Artificial intelligence may be able to estimate retinal thickness from standard fundus photographs, generating thickness maps with mean relative relative differences of approximately 5% relative to optical coherence tomography measurements and potentially extending the structural information available from diabetic retinopathy screening images, according to a comparative study published in Scientific Reports.  

For the study, researchers developed and evaluated a foundation model-based approach that predicts total retinal thickness (TRT) maps directly from color fundus photographs (CFPs) without image registration or other preprocessing steps. 

The study used RETFound, a vision transformer foundation model trained on retinal images, as the feature extraction backbone and paired CFP and optical coherence tomography (OCT) data from the Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI) dataset.  

The final analysis included 1,426 eyes from 807 participants, with paired CFP and OCT images obtained from two optical coherence tomography devices: the Topcon Maestro2 and the Topcon Triton.

Researchers also used 300 eyes from the OCTA-500 dataset to train a retinal segmentation model and 28,020 fundus photographs from EyePACS to fine-tune RETFound and reduce device-related bias.  

Three decoder training strategies for TRT estimation were compared: device-specific Individual Model, Combination Model trained using CFPs from both devices, and Random Selection Model using randomly selected thickness maps from either device.  

Both the Individual and Combination models demonstrated low error relative to their device-specific reference standards, with relative differences of approximately 5%. Mean signed differences ranged from approximately -1 μm to 2.4 μm across Maestro2 and Triton inputs.

In contrast, the Random Selection Model showed biases of -19 μm for Maestro2 images and 15 μm for Triton images when evaluated against device-specific references. However, when compared with an averaged reference map derived from both devices, the Random Selection Model yielded signed differences of approximately -2 μm for both image sources, suggesting convergence toward a device-independent representation of retinal thickness.  

Relative differences for the Random Selection Model were similar for Maestro2 inputs and Triton inputs. For Triton images, the Random Selection Model achieved significantly lower relative error than either the Individual or Combination models. Although the Random Selection Model produced the most device-independent estimates, the Individual and Combination models maintained the lowest bias when compared with device-specific OCT reference standards.  

Sector-level analyses using the Early Treatment Diabetic Retinopathy Study (ETDRS) grid showed that Random Selection Model predictions aligned more closely with the inter-device average retinal thickness map than with device-specific reference measurements. Significant differences were observed in most sectors, except the outer inferior and outer temporal sectors, which showed no significant differences.  

Qualitative analyses demonstrated that the Random Selection model generally reproduced spatial patterns of retinal thickness in both left and right eyes. However, performance declined in eyes with macular edema because the small number of cases limited the model's ability to learn edema-associated features.

Another notable finding was substantial variability between OCT platforms. Across all ETDRS sectors, retinal thickness measurements from Maestro2 were, on average, more than 30 μm greater than those from Triton, underlining the challenge of establishing a uniform retinal thickness reference standard across imaging devices.  

Several limitations were noted. The authors cautioned that the approach was not designed for direct DR or DME diagnosis and requires prospective validation before clinical use. Macular edema cases were underrepresented—with qualitative examples suggesting reduced accuracy in eyes with macular edema, especially central foveal edema—preventing meaningful quantitative subgroup analyses. The model may also be sensitive to variations in fundus image color and retinal pigmentation, and the AI-READI dataset lacked race and pigmentation metadata needed to evaluate those effects. The findings also were limited to the Maestro2 and Triton imaging systems and may not generalize to other retinal imaging platforms. In addition, the averaged retinal thickness map used for evaluation was designed to reduce device-related bias but does not represent a definitive ground truth.  

The researchers reported no competing interests. 

(Editor's note: The study manuscript will undergo further editing before final publication, and there may be errors present that affect the content. All legal disclaimers apply.)

Expert Insight 

AACE Endocrine AI invited study researcher Jui-Kai Wang (far right in image below), PhD, assistant professor in the Department of Ophthalmology at the University of Texas Southwestern Medical Center in Dallas, to elaborate on the findings.  Here's what Dr. Wang said about the research.

Why does this study matter?  

It explores a new direction for artificial intelligence (AI) in ophthalmic imaging: using latent representations to bridge information across imaging modalities. Color fundus photography remains the most widely available retinal imaging tool, while optical coherence tomography (OCT) provides valuable structural information but is not always accessible. We wanted to investigate whether AI could learn the hidden relationships between these modalities and use information from one to estimate information from the other. 

A useful analogy is how generative AI can create a realistic headshot from a photograph taken from a different angle. Similarly, our model learns features from fundus photographs that can be translated into retinal thickness information normally obtained from OCT. Our group previously demonstrated a related concept in optic disc swelling, but that approach required image registration and substantial preprocessing. In this study, we achieved an end-to-end framework and extended the concept to diabetes, a condition that affects a much larger patient population and has broad public health relevance. 

What data surprised you the most?  

First, we were impressed by the quality of the AI-READI dataset, one of the NIH Bridge2AI initiatives led by Aaron Lee, MD, and colleagues. This dataset is exceptionally well organized and includes high-quality multimodal imaging data from patients with diabetes, along with carefully curated clinical information. Having access to paired fundus photographs and OCT scans at this scale created a unique opportunity to explore questions that would have been difficult to study otherwise. 

The second came from the results themselves. We started with the hypothesis that fundus photographs and OCT images may share underlying biological information, even though each modality captures that information differently. Seeing the model learn from the superficial appearance of the retina in a fundus photograph and then estimate retinal thickness patterns that normally require OCT was fascinating. The findings suggest that AI can uncover relationships across imaging modalities that are not immediately apparent to human observers, opening new possibilities for multimodal medical imaging research. More broadly, the results support the idea that different imaging modalities may contain complementary manifestations of the same underlying biology, and that AI can help bridge those representations in meaningful ways. 

How might the findings influence endocrinology clinical practice?  

Many endocrine and metabolic diseases, including diabetes, have measurable effects on the retina. Advances in retinal imaging and AI are allowing us to extract increasingly rich information from eye images, not only about ocular health but also about systemic disease. Our study suggests that AI can identify relationships between different retinal imaging modalities and potentially recover clinically meaningful structural information from more accessible imaging data. In the long term, this type of work may help expand the role of retinal imaging in risk assessment, disease monitoring, and population screening. More broadly, it supports the idea that the eye can serve as a noninvasive platform for studying systemic diseases and their progression.