decreases as the testing demand increases and
worldwide. The widely followed lab testing (RT- PCR Test) cannot identify the presence of virus from chest part [3]. A
for the detection of this disease has various limitations with radiographic images direction is also providing a better
high cost and take long time to provide the outcome. As a support with existing tools that help to take these images as
result, diverse technologies that permit for the quick and Ultrasound, CT scan, X-ray etc. Especially the major focus
accurate finding of the infection can provide much required is kept on the Chest X-ray images because of its advantages
assistance to medical management. In recent studies, gained over other radiology images.
radiological imaging techniques, such images convey
important information about this virus. Advanced Deep The Chest X-ray imaging cab be commonly used to
learning (DL) techniques combined with the radiology images diagnosed for the abnormalities of the disease. These
can aid in the correct diagnosis of the virus, as well as defeat images are penetrating radiation of electromagnetic form,
the problem of insufficient expert physicians in rural areas. In
get passed through human body for getting the internal
this work, aimed at presenting a DL based – Convolutional
details of problem. As its also playing a major help in quick
neural network (CNN) model for the automatic detection of
diagnosis of coronavirus, whereas x-ray includes several
the coronavirus from X-ray images of chest. The Kaggle
benefits of easy availability of facility, reasonable cost and
dataset available publicly of total 42330 images from 4-
take less time for giving result. Also, a great help is
categories are used. The experiment produced the accuracy of
88.53% and 86.19% for training and validation, which is
provided for the rural areas, where testing facility is rare
better result for the highest number of radiographic images in
and includes high cost. The technological support for the
comparison to existing work. healthcare department has been very beneficial in last
decade, especially on the diagnosis point, Deep learning
Keywords-COVID-19, Deep Learning, CNN, Chest X- rays has overcome the various limitations. The neural network
has high potential in Medicare sector, mainly the image
I. INTRODUCTION processing for medical analysis and identifications of
various problems [2][7].
Since its debut in city of a China, named Wuhan in
December 2019, the pandemic (COVID-19) has infected
CNN is considered as one of the promising algorithms of
billions of people [1]. The virus caused a rapid spread of
deep learning, built up of several layers, provides an
infection. A great deal of research was done to determine
accuracy in the image recognition and classification [11].
the type of virus that caused COVID-19 sickness, and it was
The model is built based on CNN by changing layers,
determined that it belong to the influenza virus family. The
neurons and can perform differently for different created
condition of MERS (Middle East Respiratory Syndrome) is
models with good support of Result. In our work, have
originated by large family of respiratory viruses. The SARS
come up with a model based on Deep-CNN for analysing
(severe acute respiratory syndrome)-coronavirus (CoV)
the images of X-ray for the prediction of coronavirus.
and MERS-CoV [3]. The novel SARSCoV2 virus is
Section 2 explained about related work theory, followed by
capable of infecting humans’ pneumonia caused by a virus.
Section 3 gives a detailed information of Dataset used.
In some states, the population has had a particularly high
Section 4 presented the proposed approach of work in paper
death rate. The demise every day, the death toll rises around
along with representation of experiment observance in
the word.
Section 5. Last section is the conclusion presenting
proposed work.
The global healthcare authorities report, issued and
updated, tells the seriousness of the pandemic and people
II. BACKGROUND
in millions of numbers get affected globally. The
The World Health Organization has declared a Pandemic
identification of the virus research has been started by most
on 11th March 2020 named as COVID-19 and it has spread
countries, among them RT-PCR is widely preferred choice
for its detection and still after 2 years most population is rapidly in various countries worldwide, a situation never
using the approach, but still it has several limitations of faced before. The disease has its symptoms almost similar

