A Fully Automated Segment-Specific CCTA Plaque Quantification Workflow using Deep Learning

Coronary computed tomography angiography (CCTA) remains essential for diagnosing coronary artery disease, while the focus has moved from stenosis to plaque burden. Current workflows either need manual adjustments that limit consistency and efficiency, or they lack segment-specific precision. This study evaluated the effectiveness and reliability of a fully automated deep learning-based plaque quantification method using CCTA. DVPlaque workflow, a fully automated deep learning model, was developed to quantify segment-specific plaque volume. A total of 271 CCTA scans were included. Plaque quantification was performed by the expert and DVPlaque model, respectively. Reliability was measured through interclass correlation coefficient (ICC) analysis and Bland-Altman plots. DVPlaque workflow cost less time than expert measurement (p < 0.0001). Among these segments of major coronary arteries, proximal left ascending artery (LAD) had the highest prevalence of plaque (97.0%). Among these coronary arteries, the volumes of total plaque (TP), calcified plaque (CP), noncalcified plaque (NCP), and low-density NCP were the highest in LAD (all p-values < 0.05). On a per-patient level, the ICCs for TP, CP, and NCP volume were 0.962, 0.985, and 0.932. The corresponding biases and 95% limits of agreement were - 25.39 [ - 261.51, 210.73] mm³, 11.88 [ - 52.78, 76.55] mm³, and - 37.27 [ - 255.88, 181.33] mm³. For TP, CP, and NCP, per-vessel ICCs were all > 0.90, and per-segment ICCs were > 0.90 except for mid-LAD NCP (0.887). DVPlaque scheme was capable of quantifying plaque volume with excellent efficiency. The automated quantification of CP and NCP exhibited strong reliability even on a per-segment level, facilitating clinical application.