Point Cloud Tooth Segmentation Based on Boundary Feature Enhancement

Currently, the acquisition of 3D dental models using intraoral scanners has been widely applied in clinical oral medicine. However, accurately segmenting teeth from the dental model requires addressing the issues of morphological similarity between teeth and abnormalities in misshapen teeth. Since the geometric shapes of adjacent teeth are very similar, the subtle morphological variations and positional differences of each tooth may also affect finegrained segmentation, and most oral patients have dental deformities, such as displacement, crowding, and tooth loss. To address the above issues, this paper proposes a segmentation network model that enhances boundary features, which mainly consists of three parts: 1. Downsampling and feature extraction module: By introducing a boundary-aware downsampling module, the number of sampling points near the boundary is increased to improve the precision of tooth instance labels; by introducing “point curvature” information, the feature representation of boundary points is enhanced, further improving segmentation precision. 2. Point grouping module, which performs semantic segmentation on the point cloud and then obtains the tooth instance labels of points through clustering. 3. Cropping and correction module, which crops around the predicted tooth instance center points and outputs tooth-gum masks to further refine the tooth instance labels. According to the experimental results, the segmentation network proposed in this paper can perform high-precision and stable segmentation on teeth.