arXiv：cs.LG（机器学习，全量分类）· Daniel Saragih·· 17 小时前AI 评分34
EAMS：基于等变网格网络的解剖网格分割框架
Augmented Equivariant Mesh Networks for Anatomical Segmentation
AI 导读
研究者提出 EAMS——基于等变网格神经网络（EMNN）的等变解剖网格分割器，在三个临床领域、四种数据集设置下评估，覆盖边、顶点和面级监督。该框架参数量不足 2M，在颅内动脉瘤和口内分割上未扰动输入时与专用基线相当，且在几何扰动下保持稳定；现有方法在 3D-IOSSeg 上 40° 旋转时 IoU 下降 28-30 点。该工作已被 NeurIPS 2026 接收。
正文
Abstract:Anatomical mesh segmentation requires models that operate directly on irregular surface geometry while remaining robust to changes in coordinate pose across meshes of varying resolution. Existing task-specific mesh and point-cloud methods are not equivariant, and can degrade sharply under test-time perturbation, for example dropping by 28-30 IoU points on 3D-IOSSeg at $40^\circ$ rotation even with matched training augmentation. We present EAMS, an Equivariant Anatomical Mesh Segmentor built on Equivariant Mesh Neural Networks (EMNN), and evaluate it in four dataset settings across three clinical application areas, spanning edge-, vertex-, and face-level supervision. We combine intrinsic mesh descriptors with anatomy-aware priors, including PCA-derived frames for dental arches and liver surfaces, and augment message passing to provide lightweight global context. Across intracranial aneurysm and intraoral segmentation, EAMS variants are competitive with specialized baselines on unperturbed inputs while remaining stable under geometric perturbations, and on liver surfaces they expose a favorable trade-off between canonical-pose accuracy and rotation robustness. These results show that a lightweight ($<2$M parameters) equivariant framework can deliver robust anatomical mesh segmentation across diverse label types.
| Comments: | Accepted as a conference paper to NeurIPS 2026. 30 pages, 7 figures, 21 tables |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG) |
| Cite as: | arXiv:2605.08172 [cs.CV] |
| (or arXiv:2605.08172v2 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2605.08172 arXiv-issued DOI via DataCite |
Submission history
From: Daniel Saragih [view email]
[v1]
Mon, 4 May 2026 23:54:07 UTC (35,336 KB)
[v2]
Wed, 30 Sep 2026 20:37:46 UTC (39,846 KB)
来源：arXiv：cs.LG（机器学习，全量分类） · arxiv.org