Skip to main content

  • Deep Learning Can Predict Glaucoma Before Its Onset

    By Lynda Seminara
    Selected By: Henry D. Jampel, MD, MHS
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology Glaucoma, July/August 2020

    Download PDF

    Deep learning (DL) has shown promise for automated assessment of glaucoma from fundus photographs after disease onset. Thakur et al. tackled a more challenging question: Does the integra­tion of DL models into portable fundus cameras help identify glaucoma before its onset? They found that these models consistently predicted the disease sev­eral years before clinical manifestations were apparent.

    This prospective longitudinal study included 66,721 fundus photographs of 1,636 participants (3,272 eyes) of the prospective multicenter Ocular Hypertension Treatment Study (OHTS). At baseline, patients had a normal-appearing optic disc and normal visual field. Ocular measurements and fundus photographs were collected annually for 16 years during the OHTS and were examined by two independent readers. Any observed abnormalities prompted retesting and confirmation by an end-point committee.

    Using these photographs, the authors generated datasets to develop three DL models. The first classified the images as glaucomatous or nonglaucomatous according to gradient-weighted class activation maps. The other two models were trained via transfer learning to predict glaucoma in two time periods before disease onset. The models were validated using 85% of the fundus photographs and were retested on the remaining 15%. Primary outcome measures were accuracy and area under the receiver-operating characteristic curve (AUC).

    At study end, the AUC of the DL model for diagnosing glaucoma was 0.95. The AUC for predicting glaucoma development one to three years prior to onset was 0.88; that for predicting it four to seven years beforehand was 0.77.

    These findings suggest that DL mod­els are sensitive enough to identify pre­clinical signs of glaucoma from baseline fundus photographs, thus offering a simple, inexpensive, portable screening method to complement routine as­sessments. The authors cautioned that the models were less accurate for eyes without apparent glaucomatous optic neuropathy.

    The original article can be found here.