Mesh Segmentation for Individual Teeth Based on Two-Stream GCN With Self-Attention

Abstract

Dental triangular mesh is widely used in computer-aided oral medicine. Different from regular data such as digital images, the structure of triangular mesh is more complex, and traditional operations such as convolution cannot be directly applied. Therefore, the segmentation of patients’ personalized teeth from mesh through deep learning is a hot topic in the current research field. Recently, a method named TSGCN presented the idea of learning coordinate features and normal features through a two-stream architecture based on Graph Convolutional Networks (GCN), which further improved the performance compared with other methods. However, its ability to extract and process global features still can be strengthened. To this end, a method named TSGCN-SA is proposed, whose core idea is to introduce the self-attention (SA) mechanism into TSGCN. Specifically, two SA modules are introduced, the first one is used to improve the global feature extraction ability in the coordinate stream. The second one plays an important role in the adaptive contribution adjustment of each stream during the feature fusion. Experiments based on the public dataset named 3DTeethSeg show that TSGCN-SA is superior to SOTAs in terms of segmentation performance due to the proposed SA modules, and the proposed method is competent in the task of individual tooth mesh segmentation.

Keywords

• Tooth
• triangular mesh
• deep learning
• self-attention
• segmentation