Coronary Artery Segmentation with Partial Annotations in Coronary CT Angiography Images

Heart disease remains a leading cause of mortality worldwide, underscoring the importance of monitoring coronary arteries, the blood vessels supplying the heart, to prevent cardiovascular events. Although some automation in coronary artery extraction from coronary CT angiography (CCTA) has been achieved, manual adjustments are often necessary, imposing a significant burden on physicians and technicians. Consequently, the development of a fully automated coronary artery segmentation method is highly desirable. However, the annotation required for supervised learning is both labor-intensive and time-consuming. In this study, we present a method for coronary artery segmentation from CCTA that leverages partial annotations acquired during routine clinical practice. Utilizing these routinely obtained annotations has significantly reduced the annotation cost. To enable the extraction of unannotated vessels during training with partial annotations, we introduce a coronary segmentation network that incorporates recall loss, distance loss, and adversarial loss. The experimental results confirm that our proposed method achieves high segmentation accuracy.