978-1-6654-8004-8/22/$31.00 ©2022 IEEE 447

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.
1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-
2022)
to Pneumonia [1]. The detection can be done from the best result is provided by MobileNet. Total 1728 X-ray
radiology imaging tests, widely using Chest X-rays and CT images from Kaggle and GitHub is used in 80:20. Sarki et
(Computed Tomography) Scan, considered as the al. [11] proposed a deep learning-based system for the
alternative approach to RT-PCR test [2]. classification and experimented on the limited dataset from
open source and given an accuracy of 93.75%.
As these radiology images are processed and tested with
the advancement of technological support using machine III. DATASET
learning algorithms, based on DL approach using CNN
Architecture, which provides an efficient result in diagnosis The Chest X-ray images dataset used, has taken from
of various types of disease from medical images, as this Kaggle Repository, that has been prepared by the team of
algorithm is best suited for the classification of images [3]. researchers in collaboration with medical doctors from
As shown below in Table I various disease are diagnosed various online sources like SIRM, GitHub, etc. Published
with the DL support through radiographic pictures. papers, Padchest dataset, Radiology society of north
America (RSNA), Pneumonia dataset [18][19][20]. The
Table I. RELATED WORK ON VARIOUS DISEASE USING CNN Experiment dataset Contains a total number of 42330
Disease Reference Year Images Methodology images of 4-categories COVID-19, non- COVID-19 (Lung
Liver Cancer Yasaka et 2018 MRI Deep-CNN opacity & Viral Pneumonia), Normal, as presented below
[12] “Fig.1” presents the number of images per category. 70:30
Xin Dong 2020 CT Scan Hybridized Fully-
al. [13] CNN
is the training and testing dataset divided, where 29632
Pneumonia V.Sirish et 2020 Chest X- training images and 12698 testing images. “Fig. 2”
[14] ray CNN represents the various classes of X-ray images of Chest.
Breast Cancer Simon et 2020 MG Technique
[15] images
Brain Tumor Kavita et 2021 MRI No. of images various categories
[16]
Mohd. Arfi 2022 MRI 12000
al. [17] 10000
8000
6000
4000
Similarly, Deep Learning architectures also playing a 2000
0
major role in identifying the COVID-19 from chest part and
a great help will be provided in tackling with this pandemic.
Here, the related work as follows. Fatema et al. [4]
developed a CNN model in matlab and converted it to the
TensorFlow lite. The proposed model can be used on
android application for the Covid-19 prediction from Fig. 1. No. of images for various categories
radiology images. Total 1184 X-ray images are
experimented and giving an accuracy of 98.65% for 2-
category images. Areej A. et al. [5] applied CheXNet
algorithm mechanism to the CNN model and total 550 X-
ray images are operated for the prediction and attains an
accuracy of 89.7%. Shrinjal et al. [6] proposed work based
on Deep Learning approach and provided 87% accuracy for
multi-classification for the dataset that has been used from
open source of X-ray images of chest. Chakrovorti et al. [7]
(A) (B)
also presented the CNN method based on TensorFlow, for
total 1545 X-ray images of 2-classes giving their best
accuracy of 95%. Shadin et al. [8] have applied the transfer
learning Inception V3 and CNN architecture giving an
accuracy of 85.41% and 79.74% respectively and planning
to used their model for a large set of COVID-19 data.
Shivani et al. [9] a method is presented on Deep CNN
approach and experimented on CT scan and X-ray images (C) (D)
of 3- categories giving an accuracy of 94% training and
96% testing. This model is considered as less complex in Fig. 2. Different Categories of Chest X-ray images (A) COVID-19, (B)
Lung opacity, (C) Normal, (D) Viral Pneumonia [20]
comparison to others. Rawat et al. [10] proposed a COVID-
19 detection system with the help of 4 -CNN architecture –
Xception, InceptionV3, MobileNet, DenseNet121 where

448

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.
1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-
2022)
IV. PREPROCESSING AND
ARGUMENTATION

In the pre-processing point, all the dataset images of X-

ray are rescaled to 128x128 pixels and performed the
Normalization, in which the original image is converted to
Grayscale (1/255) 2D form. This step is the major
requirement for providing the quality images by vanishing
the undesired distortions and work for the required features,
helps in analysing an image in a better way. The maximum
and minimum pixel value taken is 2.0651 and -0.4842
respectively.

