Semi-supervised medical image segmentation network based on mutual learning.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Deep Learning

Background:

• Semi-supervised learning is crucial for medical image segmentation with limited labeled data.
• Overfitting and cognitive bias in models lead to error accumulation and performance degradation.
• Existing methods struggle to correct amplified errors during neural network training.

Purpose of the Study:

• To develop a novel learning strategy to enhance medical image segmentation accuracy.
• To mitigate the continuous accumulation of erroneous knowledge in semi-supervised segmentation models.
• To improve the reliability of automated medical image analysis.

Main Methods:

• Proposed a semi-supervised medical image segmentation network named MLNet, utilizing a mutual learning (ML) approach.
• Employed a teacher-student network architecture where models learn collaboratively by updating parameters.
• Introduced an image partial exchange (IPE) algorithm to minimize erroneous information and preserve contextual integrity.

Main Results:

• Achieved a 9.28% improvement in Dice coefficient on the ACDC dataset (10% labeled data), reaching 89.48%.
• Demonstrated strong performance on the BraTS2019 dataset (10% labeled data) with a Dice coefficient of 84.56%.
• Outperformed comparative methods in both accuracy and reliability across multiple metrics.

Conclusions:

• The proposed MLNet effectively alleviates error accumulation in semi-supervised medical image segmentation.
• Experimental results confirm the approach's optimal performance and reliability compared to existing methods.
• The study highlights the potential of mutual learning and targeted perturbation for robust medical image segmentation.