Automated deep learning-radiomics pipeline for non-calcified coronary plaque detection using non-contrast calcium...

Coronary artery calcium score (CACS) quantifies calcification to assess coronary artery disease (CAD), but it provides insufficient warning for low-attenuation non-calcified plaques. This study proposes and validates an automated pipeline that combines deep learning and radiomics for efficient detection of non-calcified plaques in the left anterior descending artery (LAD) and right coronary artery (RCA) using non-contrast CACS.

Patients undergoing coronary CT angiography for suspected CAD from two medical sites were retrospectively enrolled and categorized into lesion and control groups. LAD and RCA vessels on CACS images from the development set were manually annotated to train deep learning-based segmentation models for automated coronary segmentation and subsequent pericoronary adipose tissue (PCAT) extraction. Radiomics models were built for LAD and RCA using three regions of interest-coronary artery, PCAT, and their combination-based on the training set. Model performance was evaluated across all datasets using receiver operating characteristic analyses, and DeLong tests were applied for pairwise comparisons.

The SegResNet models achieved optimal performance in coronary segmentation. Radiomics models for predicting non-calcified plaques demonstrated moderate to good vessel-level diagnostic performance, with areas under the curve (AUCs) ranging from 0.700 to 0.855 across datasets, encompassing separate LAD and RCA models and all ROI strategies. The coronary artery and combined-region models generally outperformed or matched the PCAT model, with comparable AUCs between them in most settings.

The automated pipeline enables efficient detection of non-calcified coronary plaques in CACS, with combined-region models showing promise for future use. The approach may facilitate further research and support the clinical translation of chest CT for large-scale CAD screening.