Enhancing coronary artery calcium scoring via multi-organ segmentation on non-contrast cardiac computed tomography

Coronary artery calcium scoring is often considered a largely solved problem within medical artificial intelligence. However, this paper shows that meaningful improvements in accuracy and interpretability are still possible. We propose a hierarchical approach that leverages deep neural networks to segment heart structures in non-contrast CT scans. Using spatial information from these anatomical structures, we apply an intuitive distance-based methodology to identify coronary calcifications. By shifting the focus from direct pathology detection to anatomical understanding, the proposed algorithm achieves high accuracy while providing improved interpretability. We evaluate the method on a multi-vendor dataset of 122 CT scans, achieving a volume-weighted DICE score of 0.987 and an intraclass correlation coefficient of 1.00 for total Agatston scores. Results from the orCaScore challenge further confirm that the method surpasses current state-of-the-art approaches, achieving an inter-observer-level weighted DICE score of 0.980. Qualitative analyses demonstrate additional practical value, including labeling coronary artery calcifications, identifying aortic calcifications, and filtering false positive detections caused by noise.