Co-training semi-supervised medical image segmentation based on pseudo-label weight balancing

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised segmentation methods face challenges with pseudo-label information complementarity and quantity-quality imbalance.
• Existing methods often suffer from sub-network consensus and ignore valuable pseudo-labels, hindering performance.

Purpose of the Study:

• To propose a semi-supervised model for medical image segmentation that addresses pseudo-labeling challenges.
• To improve segmentation accuracy and reduce dependence on labeled data through weight balancing and co-training strategies.

Main Methods:

• Developed a novel framework, SCMT (Semi-supervised Co-training Mean Teacher), integrating a truncated Gaussian function weight balancing method.
• Employed a uniform alignment strategy to manage pseudo-label imbalance across different classes.
• Incorporated knowledge refinement and a co-training mean teacher model to enhance pseudo-label quality and model stability.

Main Results:

• Achieved superior performance on LA and Pancreas-CT datasets using only 10%/20% labeled data.
• Demonstrated significant improvements in Dice similarity coefficient, Jaccard index, Hausdorff distance, and average symmetric surface distance.
• Outperformed existing semi-supervised segmentation methods, showing consistent and stable prediction results.

Conclusions:

• The proposed SCMT model effectively addresses limitations in current semi-supervised segmentation.
• SCMT accurately captures object contours and details without shape constraints, offering highly accurate and stable segmentation.
• The framework shows advantages in surface segmentation and handling complex structures, outperforming other methods.