EAMS: Lightweight equivariant mesh segmentor withstands 40° tilt without IoU drop

A new paper accepted at ICML 2026 workshops introduces EAMS (Equivariant Anatomical Mesh Segmentor), a unified framework for anatomical mesh segmentation built on Equivariant Mesh Neural Networks (EMNN). Unlike existing task-specific methods that degrade sharply under pose perturbations—dropping 25–26 IoU points on intraoral scans at just 40° tilt—EAMS remains robust. The model uses intrinsic surface descriptors (like heat kernel signatures) and anatomy-aware priors (PCA-derived frames for dental arches and liver surfaces) combined with lightweight global context via augmented message passing. With fewer than 2M parameters, it handles vertex-, edge-, and face-level supervision across diverse tasks including intracranial aneurysm and intraoral segmentation.

A key finding is that strict roto-translational equivariance is not always beneficial. On the liver dataset, where anatomical landmarks are subtle creases, standard non-equivariant baselines can exploit raw coordinates to resolve left-right and front-back ambiguities. The equivariant network, which is blind to absolute space, struggles with these asymmetries. This trade-off suggests that hybrid approaches—combining equivariance with limited spatial cues—may yield better real-world performance. Overall, EAMS demonstrates that a single lightweight equivariant architecture can replace multiple anatomy-specific models while improving robustness to patient pose variations.