Researchers have developed a novel artificial intelligence-assisted imaging technique that may transform how high-resolution retinal images are captured in clinical settings, according to a study published in npj Artificial Intelligence this month.

The study, titled “Artificial intelligence-assisted retinal imaging enables dense pixel sampling from sparse measurements,” introduces a deep learning framework that enables adaptive optics optical coherence tomography (AOOCT) systems to produce high-quality images using significantly less data than traditionally required.

High-resolution imaging of the retina typically depends on dense pixel sampling. While dense sampling yields detailed imagery, it comes at a cost: slow acquisition speeds, large data burdens and motion artifacts resulting from eye movements during scans. To overcome these barriers, the research team led by scientists including Johnny Tam, PhD, implemented a cutting-edge AI model known as a residual in residual transformer generative adversarial network (RRTGAN). This model works alongside traditional AOOCT hardware to enhance the resolution of sparsely sampled images, effectively “filling in” missing information with high fidelity.

According to the study results, RRTGAN restored dense image quality using only 25% of the data typically required for high resolution. Compared with conventional AI reconstruction methods such as ESRGAN and SwinIR, RRTGAN produced sharper visualizations of retinal cone photoreceptors—microscopic sensory cells critical for vision—closely matching fully sampled reference images.

The study found that images enhanced by RRTGAN successfully preserved structural details such as cone cell spacing and contrast, making the AI-augmented technique more than just visually convincing, delivering scientifically robust data suitable for analyzing cellular changes in the retina.Nature

According to the team and supporting commentary from the National Eye Institute, integrating AI with advanced retinal imaging has the potential to bring cellular-level ophthalmic diagnostics into routine clinical use, improving early detection of retinal disease and guiding treatment decisions.

While the current study focused on healthy eyes, the researchers note that the next step will be to validate RRTGAN’s performance in patients with retinal pathology. Successful translation could pave the way for AI-supported imaging tools that not only accelerate image acquisition, but also unlock new possibilities in disease diagnosis and monitoring at the cellular level.

Reference

Tam, J., et al. Artificial intelligence-assisted retinal imaging enables dense pixel sampling from sparse measurements. npj Artificial Intelligence (2025). https://doi.org/10.1038/s44387-025-00038-2