Interpretable Protocol: A Novel Learning Strategy for COVID-19 Diagnosis on Chest- X-Ray Images

Abstract

Deep learning-based models (e.g., ResNet18 and ResNet50) have been employed for detecting COVID- 19 on chest-X-ray (CXR) images which have been reported to achieve good accuracy. Although these models have the ability to correctly classify the CXR images from the dataset COVID-Xray-5k into COVID- 19 and non-COVID-19 classes, our investigation shows for the first time that the good results obtained by existing methods may come from regions out of the lungs area on the images from this dataset. It is an unsolved problem that the regions used to make such decisions are automatically located in the lung area where the evidence of COVID-19 is to be found. To this end, this paper proposes an interpretable protocol for COVID-19 detection on CXR images. The proposed protocol not only improves the prediction performance but more importantly increases interpretability and trust of such prediction without the need of any region annotation. The proposed protocol is a learning strategy that can be applied to any convolutional neural networks (CNN) models. The experimental results demonstrate the superiority of the proposed strategy over the related state-of-the-arts. 

https://doi.org/10.59200/ICONIC.2024.004