Unsupervised Domain Adaptation in Biomedical Images Segmentation With Guided Diffusion Generative Prior.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semantic segmentation lacks large-scale datasets, hindering progress.
• Foundation models like Segment Anything Model (SAM) offer broad segmentation capabilities.
• SAM struggles with specific regions, especially in biomedical images, necessitating Unsupervised Domain Adaptation (UDA).

Purpose of the Study:

• To develop an effective UDA strategy for biomedical image segmentation overcoming significant domain shift.
• To leverage generative priors from diffusion models and information from SAM for enhanced segmentation accuracy and robustness.

Main Methods:

• Proposed a UDA strategy using a segmentation diffusion model to learn source mask probability distribution.
• Integrated SAM's raw segmentation outputs as supplementary inputs for adaptation and robustness.
• Evaluated the method on diverse biomedical datasets (mitochondria, endoplasmic reticulum, brain tumors) across 10 adaptation scenarios.

Main Results:

• The proposed method significantly outperformed state-of-the-art UDA techniques in various adaptation scenarios.
• Ablation studies confirmed the critical contribution of each component within the proposed strategy.
• Demonstrated improved accuracy and robustness in biomedical image segmentation.

Conclusions:

• The novel UDA approach effectively addresses the domain shift challenge in biomedical image segmentation.
• Combining diffusion models with SAM provides a powerful framework for accurate and robust segmentation.
• The method shows great promise for advancing medical image analysis applications.