Generative AI is learning to clean up, create, and even translate medical eye images. A new review explains the promise—and the pitfalls.

A new review in Survey of Ophthalmology says generative artificial intelligence (AI) is rapidly reshaping eye care by improving the images doctors use to spot disease. The authors examined 40 studies of “generative adversarial networks” (GANs)—AI models that can create realistic pictures—and outline how they’re being used to enhance image quality, detect disease, and generate synthetic data for training when real examples are scarce.  

What GANs can do today

• Clean up scans: GANs can remove shadows and other artifacts from optical coherence tomography (OCT) images—the high‑resolution scans used to inspect the retina—making subtle disease changes easier to see.  
• Create images doctors usually need dye or extra machines for: Models can translate a standard retinal photograph into a convincing fluorescein angiogram (an invasive test that uses injected dye) or other specialized views, potentially reducing patient burden and speeding diagnosis.  
• Spot key structures in noisy data: In head‑to‑head examples in the paper, a GAN segmented delicate corneal nerves more accurately than a popular AI baseline (U‑Net) when images were noisy—important for conditions like diabetic eye disease. See Fig. 6 in the review.  

Why it matters

Sharper, more informative images can help clinicians detect problems earlier (from diabetic retinopathy to glaucoma), monitor treatment, and train new algorithms and clinicians—especially for rare conditions where real‑world data are limited. A schematic in the paper shows three main GAN families—unconditional (synthetic image generation), conditional (style transfer), and image‑to‑image translation—each geared to different clinical tasks. See Fig. 8 in the review.  

Not ready for autopilot

The authors stress real risks. GANs can introduce subtle errors (so‑called “mode collapse,” spatial distortions, or misleading features). Models may also behave differently across clinics because of lighting, camera settings, or patient populations (“domain shift”). And because GANs make images that look real, there’s a clear need to guard against misuse and require strict validation before clinical deployment.  

The bottom line

Generative AI could transform ophthalmic imaging—cleaning up scans, synthesizing hard‑to‑obtain views, and boosting early detection—if the field pairs innovation with robust guardrails and transparency. The review’s take: progress is accelerating, but careful engineering, ethics, and clinical trials must come first.  

Source: Waisberg, E., Ong, J., Kamran, S. A., Masalkhi, M., Paladugu, P., Zaman, N., Lee, A. G., & Tavakkoli, A. (2025). Generative artificial intelligence in ophthalmology. Survey of ophthalmology, 70(1), 1–11. https://doi.org/10.1016/j.survophthal.2024.04.009