Leveraging labelled data knowledge: A cooperative rectification learning network for semi-supervised 3D medical image...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning in 3D medical image segmentation leverages limited labeled data with abundant unlabeled data.
• Effective utilization of unlabeled data is crucial for improving segmentation accuracy.
• Consistency learning strategies rely heavily on the quality of pseudo-labels.

Purpose of the Study:

• To introduce a novel methodology for generating high-quality pseudo-labels for semi-supervised 3D medical image segmentation.
• To enhance the effectiveness of consistency learning strategies by improving pseudo-label accuracy.
• To improve the overall performance of 3D medical image segmentation models.

Main Methods:

• Cooperative Rectification Learning Network (CRLN) for adaptive pseudo-label rectification using class prototypes.
• Dynamic Interaction Module (DIM) for prototype-feature interactions across multiple resolutions.
• Cooperative Positive Supervision (CPS) to optimize uncertain and unassertive class representations.

Main Results:

• The proposed methodology significantly improves the quality of pseudo-labels.
• The method demonstrates superior performance compared to existing semi-supervised segmentation techniques.
• Extensive experiments on three public datasets validate the effectiveness of the approach.

Conclusions:

• The developed method offers a robust solution for semi-supervised 3D medical image segmentation.
• High-quality pseudo-label generation is key to effectively utilizing unlabeled data.
• The integration of CRLN, DIM, and CPS enhances segmentation accuracy in uncertain regions.