Utility of an automated deep learning tool in a low-income country for retinopathy of prematurity

Abstract

Purpose : Retinopathy of prematurity (ROP) is the leading cause of preventable blindness worldwide especially in low- and middle-income countries (LMICs) where challenges in ROP include lack of resources and specialists. Artificial intelligence (AI), shown to be reliable and accurate in ROP, has the potential to be utilized as a tool in LMICs. The purpose of this preliminary retrospective study is to evaluate the utility of an automated deep learning tool (i-ROP DL) for ROP in a LMIC.

Methods : Clinical data and fundus images were obtained from Nepal as part of an ROP screening program. An average of 5 images were obtained per eye per examination using the Forus 3nethra neo. Images were excluded if there was an incomplete clinical diagnosis, no corollary clinic visit, or discrepancy in the image file identifier or demographic data. The i-ROP DL system was used to analyze each image set and a quantitative severity score on a scale from 1 through 9 was generated. Indirect ophthalmoscopy was performed by local ROP specialists and ROP was classified as none, mild, type 2, type 1, or post-treatment ROP. Means and frequencies were calculated. Correlation of i-ROP score vs clinical disease severity and sensitivity/specificity analyses were performed.

Results : There were a total of 440 infants in the study with 15467 eye examinations. Birth weight and gestational age of infants screened ranged from 500 to 4000 grams and 23 to 40 weeks, respectively. I-ROP severity score ranged from 1.0 to 8.93. Median i-ROP score was 2.25 (Interquartile range [IQR] 1.32-4.00). Median i-ROP score for the post-treatment category was 3.71 (IQR 2.09-5.39). For Type 1 ROP, median i-ROP score was 5.76 (IQR 4.07-7.74). For Type 2 ROP, median i-ROP score was 3.54 (IQR 1.93-5.13). For mild ROP, median i-ROP score was 3.21 (IQR 1.79-4.47). For no ROP, median i-ROP score was 2.03 (IQR 1.27-3.60). Figure 1 displays a graphical distribution of i-ROP score and severity of disease. At a cutoff score of 3, there was 86% sensitivity and 66% specificity for treatment-requiring ROP (TR-ROP). Area under the curve was 0.86 for TR-ROP.

Conclusions : This novel study demonstrates a proof of concept that the i-ROP score can potentially quantify severity of ROP based on Forus 3nethra neo fundus images in Nepal. Future analyses are required to further assess image quality and discrepancies which could improve the results to be considered for clinical use.

This is a 2020 ARVO Annual Meeting abstract.