Multi-Scale Bidirectional Enhancement Network For 3D Dental Model Segmentation

Given the importance of 3D sensors, a fine-grained analysis on 3D dental model is an important task in computer-aided orthodontic treatment planning. Particularly, the real 3D dental model can intuitively show the shape and morphology of the tooth, but due to the irregularity of the 3D tooth data, it poses a challenge for accurate tooth segmentation. In this work, with the mesh data as input, we propose an end-to-end deep neural network, called MBESegNet, for accurate tooth segmentation on 3D dental models. On the one hand, to reduce the ambiguity of the mesh feature representation near the tooth boundary, MBESegNet learns the local context by enhancing the geometric and semantic features with a bidirectional and symmetric structure. On the other hand,MBESegNet hierarchically captures the multi-scale contextual features from different scales and represent the feature map following a coarse-to-fine feature fusion strategy for accurate tooth segmentation. The experimental results demonstrate that our approach achieves competitive performance against state-of-the-art 3D shape segmentation methods.