Dynamic domain generalization for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Domain Generalization-based Medical Image Segmentation (DGMIS) aims to improve model robustness on unseen data.
• Existing DGMIS methods often use static models and global style transformations, limiting adaptation to target domain variations.
• Challenges include lack of dynamic adaptation and insufficient capture of local image details.

Purpose of the Study:

• To propose a Dynamic Domain Generalization (DDG) method for medical image segmentation.
• To enhance model generalization capability on unseen target domains through dynamic parameter adjustment and style simulation.
• To address limitations of static models and global-only style augmentation in current DGMIS.

Main Methods:

• Developed a Dynamic Position Transfer (DPT) module to decouple static and dynamic model parameters, incorporating positional encoding for adaptation.
• Introduced a Global-Local Fourier Random Transformation (GLFRT) module to capture both global and local style information, enhancing sample diversity.
• Utilized a random style selection strategy within GLFRT to balance diversity and computational cost.

Main Results:

• The proposed DDG method demonstrated superior performance over state-of-the-art approaches on multiple public medical image datasets.
• Achieved average Dice score improvements of 0.58% on Fundus, 0.76% on Prostate, and 0.76% on SCGM datasets.
• Experimental validation confirmed the effectiveness of dynamic adaptation and global-local style simulation.

Conclusions:

• The DDG method significantly improves the generalization capability of medical image segmentation models.
• Dynamic parameter adjustment and integrated global-local style transformations are key to robust segmentation across domains.
• The approach offers a promising solution for deploying medical image segmentation models in diverse clinical settings.