Purchase this article with an account.
Haslina Hamzah, Gilbert Lim, Quang duc nguyen, Baskaran Mani, Wynne Hsu, Mong Li Lee, Ching-Yu Cheng, Tien Y Wong, Daniel Ting; Artificial Intelligence using Deep Learning System for Glaucoma Suspect Detection. Invest. Ophthalmol. Vis. Sci. 2018;59(9):4074.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deep learning system (DLS) is a novel machine learning technology to screen for glaucoma suspect (GS), a major cause of blindness
We evaluated the diagnostic performance of a DLS for GS using multi-ethnic cohorts with 162,494 retinal images. We trained the DLS using 125,189 images among diabetes patients attending a Singapore national DR screening program (SiDRP), 3 population-based studies and 1 clinic-based cohorts and validated it using 7,326 images of patients from a population-based study with reference to glaucoma specialists and professional graders’ grading. We calculated area under curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Cohen’s Kappa, and repeatability (same image tested twice) of DLS in detecting GS (vertical cup to disc ratio of 0.8 or greater and other glaucomatous disc features).
Using the clinical validation datasets (7,326 images, 5,090 eyes, 2,575 patients), the AUC was 0.949 (95%CI 0.942, 0.957). The sensitivity and specificity were 90.3% (88.5-91.9) and 83.0% (81.8, 84.2), respectively. The PPV and NPV were 62.4 (60.1, 64.7) and 96.5 (95.8, 97.1), respectively. Both DLS and human graders had moderate inter-rater agreement of 0.64 (0.61, 0.66), with a repeatability of 100%.
Deep learning technology had clinically acceptable diagnostic performance for GS screening in multi-ethnic populations, highlighting its significant public health potential worldwide.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only