Abstract
This study presents a hierarchical Swin Transformer–based framework for automated segmentation of cerebrovascular structures using multimodal magnetic resonance imaging. The proposed architecture integrates patch partitioning, linear embedding, hierarchical windowed self-attention, and a multilevel encoder–decoder design to address the inherent challenges of vascular segmentation, including irregular morphology, small-caliber vessel visibility, and intensity variability across MRI modalities. A multimodal fusion module enhances the ability to capture complementary anatomical and vascular information, while skip-connected decoding ensures the preservation of fine-grained spatial features essential for accurate vessel reconstruction. The model was evaluated using a combination of open-access datasets and demonstrated superior performance across multiple quantitative metrics, achieving higher Dice similarity, precision, sensitivity, and specificity compared to existing state-of-the-art methods. Qualitative analysis further revealed accurate recovery of major arterial pathways, distal branches, and complex vascular topologies, confirming the model’s robustness in both global and localized segmentation tasks. The results highlight the discriminative strength of hierarchical attention mechanisms and emphasize their role in improving cerebrovascular characterization. Overall, the proposed framework offers a reliable and anatomically coherent approach for vascular segmentation, with strong potential for integration into clinical neuroimaging workflows and advanced cerebrovascular research applications.