Certainty-Guided Cross Contrastive Learning for Semi-Supervised Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation due to limited labeled data.
• SSL performance typically trails fully supervised methods requiring complete datasets.
• Bridging this performance gap is essential for practical clinical applications.

Purpose of the Study:

• To develop a novel SSL framework that significantly narrows the performance gap between SSL and fully supervised methods.
• To achieve high accuracy in medical image segmentation using substantially less labeled data.
• To enhance the robustness of SSL against inaccurate pseudo-labels and class imbalance.

Main Methods:

• A novel SSL framework employing a knowledge exchange process between two networks.
• A certainty-guided contrastive learning strategy to mitigate pseudo-label inaccuracies and class imbalance.
• Cross-supervised contrastive learning across multiple scales for hierarchical feature learning.
• Efficient contrastive learning via novel sampling strategies and a negative memory bank.

Main Results:

• The proposed framework achieves state-of-the-art results on three challenging medical image segmentation benchmarks.
• Demonstrated significant performance improvement using less than a quarter of the labeled data compared to traditional methods.
• Showcased improved accuracy when integrated with diverse SSL frameworks.

Conclusions:

• The novel SSL framework effectively narrows the gap between semi-supervised and fully supervised medical image segmentation.
• The certainty-guided contrastive learning and multi-scale approach are key to achieving high accuracy and robustness.
• The method offers an efficient and accurate solution for medical image segmentation with limited labeled data.