Comprehensive Summary
This study provides a general and broad examination of generative artificial intelligence (AI) in ophthalmology, assessing current applications, model types, and limitations to potential implementation in clinical settings. The authors conducted a scoping review of 40 primary studies published as of December 2024 obtained from MEDLINE, Embase, and Web of Science. They identified five main application categories—data augmentation, predictive modelling, image enhancement, segmentation, and education/interpretability. The majority of studies in the early literature were found to utilize generative adversarial networks (GANs), while more recent studies are shifting towards models with diffusion or hybrid architectures. The review also discusses important limitations, including limited dataset variability, the risk of synthetic artifacts, uncertainty in regulation, and poor generalizability of models to clinical settings in the real world.
Outcomes and Implications
The authors contend that generative AI has the potential to drive meaningful advances in the field of ophthalmology by improving the quality of imaging, creating approaches to trainer data, and supporting early detection of disease or disease progression. Clinically, generative AI tools may assist with improving accuracy in diagnosis, and informing personalized care, particularly in the areas of retinal and optic nerve disorders. Still, most of these applications remain at the proof-of-concept stage, and currently have minimal prospective real-world validation. The authors conclude the paper by stating that generative AI has the overall potential to be transformative, but that the clinical utilization of generative AI tools will depend on rigorous validation, explainability structures, and clarity in regulation prior to implementation in the clinical setting.