Accessing Refractive Errors via Eccentric Infrared Photorefraction Based on Deep Learning

Abstract

Eccentric infrared photorefraction is an attractive vision screening method which is widely used for uncooperative subjects, such as infants and toddlers. Unlike conventional slope-based photorefraction, a deep neural network is used to predict refractive error in this study. Total 1216 ocular image were collected by a homemade photorefraction device, whose corresponding refractive error was measured by a commercial autorefractor device, to create a series of dataset for our deep neural network. The mean squared error of the preliminary result is +/- 0.9 diopter, which indicates its feasibility and can be improved with bigger database.