COVID-19 detection from chest X-Ray images
using ensemble of CNN models
Sagar Deep Deb
Dept. of Electrical Engineering
Indian Institute of Technology Patna
Patna, India
sagardeepdeb@gmail.com
Rajib Kumar Jha
Dept. of Electrical Engineering
Indian Institute of Technology Patna
Patna, India
jharajib@iitp.ac.in
Abstract—The year 2020 will certainly be remembered for
the outbreak of COVID-19 pandemic. With the first case being
reported in December 2019, the SARSCoV2 virus has proved to
be one of the most deadly virus which has affected the human
civilization. Relatively high contagious rate and asymptomatic
patients also being carrier of the virus makes it more dangerous.
The only way to get a control on the outbreak is rapid testing.
With the present testing mechanisms being costly and time
consuming the end of this pandemic doesn’t seems near. These
challenges motivates us to come up with a system which can be
effective in testing large population size and at the same time be
less time consuming. We have proposed a Deep Convotuional
Neural Network based ensemble architecture for extracting
features from Chest X-Ray images and later classifying them into
three categories namely- Community Acquired Pneumonia(CAP),
Normal and COVID-19. We have shown that applying such tech-
nique can give better performance. Our ensemble network uses
three pre-trained DCNN networks namely- NASNet, MobileNet
and DenseNet. The low level features extracted from the three
DCNN architectures are later concatenated and applied to a
classifier for final classification. We have achieved an accuracy
of 91.99% which is slightly better than the state of the art
performances.
Index Terms—COVID-19, Chest X-Ray, DCNN, ensemble
model
I. I NTRODUCTION
The first case of COVID-19 was detected in the Wuhan
province of China back in December 2019 and started spread-
ing all across the world since then. The World Health Orga-
nization or WHO has declared it a pandemic in 11th March
2020. Today on October the 12th we have 7.18 million cases
and 110,135 people have succumbed to the deadly virus world
wide. The spread of this disease seems uncontrollable even
after different part of the world had been under harsh lock
down at different point of time. The R0 value is a parameter
which measures how contagious an infectious disease is and
thus it is one of the important scalar to study how the
contagious disease has progressed. R0 can be used to find out
the transmission and decline rate of that particular disease.
Data suggests that the R0 value which changes with geogra-
phy, population density and other important social measures
followed in a community is higher than the deadly virus
outbreak of 1918. The rapid spread of a disease of this kind
© IEEE 2021. This article is free to access and download, along with rights for full text and data mining, re-use and analysis.
demands a highly accurate point-of-care COVID-19 screening
[1]. The gold standard screening approach based on Reverse
Transcription Polymerase Chain Reaction (RT-PCR) shows
good accuracy but is subject to considerable cost and slow
turnover time constraints, rendering it not scalable to the ever-
increasing population at risk. [2]. Though the countries are
giving most importance to testing but again the large scale
testing is somewhat impossible with the given high cost and
non scalable testing equipment.
Researchers in the field of bio-medical image processing
and Artificial Intelligence have tried to find a solution with the
available resources. [3]–[5] have displayed great potential with
Computed Tomography (CT) images for detection of COVID-
19. The wide availability of Chest X-Ray (CXR) in diverse
health care settings makes it an attractive option for rapid,
accurate yet inexpensive point-of-care screening in primary
care clinics.
II. RELATED WORKS
Convolutional Neural Networks(CNN) has time and again
proved to be the best machine learning algorithm for image
classification and recognition task. For almost all bio-medical
image classification problems researchers have used CNN.
Like for lung cancer [7], blood cancer [8] and breast cancer
[9] researchers have reported state of the art performance
using CNN structures. Detection and classification of COVID-
19 from both Chest X-Ray and CT images being an image
classification task, also attracted researchers to train deep
learning models on the largely public datasets available. Li et.
al. [1] presented COVID-MobileXpert, a DCNN based mobile
application that can used for point-of-care COVID-19 screen-
ing from given noisy snapshots of chest X-ray image. Goodwin
et. al. [10] have used twelve most common pre-trained DCNN
structures to compare the results obtained for classification of
Chest X-Ray images into healthy, Community acquired pneu-
monia(CAP) and COVID-19. The highest accuracy obtained
by them is 88.4% using Densenet201 architecture. In [11]
the researchers have used inception migration-learning model
to detect COVID-19 from 453 CT images and achieved an
accuracy of 73.1%. [12] have also used Chest CT images for
detection of COVID-19. They have reported an AUC of 0.96.
2020 International Conference on Power, Instrumentation, Control and Computing (PICC)
2020 International Conference on Power, Instrumentation, Control and Computing (PICC) | 978-1-7281-7590-4/20/$31.00 ©2020 IEEE | DOI: 10.1109/PICC51425.2020.9362499