Accepted for/Published in: JMIR Medical Informatics
Date Submitted: Sep 12, 2019
Date Accepted: Mar 20, 2020
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
A deep learning-based prediction of refractive error using photorefraction images captured by a smartphone for children: Model Development and Validation Study.
ABSTRACT
Background:
The prediction of refractive error for children is crucial to detect amblyopia due to refractive error, which leads to permanent visual impairment and is potentially curable if detected early. Therefore, various equipment has been adopted to screen patients and detect risks for amblyopia more easily and widely. For efficient screening, the easy accessibility of the screening tool and the accuracy of a prediction algorithm are most important.
Objective:
This study aims to develop an automated deep learning-based system to predict a range of refractive error in children with eccentric photorefraction images captured by a smartphone, which has easy accessibility and high accuracy.
Methods:
We collected eccentric photorefraction images using a smartphone (LGM-X800K, LG Electronics Inc., Seoul, Korea) at a 1-meter distance from children (mean age of 4.32±1.87) in a dark room without cycloplegia. The acquired 305 photorefraction images were divided into seven classes according to the spherical values measured by the cycloplegic refraction (≤-5.0 D, > -5.0 and ≤ -3.0 D, >-3.0 and ≤ -0.5 D, > -0.5 and < +0.5 D, ≥+0.5 and < +3.0 D, ≥+3.0 and <+5.0 D, and ≥+5.0 D). For our deep learning model, we designed and trained a Residual Network to classify photorefraction images into the most probable class of refractive error using photorefraction images.
Results:
We used the 5-fold cross validation to evaluate the performance of our deep learning model. Our deep learning models resulted in the overall accuracy of 81.6% (249/305) and the following accuracy for each class of refractive error: 80.0% in ≤ -5.0 diopters (D), 77.8% in > -5.0 and ≤ -3.0 D, 82.0% in >-3.0 and ≤ -0.5 D, 83.3% in > -0.5 and < +0.5 D, 82.8% in ≥+0.5 and < +3.0 D, 79.3% in ≥+3.0 and <+5.0 D, and 75.0% in ≥ +5.0 D.
Conclusions:
The result demonstrated that our deep learning-based system achieved considerably accurate performance despite the insufficient size of the photorefraction image dataset. The current study showed the possibility of developing precise and accessible smartphone-based prediction systems for refractive error using deep learning with a further collection of pediatric photorefraction images.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.