Vision Transformers with Augmented Dataset using DDPM to detect Parkinson's Disease

Denoising Diffusion Probabilistic Models (DDPM) have demonstrated their prowess in generating photorealistic and high-fidelity images, primarily in the context of images and videos. However, their potential to enhance downstream tasks like medical image classification remains underexplored. This study investigates the potential of utilizing synthetic data generated by DDPM for Parkinson's disease detection from MRI data. By training a range of Transformer models – ViT, LeViT, MaxViT, Swin Transformer, and DeiT using the synthetic data, we aim to elevate the performance of image classifiers. Our research addresses the gap between the potent image generation capabilities of DDPM and the enhancement of image classifiers. Integrating synthetic data, our aim is to enhance the classification accuracy of advanced Vision Transformer models, bolstering Parkinson's disease detection. This exploration reveals the symbiotic potential between generative models and classifiers, promising improved disease detection outcomes from medical images. Conducting a comprehensive evaluation across various Transformer architectures, we evaluate the impact of synthetic data on classifier performance. These findings illuminate the feasibility of integrating DDPM-generated data into the image classifier training process, advancing both image generation and classification realms in medical imaging.