qpuchen/nnUNet_att_position_correction

What it solves

This project addresses the challenge of 3D tooth segmentation in medical imaging, specifically focusing on improving accuracy when labeled data is scarce. It provides a method to leverage unlabeled data through semi-supervised learning and corrects abnormal segmentation points to refine the final output.

How it works

The system follows a multi-stage pipeline:

1. Initial Training: A standard nnU-Net is trained on available labeled data using 5-fold cross-validation.
2. Pseudo-Labeling: The trained model generates pseudo-labels for unlabeled data. These labels are iteratively refined and filtered by removing low-quality labels with high variance across different rounds.
3. Enhanced Architecture: The project introduces nnUNet-att, which incorporates axial attention and positional correction into the network architecture to improve segmentation precision.
4. Post-Processing: A dedicated position correction module is applied to fix abnormal segmentation points in the final results.

Who it’s for

This tool is designed for medical imaging researchers and practitioners working on dental 3D segmentation and those interested in semi-supervised learning for medical image analysis.

Highlights

• Semi-Supervised Approach: Uses an iterative pseudo-labeling process to utilize unlabeled datasets.
• Axial Attention: Enhances the standard nnU-Net architecture for better feature capture.
• Positional Correction: Includes a specific module to correct segmentation errors based on spatial positions.
• Proven Performance: Ranked 5th in the MICCAI 2023 STS challenge with a score of 81.47% on unseen test data.