Deep learning algorithm's prediction of RNFL thickness gauges risk for glaucoma conversion
An OCT-trained deep learning algorithm’s predictions of retinal nerve fiber layer thicknesses from fundus photographs effectively predicted future development of visual field defects in eyes of glaucoma suspects.
Researchers in a retrospective cohort study used a machine-to-machine (M2M) OCT-trained deep learning algorithm to predict retinal nerve fiber layer (RNFL) thickness measurements from fundus photographs. The model was applied to fundus photographs of glaucoma suspect eyes followed over time. Researchers investigated the efficacy of using the M2M-predicted RNFL thicknesses to detect glaucomatous damage before the appearance of visual field defects and evaluated the deep learning predictions of RNFL thickness and longitudinal changes to predict the risk for glaucoma conversion.
“An AI algorithm can be trained to predict accurate RNFL thickness measurements from simple color fundus photographs, and such measurements are predictive of future visual field defects in glaucoma suspects. As such, these measurements could help clinicians decide which suspects are at higher risk for developing glaucoma and may require treatment or closer monitoring,” study co-author Felipe A. Medeiros, MD, PhD, told Healio/OSN.
Previous research relied on subjective assessments of fundus photographs by clinicians, which have low reliability. The AI tool developed in this study provides an objective method to quantify glaucomatous damage on fundus photos, Medeiros said.
The study included 1,072 eyes of 827 glaucoma suspects, with a mean 4.2 fundus photographs per eye available during follow-up. A total of 196 eyes (18%) converted to glaucoma during follow-up (converters). A mean 9.8 and 9.6 visual fields were tested per eye during follow-up in converters and non-converters, respectively.
The M2M mean baseline predicted RNFL thickness from fundus photographs was 88.7 µm for converters and 92.1 µm for non-converters. The mean rate of change for predicted RNFL thickness was faster for converters at –1.02 µm per year compared with –0.67 µm per year for non-converters, which was a statistically significant difference (P < .001).
A univariable analysis showed a 71% increase in the hazard of developing visual field defects for each 10 µm thinner predicted baseline RNFL thickness. Additionally, each 1 µm per year faster decrease in predicted RNFL thickness was associated with about a twofold increase in developing visual field loss (P < .001).
Researchers found that eyes were significantly more likely to convert to glaucoma if they had lower M2M predictions of RNFL thickness at baseline and faster declines in thickness over time.
“The performance was very similar to that of measurements obtained by much more sophisticated devices such as OCT. Therefore, the use of AI enables one to get predictions from photographs which rival those obtained from OCT in terms of detecting glaucomatous damage in suspects. This is an outstanding result given that photographs are a much cheaper and easily available test,” Medeiros said.
This method holds promise in population-based screening for glaucoma, as well as opportunistic screening in a primary care setting. Ophthalmologists who do not have access to OCT instruments may also find the method to be of value, Medeiros said.
Despite its accuracy, this method based on simple fundus photographs should not be expected to replace OCT. However, in the absence of OCT or when its use is impractical, the method could achieve great results, he said.
“It is important to note, however, that although the performance of the AI method was very promising, the method still needs validation in other populations,” Medeiros said.