Medical image generation using artificial intelligence, especially Generative Adversarial Networks (GANs), has become a powerful solution to address challenges of limited data, class imbalance, and privacy restrictions in clinical imaging. While several GAN-based strategies have been proposed, earlier reviews often describe models in isolation without linking them clearly to medical datasets, evaluation measures, or diagnostic impact. This review focuses on GAN-based augmentation methods applied to medical imaging, systematically comparing prominent GAN families such as DCGAN, Pix2Pix, CycleGAN, StarGAN, and DualGAN across commonly used datasets, including BraTS, ISIC, DRIVE, and ADNI. Reported outcomes are synthesised using image-quality metrics (SSIM, PSNR, FID, LPIPS) and task-based measures (accuracy, sensitivity, Dice score). Findings suggest Pix2Pix frequently improves accuracy and SSIM on MRI and dermoscopy images, CycleGAN enhances sensitivity in retinal tasks, while newer models like StyleGAN and diffusion-based approaches achieve stronger perceptual fidelity. The review concludes that GAN-driven augmentation provides measurable benefits but is highly task- and dataset-dependent, emphasizing the future promise of hybrid GAN–diffusion pipelines and explainability tools for clinical adoption.
Dash et al. (Tue,) studied this question.