MICCAI STS 2024 challenge: Semi-supervised instance-level tooth segmentation in panoramic X-ray and CBCT images

Area of Science:

• Medical Image Analysis
• Dental Imaging
• Artificial Intelligence in Healthcare

Background:

• Automated tooth segmentation in dental imaging (Orthopantomograms - OPGs, Cone-Beam Computed Tomography - CBCT) is crucial for diagnostics.
• Manual instance-level annotation for creating large datasets is labor-intensive, leading to data scarcity.
• Semi-supervised learning (SSL) presents a promising solution to address data limitations in medical image segmentation.

Purpose of the Study:

• To benchmark and advance semi-supervised learning (SSL) methods for instance-level tooth segmentation in dental imaging.
• To organize the 2nd Semi-supervised Teeth Segmentation (STS 2024) Challenge to foster research and development in this area.
• To provide a large-scale, publicly available dataset for OPG and CBCT images with instance-level annotations.

Main Methods:

• Organized the STS 2024 Challenge, providing a dataset of over 90,000 2D OPG and 3D CBCT images with annotations.
• Evaluated valid, open-source, deep learning-based SSL submissions from participating teams.
• Utilized hybrid semi-supervised frameworks combining foundational models (e.g., SAM) with multi-stage refinement pipelines.

Main Results:

• The STS 2024 Challenge attracted 114 (OPG) and 106 (CBCT) registered teams.
• Winning SSL models significantly outperformed a fully-supervised nnU-Net baseline.
• Top methods achieved substantial improvements: >44 percentage points in Instance Affinity (IA) for OPGs and 61 percentage points in Instance Dice for CBCT.

Conclusions:

• Semi-supervised learning is highly effective for complex, instance-level medical image segmentation tasks with limited labeled data.
• Hybrid SSL approaches leveraging foundational models and multi-stage refinement pipelines show the most promise.
• The challenge dataset and submitted code are publicly available on GitHub, promoting transparency and reproducibility in dental AI research.