Augmented Equivariant Mesh Networks for Anatomical Mesh Segmentation

Triangle meshes are a structured yet underexplored clinical data modality: as the native output of many surgical-planning, dental, and vascular reconstruction pipelines, they carry rich surface topology that flat point clouds and volumetric grids cannot directly represent. Robust segmentation on such meshes requires models that respect arbitrary patient pose and varying mesh resolution---properties that current task-specific methods lack. We present EAMS, a lightweight ($<$2M parameter) equivariant mesh segmentation framework augmented with anatomy-aware intrinsic features and two long-range context modules. Evaluated across three clinically distinct tasks spanning vertex-, face-, and edge-level supervision: intracranial aneurysm, intraoral tooth, and liver-surface segmentation, EAMS variants are competitive with strong task-specific baselines on canonical inputs while delivering markedly stronger robustness under test-time geometric perturbation. These results position anatomical meshes as a practically important structured-data regime for healthcare ML and demonstrate that compact equivariant models can achieve robustness without task-specific architecture redesign.