Few sampling meshes-based 3D tooth segmentation via region-aware graph convolutional network

Precise segmentation of teeth from intraoral scanner images is crucial for computer-assisted orthodontic treatment planning, yet current segmentation quality often falls below clinical standards due to intricate tooth morphology and blurred gingival lines. Previous deep learning-based methods typically focus on localized tooth information, emphasizing detailed relations between each tooth while disregarding the holistic information of tooth models. Furthermore, unique geometric information such as the centroid position of teeth remains underutilized. To address these issues, we propose a Region-Aware Graph Convolutional Network (RAGCNet) for 3D tooth segmentation, which is capable of effectively handling both local fine-grained details and global holistic feature with few sampling meshes. Specifically, considering the differences in intraoral scanning accuracy, we sample central meshes using an improved Farthest Point Sampling (FPS) algorithm, and then aggregate the information of neighbor meshes using the K-Nearest Neighbor (KNN) method. Meanwhile, a specially designed Region-Aware Module (RAM) via attention mechanism is proposed for feature extraction and fusion. Additionally, we propose a novel Centroid Loss based on tooth centroid coordinates to impose additional constraints on segmentation results. Evaluation on real datasets with 3D intraoral scanner-acquired tooth mesh models demonstrates that RAGCNet outperforms other SOTA methods in 3D tooth segmentation.