Exploring Feature Representation Learning for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning is crucial for medical image segmentation due to limited labeled data.
• Existing methods often rely on direct prediction regularization, which can be suboptimal.
• Feature representation learning offers a promising alternative for enhancing segmentation performance.

Purpose of the Study:

• To develop a novel two-stage framework for semi-supervised medical image segmentation.
• To improve segmentation accuracy by focusing on feature space regularization.
• To introduce a stage-adaptive contrastive learning approach for better feature representation.

Main Methods:

• A two-stage framework employing contrastive learning for feature representation.
• Boundary-aware contrastive loss in the first stage utilizing labeled images.
• Prototype-aware contrastive loss in the second stage using labeled and pseudo-labeled images.
• An aleatoric uncertainty-aware method for generating high-quality pseudo labels.

Main Results:

• The proposed method significantly enhances medical image segmentation performance.
• Achieved state-of-the-art results on three public medical image segmentation benchmarks.
• Demonstrated the effectiveness of feature space regularization over direct prediction supervision.

Conclusions:

• The proposed two-stage framework offers an effective approach to semi-supervised medical image segmentation.
• Feature representation learning through contrastive methods is key to improving segmentation.
• Aleatoric uncertainty-aware pseudo-labeling enhances the robustness and accuracy of the segmentation framework.