Bootstrapping Semi-supervised Medical Image Segmentation with Anatomical-Aware Contrastive Distillation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Contrastive learning is effective for medical image segmentation with limited annotations.
• Existing methods often assume balanced class distributions, which is unrealistic for medical data.
• Imbalanced medical data leads to blurry contours and mislabeled rare objects.

Purpose of the Study:

• To develop a novel framework, ACTION (Anatomical-aware Contrastive Distillation), for semi-supervised medical image segmentation.
• To address the challenges posed by imbalanced class distributions in medical imaging.
• To improve segmentation accuracy, especially for rare anatomical structures.

Main Methods:

• ACTION employs an iterative contrastive distillation algorithm with soft negative labeling.
• It captures semantically similar features from negative samples to enhance data diversity.
• The framework integrates global semantic relationships and local anatomical features with minimal memory overhead.
• Anatomical contrast is introduced by sampling hard negative pixels to refine segmentation boundaries.

Main Results:

• ACTION significantly outperforms existing state-of-the-art semi-supervised methods on benchmark datasets.
• The method demonstrates improved accuracy in segmenting rare objects.
• Smoother segmentation boundaries and more precise predictions are achieved.

Conclusions:

• ACTION provides an effective solution for semi-supervised medical image segmentation, particularly in imbalanced scenarios.
• The anatomical-aware contrastive distillation approach enhances segmentation performance.
• This framework offers a promising direction for improving medical image analysis with limited labeled data.