Neighborhood Saliency and Geometric Structure Enhanced Network for Intracranial Aneurysm Segmentation in Point Clouds

ABSTRACT Intracranial aneurysm (IA) is a potentially fatal cerebrovascular disease, whose rupture often results in acute subarachnoid hemorrhage, posing a severe threat to neurological function. Diagnostic accuracy for IAs is markedly improved in 3D surface analysis, where point cloud–based segmentation has shown superior performance. However, the complex vascular geometry and the underutilization of critical feature information continue to impede the accurate delineation of aneurysms and normal vessels. Therefore, we introduce the Neighborhood Saliency and Geometric Structure Enhanced Network (NSGSE), a point‐cloud framework dedicated to segmenting IAs. Unlike conventional frameworks that treat all points equally, NSGSE's neighborhood‐saliency‐enhanced encoder aggregates each token with its nearest neighbors and applies compact local transformations, thereby capturing critical geometric cues. Furthermore, NSGSE's Mamba‐based decoder leverages a static k‐NN graph with an Implicit Communication Module to realize implicit communication among adjacent patches, thereby improving the modeling of complex geometric structures. To ensure geometry‐stable representations, NSGSE introduces rotation‐invariant orientation encodings (RI‐OE) and rotation‐invariant positional encoding (RI‐PE) that jointly encode local orientations and patch centers for consistent latent spaces. On the IntrA dataset, NSGSE achieved an IoU of 97.62% for vessels and 84.35% for aneurysms, outperforming existing methods. We also evaluated NSGSE on 3D shape datasets, attaining an accuracy of 93.64% on ModelNet40 and a mean IoU of 84.8% on ShapeNetPart, demonstrating its broad applicability. Our code is available at https://github.com/xxx‐xyw/NSGSE .