Tooth Defect Segmentation in 3D Mesh Scans Using Deep Learning

Computer-aided systems are widely used in digital dentistry to help human experts for accurate and efficient diagnosis and treatment planning. In this paper, we study the problem of tooth defect segmentation in 3-Dimensional (3D) mesh scans, which is a prerequisite task in many dental applications. Existing models usually perform poorly in this task due to the highly imbalanced characteristic of tooth defects. To tackle this issue, we propose a novel triple-stream graph convolutional network named TripleNet to learn multi-scale geometric features from mesh scans for end-to-end tooth defect segmentation. With predefined geometrical features as inputs and a focal loss for training guidance, we achieve state-of-the-art performance on 3D tooth defect segmentation. Our work exhibits the great potential of artificial intelligence for future digital dentistry.