.png){kind=icon}
.png){kind=icon}
Return
Disentangled generative uncertainty-aware multi-modal diffusion segmentation of medical images
.png){kind=icon}
PRE
AI
D
S
M
DOI：10.1016/j.media.2026.104122.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Generative AI for Trustworthy Medical Segmentation: We introduce D-GUMM-DS, a novel framework that uniquely applies Denoising Diffusion Probabilistic Models (DDPMs) to inherently integrate uncertainty quantification into multi-modal medical image segmentation. • Clinically Meaningful Uncertainty: The framework produces reliable, pixel-wise, and global uncertainty maps directly from the divergence of generated segmentation samples, addressing a key barrier to clinical adoption. • Adaptive Multi-Modal Fusion: A novel disentangled, uncertainty-aware fusion mechanism dynamically combines information from diverse imaging modalities, resolving conflicts and leveraging complementary strengths. • Superior Performance and Interpretability: Our approach achieves superior segmentation accuracy while providing well-calibrated and clinically interpretable uncertainty information, significantly enhancing trust in AI-driven medical image analysis. • Beyond Post-Hoc Methods: D-GUMM-DS moves past traditional post-hoc uncertainty methods and computationally expensive approximations by embedding the probabilistic nature of generative models directly into the pipeline.
Keywords:
Generative AI
Medical image segmentation
Uncertainty quantification
Diffusion models
Multi-modal fusion
Journal
IF:
11.8
Papers:
3.7K
Citations:
2.4W
Organization
Cited Papers