EyeScreen: Development and Potential of a Novel Machine Learning Application to Detect Leukocoria

Alec Bernard; Shang Zhou Xia; Sahal Saleh; Tochukwu Ndukwe; Joshua Meyer; Elliot Soloway; Mandefro Sintayehu; Blen Teshome Ramet; Bezawit Tadegegne; Christine Nelson; Hakan Demirci

doi:10.1016/j.xops.2022.100158

EyeScreen: Development and Potential of a Novel Machine Learning Application to Detect Leukocoria

Ophthalmol Sci. 2022 Apr 15;2(3):100158. doi: 10.1016/j.xops.2022.100158. eCollection 2022 Sep.

Authors

Affiliations

¹ School of Medicine, University of Michigan, Ann Arbor, Michigan.
² Department of Electrical Engineering & Computer Science, College of Engineering, University of Michigan, Ann Arbor, Michigan.
³ Department of Ophthalmology, St. Paul's Hospital, Addis Ababa, Ethiopia.
⁴ Department of Pediatrics, St. Paul's Hospital, Addis Ababa, Ethiopia.
⁵ Kellogg Eye Center, Department of Ophthalmology and Visual Sciences, University of Michigan Ann Arbor, Michigan.

Abstract

Purpose: Early diagnosis and treatment of retinoblastoma are of paramount importance for a positive clinical outcome. The most common sign of retinoblastoma is leukocoria, or white pupil. Effective, easy-to-perform, community-based screening is needed to improve outcomes in lower-income regions. The EyeScreen (developed by Joshua Meyer from the University of Michigan) Android (Google LLC) smartphone application is an important step toward addressing this need. The purpose of this study was to examine the potential of the novel use of low-cost technologies-a cell phone application and machine learning-to identify leukocoria.

Design: A cell phone application was developed and refined with the feedback from on-site, single-population use in Ethiopia. Application performance was evaluated in this technology validation study.

Participants: One thousand four hundred fifty-seven participants were recruited from ophthalmology and pediatric clinics in Addis Ababa, Ethiopia.

Methods: Photographs obtained with inexpensive Android smartphones running the EyeScreen Application were used to train an ImageNet (ResNet) machine learning model and to measure the performance of the app. Eighty percent of the images were used in training the model, and 20% were reserved for testing.

Main outcome measures: Performance of the model was measured in terms of sensitivity, specificity, receiver operating characteristic (ROC) curve, and precision-recall curve.

Results: Analyses of the participant images resulted in the following at the participant level: sensitivity, 87%; specificity, 73%; area under the ROC curve, 0.93; and area under the precision-recall curve, 0.77.

Conclusions: EyeScreen has the potential to serve as an effective screening tool in the areas of the world most affected by delayed retinoblastoma diagnosis. The relatively high initial performance of the machine learning model with small training datasets in this early-phase study can serve as a proof of concept for future use of machine learning and artificial intelligence in ophthalmic applications.

Keywords: Application; Machine learning; ROC, receiver operating characteristic; Retinoblastoma; Screening.