Deep learning matches experts in scoring aortic valve calcium

In nongated chest CT scans from patients across 33 sites in the United States and Brazil, an automated deep learning algorithm accurately quantified aortic valve calcification. Researchers trained a convolutional neural network on 1,446 scans, validated it on 361, and evaluated performance on a separate holdout set of 239 scans collected between 2021 and 2024. Manual segmentations independently verified by at least two board-certified radiologists served as reference standards. The algorithm-estimated Agatston scores correlated closely with expert assessments (r = 0.99; 95% CI: 0.98-0.99; P < 0.001) with minimal bias on Bland-Altman analysis (mean difference: 5.2 AU). When detecting high calcium burden linked to moderate-to-severe aortic stenosis, the algorithm showed 0.92 sensitivity and 0.98 specificity, with consistent performance across demographic subgroups and technical CT parameters.