Page 44 - AIH-2-1

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

to 0.01) to various network layers. This nuanced the decoupling of physical examination of the patient and
approach enhances both the low-level feature extraction analytical pathology leads to an effective and modular
and high-level abstraction capabilities of the models. approach. This is likely to significantly enhance detection
(3) One-cycle policy implementation. We integrated a one- speed, accuracy, and sensitivity, expected to form the
cycle training policy that dynamically adjusts learning fundamental cornerstone that will be pivotal for an
rates across epochs. This method has been shown to extensive digital architecture to safeguard against many
significantly improve model convergence, stability, future pandemics (to be elaborated in the follow-up article).
and overall performance. Furthermore, these models help in the early screening of
(4) Comparative analysis of advanced architectures. While suspects in remote places in countries where the health
individual CNN architectures have previously been care providers as well as resources (such as RT-PCR kits
applied to COVID-19 detection, our study provides and CT scan machines) are limited.
a comprehensive comparison of EfficientNet, ResNet,
and SeResNext using these advanced training Acknowledgments
strategies. This comparison offers valuable insights The authors acknowledge the sincere help from the
into the relative strengths of these architectures for authorities of VARCoE, IIT BBS, and SCBMCH for the
this specific task. research support and encouragement.
(5) Reproducibility and benchmarking. By using a
publicly available dataset and clearly documenting Funding
our methodologies, we provide a robust benchmark None.
for future studies in medical image analysis, which
extends beyond the scope of COVID-19 detection. Conflict of interest
These methodological innovations collectively enhance The authors declare no conflicts of interest.
the robustness, accuracy, and generalizability of CNN
models for medical image analysis. While the immediate Author contributions
application to COVID-19 may seem less pressing now, the Conceptualization: Ajay Kumar Gogineni, Kisor Kumar
techniques we developed have broader implications for Sahu
improving deep learning approaches in medical imaging Formal analysis: All authors
across various conditions that might be very useful in our Investigation: Ajay Kumar Gogineni, Madapathi Hitesh
fight against future pandemics. Methodology: Ajay Kumar Gogineni, Kisor Kumar Sahu

Among the models implemented in the present study, Writing – original draft: All authors
ResNet and DenseNet have achieved more than 94% Writing – review & editing: All authors
accuracy. This is far superior to the typical sensitivity of 70
– 80% for RT-PCR. Our results indicate that EfficientNet Ethics approval and consent to participate
is best at classifying normal images, and SeResNext is Not applicable.
best at classifying pneumonia. ResNet performs best for
classifying images pertaining to COVID-19. While the Consent for publication
accuracy of the present method is expected to get better Not applicable.
with increasing usage, which is an inherent feature of
artificial intelligence, there is no such chance for RT-PCR, Availability of data
since this traditional method is not a smart protocol. The
model is able to learn the inherent features of pneumonia Data are available at the following resources:
and COVID-19 from a relatively small dataset. The (1) Cohen JP, Morrison P, Dao L. COVID-19 Image Data
Collection. arXiv.org. Accessed April 9, 2021. https://
performance of the model can be improved further by doi.org/10.48550/arXiv.2003.11597
collecting data from diverse geographical regions. This will (2) Yang X, He X, Zhao J, Zhang Y, Zhang S, Xie P. COVID-
also improve the generalizability of the model.
CT-Dataset: A CT Scan Dataset about COVID-19.
We strongly believe that this ML-aided diagnostic arXiv:200313865 [cs, eess, stat]. Published online June
protocol can help in detecting individuals suspected of 17, 2020. http://arxiv.org/abs/2003.13865
carrying infections with greater speed and accuracy, and (3) Cohen JP. ieee8023/covid-chestxray-dataset. GitHub.
more importantly, it charts out the blueprint to rapidly Published June 10, 2020. https://github.com/ieee8023/
develop a new med-tech protocol for quick screening of covid-chestxray-dataset
future pandemics. It is pertinent to point out here that (4) Chest X-Ray Images (Pneumonia). www.kaggle.

Volume 2 Issue 1 (2025) 38 doi: 10.36922/aih.2888