Dual-channel graph neural network for dental 3D point cloud model segmentation

By using 3D teeth segmentation via Intraoral scan, digitized dentistry becomes possible and easy. 3D instance segmentation is complex due to the complicated architecture of intraoral scan teeth, lack of boundary between adjacent teeth, and point cloud data noise. Primary methods cannot simultaneously obtain local and global connection data to the dental arches in areas with overlapping archform and crowded arch dimensions, resulting in the errors that are found there. To address the multi-scale representation gap and overcome the limitation of single-scale feature learning, dual-Graph Fusion Network (DGF-Net) can be used to reduce the gap between multi-scale and single-scale graph feature learning. A key feature of DGF-Net is its dual-graph design, which comprises of two channels for graph processing. Global graph channels are active in collecting topology and semantics, while local graph channels are active in collecting high resolution geometry. Under a trained fusion module both channels provide different feature detail information which is fused together. On comprehensive clinical dentistry point cloud dataset, the DGF-Net achieved 98.54% accuracy in segmentation and 97.89% Dice coefficient, outperforming state-of-the-art approaches. Through the accurate and stable nature of the DGF-Net model, various dental treatment scenarios can be efficiently realized, such as automated orthodontic simulation or precision manufacturing of patient-specific surgical guides, which enhances the efficiency and personalized nature of dental treatment.