Topology-Constrained Quantized nnUNet for Efficient and Anatomically Accurate 3D Tooth Segmentation

Researchers propose a topology-constrained 8-bit quantized nnUNet that preserves tooth count, adjacency relationships, and cavity integrity during 3D segmentation by integrating tooth-specific topological losses (connected-component analysis, adjacency consistency, hole detection) into quantization-aware training, reducing model size while maintaining anatomical accuracy without architectural modifications.