Geo-Net: Geometry-Guided Pretraining for Tooth Point Cloud Segmentation.

Area of Science:

• Computer Vision
• Medical Imaging
• Orthodontics

Background:

• Accurate 3D tooth segmentation is crucial for orthodontic applications.
• Supervised learning methods for segmentation require large, labor-intensive annotated datasets.
• Existing methods struggle with the scale and annotation cost of labeled data.

Purpose of the Study:

• To develop a self-supervised pretraining framework (Geo-Net) to improve 3D tooth point cloud segmentation.
• To leverage large-scale unlabeled data to reduce reliance on annotated datasets.
• To enhance segmentation performance by incorporating geometric information.

Main Methods:

• Proposed a self-supervised pretraining framework, Geo-Net, based on scalable masked autoencoders.
• Introduced curvature-aware patching algorithm (CPA) to assemble informative patches guided by point curvatures.
• Developed scale-aware reconstruction (SCR) for multiple reconstructions across network layers to enhance scale-aware modeling.

Main Results:

• Geo-Net significantly boosted segmentation performance on 3D tooth point clouds.
• The framework achieved superior mean Intersection of Union (mIoU) compared to supervised methods with equivalent labeled data.
• Self-supervised pretraining with unlabeled data improved segmentation accuracy.

Conclusions:

• Geo-Net effectively leverages large-scale unlabeled data for 3D tooth segmentation.
• The proposed geometric-guided pretraining enhances the capacity of segmentation models.
• This approach offers a scalable and efficient solution for orthodontic applications requiring precise tooth delineation.