Deep learning based automatic quantification of aortic valve calcification on contrast enhanced coronary CT angiography

Abstract

Quantifying aortic valve calcification is critical for assessing the severity of aortic stenosis, predicting cardiovascular risk, and guiding treatment decisions. This study evaluated the feasibility of a deep learning-based automatic quantification of aortic valve calcification using contrast-enhanced coronary CT angiography and compared the results with manual calcium scoring. A retrospective analysis of 177 patients undergoing aortic stenosis evaluation was conducted, divided into a development set (n = 97) and an internal validation set (n = 80). The DeepLab v3 + model segmented the ascending aorta, and the XGBoost model refined the aortic valve region using representative attenuation values. Calcifications were identified with a tailored threshold based on these values and quantified using a weighted scoring method analogous to the Agatston score. The automated method showed excellent agreement with manual Agatston scores derived from non-contrast CT (Pearson correlation coefficient = 0.93, 95% confidence interval [CI]: 0.89–0.95, p < 0.001, concordance correlation coefficient = 0.92, 95% CI: 0.87–0.95). For classifying severe aortic stenosis, defined by calcium scores exceeding 2000 for men and 1300 for women, the approach achieved a sensitivity of 88.6%, specificity of 91.1%, and overall accuracy of 90.0%. This deep learning model provides automated aortic valve calcification quantification with high accuracy on enhanced CT. This approach offers an alternative for measuring aortic valve calcium when non-contrast CT is unavailable, with the potential to reduce reliance on non-contrast CT, minimize operator dependency, and lower patient radiation exposure.