Evaluation of an Artificial Intelligence System (i-ROP DL) for Retinopathy of Prematurity Screening in Nepal using the Forus 3nethra neo and in Mongolia using the Retcam Portable^®

Abstract

Purpose : Our group has demonstrated the performance of the i-ROP Deep Learning (DL) system for retinopathy of prematurity (ROP) telemedicine screening programs in Nepal and Mongolia. The purpose of this study is to evaluate the performance of i-ROP DL in Nepal and Mongolia when increasing the number of infants included and conducting further analysis by country and fundus camera.

Methods : This retrospective study evaluated prospectively collected data from ROP screening programs in Nepal and Mongolia. Birth weight, gestational age, and ROP severity based on the International Classification of ROP (ICROP) guidelines were recorded. Fundus images were obtained with the Forus 3nethra neo in Nepal and the RetCam^® Portable in Mongolia. The i-ROP DL system, previously trained on RetCam^® images, was used to identify posterior pole images in the dataset and subsequently generated a mean vascular severity score (1-9). Analysis of variance was used to compare vascular severity scores to ROP categories by country. A p value ≤ 0.05 was considered statistically significant. Stata MP 13, SAS, and R statistical software programs were used.

Results : There were a total of 377 patients and 860 exams in Nepal and 321 patients and 917 exams in Mongolia. Average birth weight and gestational age were lower in Mongolia compared to Nepal (both p < 0.001). Overall disease prevalence of treatment-requiring ROP was 14% in Mongolia vs 2% in Nepal (p < 0.001). See Table 1. The ROP vascular severity score was higher in Mongolia compared to Nepal at the population level (3.5 vs. 2.5, p < 0.002). Figure 1 displays the vascular severity scores overall and by category for Nepal and Mongolia.

Conclusions : In Mongolia and Nepal, the AI-generated vascular severity scores correspond to ICROP disease severity using two different camera systems, despite the i-ROP DL system not being previously trained on Forus 3nethra neo images. Vascular severity at the population level may be a useful tool for epidemiologic assessment of geographic and temporal variations in ROP severity.

This is a 2021 ARVO Annual Meeting abstract.

 

Demographic data and rates of ROP severity by category in Nepal and Mongolia.

 

Retinopathy of prematurity (ROP) vascular severity scores in Nepal and Mongolia. The scores are presented for overall ROP severity and by category. Asterisks denote higher vascular severity scores.