DiffMIC-v2: Medical Image Classification via Improved Diffusion Network.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Denoising Diffusion Models (DDMs) excel in generative tasks but are underutilized in medical diagnosis.
• Existing methods struggle with noise and perturbations in medical image representations.

Purpose of the Study:

• To propose DiffMIC-v2, a diffusion-based network for general medical image classification.
• To enhance medical image analysis by eliminating noise and improving representation learning.

Main Methods:

• Developed DiffMIC-v2, a diffusion-based network for medical image classification.
• Implemented an improved dual-conditional guidance strategy for enhanced regional attention.
• Introduced a novel Heterologous diffusion process for efficient latent space learning.

Main Results:

• DiffMIC-v2 demonstrated superior performance on four diverse medical classification tasks (chest X-ray, ultrasound, dermatoscopic, fundus images).
• The model significantly outperformed state-of-the-art methods in multi-class and multi-label classification.
• Achieved higher runtime efficiency and accuracy with fewer iterations compared to previous models.

Conclusions:

• DiffMIC-v2 shows significant effectiveness and universality for medical image classification.
• The proposed model offers a promising direction for applying diffusion models in medical diagnosis.
• The advancements in guidance and diffusion processes contribute to robust medical image analysis.