Background
Medical imaging is a cornerstone of modern clinical medicine, supporting diagnostic assessment, therapeutic planning, and prognostic evaluation. However, real-world imaging is often constrained by practical limitations, including scarce and heterogeneous data, acquisition trade-offs that can introduce noise, artifacts, and loss of resolution, and incomplete longitudinal follow-up. These challenges motivate methods that can synthesize missing data, harmonize heterogeneous inputs, and augment incomplete datasets, which has become a key clinical driver for bringing generative AI into medical imaging workflows.

Research Progress
In their Research review article, “Generative Artificial Intelligence in Medical Imaging: Foundations, Progress, and Clinical Translation,” Hairong Zheng and Shanshan Wang (Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology,
Chinese Academy of Sciences, Shenzhen, China.) et al. provide a comprehensive overview of recent advances in generative modeling for medical imaging.

The review summarizes how generative AI is enabling capabilities such as data synthesis, image enhancement, modality translation, and spatiotemporal modeling. It covers a broad range of generative approaches, including GANs, VAEs, diffusion models, and emerging multimodal foundation architectures, and also discusses sequence modeling methods such as Transformers, Mamba, and autoregressive models.

A defining strength of the review is its clinic-facing, workflow-first organization. Rather than presenting generative methods only by architecture, the authors map how these approaches are being studied along the full imaging pathway, from acquisition and reconstruction to cross-modality synthesis, and onward to applications relevant to diagnosis, treatment, and prognosis. Framed this way, generative AI is discussed as a means to help clinicians see clearer under constrained scans, see more completely when modalities are missing, update information faster in time-sensitive settings, and model changes over time for longitudinal assessment.
To support more consistent benchmarking and clinical translation, the review introduces a three-level evaluation framework that considers pixel-level fidelity, feature-level realism, and task-level clinical relevance. In addition, the paper summarizes public datasets and benchmarking resources used in the field, and provides an expanded dataset list in the supplementary materials to facilitate reproducible development and evaluation.

Challenges and Boundaries
The authors identify several obstacles to real-world deployment. These include limited generalization under domain shift, risks of hallucinated or unreliable features, ongoing data scarcity and privacy concerns, and stringent regulatory and ethical constraints. They also note clinical considerations around output reliability, transparency for decision-making, and integration into established workflows.

Future Prospects
Looking ahead, the review discusses the convergence of generative AI with large-scale foundation models, highlighting how this synergy may support the next generation of scalable, reliable, and clinically integrated imaging systems. Beyond current foundation-model paradigms, the authors point to longer-term directions toward world models and digital twins that can simulate physiological processes and individualized disease trajectories in a dynamic and interpretable manner. They emphasize the role of multimodal information, including combinations of imaging, text, and genomics, to enable more holistic and personalized understanding of health and disease.

To realize this vision, the review highlights the need for improved robustness to data heterogeneity, incorporation of anatomical and physiological priors, uncertainty quantification, explainable design to foster clinical trust, and more efficient inference to support time-sensitive applications. It also stresses the importance of trustworthy and well-governed synthetic imaging, including hallucination detection, ethical oversight, and standardized evaluation, alongside multi-institutional foundation models and practical integration into clinical systems.

The complete study is accessible via DOI:10.34133/research.1029