High Technology Letters ISSN NO : 1006-6748

Performance Analysis of Various Deep Learning Algorithms in COVID-19

Detection with Chest X-RAY images.
A. Naga Kalyani
BVRIT HYDERABAD College of Engineering for Women
Computer Science & Engineering

L.Lakshmi
BVRIT HYDERABAD College of Engineering for Women
Computer Science & Engineering

Abstract

Corona Virus is one of the large families of viruses causing infection vacillating from cold to
more serious illnesses such as respiratory problems. This virus was identified in Wuhan,
China named COVID-19. To forecast the spread of the coronavirus deep learning algorithms
for image classification are widely used. This paper provides useful insights into the analysis
of various deep learning algorithms used in COVID-19 detection with chest X-Ray images.
We have used the Kaggle chest X-Ray data set for analysis of COVID positive or negative.
The deep learning model used here is VGG16 with filters of size 3 x 3 with the stride of 1 and
the same padding and all the max pool layers have the kernel of size 4 X 4 with the stride of
2. The model is trained and tested using SGD, ADAM, and ADAMAX optimizers with
ReLU and ELU activation functions. The experimental results revealed that ADAM and
ADAMAX optimizer with ELU activation function produced 100 percent results.

Keywords:- Deep learning, activation functions, optimizers, ADAM, SGD, ADAMAX, X-

Ray images.

I. INTRODUCTION
Covid-19 also known as Coronavirus outburst was identified in December 2019. The
epicenter of the virus was found in Wuhan City, China [1]. The virus was spread across the
globe and affected many parts of the USA, Russia, Europe, Asia, Australia, and some parts of
Africa. Millions of people tested positive and lakhs of people lost their life.

The virus mainly affects the respiratory system of the human body and it spread with tiny
droplets formed during talking, sneezing, and coughing. The virus mainly enters the human
body through mouth and nose oozing. The people may breathe in those droplets and may get
infected [2]. Test, Isolate, and trace are the only strategy suggested by the World Health
Organization to deal with this pandemic. Experts have time and again mentioned that testing
is the only way to know the status of the pandemic. Testing is essential to know who, why,
where, and how many are infected by the coronavirus.

In the case of testing, if two people get infected and they have not been diagnosed and
quarantined then they may spread the virus potentially to 800 people [3]. If they are tested
and quarantined then they may spread the virus potentially to 5 people. Covid-19 testing is
done in two different approaches. Firstly, the antibody test and secondly virus content test. In

Volume 26, Issue 9, 2020 1160 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

some parts of the world due to lack of adequate access to testing, experts have tried to use CT
scan images or X-ray images of the chest to diagnose COVID-19 efficiently compared with
human body thermal scanning [4]. The proposed work mainly focuses on the processing of
the chest x-ray images to identify if COVID is positive or negative.

In recent years, due to the great improvement in image processing and classification, due to
advances in deep learning algorithms and their ability to process large data sets [5]. So we
can use deep learning algorithms to process the COIVD X-ray dataset and efficiently detect it
is positive or not. The symptoms COVID are similar to pneumonia, as both affect the
respiratory system of the human body, so for the detection of COVID, the best way is to use
X-ray testing. This paper presents a comparative study of various deep learning used in X-ray
based classification of COVID positive or negative patients.

II. Related Works

Makris et al., 2020 proposed a pre-trained CNN model using chest X-ray images to identify
the COVID, infected patients. They have created a dataset with X-ray images of both COVID
positive and negative patients. They have used transfer learning for transferring knowledge
from a pre-trained model to the models to be trained. They have achieved good accuracy and
precision results but they have used a small dataset and it can be used with more
architectures.

The study idea of coronavirus, its symptoms, and various methods of testing are well
described in by Lu et al., 2020. In this paper, they have studied and experimented testing for
coronavirus using chest x-ray images or chest tomography CT images are routine testing
tools for pneumonia for fast diagnosis.

Lowe, 1999 developed a visible recognition system has been established that uses native
options that square measure invariant in terms of rotation, location, and, in part, changes in
illumination. The experimental results revealed that it performs tough object recognition with
partially cluttered images with less computation time and with these efforts presently, the
primary real-world face recognition program was place operative in 2001.

Considering the large rate of COVID suspects and very few trained radiologists, Kanne et al.,
2020 proposed some methods for identification of such abnormalities can assist the
diagnostic procedure and improve the rate of early diagnosis effectively. Computer Vision,
Machine, and Deep Learning are powerful tools in such applications. Whereas not centered
on deep learning, the algorithmic program has learned that options might facilitate to spot
faces. Once the realm of laptop vision began to develop, the cluster felt associate pressing
want for a regular image dataset and customary assessment measures to live the success of
their models.

