Hierarchical Self-Supervised Learning for 3D Tooth Segmentation in Intra-Oral Mesh Scans

Area of Science:

• Computer Vision
• Digital Dentistry
• Machine Learning

Background:

• Accurate 3D tooth and gingiva segmentation from intraoral scans (IOS) is crucial for digital dentistry applications like orthodontics.
• Current deep learning methods for 3D tooth segmentation often require large, meticulously labeled datasets, which are costly and time-consuming to create.
• The need for efficient methods to leverage abundant unlabeled IOS data is apparent.

Purpose of the Study:

• To develop a novel self-supervised learning framework (STSNet) to enhance 3D tooth segmentation performance.
• To reduce the dependency on large-scale labeled datasets by utilizing extensive unlabeled IOS data.
• To demonstrate the effectiveness of unsupervised pre-training for improving segmentation accuracy and reducing annotation burden.

Main Methods:

• Proposed a two-stage training framework: unsupervised pre-training followed by supervised fine-tuning.
• Introduced three hierarchical contrastive losses (point-level, region-level, cross-level) for unsupervised representation learning.
• Utilized augmented views of IOS meshes to learn robust features from unlabeled data.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 89.88% with the same amount of annotated samples compared to supervised methods.
• Demonstrated superior performance gains when using limited labeled data.
• Showcased comparable or better performance using only 40% of annotated samples versus fully supervised baselines.

Conclusions:

• STSNet is the first unsupervised pre-training approach for 3D tooth segmentation, significantly boosting performance.
• The framework effectively reduces the need for extensive manual annotation and verification in digital dental workflows.
• Self-supervised learning holds strong potential for advancing 3D intraoral scan analysis and applications.