Robust cross center coronary artery segmentation on computed tomography angiography with limited local annotations

Background Coronary CT angiography (CCTA) enables non-invasive assessment of coronary artery disease. Quantitative CCTA analyses, including plaque and radiomics-based biomarkers, require accurate coronary artery segmentation. However, manual segmentation is time-consuming and limits large multi-center studies. Automated models often perform well only on data acquired with similar scanners and protocols and may lose accuracy when applied across centers. We therefore tested how well segmentation models transfer between centers with minimal local adaptation effort. Methods Two deep learning models—a conventional 3D convolutional neural network (CNN: 3D U-Net) and a Transformer-based architecture (TransUNet)—were first trained on a large public CCTA dataset (n=1000) and internally validated. The pre-trained models were then applied to two independents external CCTA cohorts (total n=187, Center A= 93, and Center B= 94) acquired with different scanners and protocols (Table 1). Performance was evaluated under two conditions: (1) direct transfer, where models were applied without center-specific retraining; (2) minimal adaptation, where a small subset per center (30–37 manually segmented cases; 80/20 train/validation) was used for fine-tuning. Fine-tuned models were evaluated on held-out annotated scans and subsequently used to segment the remaining cases in each cohort. Results On the original public test set, both models showed similar baseline performance (Dice 0.767±0.066 for 3D U-Net; 0.764±0.055 for TransUNet, Table 2. With direct transfer to external cohorts, accuracy declined substantially (Dice 0.45–0.63), reflecting limited cross-center generalizability After fine-tuning using ~34 local annotations, segmentation performance improved markedly at both centers. TransUNet achieved higher Dice scores than 3D U-Net (Center A: 0.785±0.043 vs 0.749±0.077, Center B: 0.732±0.069 vs 0.701±0.130;) and higher precision, while recall remained high (0.83–0.91). Conclusions Approximately 30 manually annotated CCTA scans per center were sufficient to restore reliable coronary artery segmentation despite substantial protocol differences. Transformer-based segmentation provided higher precision and more consistent performance than a conventional 3D CNN. These findings support scalable multi-center quantitative CCTA analysis with limited additional manual effort.