TeethGNN: Semantic 3D Teeth Segmentation With Graph Neural Networks.

Area of Science:

• Computer Vision
• Medical Imaging
• Computational Geometry

Background:

• Accurate 3D tooth segmentation is crucial for dental diagnostics and treatment planning.
• Existing methods often struggle with noise, artifacts, and complex dental anatomies like malocclusion.
• Convolutional Neural Network (CNN) based approaches typically require converting irregular mesh data into regular grids, losing valuable geometric information.

Purpose of the Study:

• To develop a novel 3D tooth segmentation method utilizing graph neural networks (GNNs) for direct processing of mesh data.
• To improve robustness against scanning noise, foreign objects, and severe malocclusion.
• To overcome limitations of previous methods in handling crowded teeth and incomplete segmentation.

Main Methods:

• Developed TeethGNN, a GNN-based approach for segmenting 3D tooth models represented as meshes.
• Explored the dual space of mesh data to leverage its non-uniform structure for feature learning.
• Designed a two-branch network predicting segmentation labels and offsets from tooth centroids, followed by clustering for improved separation.

Main Results:

• Achieved state-of-the-art performance in 3D tooth segmentation.
• Demonstrated superior quantitative and qualitative results compared to existing methods.
• Successfully separated adjoining teeth and corrected incomplete segmentations.

Conclusions:

• TeethGNN offers an effective solution for precise 3D tooth segmentation directly on mesh data.
• The GNN approach eliminates the need for manual feature extraction and accelerates inference.
• This method shows significant potential for advancing computer-aided dentistry.