Page 36 - AIH-2-1

Artificial Intelligence in Health Deep learning on chest X-ray and CT for COVID-19

2 School of Minerals, Metallurgical and Materials Engineering, Indian Institute of Technology, Bhubaneswar, Odisha, India
3 Department of Pulmonary Medicine, Srirama Chandra Bhanja Medical College and Hospital, Cuttack, Odisha, India
4 Centre of Excellence for Novel Energy Materials (CENEMA), Indian Institute of Technology, Bhubaneswar, Odisha, India
5 Virtual and Augmented Reality Centre of Excellence, Indian Institute of Technology, Bhubaneswar, Odisha, India

1. Introduction collapse of the lungs caused by COVID-19) can be partially
reversed by maintaining continuous positive airway
In December 2019, clusters of pneumonia-like cases of pressure, which forces the collapsed alveoli to remain open
unknown origin were first reported in the city of Wuhan with oxygen-rich air. A far worse situation demands the
in China. Upon investigations, it was found that the use of ventilators, where breathing is assisted with a life-
disease was caused by a new type of single-stranded RNA support device. Thus, since it is a common symptom across
virus, which was officially named by the World Health various respiratory diseases, depleting oxygen saturation
Organization (WHO) as severe acute respiratory syndrome cannot be a definitive indicator of COVID-19 infection.
coronavirus 2 (SARS-CoV-2). The disease caused by
1,2
SARS-CoV-2 was subsequently named COVID-19. The To arrest the spread of the disease, affected countries
rate of spread and severity of COVID-19 all around the have adopted reverse-transcription polymerase chain
world forced the WHO to declare it as a pandemic on reaction (RT-PCR) tests as the gold-standard diagnostic
March 11, 2020. It had caused unparalleled disruptions method to detect COVID-19 infection and isolate the
in the post-internet modern era of global civilization and, infected individuals as early as possible to limit further
in a way, demonstrated the serious shortcomings and transmission of the infection. Despite having many positive
vulnerabilities of traditional approaches for tackling such aspects, this testing method requires the setting-up of
2
issues. It was depicted that, although we had the bits and entirely new facilities with specially trained personnel for
pieces to construct a robust technology-armed first line sample collection and analysis. Such facilities are largely
of defense against global pandemics based upon smart authority-regulated and many underdeveloped countries
integration of the new age tools, we did not invest enough find it difficult to procure sufficient numbers of test kits.
to construct the requisite technological architecture to Moreover, it is time-consuming and sometimes gives
realize this. Therefore, various countries attempted to do non-negligible false negative (infected person is tested
what they could do best under the prevailing situations. negative) as well as false positive (uninfected person is
Some implemented nationwide lockdowns, limiting tested positive) results.
movement of the population, commanding social Machine learning (ML) has made a significant impact
distancing, and expanding the consciousness of cleanliness in many disciplines of science and technology. 4-11 An ML
and good hygiene as a range of preventive measures against tool, based on chest X-ray and/or computed tomography
the pandemic; an early analysis of such measures has been (CT) scan images of COVID-19 suspected individuals,
reported in the literature. 3 can be very attractive in this scenario, primarily because
SARS-CoV-2 can cause symptoms of fever, fatigue, of the very limited resources required and short diagnosis
headache, cough, sore throat, myalgia (muscle pain), time. Computer-aided biomedical image analysis might
anosmia (loss of smell), and other respiratory symptoms. turn out to be an additional useful tool to assist medical
Most people recover from the disease without requiring practitioners in correct and quick decision-making.
7
any special treatment but those who have comorbid For example, Habib et al. introduced a novel modified
medical conditions such as chronic kidney disease, lightweight SqueezeNet (SQN-MF) model (demonstrated
autoimmune disease, cancer, heart conditions, obesity, in non-medical application) coupled with continuous
diabetes, and respiratory disease are more prone to develop wavelet transformation for converting acoustic emission
serious illness. COVID-19 causes serious damage to the signals into two-dimensional (2D) images, achieving 100%
3
lungs, causing oxygen deficiency in the patient, a condition classification accuracy (surpassing traditional techniques
referred to as hypoxia. This can be easily identified by by 20.8%). The lightweight model (0.5 MB) is suitable for
lower oxygen saturation level, typically less than 94%. field programmable gate array implementation, enabling
Infections due to bacterial pneumonia and pulmonary real-time monitoring. Their success in achieving high
embolism present similar characteristics. Furthermore, accuracy with a memory-efficient model demonstrates the
other underlying conditions such as chronic obstructive potential for similar approaches in medical diagnostics
pulmonary disease typically complicate the identification that will help to build reliable, rapid screening protocols for
of COVID-19. Atelectasis (indicating a partial or complete infectious diseases like COVID-19. This kind of lightweight

Volume 2 Issue 1 (2025) 30 doi: 10.36922/aih.2888