DCDiff: Dual-Granularity Cooperative Diffusion Models for Pathology Image Analysis.

Area of Science:

• Digital Pathology
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Whole Slide Images (WSIs) are crucial for medical diagnosis and treatment, with deep learning methods widely used for classification.
• Current deep learning approaches for WSI analysis often overlook contextual information by treating image regions as isolated, limiting accuracy.

Purpose of the Study:

• To propose a novel Dual-Granularity Cooperative Diffusion Model (DCDiff) for precise Whole Slide Image classification.
• To enhance WSI analysis by incorporating contextual information through a dual-granularity approach.

Main Methods:

• Developed a cooperative forward and reverse diffusion strategy using fine- and coarse-granularity levels to improve context awareness.
• Introduced a coupled U-Net with a Fine- and Coarse-granularity Cooperative Aware (FCCA) model for dual-granularity denoising and information exchange.
• Extracted cooperative diffusion features enabling cross-sample perception from reconstructed training sample distributions.

Main Results:

• The DCDiff model demonstrated superior performance compared to state-of-the-art methods on three public WSI datasets.
• The cooperative diffusion features effectively captured cross-sample relationships, leading to improved classification accuracy.
• The dual-granularity approach successfully integrated contextual information, overcoming limitations of existing methods.

Conclusions:

• The proposed DCDiff model offers a significant advancement in Whole Slide Image classification by effectively utilizing contextual information.
• The dual-granularity cooperative diffusion strategy provides a robust framework for analyzing complex medical images.
• The findings suggest DCDiff has strong potential for improving diagnostic accuracy in digital pathology.