S2CAC: Semi-supervised coronary artery calcium segmentation via scoring-driven consistency and negative sample boosting

Coronary artery calcium (CAC) scoring plays a pivotal role in assessing the risk for cardiovascular disease events to guide the intensity of cardiovascular disease preventive efforts. Accurate CAC scoring from gated cardiac Computed Tomography (CT) relies on precise segmentation of calcification. However, the small size, irregular shape, and sparse distribution of calcification in 3D volumes present significant challenges for automated CAC assessment. Training reliable automatic segmentation models typically requires large-scale annotated datasets, yet the annotation process is resource-intensive, requiring highly trained specialists. To address this limitation, we propose S${}^2$CAC, a semi-supervised learning framework for CAC segmentation that achieves robust performance with minimal labeled data. First, we design a dual-path hybrid transformer architecture that jointly optimizes pixel-level segmentation and volume-level scoring through feature symbiosis, minimizing the information loss caused by down-sampling operations and enhancing the model’s ability to preserve fine-grained calcification details. Second, we introduce a scoring-driven consistency mechanism that aligns pixel-level segmentation with volume-level CAC scores through differentiable score estimation, effectively leveraging unlabeled data. Third, we address the challenge of incorporating negative samples (cases without CAC) into training. Directly using these samples risks model collapse, as the sparse nature of CAC regions may lead the model to predict all-zero maps. To mitigate this, we design a dynamic weighted loss function that integrates negative samples into the training process while preserving the model’s sensitivity to calcification. This approach effectively reduces over-segmentation and enhances overall model performance. We validate our framework on two public non-contrast gated CT datasets, achieving state-of-the-art performance over previous baseline methods. Additionally, the Agatston scores derived from our segmentation maps demonstrate strong concordance with manual annotations. These results highlight the potential of our approach to reduce dependence on annotated data while maintaining high accuracy in CAC scoring. Code and trained model weights are available at: https://github.com/JinkuiH/S2CAC