Joint Geometric Self-Attention and Boundary-Aware Search for High-Precision Intracranial Aneurysm Mesh Segmentation

Shuang Li, Fuhao Zhang, Jialin Zhang, Ling Wang, Hui Zhou, Dapeng Chen, Jinshan Tang, Jingfeng Jiang, Nan Mu

Abstract: Accurate segmentation of intracranial aneurysms (IAs) is critical due to their high rupture risks. Traditional voxel methods struggle to preserve surface details and maintain topological consistency. In contrast, mesh segmentation efficiently represents complex geometries using vertices, edges, and faces, but still faces challenges in modeling global context and refining ambiguous boundaries. We propose a novel mesh segmentation framework, GSBS-MSeg, composed of two main components: a geometric self-attention module and a boundary-aware search (BS) module. The geometric self-attention module captures long-range geometric dependencies, enhancing the global context and integrating both local and global features for improved segmentation accuracy. The BS module refines mesh boundaries and ensures topological consistency, significantly improving local boundary precision. GSBS-MSeg was evaluated using five-fold cross-validation on two publicly available datasets. Compared to existing methods, GSBS-MSeg improves aneurysm segmentation accuracy by 45.6%, total accuracy by 11.96%, and vessel segmentation accuracy by 7.73%. These results demonstrate its effectiveness and superior performance for mesh-based segmentation tasks.

External IDs:dblp:conf/smc/LiZZWZCTJM25