Domain-Generalized Discrete Diffusion Model for Cross-Domain Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Domain shift significantly degrades medical image segmentation model performance on unseen data.
• Variations in acquisition protocols and modalities cause domain shift, limiting model generalizability.

Purpose of the Study:

• To develop a novel diffusion-based generative model for single-source domain generalization in medical image segmentation.
• To enhance the robustness and cross-domain performance of segmentation models.

Main Methods:

• Introduced the Domain-Generalized Discrete Diffusion Model for Segmentation (DG-DDM-Seg).
• Employed robust feature extraction from conditional images to ensure domain independence.
• Utilized a two-path reverse diffusion process with robust features and pseudo-labels for training.

Main Results:

• DG-DDM-Seg achieved state-of-the-art performance in cross-domain medical image segmentation.
• Demonstrated effectiveness across domain shifts in modality, sequence, and site.
• The model generates discrete conditional distributions of segmentation masks for unseen domains.

Conclusions:

• DG-DDM-Seg effectively addresses the domain shift problem in medical image segmentation.
• The proposed methods enhance domain independence and cross-domain segmentation accuracy.
• The diffusion-based approach offers a promising direction for generalized medical image analysis.