Disentangled generative uncertainty-aware multi-modal diffusion segmentation of medical images

Area of Science:

• Artificial Intelligence in Medicine
• Medical Image Analysis
• Deep Learning

Background:

• Clinical integration of deep learning for medical image segmentation is hindered by lack of transparency and uncertainty quantification (UQ).
• Multi-modal imaging presents challenges in data integration and robust uncertainty estimation for trustworthy AI.
• Clinicians require precise segmentations and explicit confidence measures for high-stakes medical decisions.

Purpose of the Study:

• To introduce a novel framework, Disentangled Generative Uncertainty-Aware Multi-Modal Diffusion Segmentation (D-GUMM-DS), for robust multi-modal medical image segmentation.
• To leverage Generative AI (GenAI) and Denoising Diffusion Probabilistic Models (DDPMs) for inherent uncertainty quantification.
• To develop an uncertainty-aware fusion mechanism for intelligent integration of multi-modal features.

Main Methods:

• Utilized Denoising Diffusion Probabilistic Models (DDPMs) to learn data distributions and generate multiple plausible segmentations.
• Developed a disentangled, adaptive, and uncertainty-aware fusion mechanism for multi-modal feature integration.
• Derived pixel-wise and global uncertainty estimates by analyzing divergence among generated segmentation samples.

Main Results:

• Achieved highly accurate and robust segmentations in multi-modal medical imaging.
• Provided well-calibrated and clinically interpretable pixel-wise and global uncertainty maps.
• Demonstrated enhanced trust and decision support in AI-driven medical image analysis.

Conclusions:

• The proposed generative paradigm offers a principled approach to trustworthy AI in medical image segmentation.
• Direct integration of GenAI's probabilistic nature overcomes limitations of post-hoc UQ methods.
• D-GUMM-DS effectively addresses data heterogeneity and enhances clinical utility through reliable uncertainty estimation.