Hyper parameter tuning.

Rapid and accurate detection of coronary plaques on CCTA is critical for timely CAD diagnosis but is limited by reader workload and interobserver variability. Our objective was to evaluate the effectiveness of machine learning (ML) based on automated segmentation of coronary plaques in coronary computed tomography angiography (CCTA). We retrospectively analyzed CCTA scans from 1,642 patients (4,711 training vessels, 1,112 test vessels, plus 1,613 negative vessels) using an nnU-Net 3D full resolution architecture. Hyperparameters (batch size, learning-rate scheduler, epochs) were optimized on a 10% subset of the training dataset, and the final model was trained with 5-fold cross-validation on positive cases. Test performance was assessed at the plaque, vessel, and exam levels. Of 2,090 ground-truth plaques, the model achieved 1,772 TP for 84.8% sensitivity (95% CI 83.2–86.3%), 382 FP for 82.3% precision (95% CI 80.6–83.8), and a median Dice score of 0.88 (95% CI 0.87–0.89). At the vessel level, sensitivity was 94.7% (95% CI 93.2–95.9%) and specificity was 84.9% (95% CI 83.0–86.5%). At the examination level, sensitivity was 97.4% (95% CI 94.5–98.8%) and specificity was 79.7% (95% CI 76.5–82.6%). Detection of small, non-calcified plaques remained challenging, and false positives were often due to motion or step artifacts. Our model showed promising results in detecting coronary plaques in CCTA examinations, particularly for medium to large plaques and high negative predictive value for ruling out significant CAD, offering potential to streamline radiology workflows. Future work will focus on small-plaque detection, artifact robustness, and multi-center validation to enhance clinical utility.