Classifying Coronary Artery Calcification Risk using Radiomics from Non-Gated CT Scans

Opportunistic coronary artery calcification (CAC) scoring using chest computed tomography (CT) scans is a valuable tool for predicting cardiovascular event risk. Despite over 15 million chest CT scans performed annually, routine CAC scoring remains limited due to the manual labor involved. Although deep learning algorithms have shown promise in segmenting CAC lesions for risk evaluation, widespread adoption is still lacking because of the costs associated with these technologies and the difficulties in integrating them into existing clinical workflows. In this study, we propose a pipeline for CAC segmentation and risk classification. We first train a segmentation model on gated CT scans and subsequently fine-tune it using a smaller dataset of non-gated scans. We employ TotalSegmentator to cone down non-gated scans, simulating the field of view typical of gated CT scans. Radiomic features extracted from the segmented CAC lesions are then utilized to classify CAC risk scores. Our pipeline streamlines the CAC severity classification process using non-gated scans, achieving a Cohen's Kappa of 0.68 and an accuracy of 81%. The results highlight the benefits of our approach for more effective CAC risk evaluation in comparison to benchmark models.