Confidence-guided mask learning for semi-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning utilizes limited labeled data and abundant unlabeled data for model training.
• Current methods often suffer from confirmation bias, limiting performance in precise segmentation tasks.
• Addressing confirmation bias is crucial for advancing semi-supervised medical image segmentation.

Purpose of the Study:

• To propose a novel Confidence-Guided Mask Learning (CGML) method for semi-supervised medical image segmentation.
• To enhance feature representation learning and model discrimination in uncertain regions.
• To generate more reliable segmentation results by mitigating confirmation bias.

Main Methods:

• Introduced an auxiliary generation task with mask learning for reconstructing masked images.
• Developed a confidence-guided masking strategy to improve discrimination in uncertain areas.
• Implemented a triple-consistency loss for consistent predictions across original, masked, and reconstructed images.

Main Results:

• The proposed CGML method demonstrated significant improvements in performance.
• Experiments on two datasets confirmed the effectiveness of the novel approach.
• The method successfully enhanced feature representation and model discrimination.

Conclusions:

• Confidence-Guided Mask Learning (CGML) offers a promising solution for semi-supervised medical image segmentation.
• The integration of mask learning and confidence guidance effectively addresses confirmation bias.
• The proposed approach achieves remarkable performance, advancing the field of medical image analysis.