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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Disease Research

Background:

• Coronary artery calcium (CAC) scoring is crucial for cardiovascular disease (CVD) risk assessment, guiding preventive strategies.
• Accurate CAC scoring from cardiac CT requires precise calcification segmentation, which is challenging due to calcification's characteristics.
• Manual annotation for training automated segmentation models is labor-intensive and requires specialized expertise.

Purpose of the Study:

• To develop a semi-supervised learning framework (S²CAC) for robust CAC segmentation with minimal labeled data.
• To enhance the accuracy of automated CAC scoring and reduce reliance on extensive manual annotations.

Main Methods:

• Proposed S²CAC, a semi-supervised framework utilizing a dual-path hybrid transformer for joint pixel-level segmentation and volume-level scoring.
• Introduced a scoring-driven consistency mechanism to leverage unlabeled data through differentiable score estimation.
• Developed a dynamic weighted loss function to effectively incorporate negative samples (no CAC) and prevent model collapse.

Main Results:

• S²CAC achieved state-of-the-art performance on two public gated CT datasets compared to baseline methods.
• Segmentation maps generated high concordance with manual annotations for Agatston scores.
• Demonstrated robust performance with significantly reduced labeled data requirements.

Conclusions:

• S²CAC offers a promising approach to improve automated CAC segmentation and scoring accuracy.
• The framework effectively reduces the dependency on large annotated datasets, making CAC assessment more efficient.
• This method has the potential to enhance cardiovascular disease risk stratification and preventive care.