Authors:
Srishti Lodha, Harsh Malani, Arvind Kumar Bhardwaj

Addresses:
1,2, School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, Tamil Nadu, India. 3Department of Information Technology, Capgemini, Houston, Texas, United States of America. shrishti2k1@gmail.com1, harsh13092001@gmail.com2, arvind.bhardwaj@capgemini.com3  

Abstract:

Pneumonia, a severe and life-threatening result of bacterial or viral infection, can inflate air sacs in the lungs. The disease, which can target young and old people, is common in several countries. Generally, blood tests, pulse oximetry, sputum tests, CT scans, and chest X-rays are used to diagnose Pneumonia. Deep Learning (DL) models can be excessively helpful in analyzing the results of these tests. Over the past few years, several studies have suggested the implementation of different DL architectures for Pneumonia detection. However, these incorporate many trainable parameters for feature extraction from images, leading to a significantly high training time and resource consumption. Moreover, convolutions become monotonous after a certain number of layers, making it extremely difficult to improve the accuracy. In this research, we use Vision Transformers (ViT) for Pneumonia detection, an image classification architecture developed by modifying transformers in 2021. To our knowledge, ViT has only been implemented in one study before this research for Pneumonia diagnosis. Our approach outperformed all existing research and state-of-the-art architectures in this domain regarding all performance metrics and training time and recorded a validation accuracy of 98.18%. We also compare our model’s performance with other tuned DL models (CNN) and analyze the performance gap.  

Keywords: Performance Evaluation; Vision Transformers; Diagnosis of Pneumonia; CT Scan and Chest X-rays; Deep Learning (DL) Models; Architectures for Pneumonia; Blood Tests.

Received on: 04/11/2022, Revised on: 03/01/2023, Accepted on: 09/02/2023, Published on: 25/02/2023

FMDB Transactions on Sustainable Computing Systems, 2023 Vol. 1 No. 1, Pages: 21-31

• Views : 216
• Downloads : 17

Download PDF