Document Type

Project

Lead Author Type

CIS Masters Student

Advisors

Dr. Greg Wolffe, wolffe@gvsu.edu

Embargo Period

3-8-2021

Keywords

Deep Learning, Computer Vision, COVID-19, Chest X-rays

Abstract

Over the past year, COVID-19 has affected countries world-wide. COVID-19 detection tests have allowed us to control the spread of the disease; however, COVID-19 testing kits are highly specialized and difficult to procure in quantity. X-rays, on the other hand, have broad clinical usage and therefore tend to be readily available. As a result, radiologists have begun using chest X-rays to diagnose COVID-19 in patients with respiratory distress. The goal of this research project was to demonstrate that deep learning can be used to automatically detect COVID-19 from chest X-rays. Automated detection with deep learning models could help make X-ray diagnosis even more efficient, relieving the burden on radiologists, especially those not specifically trained in detecting COVID-19. However, training deep learning models for computer vision tasks requires a large amount of labeled data, which does not yet exist for COVID-19 chest X-rays. Therefore, this project utilized transfer learning to adapt a pre-trained ResNet50 model for the COVID-19 classification task. Results demonstrated that this approach can successfully classify chest X-rays (normal versus viral pneumonia versus COVID-19) with 94% overall accuracy. In the binary classification problem (COVID-19 versus other), this approach could classify chest X-rays with an even higher accuracy of 98%. In this project, I also implemented gradient class activation maps to visualize different convolutional layers of the model. These visualizations highlight those areas within each X-ray the model found most significant, providing a degree of interpretability to the classifications. In summary, this research project established the viability of using deep learning for detecting respiratory conditions via computer vision, despite the current limited availability of data.

Share

COinS