Abstract
INTRODUCTION: This study developed and validated a deep learning-based framework to detect and quantify coronary artery stenosis, vulnerable plaque, and calcification in Curved-Multiplanar Reconstruction (cMPR) images to support clinical decision-making in coronary artery disease (CAD). METHODS: We analyzed 1,715 patients (5,112 vessels) from 2014 to 2022. Each vessel was reconstructed into a cMPR scan consisting of 13 sequential cross-sectional slices. Using a 2D nnU-Net framework, we developed a Stenosis Segmentation Model (SSM) and a Vulnerable Plaque Segmentation Model (VPSM). A time-independent test set (n = 824 patients, 2,437 vessels) was used for unbiased evaluation. Performance was assessed using Dice Similarity Coefficient (DSC), Positive Predictive Value (PPV), and Negative Predictive Value (NPV). For stenosis quantification, Mean Absolute Error (MAE) and Bland-Altman analysis were employed. RESULTS: The SSM achieved a vessel-level sensitivity of 0.84 and a high NPV of 0.98. The MAE for diameter stenosis was 12.4%, with a mean bias of +1.2% in Bland-Altman analysis, demonstrating robust agreement with expert references across the full test spectrum. For vulnerable plaque detection, the VPSM achieved a sensitivity of 0.80 and an NPV of 0.97 at the slice level. Calcification assessment showed high inter-rater reliability (ICC = 0.84) and substantial agreement with expert visual scoring (Kappa = 0.76). CONCLUSION: The proposed automated analysis demonstrated high diagnostic reliability, particularly in its negative predictive power, making it a powerful non-invasive tool for CAD screening. By providing objective quantification of stenosis, calcification, and vulnerable plaques, this method offers a significant advancement in standardized cMPR evaluation in clinical environments.