In augmentation, the changes are made in the training set,

in which the single image can be presented in different
ways to explore it widely. So, when the data is provided to
the network framework can avoid the problem of
overfitting and increase dataset. The following steps are Fig. 3. The proposed workflow
performed in augmentation as randomly rotation, shifting
horizontally and vertically and flipping of an image. A. Deep Learning

V. PROPOSED WORK Deep Learning is a part of Machine Learning Technique,

which works based on Artificial Neural Network, after
The following “Fig. 3” represents the proposed taking the idea from human brain process, that how from
workflow of our paper, which is followed as pre- learning and previous experience, the work or problem can
processing the dataset and divided it into training and be solved efficiently, improves quality and reduce the time
testing, trained the network on training dataset and perform to perform any work. The prepared model based on deep
testing on test data for trained CNN model. Further learning are widely preferred worldwide for certain tasks as
obtained the classification result for the prediction of Segmentation, Detection, and Classification. The radiology
COVID-19 with other categorical images. medical images like MRI, X-rays, CT scan, etc. are
currently operated widely with these types of models. It has
been seen in last decade, how DL has played an important
role in medical field especially Diagnosis of certain types
of disease and overcome the various limitations.

B. CNN

CNN is considered as the major approach for the image

recognition and classification in DL. It has performed better
in computer vision events and many competitions of image
analysis. The detailed process is based on the multiple
feature extraction at various stages that help to learn the
features automatically from provided dataset. Majorly there
are 3 layers and various other important components as
Activation function, Normalization, Dropout. The below
“Fig. 4” represent the presented CNN architecture. All
these terms are further explained as-

449

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.
1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-
2022)

Fig. 4. CNN architecture

• Convolutional Layer- This layer is the first and to the dense layer. The Table II presents the
base level of the Neural Network, in which the applied CNN layer-wise.
extraction work of the features is performed from
TABLE II. PROPOSED CNN LAYER-WISE ARCHITECTURE
the 2-D form of an original image, with the
support of Kernel or Feature Detector. The size of Layer Output Shape Parameter
the detector is either 5x5 or 3x3. The activation input layer (None, 128, 128, 3) 0
conv2d_3 (None, 126, 126, 3) 896
function is applied to each value of feature map. max_pooling2d_3 (None, 63, 63, 32) 0
The obtained output is passed as input for the next dropout_4 (None, 63, 63, 32) 0
layer. Relu Activation function is used after the conv2d_4 (None, 61, 61, 64) 18496
layer to maintain the non-linearity of the data. max_pooling2d_4 (None, 30, 30, 64) 0
dropout_5 (None, 30, 30, 64 0
• Pooling layer- As the Feature extraction is done in conv2d_5 (None, 28, 28, 128) 73865
the previous film, this layer helps in decreasing the max_pooling2d_5 (None, 14, 14, 128) 0
dimensionality of data. The main aim is extracting dropout_6 (None, 14, 14, 128) 0
flatten_1 (None, 25088) 0
the features combination which can reduce the
dense_2 (None, 256) 6422784
chance of overfitting, highlights the important dropout_7 (None, 256) 0
information and network complexity can be dense_3 (None, 4) 1028
regulated. Our model has used the Maxpooling
formulation with 2x2 dimension and further VI. EXPERIMENT AND RESULT
makes the connection to dense layer.
• Dense layer- This layer is also termed as fully- The experimentation work was performed for the
connected layer and similar to output layer, in training and validation set for 20 consecutive epochs. It was
which the input of each film is connected to all observed that the proposed model has given an 88.53% and
activation unit of the preceding layer. The result is 86.19% accuracy for the training and validation
provided for the final classification, using the respectively, whereas the generated loss for training and
SoftMax function for the prediction of the given validation is 0.3087 and 0.3956 respectively. As shown in
dataset values. Table III. The below “Fig. 5” represents the epoch v\s
• Dropout- This is a training strategy in which accuracy and epoch v\s loss graph for training and
randomly selected neurons are rejected. The validation.
chance for training a particular vertex in a layer,
TABLE III. EXPERIMENT RESULT
where no dropout implies as 1.0 and 0.0 denotes
no outputs from the layer, is the default Values Training Validation
interpretation of the dropout hyperparameter. In a Accuracy 0.8853 0.8619
Loss 0.3087 0.3956
hidden layer, a desirable value for dropout is
between 0.5 and 0.8.
• Flatten- The layer helps to covert the obtained 2-
D arrays input into vector form. The obtained
input is feature maps (pooled). To identify an
image, the flattened matrix is provided as an input

450

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.
1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-
2022)

