Deep Learning Algorithm Predicts Glaucoma Based on Fundus Photography

View inside human eye disorders – showing retina, optic nerve and macula .
The technology has the potential to help steer treatment decisions.

Longitudinal changes in the optical coherence tomography (OCT)-trained deep learning algorithm’s predictions of retinal nerve fiber layer (RNFL) thickness measurements based on fundus photographs can predict risk of future glaucomatous conversion in suspect eyes, researchers found in a study published in American Journal of Ophthalmology. The technology synthesized the information from the photos into predictions of future visual field defects.

“This method can be potentially used to develop a cheap and accessible glaucoma screening method as well as to monitor risk of glaucomatous conversion in eyes suspected of the disease,” the study says. With the ability to obtain fundus photos remotely, this predictive technology could be applied in settings where OCT imaging is impractical, according to the researchers.  

The investigators reviewed data from glaucoma suspects enrolled in the Duke Glaucoma Registry (DGR). All eyes had normal standard automated perimetry (SAP) tests at baseline. Follow-ups with each patient included taking additional SAP and fundus photographs, until development of glaucomatous visual field defects. The researchers only analyzed photographs taken before confirmed glaucomatous conversion. Eyes that had at least 2 consecutive abnormal visual field tests during follow-up were labeled “conversors” (indicating glaucomatous conversion). The investigators trained and validated the algorithm to obtain RNFL thickness predictions for each fundus photograph from the study cohort.

The study included 1072 eyes of 827 glaucoma suspect patients. A total of 196 eyes converted to glaucoma during follow-up, with a mean time of approximately 4.4 years among the conversors and 6.3 years for the non-conversors. The mean baseline-predicted RNFL thickness from fundus photos was approximately 88.7 µm for conversors and 92.1 µm for non-conversors (P <.001).

Baseline and rate of change of predicted RNFL thickness significantly predicted conversion to glaucoma, according to the investigation.

Each 10 µm thinner predicted baseline RNFL thickness carried a hazard ratio (HR) of 1.71 (95% CI: 1.48 to 1.98; P <.001) and each 1 µm/year faster decrease in predicted RNFL thickness was associated with about twice the hazard of developing visual field loss (HR: 2.10; 95% CI: 1.44 to 3.06; P <.001). The univariable model had a modified R2 index of 34% (95% CI: 25% to 44%).

Adjusting for baseline age, sex, race, central corneal thickness (CCT), baseline pattern standard deviation (PSD), and mean intraocular pressure (IOP), each 10 µm thinner predicted baseline RNFL thickness carried a 56% increase in HR (95% CI: 1.33 to 1.82; P <0.001), and each 1 µm/year faster decrease in predicted RNFL thickness was associated with a 99% increase in hazard (HR: 1.99; 95% CI: 1.36 to 2.93; P <0.001). The multivariable model had a modified R2 index of 51% (95% CI: 41% to 62%).

While the algorithm’s predictions from fundus photographs have previously correlated well with actual OCT measurements, the value in predicting glaucomatous damage before visual field defects appear has not yet been established, the researchers said.

Limitations of the study included the lack of specificity and control over the treatment regimen of the cohort and use of different cameras to acquire photos.

Disclosure: One study author declared affiliations with the biotech or pharmaceutical industry. Please see the original reference for a full list of authors’ disclosures.


Lee T, Jammal AA, Mariottoni EB, Medeiros FA. Predicting glaucoma development with longitudinal deep learning predictions from fundus photographs. Am J Ophthalmol. Published online January 7, 2021. doi:10.1016/j.ajo.2020.12.031