Diffusion Model-based Medical Image Generation as a Potential Data Augmentation Strategy for AI Applications.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical artificial intelligence (AI) applications require large, diverse datasets for robust training.
• Access to real-world clinical data can be limited by privacy concerns and cost.
• Data augmentation strategies are crucial for enhancing AI model performance with limited datasets.

Purpose of the Study:

• To explore a generative image synthesis method using diffusion models for medical data augmentation.
• To evaluate the efficiency and cost-effectiveness of diffusion models in low-resource computing environments.
• To identify optimal training parameters for high-fidelity synthetic medical image generation.

Main Methods:

• Utilized the MedMNIST v2 dataset for training under low-performance computing conditions.
• Developed an annotated diffusion model to synthesize new medical images based on existing data characteristics.
• Performed quantitative evaluations using loss function gradient descent and Fréchet Inception Distance (FID) with varying loss functions and feature vector dimensions.

Main Results:

• The diffusion model successfully generated medical images with similar styles but varied anatomical details compared to original data.
• The L2 loss function with a feature vector dimension of 64 yielded the best FID score of 0.85.
• The Huber loss function demonstrated enhanced model robustness, though with a higher FID of 15.2 at a feature vector dimension of 2048.

Conclusions:

• Diffusion model-based medical image synthesis is a viable augmentation strategy for AI, especially when real data is scarce.
• Optimal training parameters, including loss function choice and feature vector dimensionality, significantly impact synthetic image quality.
• Further research should focus on applying these models to more complex medical datasets and clinical scenarios.