Enhancing Coronary Artery Calcium Scoring via Multi-Organ Segmentation on Non-Contrast Cardiac Computed Tomography

TL;DR

This work tackles coronary artery calcium scoring on non-contrast CT by embedding anatomical context to enhance both accuracy and interpretability. It introduces an end-to-end anatomically-informed pipeline that segments cardiovascular structures, localizes calcifications to ostia, filters false positives, and labels calcifications across four CA territories, yielding interpretable maps and per-vessel scores $AS$. Validated on a large, multi-center NC CT dataset (534 scans) with expert-ground truth, the method achieves inter-observer level accuracy and outperforms prior automated approaches in the orCaScore Grand Challenge, while maintaining fast inference (~6 minutes on an NVIDIA A100). The approach provides per-vessel CAC measures and transparent anatomical explanations, supporting more precise risk stratification and clinically actionable insights for cardiovascular care.

Abstract

Despite coronary artery calcium scoring being considered a largely solved problem within the realm of medical artificial intelligence, this paper argues that significant improvements can still be made. By shifting the focus from pathology detection to a deeper understanding of anatomy, the novel algorithm proposed in the paper both achieves high accuracy in coronary artery calcium scoring and offers enhanced interpretability of the results. This approach not only aids in the precise quantification of calcifications in coronary arteries, but also provides valuable insights into the underlying anatomical structures. Through this anatomically-informed methodology, the paper shows how a nuanced understanding of the heart's anatomy can lead to more accurate and interpretable results in the field of cardiovascular health. We demonstrate the superior accuracy of the proposed method by evaluating it on an open-source multi-vendor dataset, where we obtain results at the inter-observer level, surpassing the current state of the art. Finally, the qualitative analyses show the practical value of the algorithm in such tasks as labeling coronary artery calcifications, identifying aortic calcifications, and filtering out false positive detections due to noise.

Figures (6)

• Figure 1: Overview of the proposed method. The colors correspond to specific types of modules (segmentation, localization, labeling, filtering and scoring).
• Figure 2: Graphical overview of the main steps of the proposed algorithm.
• Figure 3: Visual summary of our method to obtain labeled CA calcifications: (a) three levels of calcification proposals: thresholding at 130 HU (yellow, translucent voxels), filtering using pericardium segmentation mask (blue, translucent voxels), filtering using CA segmentation mask (green voxels), and (b) labeling of calcifications based on their proximity to the CA areas (green---RCA, red---LM, yellow---LAD, cyan---LCx) and removal of aortic calcifications (purple, translucent voxels).
• Figure 4: The experimental results (a) DICE, sensitivity and specificity---mean, and the mean weighted by their volume in mm$^3$ (Weighted), and by the number of calcifications in GT (Count), with (b) distribution of metrics computed on complete sets of GT, and (c) distribution of metrics for CTs where both types of GT (our and orCaScore) are available. The metrics that outperform Observer 1 are bold.
• Figure 5: Agatston score (AS) correlation plot for our test set: (a) for all samples, (b) for AS $<$ 500 Agatston units (AU).