AI BriefingEAMS Unveiled for Medical 3D Mesh Segmentation
Key point
A medical 3D mesh segmentation model called EAMS is proposed, robust to rotation and deformation.
Details
Existing mesh- and point cloud-based segmentation methods lack Equivariance, causing performance to degrade sharply with changes in patient posture or mesh resolution. For example, in intraoral scan data, a tilt of just 40 degrees can cause IoU to drop by 25 to 26 points.
EAMS (Equivariant Anatomical Mesh Segmentor) is built on top of EMNN (Equivariant Mesh Neural Networks) and shows stable performance across diverse clinical tasks (edge, vertex, face-level supervision).
Key features are as follows:
- Combines Intrinsic mesh descriptors with anatomical priors (such as PCA-based frames for dental arches and liver surfaces).
- Augments message passing to provide lightweight global context.
- Uses fewer than 2M lightweight parameters, enabling it to handle various supervision types without separate architectures.
Experimental results show that EAMS maintains performance on par with existing specialized models in intracranial aneurysm and intraoral segmentation tasks, while demonstrating very high stability against geometric deformation.