As X-ray imaging systems are cheaper than CT scan systems Das et al., 2020 proposed a
deep learning-based CNN model known as Truncate Inception Net to analyze COVID
positive chest X-rays. They have used TB positive, pneumonia positive, COVID positive, and
healthy person data set for analysis. Their system provides a very good accuracy of 99.96 in

Volume 26, Issue 9, 2020 1161 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

classifying the data of persons with these health problems. This model is evaluated on a
dataset of people in a particular region and a model can be tested on cross-population data set
for better evaluation of the system.

III. Chest X-Ray Dataset:

In this pandemic situation, it is very hard to find the dataset for outbreaks like covid19
especially when the entire world is looking more to test through collecting blood samples or
through thermal imaging instead of X-rays to detect COVID positive patients. The predefined
chest x-ray image dataset is extracted from Kaggle which is contributed by (Cohen et al.,
2020). It contains chest X-ray images of COVID positive patients and negative patients. After
separating the noisy images from the dataset with two labels, normal and COVID with 10000
images in each label is extracted.

IV. Proposed System

The approach for classification has been divided into four main steps. Load the data,
preprocess and split: Load the image dataset preprocess the image by converting to RGB
channel ordering and, resizes the 224X224 and then perform the split. The dataset is split into
a test and train set where 80% images are provided for training and 40% images for testing.
Configure and build the model: After splitting the dataset use the predefined network model
and experiment for performance analysis of optimizers and activation functions. The
architecture proposed system is shown in Fig. 1

Fig 1: The architecture of VGG16 model

1. Network model: VGG16 has 16 convolutional neural networks that are 16 layers
deep. In the VGG architecture, all the convolutional layers use filters of size 3 x 3
with a stride of 1 and the same padding, and all the max pool layers have the kernel of
size 4 X 4 with the stride of 2.

2. Activation functions: These functions will decide whether to pass a signal to the next
layer or not. The activation functions used widely on CNN are ReLU and ELU.

3. Optimizers: While training a Neural Network we aim to minimize the loss function
and by updating the weights and make our predictions as accurate as possible. In the
paper, we used SGD, ADAM, ADAMAX.

Volume 26, Issue 9, 2020 1162 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

4. Training and Evaluation of the model: The VGG16 model is trained for 25 epochs at a
learning rate of le-2, loss function as binary_cross entropy, batch-size of 8, and a
Dropout rate of 0.001. Implementing hyperparameters improves the performance of
the model.

V. Experimental and Result

For experimental evaluation, we are using the Kaggle dataset with VGG16 by using different
activation functions and optimizers. Table 1 represents the loss and accuracy of the proposed
model during training and testing.

Activation Optimizer Train Test Loss Train Acc. Test Acc.

Loss
ReLU SGD 58% 56% 90% 90%
ELU SGD 5% 5% 77% 90%
ReLU ADAM 2% 18% 100% 90%
ELU ADAM 0.8% 0.1% 100% 100%
ReLU ADAMAX 6% 10% 100% 100%
ELU ADAMAX 3% 0.9% 100% 100%
Table1: Loss and Accuracy obtained various models with different activation functions

ReLU_SGD loss and accuracy graph shown in Fig 2, represents the loss and accuracy using
the SGD classifier with the ReLU activation function. We have evaluated the model for 25
epochs the model attained train and test loss of 58 and 56 percent each. The accuracy of the
model is recorded as 50 percent in the initial epochs but subsequently achieved 90 percent
train and test accuracy which is a sensibly good accuracy.

Fig 2: Loss and Accuracy obtained using SGD optimizer with ReLU activation function

ELU_SGD loss and accuracy graph shown in Fig 3, represents the loss and accuracy using
the SGD classifier with ELU activation function. We have evaluated the model for 25 epochs
the model attained train and test loss of 5 percent each. The accuracy of the model is recorded
as 45 percent in the initial epochs but subsequently achieved 77 percent train and 90 percent
test accuracy which is a sensibly good accuracy.

Volume 26, Issue 9, 2020 1163 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

Fig 3: Loss and Accuracy obtained using SGD optimizer with ELU activation function

ReLU_ADAM loss and accuracy graph is shown in Fig 4, represents the loss and accuracy
using the ADAM classifier with ELU activation function. We have evaluated the model for
25 epochs the model attained train and test loss of 2 and 18 percent each. The accuracy of the
model is recorded as 48 percent in the initial epochs but subsequently achieved 100 percent
train and 90 percent test accuracy.

Fig 4: Loss and Accuracy obtained using ADAM optimizer with ReLU activation function

ELU_ADAM loss and accuracy graph is shown in Fig 5, represents the loss and accuracy
using the ADAM classifier with ELU activation function. We have evaluated the model for
25 epochs the model attained train and test loss of 0.8 and 0.1 percent each. The accuracy of
the model is recorded as 55 percent in the initial epochs but subsequently achieved 100
percent train and test accuracy.

