Generative AI for Diagnostic Medical Imaging: A Review

This review provides a comprehensive analysis of recent advancements in generative deep learning (DL) models applied to diagnostic medical imaging, emphasizing their transformative potential in enhancing diagnostic accuracy, reducing radiation exposure, and improving data handling. We explore the architectures, applications, and unique contributions of generative adversarial networks (GANs), autoencoders (AEs), diffusion models, and transformer-based models. The key areas include synthetic data generation for training, text-to-image and image-to-text translation for interpretability, and image-to-image enhancement across imaging modalities. We designed different pipeline architectures presenting basic and advanced generative models specifically designed for medical imaging applications. These include enhanced GAN configurations, such as the multi-layer ML-C-GAN and Temporal-GAN for time-sequenced medical images, and specialized AE-GAN hybrids such as Atten-AE and M3AE, which combine attention modules and language encoding for text-to-image and image-to-text translation. Each pipeline uniquely addresses challenges in synthetic image quality, temporal progression, and accurate caption generation, showcasing their capacity to produce clinically relevant, high-fidelity images across modalities. The discussion highlights these architectural innovations, emphasizing their role in enhancing image synthesis, diagnostic reporting, and patient-specific image interpretation within medical imaging. The review concludes by identifying future directions to refine generative models for clinical applications, ultimately aiming to facilitate more accurate, accessible, and personalized patient care.