Reproduce “STS MICCAI 2023 Challenge: Grand challenge on 2D and 3D semi-supervised tooth segmentation” on OpenResearch

About this paper

Yaqi Wang, Yifan Zhang, Xiaodiao Chen, Shuai Wang, Dahong Qian, Fan Ye, Feng Xu, Hongyuan Zhang and 19 more

The STS MICCAI 2023 Challenge provided standardized datasets and evaluation for semi-supervised tooth segmentation in 2D panoramic X-rays and 3D CBCT volumes. Top-performing methods achieved Dice scores exceeding 93% for 2D segmentation and over 84% for 3D segmentation, demonstrating the effectiveness of leveraging limited labeled data with extensive unlabeled data.

Key insights

• Combining multiple semi-supervised learning strategies, such as pseudo-labeling, mean-teacher, and consistency regularization, significantly improves tooth segmentation accuracy.
• Domain-specific customizations to established architectures (like U-Net and nnU-Net) and robust data augmentation are crucial for handling dental image complexities.
• Multi-stage training approaches, often incorporating initial supervised learning followed by semi-supervised learning and fine-tuning, proved effective, especially for 3D segmentation.

Summary by alphaXiv.