Refined annotations and specialized augmentation for minor epicardial arteries in coronary angiography

AI-based analysis of coronary angiography (CAG) has advanced rapidly in recent years. However, public CAG datasets are not only limited in number but also frequently lack annotations for small vessels, as these are not primary treatment targets and their annotation is technically challenging. Further progress in CAG segmentation requires both larger datasets and precise labeling of small-caliber vessels. We developed a CAG segmentation framework that incorporates refined annotations, CAG-specific augmentation, and a small-vessel-focused evaluation metric. The refined dataset (fine-ARCADE dataset) comprises 196 right coronary artery images with additional annotations of minor epicardial arteries. Copy-paste augmentation was modified for vessel images by matching intensities, adjusting vessel overlap, and varying branching complexity to enhance data diversity. We fine-tuned SAM-Med2D on public and augmented CAG datasets for 50 epochs. Segmentation performance was evaluated using three metrics: \(\text {DSC}_{\text {overall}}\) (full vessel tree), \(\text {DSC}_{\text {minor}}\) (minor epicardial arteries), and \(\text {DSC}_{\text {major}}\) (major epicardial arteries). Fine-tuned models outperformed the pretrained baseline across all public CAG datasets. CAG-specific augmentation consistently improved \(\text {DSC}_{\text {overall}}\) by up to \(7.35\%\) , and \(\text {DSC}_{\text {minor}}\) by up to \(11.45\%\) , with the best gains observed in the ARCADE and XCAD datasets. These results suggest that while existing public CAG datasets provide limited assessment of small-vessel segmentation, refined annotations enable more accurate evaluation, and CAG-specific augmentation effectively enhances model performance in minor epicardial regions.