COVID-19 PNEUMONIA CHEST X-RAY PATTERN SYNTHESIS BY STABLE DIFFUSION.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• COVID-19 pneumonia frequently causes bilateral lung edema, a critical finding in chest X-rays.
• Synthesizing realistic medical images is crucial for augmenting datasets and training AI models.

Purpose of the Study:

• To fine-tune a stable diffusion model for generating high-resolution (512x512) chest X-ray images depicting bilateral lung edema.
• To evaluate the effectiveness of a class-specific prior preservation strategy in image synthesis.
• To compare the performance of the proposed method against conventional techniques like WGAN and DDIM.

Main Methods:

• Fine-tuning a stable diffusion model using a class-specific prior preservation strategy.
• Utilizing 300 positive and 400 negative chest X-ray images from the MIDRC dataset.
• Synthesizing images with bilateral lung edema indicative of COVID-19 pneumonia.
• Comparing synthetic image quality using Frechet Inception Distance (FID) and Kernel Inception Distance (KID).
• Evaluating classification performance using a trained Vision Transformer (ViT).

Main Results:

• The fine-tuned stable diffusion model achieved superior image synthesis quality, evidenced by FID of 9.2158 and KID of 0.0818.
• The synthetic images demonstrated significantly better performance compared to those generated by WGAN and DDIM.
• A Vision Transformer (ViT) achieved high classification accuracy (0.9975), precision (1.0), and recall (0.9950) using the synthetic images.

Conclusions:

• Stable diffusion models, when fine-tuned with prior preservation, can effectively synthesize high-quality, high-resolution chest X-ray images.
• This technique requires a limited number of real images and text prompts for guidance.
• The synthesized images are suitable for training AI models, improving diagnostic capabilities for conditions like COVID-19 pneumonia.