Deep Learning-Based Detection of Papilledema on Retinal Photographs From Handheld Cameras: A Prospective Study.

Abstract

Background: Papilledema and other optic neuropathies are critical findings in neuro-ophthalmology that require timely and accurate diagnosis. This study evaluates the performance of a deep learning system (DLS) to identify papilledema and other optic neuropathies, when applied to a large dataset of retinal photographs acquired prospectively with a handheld nonmydriatic camera in a neuro-ophthalmology department.

Methods: International multiethnic, multicenter study including 20,533 retinal photographs (10,647 patients) from 31 centers worldwide. The training and internal validation datasets consisted of 18,981 mydriatic retinal photographs (9,830 patients) from 22 countries. The external testing dataset included 1,552 prospectively collected retinal photographs (817 patients) acquired with a handheld camera (Aurora, Optomed, Finland). The DLS segmented the optic disc and peripapillary region, then classified each photograph as 1/"papilledema," 2/"other" optic disc abnormalities (i.e., swelling because of other causes, atrophy, etc.), or 3/"normal." The performance of the DLS in classifying optic disc appearance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Diagnostic outcomes were evaluated at the eye level and at the patient level.

Results: The DLS achieved an accuracy of 99.5% (95% confidence interval [CI], 99.1-99.8), a sensitivity of 81.0% (95% CI, 64.1-97.7), a specificity of 99.7% (95% CI, 99.5-99.9), and an AUC of 98.3% (95% CI, 97.6-99.9) for differentiating papilledema from "others" and healthy controls.

Conclusions: A DLS trained on a large dataset of mydriatic photographs achieved excellent diagnostic performance for detection of papilledema and other optic disc abnormalities when applied to nonmydriatic retinal photographs acquired with a handheld camera in real life conditions.