FEU-Diff: A Diffusion Model With Fuzzy Evidence-Driven Dynamic Uncertainty Fusion for Medical Image Segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Diffusion models are emerging generative frameworks for medical image segmentation.
• Current methods struggle with adaptive fusion of conditional priors and denoised features.
• Existing approaches lack explicit modeling of pixel-level uncertainty, leading to detail loss.

Purpose of the Study:

• To introduce FEU-Diff, a novel diffusion-based segmentation framework.
• To address limitations in adaptive fusion and uncertainty modeling in diffusion models.
• To improve accuracy and structural detail preservation in medical image segmentation.

Main Methods:

• FEU-Diff integrates fuzzy evidence modeling and uncertainty fusion (UF).
• A fuzzy semantic enhancement (FSE) module models pixel-level uncertainty using Gaussian membership functions and fuzzy logic.
• An evidence dynamic fusion (EDF) module uses Dirichlet distribution for adaptive feature fusion and UF quantifies prediction discrepancies.

Main Results:

• FEU-Diff outperforms state-of-the-art methods on four public datasets.
• Achieved average improvements: 1.42% Dice Similarity Coefficient (DSC), 1.47% Intersection over Union (IoU).
• Reduced 95th percentile Hausdorff distance (HD95) by 2.26 mm and generated interpretable uncertainty maps.

Conclusions:

• FEU-Diff effectively improves medical image segmentation accuracy and completeness.
• The framework's fuzzy evidence and uncertainty fusion mechanisms enhance boundary delineation and detail preservation.
• FEU-Diff offers enhanced clinical interpretability through generated uncertainty maps.