Fig. 5. Epoch v/s accuracy and loss graph

The achievement of the classification model is observed TABLE 1V. PERFORMANCE METRICS
on following terms presented in below Table IV. with Performance Values Formulation
respect to COVID-19. They are specificity, sensitivity, Measure
precision, F1 score, False positive rate. All the terms are Specificity 0.8292 TN/(FP+TN)
calculated on the basis of actual and predicted values based Sensitivity 0.1711 TP/(TP+FN)
on four factors as explained. True positive as TP, refers as Positive predictive 0.1599 TP/(TP+FP)
correct result for true value. True negative as TN, refers as value
correct result for false value. False positive as FP, refers as False Positive Rate 0.1707 FP/(FP+TN)
false prediction for True values. False negative as FN, F1-Score 0.1655 2TP/(2TP+FP+FN)
refers as false prediction for false values. The confusion Accuracy 0.8853 (TP+TN)/(TP+FP+FN+TN)
matrix is presented in the “Fig. 6”.

Fig. 6. Confusion matrix

VII. CONCLUSION evolution of a DL techniques to aid Medicare experts in

detecting the pandemic from radiology images and
In the Diagnosis of COVID-19 problem from various lab determining the seriousness of the infection in a short time
testing process have faced with the many problems and and at reasonable cost. We Suggested a DL-based, CNN
need to come up with different approaches of detection, model to diagnosis and classify COVID-19 instances from
Artificial intelligence experts, have pushed for the

