DiffMCG: A diffusion model with mask-conditioned guiding module for medical image classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Diffusion models show promise in medical image classification by reducing noise.
• Current models overlook crucial lesion mask information, limiting denoising and classification performance.

Purpose of the Study:

• To propose DiffMCG, a diffusion model incorporating mask-conditioned guiding for enhanced medical image classification.
• To improve denoising and classification accuracy by utilizing both image and lesion mask features.

Main Methods:

• Introduced a Mask-Conditioned Guiding (MCG) module to extract features from medical images and their masks.
• Designed a U-Net denoising network using multi-layer perceptrons (MLP) for denoising within the category label space.
• Implemented MMD regularization loss for multimodal information consistency during diffusion.

Main Results:

• Comparative and ablation experiments validated the effectiveness of the MCG module.
• The proposed DiffMCG model demonstrated superior denoising and classification performance.
• The MCG module significantly improved the utilization of structured semantic information.

Conclusions:

• DiffMCG enhances medical image classification by effectively integrating lesion mask information.
• The MCG module provides technical support for precision medical diagnostics.
• This approach advances the application of diffusion models in healthcare.