TeethNet: Dual-Stream Attention Network for 3D Tooth Segmentation

In dentistry, tooth segmentation techniques are of critical importance for clinical applications such as orthodontics, orthopedics and prosthetics. However, the task faces many challenges, such as fuzzy tooth boundaries, complex tooth morphology, and localization difficulties. Existing deep learning models often struggle to achieve both high-precision segmentation and sufficient robustness when extracting features from tooth and gingival boundaries. In view of this, we propose a new 3D tooth segmentation method——TeethNet. The core innovations of TeethNet include the multilayer attention perception module and the dual cross-attention module. The multi-layer attention perception module effectively facilitates the fusion of multi-layer features to obtain a better representation of facial features. The dual cross-attention module significantly improves the model's ability to handle complex tooth morphology and gingival boundaries by fusing 3D coordinate and normal features. TeethNet is constructed based on a dual-stream feature extraction network, where the C-stream (coordinate stream) is responsible for extracting multi-scale geometric features, while the N-stream (normal vector stream) focuses on extracting fine- grained boundary features. The multilayer attention perception module facilitates the deep fusion of low and high-dimensional features, while the dual cross-attention module further facilitates the information interaction between the two streams, enabling the model to more accurately understand the geometrical structure and spatial relationship of teeth. Experimental results on the Teeth3DS dataset show that TeethNet outperforms existing methods in 3D tooth segmentation tasks. This study provides a new effective solution for 3D tooth segmentation, which has a promising future for applications in clinical diagnosis, orthodontics and restorative dentistry.