A boundary-aware framework with adversarial training and integrated deep supervision for DSA cerebral vessel segmentation

Accurate cerebral vessel segmentation from digital subtraction angiography (DSA) remains challenging because distal vessels are thin, sparsely represented, and often exhibit weak contrast and ambiguous boundaries. Region-overlap objectives alone may not fully characterize vessel continuity, boundary fidelity, and prediction stability. This study investigated a boundary-aware framework with adversarial training and integrated deep supervision (BAIS) for DSA cerebral vessel segmentation. BAIS was instantiated on an FR-UNet-based segmentation network and combined projected gradient descent (PGD)-based adversarial training, deep supervision, centerline Dice (clDice) regularization, and boundary-aware supervision within a shared multi-output optimization framework. The method was evaluated on two DSA benchmark datasets. The official patient-level split of DIAS was used for training, validation, and testing, and DSCA was evaluated using a five-fold protocol. Performance was assessed using region-overlap, centerline, boundary, and distance-based metrics. Component-wise ablation, hyperparameter sensitivity, perturbation robustness, cross-dataset evaluation, qualitative analysis, and computational profiling were also performed. On the 20-case DIAS test set, BAIS achieved a Dice similarity coefficient (DSC) of 0.7903 and a clDice of 0.7268. On DSCA, the corresponding five-fold mean values were 0.8927 and 0.8728, respectively. The complete configuration showed competitive performance among the evaluated ablation settings, although individual components were not uniformly beneficial when used in isolation. Under the evaluated intensity, noise, and adversarial perturbations, BAIS yielded higher numerical segmentation performance than the FR-UNet baseline, while both models deteriorated under stronger perturbations. In cross-dataset evaluation, BAIS yielded higher region-overlap scores in both transfer directions, whereas topology- and boundary-sensitive metrics showed mixed changes. BAIS showed competitive in-domain segmentation performance and greater relative robustness than the FR-UNet baseline under the evaluated perturbations. However, improvements in region overlap did not consistently translate into improved topology or boundary fidelity under dataset shift. These findings support further evaluation of the framework, particularly with independent multi-center data and broader validation settings.