A novel fully automated deep learning model for coronary artery calcification detection on computed tomography

Objectives: To assess the diagnostic accuracy of a fully automated deep learning (DL) model for coronary artery segmentation and calcification detection on non-contrast, non-gated CT scans. Methods: A two-stage 3D segmentation pipeline was developed using 42 non-contrast and 27 contrast-enhanced CT scans to identify coronary artery calcifications in the right coronary artery (RCA), left anterior descending artery (LAD), and left circumflex artery (LCX). The model was trained with anatomically refined labels and region-based optimisation to improve structural coherence. Model outputs were visually assessed in a separate cohort of 100 scans by two independent, experienced observers. Segmentation and detection performance were evaluated against manually annotated reference standards using a binary analysis in 473 internal and external scans. Volumetric measurements of calcifications were compared with baseline severity gradings derived from radiologist reports. Results: Most model outputs were rated as excellent in the visual assessment, with good agreement between the outputs and manual reference standards for coronary artery segmentation (κ 0.68 to 0.81) and calcification detection (κ 0.79 to 0.85). The model accurately detected the presence of calcifications in the RCA (κ = 0.82, p < 0.001), LAD (κ = 0.93, p < 0.001), and LCX (κ = 0.82, p < 0.001). The diagnostic accuracy metrics of the model for calcification detection were: sensitivity, 95%; specificity, 98%; positive predictive value, 99%; and negative predictive value, 88%. The volume of calcification yielded by the model correlated with radiologist-reported disease severity, with regression coefficients of 28.3 for RCA, 28.7 for LAD, and 77.5 for LCX. Conclusions: The developed DL model segmented the coronary arteries, detected the presence of calcifications, and predicted disease severity with high accuracy.