A diffusion model multi-scale feature fusion network for imbalanced medical image classification research.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Medical image classification faces challenges with highly imbalanced datasets, leading to low model accuracy.
• Traditional methods struggle with the scarcity of trainable data in medical image classification tasks.

Purpose of the Study:

• To effectively address poor training outcomes caused by imbalanced class datasets in medical imaging.
• To propose a superior network framework for enhancing medical image classification accuracy.

Main Methods:

• Introduced the diffusion model multi-scale feature fusion network (DMSFF) utilizing a diffusion generation model to overcome imbalanced classes (DMOIC).
• Implemented an image augmentation strategy through cropping (IASTC) and a multi-scale feature fusion network (MSFF) for hierarchical feature utilization.
• The DMSFF network is designed to tackle small, imbalanced samples and low accuracy in medical image classification.

Main Results:

• Evaluated DMSFF on highly imbalanced datasets APTOS2019 and ISIC2018, demonstrating significant improvements.
• Achieved classification accuracy of 0.872 and 0.906, and F1 scores of 0.731 and 0.836 on the respective datasets.
• Outperformed existing classification models in both accuracy and F1 score.

Conclusions:

• The proposed DMSFF architecture surpasses current methods in medical image classification on imbalanced datasets.
• Validated the efficacy of generative model-based class balancing and multi-scale feature fusion for performance enhancement.
• The DMSFF method shows potential for broad application across various imbalanced class datasets, promising improved results.