ABSTRACT
The health crisis caused by the COVID-19 pandemic has led to unprecedented research efforts to build AI solutions that can assist healthcare systems. In this work, we propose a novel CNN-based system that detects COVID-19 infection and other pneumonia from CT scans, segments COVID-specific lesions, namely Ground Glass Opacities (GGO) and Consolidations (CL), and computes the percentage of lungs that have been affected by COVID and provides an explanation of the basis in which the diagnosis has been made through the comparison of class activation maps pertaining to the diagnosis with the segmented lesions. This can assist healthcare setups in the rapid contactless screening of COVID-19 and assess the stage and severity of the disease, while also providing some level of transparency on the rationale behind the model's decisions. Based on the initial results of the interpretation of the model's decisions, all the non-lung areas from the CT images were removed using a contour detection-based region of interest (ROI) extraction approach. This was done to prevent the model from making decisions based on details in non-lung areas, which are clinically irrelevant for COVID diagnosis. This is the first work to utilize such a contour detection-based ROI extraction approach for medical images, based on our study. The model has achieved a mean F1 score of 0.87 for multi-label classification (COVID, Common Pneumonia & Normal) and a Dice Similarity Coefficient (DSC) of 0.8066 for lesion segmentation which has exceeded the DSC achieved by 6 out of 7 lesion segmentation models referenced in our study. © 2022 IEEE.