Enhancing 3D Intra-Oral Tooth Segmentation Through Multi-Scale Graph Convolution Neural Networks and Cross-Domain Integration

Abstract

Purpose: Accurate tooth segmentation and labeling play a crucial role in digital dental diagnostic systems. Despite advancements in 3D intra-oral scanning techniques, challenges remain in handling ambiguous boundaries at the gingival junction and complex dental conditions. The aim of this study is to introduce a novel 3D tooth segmentation network that utilizes multi-scale graph convolutional neural networks and cross-domain integration to address these limitations. Methods: Our network proposes a deep learning-based dual-branch architecture to independently extract features of tooth coordinates and normal vectors. The coordinate branch captures fine-grained features across tooth structures through multi-scale graph convolution, thereby achieving accurate tooth boundary localization. The cross-domain fusion module (CFM) adaptively integrates features from the two branches through spatial attention and channel attention maps to enhance the representation of tooth boundaries and complex morphology. Results: Evaluated on a public 3D IOS dataset, our method outperforms the state-of-the-art, achieving improvements of 2.38%, 2.66%, and 2.44% in OA, mIoU, and mAcc, respectively. Conclusion: These results validate the effectiveness of our approach in enhancing the accuracy and robustness of 3D intra-oral tooth segmentation.