Skip to main content

  • Deep Learning Model Predicts Glaucomatous Changes

    By Lynda Seminara
    Selected By: Stephen D. McLeod, MD
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology, March 2021

    Download PDF

    Medeiros et al. developed and trained a deep learning algorithm to analyze fundus photographs and predict global retinal nerve fiber layer (RNFL) thickness from images obtained by spectral-domain optical coherence tomography (SD-OCT). Their model produced objective and quantitative estimates of glaucomatous changes.

    For this retrospective study, the authors used a 50% sample gathered from a large glaucoma registry that included patients with confirmed or suspected glaucoma. Participants had at least two longitudinal photographs from follow-up visits. Overall, there were 33,466 pairs of fundus photo­graphs and corresponding SD-OCT images, collected from 717 patients (1,147 eyes). Average follow-up per eye was 5.3 ± 3.3 years. The main outcome was the relationship between changes in RNFL predicted from photographs and the changes seen over time by SD-OCT.

    The mean global RNFL thickness estimated from the fundus photographs was 84.6 ± 14.4 μm, versus 84.5 ± 17.0 μm observed from corresponding SD-OCT scans, denoting a strong correlation between the two methods (R2 = 63.6%; p < .001). The average change in RNFL thickness identified by the algorithm throughout the follow-up period was –4.3 ± 5.8 μm, whereas that observed by SD-OCT was –4.8 ± 5.3 μm. There was a strong correlation between the RNFL and SD-OCT changes (r = 0.76; 95% CI, 0.70-0.80; p < .001).

    As this model provided objective and accurate estimates of RNFL thick­ness that correlated well with SD-OCT data, it may have a role in monitoring glaucoma progression, the authors said. They also noted its potential utility in settings where SD-OCT is not available or feasible.

    The original article can be found here.