An Automated Machine Learning Based Framework for Scan Level Coronary Artery Calcium Risk Stratification

Coronary artery calcium (CAC) scoring plays a crucial role in the early detection and risk stratification of coronary artery disease (CAD). Current techniques for CAC analysis rely on manual annotations by clinicians and large curated training datasets. This study introduces a fully automatic framework for CAC risk stratification that eliminates the need for manual clinical annotations. We introduce a novel pixel-attenuation-based feature extraction pipeline for automatic CAC assessment that leverages pseudo-labelling to identify regions of interest, reducing reliance on expert annotations. We compared our method against two alternative feature extraction strategies: a radiomics-based pipeline and a foundation-model-based pipeline. Our study focuses on non-contrast images from coronary computed tomography angiography (CCTA) scans. Training and testing were conducted on an in-house electrocardiogram (ECG)-gated CCTA dataset comprising 182 patients, categorized into risk groups based on CAC scores. We further investigated the impact of training exclusively on non-contrast data versus a combined dataset of contrast-enhanced and non-contrast scans, while all evaluations performed solely on non-contrast images. To assess generalizability, we applied our pipeline to the publicly available COCA-Coronary Calcium and Chest CTs dataset, demonstrating robust performance even on lower-quality, non-gated scans. The pixel-attenuation feature extraction pipeline outperformed radiomics and deep learning approaches for CAC risk classification, with accuracies of 92%, 93%, and 87%, and F1-scores of 92%, 90%, and 78% across the two-, three-, and five-class classification tasks, respectively. We present a cost-effective and time-efficient framework for early CAD patient stratification that can be implemented in clinical settings while reducing additional diagnostic procedures.