Uncertainty Global Contrastive Learning Framework for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation faces challenges with fuzzy object boundaries and limited labeled data.
• Accurate classification of segmentation boundaries is difficult due to data scarcity and overlapping object boundaries.

Purpose of the Study:

• To propose an Uncertainty Global Contrastive Learning (UGCL) framework to address fuzzy boundaries in semi-supervised medical image segmentation.
• To enhance segmentation accuracy by effectively utilizing unreliable data regions.

Main Methods:

• Developed patch filtering and classification entropy filtering methods for reliable pseudo-label generation.
• Introduced an uncertainty global contrast learning method to leverage complementary information from unreliable regions.
• Integrated consistency regularization, focusing on unreliable points for improved segmentation.

Main Results:

• The UGCL framework successfully distinguishes fuzzy boundaries and high-entropy pixels as unreliable.
• Contrastive learning and consistency regularization on uncertain points significantly improved segmentation accuracy.
• Evaluated on two public datasets, the method demonstrated substantial improvements over state-of-the-art techniques.

Conclusions:

• The proposed UGCL framework effectively handles unreliable data in semi-supervised medical image segmentation.
• Leveraging uncertainty through contrastive learning and consistency regularization enhances segmentation performance.
• The method offers a promising approach for improving medical image analysis with limited labeled data.