Exploring GANs and diffusing models for medical Image data augmentation

Abstract

The scarcity of medical images remains a major challenge in the development of accurate and reliable deep learning models for medical image analysis. This thesis explores the use of generative models specifically Generative Adversarial Networks (GANs) and Diffusion Models to synthesize realistic chest X-ray images for data`augmentation. Different generative models were implemented and evaluated such as DCGAN, WGAN-GP, ProGAN, StyleGAN3, DDPM,`and DDIM. All models were trained to generate grayscale chest X-ray images at a resolution of 64×64 pixels. The quality of`the generated images was assessed using both quantitative metrics, such as the Fréchet Inception Distance (FID), and qualitative evaluation through a Visual Turing Test (VTT) conducted by a radiologist. Results showed that diffusion models, particularly DDIM and DDPM, produced the most realistic images, in`the normal and pneumonia classes. Interestingly, DCGAN showed unexpectedly strong performance when generating Normal chest X-ray images. The findings demonstrate that generative models, especially diffusion-based models can effectively contribute to expanding medical datasets and improving deep learning performance, although computational limitations and resolution constraints remain challenges