Rethinking Semi-Supervised Medical Image Segmentation: A Variance-Reduction Perspective.

Area of Science:

• Medical Imaging Analysis
• Computer Vision
• Machine Learning

Background:

• Contrastive learning (CL) is key for medical image segmentation, enhancing visual representations by contrasting sample pairs.
• Current CL methods struggle with limited labels and distinguishing similar anatomical regions, leading to model collapse and misclassification of tail classes.

Purpose of the Study:

• To introduce ARCO, a novel semi-supervised contrastive learning framework using stratified group theory for improved medical image segmentation.
• To address the limitations of existing CL methods in handling limited labels and class imbalance in medical imaging.

Main Methods:

• ARCO employs variance-reduced estimation techniques, proven theoretically universal for variance reduction, particularly effective in pixel/voxel-level segmentation with scarce labels.
• The framework utilizes stratified group theory to enhance the sampling of dissimilar examples, mitigating model collapse.

Main Results:

• ARCO consistently outperformed state-of-the-art semi-supervised methods across eight diverse benchmarks (five medical, three semantic segmentation datasets) under various label settings.
• Augmenting existing CL frameworks with ARCO's sampling techniques yielded significant performance gains.

Conclusions:

• ARCO represents a significant advancement in semi-supervised medical image segmentation, offering a robust solution for safety-critical applications.
• The study quantifies limitations of current self-supervision objectives and highlights the benefits of variance-reduced estimation in CL for medical imaging.