RNFL Thickness Estimates Successfully Derived from Optic Disc Photos Alone

Subtle optic disc changes may go unrecognized in a routine clinical evaluation of the optic nerve, given the subjective nature of the assessment. This research proposes a future paradigm whereby fundus photography can call upon AI models to estimate RNFL thickness measures, and hence glaucoma risk, without use of OCT. Photo: Marc D. Myers, OD, and Andrew S. Gurwood, OD. Click image to enlarge. To better assess and predict if patients with ocular hypertension will develop primary open-angle glaucoma (POAG), investigators would like to translate fundus pictures into more concrete, discrete pieces of information. Pursuit of this goal started over 20 years ago with investigators taking optic disc photos from those part of the Ocular Hypertension Treatment Study (OHTS) to assess the use of predicted retinal nerve fiber layer (RNFL) thickness as a risk factor in developing POAG.Included in the OHTS population that was studied for this analysis were 1,444 participants with ocular hypertension; mean age was 56 and 57.7% of the group was female. A deep learning model trained with OCT images was used to generate predicted RNFL thickness from the 66,714 fundus photos captured during the OHTS1 and OHTS2 cohorts. Baseline predicted RNFL thickness was greater for eyes that did not convert to POAG than those that did, with values of 97.1µm and 94.1µm, respectively. It was found that, in both univariate and multivariate analysis, predicted baseline RNFL was a bellwether of conversion to POAG during follow-up per 10µm thinner in predicted RNFL. Also in both analyses, baseline age, intraocular pressure, central corneal thickness, pattern standard deviation, mean deviation and cup-to-disc ratio (CDR) remained predictors of conversion. Finally, longitudinal change in predicted RNFL (per 1µm/year faster loss) was also found as a POAG conversion predictor.Because eyes that converted to POAG had predicted RNFL loss twice as fast as those that didn’t convert and that each 1µm/year faster decrease in predicted RNFL was linked with an almost six-fold increase in conversion risk, the authors relay that these findings suggest the deep learning model “offers a valuable enhancement to traditional risk factors, improving both baseline risk prediction and longitudinal monitoring of glaucoma progression.”1The authors also point to their model being a helpful surrogate measure useful in situations when OCT is unavailable, despite the model not fully being able to replicate all information seen in OCT scans.This sentiment is echoed in an invited commentary also published to the same journal as the original study—JAMA Ophthalmology—but adds some additional consideration. The author of the commentary highlights that while converting to an objective RNFL estimate addresses measurement variability, the value it adds for increasing predictive accuracy of POAG onset may be limited by the data presented in the research.What’s more, developed healthcare systems are already increasingly adding direct RNFL assessment via OCT with relatively reproducible measurements of the RNFL, making it so that the clinical imperative to predict RNFL thickness from fundus photography is diminished when OCT is available.However, the author does agree that despite OCT being widespread in certain areas, optic disc photography is more accessible and cost-effective, especially in community-based screening initiatives or environments with limited resources. As he explains, “in such settings, a model that enhances the informational yield from fundus photographs by providing a quantitative RNFL estimate could be beneficial.”2Click here for the study and here for the commentary. 1. Liu JC, Jammal AA, Scherer R, et al. Predicting retinal nerve fiber layer thickness from Ocular Hypertension Treatment Study optic disc photographs. JAMA Ophthalmol. June 26, 2025. [Epub ahead of print].2. Yohannan J. Optic disc photographs to RNFL thickness—from subjective to objective. JAMA Ophthalmol. June 26, 2025. [Epub ahead of print].