451

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.
1st International Conference on Computational Intelligence and Sustainable Engineering Solution(CISES-
2022)
an image of Chest X-ray. Presented concept is totally Network from Chest X-ray Images." In 2021 6th International Conference
for Convergence in Technology (I2CT), pp. 1-6. IEEE, 2021.
automated, with an end-to-end system that eliminates the
necessitate for human intervention. The model has been [9] Sharma, Shivani, and Shamik Tiwari. "COVID-19 Diagnosis using X-
trained for 40000+ images of 4-categories i.e., COVID-19, Ray Images and Deep learning." In 2021 International Conference on
Viral Pneumonia, Normal, Lung opacity cases from Kaggle Artificial Intelligence and Smart Systems (ICAIS), pp. 344-349. IEEE,
2021.
repository. The dataset is divided into 70:30 for training
and validation i.e., 29632 and 12698 respectively. The [10] Rawat, Ram Murti, Shivam Garg, Naman Jain, and Gagan Gupta.
achieved accuracy for the model is 88.53% for the training "Covid-19 detection using convolutional neural network architectures
based upon chest X-rays images." In 2021 5th International Conference
phase and 86.19% for the validation set. It has been on Intelligent Computing and Control Systems (ICICCS), pp. 1070-1074.
presented that the dataset used is highest in number for this IEEE, 2021.
experiment in comparison to previous work. The selection
[11] Sarki, Rubina, Khandakar Ahmed, Hua Wang, Yanchun Zhang, and
of X- ray images is done due to its low cost from other Kate Wang. "Automated Detection of COVID-19 through Convolutional
radiographic images and fast result capacity from other lab Neural Network using Chest x-ray images." Plos one 17, no. 1 (2022):
test. Our model, based on these promising results, is an e0262052.
intriguing aid that can assist primary-care in quickly
[12] Yasaka, Koichiro, Hiroyuki Akai, Akira Kunimatsu, Osamu Abe, and
diagnosing and detecting the infection. Furthermore, we Shigeru Kiryu. "Liver fibrosis: deep convolutional neural network for
want to authenticate the presented model by using staging by using gadoxetic acid–enhanced hepatobiliary phase MR
additional number of images from various hospitals, as well images." Radiology 287, no. 1 (2018): 146-155.
as to label CT images for COVID-19 identification and [13] Dong, Xin, Yizhao Zhou, Lantian Wang, Jingfeng Peng, Yanbo Lou,
compare the output with proposed model trained on applied and Yiqun Fan. "Liver cancer detection using hybridized fully
dataset images. convolutional neural network based on deep learning framework." IEEE
Access 8 (2020): 129889-129898.
REFERENCES
[14] Sirish Kaushik, V., Anand Nayyar, Gaurav Kataria, and Rachna Jain.
"Pneumonia detection using convolutional neural networks (cnns)." In
[1] Corman, Victor M., Olfert Landt, Marco Kaiser, Richard Molenkamp,
Proceedings of First International Conference on Computing,
Adam Meijer, Daniel KW Chu, Tobias Bleicker et al. "Detection of 2019
Communications, and Cyber- Security (IC4S 2019), pp. 471-483.
novel coronavirus (2019-nCoV) by real-time RT-PCR." Eurosurveillance
Springer, Singapore, 2020.
25, no. 3 (2020): 2000045.
[15] Hadush, Simon, Yaecob Girmay, Abiot Sinamo, and Gebrekirstos
[2] LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553),
Hagos. "Breast cancer detection using convolutional neural networks."
436 (2015)
arXiv preprint arXiv:2003.07911 (2020).
[3] Corman, Victor M., Olfert Landt, Marco Kaiser, Richard Molenkamp,
[16] Bathe, Kavita and Rana, Varun and Singh, Sanjay and Singh, Vijay,
Adam Meijer, Daniel KW Chu, Tobias Bleicker et al. "Detection of 2019
Brain Tumor Detection Using Deep Learning Techniques (MAY 7, 2021).
novel coronavirus (2019-nCoV) by real-time RT-PCR." Eurosurveillance
Proceedings of the 4th International Conference on Advances in Science
25, no. 3 (2020): 2000045.
& Technology (ICAST2021), Available at SSRN:
[4] Bushra, Kanij Fatema, Md Ahamed, and Mohiuddin Ahmad. https://ssrn.com/abstract=3867216
"Automated detection of COVID-19 from X-ray images using CNN and
[17] Arif, Muhammad, F. Ajesh, Shermin Shamsudheen, Oana Geman,
Android mobile." Research on Biomedical Engineering 37, no. 3 (2021):
Diana Izdrui, and Dragos Vicoveanu. "Brain Tumor Detection and
545-552.
Classification by MRI Using Biologically Inspired Orthogonal Wavelet
[5] Magher, Ashraf Yunis. "COVID-19 Detection in X-ray Images using Transform and Deep Learning Techniques." Journal of Healthcare
CNN Algorithm." In 2020 International Conference on Promising Engineering 2022 (2022).
Electronic Technologies (ICPET), pp. 5-9. IEEE, 2020
[18] M.E.H. Chowdhury, T. Rahman, A. Khandakar, R. Mazhar, M.A.
[6] Singh, Shrinjal, Piyush Sapra, Aman Garg, and Dinesh Kumar Kadir, Z.B. Mahbub, K.R. Islam, M.S. Khan, A. Iqbal, N. Al-Emadi,
Vishwakarma. "CNN based Covid-aid: Covid 19 Detection using Chest M.B.I. Reaz, M. T. Islam, “Can AI help in screening Viral and COVID-
X-ray." In 2021 5th International Conference on Computing 19 pneumonia?” IEEE Access, Vol. 8, 2020, pp. 132665 - 132676.
Methodologies and Communication (ICCMC), pp. 1791-1797. IEEE,
[19] Rahman, T., Khandakar, A., Qiblawey, Y., Tahir, A., Kiranyaz, S.,
2021.
Kashem, S.B.A., Islam, M.T., Maadeed, S.A., Zughaier, S.M., Khan, M.S.
[7] Chakravorti, Tatiana, Vinay Kumar Addala, and J. Shivam Verma. and Chowdhury, M.E., 2020. Exploring the Effect of Image Enhancement
"Detection and Classification of COVID 19 using Convolutional Neural Techniques on COVID-19 Detection using Chest X-ray Images.
Network from Chest X-ray Images." In 2021 6th International Conference
[20] https://www.kaggle.com/datasets/tawsifurrahman/covid19-
for Convergence in Technology (I2CT), pp. 1-6. IEEE, 2021.
radiography-database.
[8] Chakravorti, Tatiana, Vinay Kumar Addala, and J. Shivam Verma.
"Detection and Classification of COVID 19 using Convolutional Neural

452

Authorized licensed use limited to: MUKESH NARUKULLAPATI. Downloaded on October 28,2023 at 05:49:12 UTC from IEEE Xplore. Restrictions apply.