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

Researchers Paarth Prasad and Ruchika Malhotra introduce a topology-constrained quantized nnUNet framework for efficient and anatomically accurate 3D tooth segmentation from dental CBCT scans. The model employs an 8-bit quantized nnUNet backbone with dynamically calibrated weights and activations to minimize precision loss during inference. To counter spatial distortion from quantization, they integrate a tooth-specific topological loss into quantization-aware training. This loss combines connected-component analysis, adjacency consistency, and hole detection penalties, preserving critical anatomical structures like tooth count, inter-tooth adjacency, and cavity integrity without modifying the network architecture.

The joint optimization harmonizes cross-entropy loss, quantization regularization, and topological constraints, enabling end-to-end training with gradient approximations for persistent homology terms. Experiments show the approach significantly reduces topological errors compared to standard quantized models, producing clinically plausible segmentations while retaining hardware efficiency via integer-only inference. This work bridges computational efficiency and anatomical precision, offering a practical deployment solution for real-world dental applications in resource-constrained environments.