Deep Learning-Driven Automated Aortic Valve Calcification Assessment in CT Calcium Score Screening Images

Aortic valve stenosis (AVS) is the most common valvular disease and the third most common cardiovascular disease, with a growing impact in the aging population. AVS culminates in heart failure if left undiagnosed or untreated. Transcatheter aortic valve replacement (TAVR) is a successful interventional procedure with reduced risk compared to surgical alternatives. As early intervention may reduce the number of patients who experience heart failure, there is a need for early detection of AVS. CT calcium score (CTCS) exams can be leveraged to opportunistically detect AVS via identification and evaluation of aortic valve calcifications. We developed methods to automatically detect aortic valve calcifications from CTCS images by identifying an aortic valve ROI and applying clinically approved Agatston (Ag) criteria to obtain calcifications. We retrospectively collected 263 patients who underwent CTCS exams, and used 70% for training, 15% for validation, and 15% for testing. Among the 40 patients from the held-out test set, we had 82.5% agreement between Agatston scores calculated from manually labeled ground truth vs. predicted cylinder masks, with only 5% of held-out cases having Agatston differences ≥ 200 Ag. This method offers promise in automated aortic valve calcification identification, paving the way for early AVS detection.