Mutual learning with reliable pseudo label for semi-supervised medical image segmentation.

Area of Science:

• Medical image analysis
• Machine learning
• Computer vision

Background:

• Semi-supervised learning reduces the need for extensive data annotation in medical imaging.
• Pseudo-labeling is a key technique, but unreliable pseudo-labels introduce noise and hinder model performance.
• Existing methods struggle with accurately identifying and utilizing reliable pseudo-labels.

Purpose of the Study:

• To develop a method for learning from reliable pseudo-labels in semi-supervised medical image segmentation.
• To address the critical questions of identifying reliable pseudo-labels and quantifying their reliability.
• To improve segmentation performance by minimizing the impact of noisy pseudo-labels.

Main Methods:

• A comparative analysis of two subnetworks was performed to assess pseudo-label reliability.
• Pseudo-label reliability was determined by comparing prediction confidence between subnetworks.
• Intra-class similarity of predicted classes was used to further assess pseudo-label reliability.

Main Results:

• The proposed approach selectively incorporates knowledge from subnetworks based on pseudo-label reliability.
• Experimental results on Left Atrium, Pancreas-CT, and Brats-2019 datasets demonstrate superior performance.
• The method effectively reduces noise from unreliable pseudo-labels, leading to improved segmentation accuracy.

Conclusions:

• Learning from reliable pseudo-labels is crucial for effective semi-supervised medical image segmentation.
• The proposed method successfully identifies and utilizes reliable pseudo-labels, achieving state-of-the-art results.
• This approach offers a robust solution for enhancing medical image segmentation with limited annotated data.