Explained Deep Learning Framework for COVID-19 Detection in Volumetric CT Images Aligned with the British Society of Thoracic Imaging Reporting Guidance: A Pilot Study.

Abstract

In March 2020, the British Society of Thoracic Imaging (BSTI) introduced a reporting guidance for COVID-19 detection to streamline standardised reporting and enhance agreement between radiologists. However, most current DL methods do not conform to this guidance. This study introduces a multi-class deep learning (DL) model to identify BSTI COVID-19 categories within CT volumes, classified as 'Classic', 'Probable', 'Indeterminate', or 'Non-COVID'. A total of 56 CT pseudoanonymised images were collected from patients with suspected COVID-19 and annotated by an experienced chest subspecialty radiologist following the BSTI guidance. We evaluated the performance of multiple DL-based models, including three-dimensional (3D) ResNet architectures, pre-trained on the Kinetics-700 video dataset. For better interpretability of the results, our approach incorporates a post-hoc visual explainability feature to highlight the areas of the image most indicative of the COVID-19 category. Our four-class classification DL framework achieves an overall accuracy of 75%. However, the model struggled to detect the 'Indeterminate' COVID-19 group, whose removal significantly improved the model's accuracy to 90%. The proposed explainable multi-classification DL model yields accurate detection of 'Classic', 'Probable', and 'Non-COVID' categories with poor detection ability for 'Indeterminate' COVID-19 cases. These findings are consistent with clinical studies that aimed at validating the BSTI reporting manually amongst consultant radiologists.