Abstract
Rupture of intracranial aneurysms is a major cause of subarachnoid hemorrhage, making accurate segmentation important for aneurysm assessment and treatment planning. However, automated segmentation in computed tomography angiography (CTA) remains challenging because of small lesion size, complex morphology, and ambiguous boundaries between aneurysms and adjacent vessels. To address these challenges, we propose HSF-Net, a Swin UNETR-based segmentation network that integrates a Hybrid-Scale Fusion (HSF) module and the Convolutional Block Attention Module (CBAM) into the decoder. Unlike conventional decoders that mainly rely on simple feature concatenation, HSF-Net follows a "fuse first, refine later" strategy, in which HSF performs input-dependent multi-scale feature fusion and CBAM subsequently refines the fused features through channel and spatial attention. On the internal independent test set, HSF-Net achieved a mean patient-level Dice score of 83.4%, outperforming the Swin UNETR baseline and obtaining the highest Dice score among the evaluated methods under the same evaluation protocol. Ablation studies confirmed the complementary contributions of HSF and CBAM. Evaluation on an independent external cohort provided preliminary evidence of its cross-center applicability. These results suggest that HSF-Net provides an effective approach for automated intracranial aneurysm segmentation in CTA images.