From minimax to ProGAN : advancing GAN Stability for synthetic X-ray generation

Abstract

One of the most impactful developments in deep learning, Generative Adversarial Networks (GANs) have revolutionized generative modeling by enabling the synthesis of realistic data across a variety of domains. Their utility spans from art and natural image generation to scientific and medical applications. Unlike classical generative models that rely on explicit density estimation or variational inference, GANs learn to generate data through an adversarial framework, allowing them to represent complex distributions with high fidelity. In this dissertation, the theory and architecture of GANs are explored in depth, beginning with their minimax formulation and progressing through improvements such as Deep Convolutional GANs (DCGANs), Wasserstein GANs (WGANs), and WGANs with Gradient Penalty (WGAN-GP). The dissertation culminates with the study of Progressive Growing of GANs (ProGAN), a technique that enables stable training for high-resolution image generation. A central application investigated in this work is the generation of synthetic medical images, specifically chest X-rays. In clinical practice, the scarcity of labeled data and the need for diverse training examples make synthetic augmentation a compelling approach. GAN-based models are evaluated for their ability to generate realistic X-ray images, with emphasis on fidelity, structure, and stability across architectures. By comparing GAN variants and examining their performance on real datasets, this dissertation highlights the practical relevance of generative models in medical imaging and their potential to support diagnostic and research workflows.