S&D Messenger: Exchanging Semantic and Domain Knowledge for Generic Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Manual labeling for medical image segmentation is time-consuming.
• Domain variations in datasets create challenges like semi-supervised medical domain generalization (Semi-MDG) and unsupervised medical domain adaptation (UMDA).
• Existing methods struggle to integrate semantic and domain knowledge effectively in semi-supervised settings.

Purpose of the Study:

• Develop a unified framework for semi-supervised medical image segmentation (SSMIS), Semi-MDG, and UMDA.
• Address the challenge of disparate knowledge sources (labeled vs. unlabeled data).
• Enhance the comprehension of both semantic and domain knowledge within a single learning process.

Main Methods:

• Introduced a Semantic & Domain Knowledge Messenger (S&D Messenger).
• Facilitated direct knowledge transfer between labeled and unlabeled datasets.
• Integrated the S&D Messenger into a pseudo-labeling approach.

Main Results:

• Achieved significant performance improvements on ten benchmark datasets.
• Demonstrated substantial gains in SSMIS (+7.5%), UMDA (+5.6%), and Semi-MDG (+1.14%) tasks.
• Outperformed state-of-the-art methods specialized for individual tasks.

Conclusions:

• The S&D Messenger provides a generic and effective solution for medical image segmentation tasks with domain variations.
• Direct knowledge delivery between labeled and unlabeled sets is crucial for mastering diverse semi-supervised scenarios.
• The proposed framework offers a versatile approach to advance medical image analysis.