Utilization of a Deep Learning Image Assessment Tool for Epidemiologic Surveillance of Retinopathy of Prematurity

Abstract

Purpose : Primary and secondary prevention for retinopathy of prematurity (ROP) are challenging in low-and middle-income countries (LMIC) like India. Primary prevention through strict oxygen regulation is known to vary between hospitals and there are no objective methods of monitoring these differences. In this study, we used a deep learning-based ROP severity score to evaluate whether differences in disease severity related to primary prevention could be identified among hospitals in the Aravind system.

Methods : As part of the ROP Eradication Save Our Sight (ROPE-SOS) program, fundus photographs were evaluated from hospitals in the ROPE-SOS network from August 2015 to October 2017. Using deep learning image analysis methods previously published, the Imaging and Informatics in ROP (i-ROP) severity score was calculated for each image on a 1-9 scale. Images without an optic nerve and of inadequate quality were automatically excluded by the i-ROP system from analysis, as were hospitals with <5 examinations in the database. The overall mean scores per infant at initial screening examination were then compared across different hospital systems using ANOVA and multivariate linear regression.

Results : 219 eye examinations from 137 infants in 14 hospital systems were analyzed (568 images). 72 subjects were male (53%), with mean±SD postnatal age of 31±19 days (range 7 to 100), birth weight 1478 ± 401g (range 770 to 3200g), and gestational age 32±2 weeks (range 25 to 37 weeks). The overall mean±SD iROP score was 2.9±1.8 (range 1.0 to 8.6). Mean iROP scores are shown in the Table, and were significantly different between hospitals (ANOVA, p<0.0001). Specifically, after controlling for postnatal age, gestational age, and birth weight, the mean iROP score at Hospital 15 was 1.7 higher than other hospitals (multivariate linear regression, p=0.04). Of note, this hospital has been independently identified by the telemedicine reader as an outlier.

Conclusions : In this study, an objective deep learning-based ROP severity score for epidemiologic surveillance of ROP identified statistically significant differences in disease burden at different hospitals within the Aravind system, independent of known risk factors for ROP. This has potential utility for disease surveillance evaluating differences in primary prevention between hospitals in LMIC.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.