A systematic review published in Artificial Intelligence & Autonomous Systems compares GANs, diffusion models, VAEs and large language models across five medical application areas. It highlights hallucination, privacy, bias, explainability and governance challenges, and calls for trustworthy, human-centered deployment of generative AI as an assistive tool for medical professionals.
With the rapid growth of medical data, uneven distribution of healthcare resources and increasing complexity of disease diagnosis and treatment, improving information utilization, supporting clinical decision-making and advancing personalized medicine have become key directions in medical artificial intelligence. Unlike traditional AI, which focuses on classification and prediction, generative AI can learn data distributions and generate new content such as text and images. It therefore shows potential in clinical decision support, medical imaging, drug discovery and medical education.
Recently, Yifei Li, Yipeng Liao, Liqun Lin and Tiesong Zhao from Fuzhou University published a systematic review in Artificial Intelligence & Autonomous Systems . The review provides a comprehensive overview of the technical foundations, application progress and future directions of generative AI in medicine. It compares four representative generative models—generative adversarial networks (GANs), diffusion models, variational autoencoders (VAEs) and large language models (LLMs)—and analyzes their probabilistic foundations, technical characteristics and applicable medical scenarios.
At the application level, the review summarizes medical advances in five areas: clinical decision support, including intelligent diagnostic reasoning, personalized treatment plan generation and automated clinical documentation; medical image synthesis, including data augmentation, cross-modal conversion and small-sample synthesis; drug discovery, including molecular generation, clinical trial simulation and patient recruitment; medical literature processing and knowledge integration; and medical education and training, including personalized learning, virtual cases and intelligent assessment feedback.
The paper also identifies multiple challenges. At the data level, medical training data are highly sensitive, structurally incomplete and insufficiently representative. At the model level, factual errors, hallucinations, black-box characteristics and bias amplification remain difficult to avoid. At the application level, clinical liability, academic writing standards and the risk of overreliance require attention. At the governance level, existing regulatory frameworks struggle to keep pace with technological development. The review further organizes these challenges across four dimensions: data, models, applications and system governance.
For future directions, the review proposes developing uncertainty quantification, explainable AI and retrieval-augmented generation (RAG), while strengthening data governance, ethical standards, regulatory systems and interdisciplinary collaboration. The authors emphasize that generative AI should currently serve as an assistive tool for medical professionals, rather than a substitute capable of making independent clinical decisions. Only by building trustworthy, equitable, explainable and accountable technological and governance systems can generative AI achieve safer, more reliable and more human-centered applications in medicine.
Li Y, Liao Y, Lin L, Zhao T. Generative artificial intelligence in medicine: a systematic review of models, applications, challenges, and future directions. Artif. Intell. Auton. Syst. 2026(2):0007, https://doi.org/10.55092/aias20260007.
Artificial Intelligence and Autonomous Systems
Literature review
Not applicable
Generative artificial intelligence in medicine: a systematic review of models, applications, challenges, and future directions
6-Nov-2026