S²Match: Revisiting Weak-to-Strong Consistency From a Semantic Similarity Perspective for Semi-Supervised Medical...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning (SSL) reduces the need for extensive labeled data in medical image segmentation.
• Existing weak-to-strong consistency frameworks show promise but suffer from neglecting contextual dependencies and semantic similarity between labeled and unlabeled data.

Purpose of the Study:

• To propose a novel SSL framework, S²Match, to address limitations in medical image segmentation.
• To improve segmentation accuracy by integrating contextual dependencies and semantic similarity.
• To enhance feature representation using multi-scale information.

Main Methods:

• Developed S²Match framework with two key designs: intra-image pair-wise affinity map for contextual dependencies and feature querying for class representation learning.
• Introduced a Spatial-aware Fusion Module (SFM) for robust feature extraction from multiple scales.
• Applied the framework to medical image segmentation tasks.

Main Results:

• S²Match demonstrated consistent improvements over state-of-the-art methods on five public medical image segmentation benchmarks.
• Achieved competitive performance in both 2D and 3D medical image segmentation tasks.
• The proposed methods effectively addressed issues of inconsistent predictions and class-distribution discrepancy.

Conclusions:

• S²Match offers an effective SSL framework for medical image segmentation.
• The integration of contextual dependencies and semantic similarity significantly enhances segmentation performance.
• The SFM module contributes to robust feature extraction, further boosting results.