Siamese evolutionary masking: Enhancing the generalization of self-supervised medical image segmentation model.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Self-supervised learning (SSL) extracts features from unlabeled medical images for segmentation.
• Variations in medical image data distribution hinder model generalizability.
• Existing SSL methods like Instance Discrimination and Masked Image Modeling have limitations in capturing both global and local information.

Purpose of the Study:

• To introduce a novel self-supervised learning framework, Siamese Evolutionary Masking (SEM), to enhance generalizability in medical image segmentation.
• To effectively combine global and local feature extraction for improved model performance.
• To address the challenge of domain shift in medical imaging datasets.

Main Methods:

• Developed the Siamese Evolutionary Masking (SEM) framework with Siamese architecture (online and target branches).
• Implemented an evolutionary masking strategy (grid to block) in the online branch for robust feature learning.
• Introduced a Switch Decoder module to align features between branches and balance global/local information.

Main Results:

• SEM demonstrated strong performance compared to other SSL methods across six diverse datasets (skin and chest X-ray).
• Achieved superior cross-dataset segmentation and generalization performance.
• Reported Dice scores of 81.8% and 91.1%, Jaccard indices of 72.2% and 84.4%, and optimal HD95% of 13.1% and 10.5% in cross-dataset tests.

Conclusions:

• The SEM framework effectively enhances the generalizability of medical image segmentation models.
• Combining global and local feature learning through SEM overcomes limitations of existing SSL approaches.
• SEM offers a promising solution for robust medical image segmentation in the presence of data variations.