Volume 26, Issue 9, 2020 1164 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

Fig 5: Loss and Accuracy obtained using ADAM optimizer with ELU activation function

ReLU_ADAMAX loss and accuracy graph is shown in Fig 6, represents the loss and
accuracy using the ADAMAX classifier with ELU activation function. We have evaluated
the model for 25 epochs the model attained train and test loss of 6 and 10 percent each. The
accuracy of the model is recorded as 50 percent in the initial epochs but subsequently
achieved 100 train and test percent accuracy.

Fig 6: Loss and Accuracy obtained using ADAMAX optimizer with ReLU activation function

ELU_ADAMAX loss and accuracy graph is shown in Fig 7, represents the loss and accuracy
using the ADAMAX classifier with ELU activation function. We have evaluated the model
for 25 epochs the model attained train and test loss of 3 and 0.9 percent each. The accuracy of
the model is recorded as 45 percent in the initial epochs but subsequently achieved 100 train
and test percent accuracy.

Volume 26, Issue 9, 2020 1165 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

Fig 7: Loss and Accuracy obtained using ADAMAX optimizer with ELU activation function

VI. CONCLUTION
The outbreak of the coronavirus has affected many people's lives in different ways. The most
common symptoms include dry cough, fever, and tiredness. It mostly affects the respiratory
system of the human body. To effectively assess whether the person affected by the
coronavirus or not we can test the X-Rays of the human chest. Nowadays most of the health
care systems using AI-driven technology, to detect the coronavirus efficiently. We have used
deep learning-based VGG16 with filters of size 3 x 3 with the stride of 1 and the same
padding and all the max pool layers have the kernel of size 4 X 4 with the stride of 2. The
model is trained and tested using SGD, ADAM, and ADAMAX optimizers with ReLU and
ELU activation functions. The experimental results revealed that ADAM and ADAMAX
optimizer with ELU activation function produced 100% results.

References
1. Gorbalenya, A.E., 2020. Severe acute respiratory syndrome-related coronavirus–The
species and its viruses, a statement of the Coronavirus Study Group. BioRxiv.
2. Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., Zhang, L., Fan, G., Xu, J., Gu,
X. and Cheng, Z., 2020. Clinical features of patients infected with 2019 novel
coronavirus in Wuhan, China. The Lancet, 395(10223), pp.497-506.
3. Wang, C., Horby, P.W., Hayden, F.G. and Gao, G.F., 2020. A novel coronavirus
outbreak of global health concern. The Lancet, 395(10223), pp.470-473.
4. Essentials for radiologists on COVID-19: an update - radiology scientific expert
panel, Kanne et al., 2020
5. Viola, P. and Jones, M.J., 2004. Robust real-time face detection. International journal
of computer vision, 57(2), pp.137-154.
6. Makris, A., Kontopoulos, I., & Tserpes, K. (2020). COVID-19 detection from chest
X-Ray images using Deep Learning and Convolutional Neural Networks. MedRxiv,
December 2019, 2020.05.22.20110817. https://doi.org/10.1101/2020.05.22.20110817.
7. Lu, H., Stratton, C. W., & Tang, Y. W. (2020). Outbreak of pneumonia of unknown
etiology in Wuhan, China: The mystery and the miracle. Journal of Medical Virology,
92(4), 401–402. https://doi.org/10.1002/jmv.25678
8. Lowe, D. G. (1999). Object recognition from local scale-invariant features.
Proceedings of the Seventh IEEE International Conference on Computer Vision,
1150–1157 vol.2. https://doi.org/10.1109/ICCV.1999.790410

Volume 26, Issue 9, 2020 1166 http://www.gjstx-e.cn/

High Technology Letters ISSN NO : 1006-6748

9. Kanne, J. P., Little, B. P., Chung, J. H., Elicker, B. M., & Ketai, L. H. (2020).
Essentials for Radiologists on COVID-19: An Update—Radiology Scientific Expert
Panel. Radiology, 296(2), E113–E114. https://doi.org/10.1148/radiol.2020200527
10. Das, D., Santosh, K. C., & Pal, U. (2020). Truncated inception net: COVID-19
outbreak screening using chest X-rays. Physical and Engineering Sciences in
Medicine, 43(3), 915–925. https://doi.org/10.1007/s13246-020-00888-x
11. Cohen, J. P., Morrison, P., & Dao, L. (2020). COVID-19 Image Data Collection.
arXiv:2003.11597 [Cs, Eess, q-Bio]. http://arxiv.org/abs/2003.11597

Volume 26, Issue 9, 2020 1167 http://www.gjstx-e.cn/