Automated Computer-Based Image Analysis in Monitoring Disease Progression for Retinopathy of Prematurity

Abstract

Purpose : To apply an automated deep learning algorithm in objectively monitoring retinopathy of prematurity (ROP) progression using a large cohort of babies born prematurely.

Methods : Images from clinical exams performed between July 2011 and December 2016 of infants in the multicenter Imaging & Informatics in ROP (i-ROP) study were reviewed. A previously validated computer-based deep learning algorithm was used to grade the exam images and a continuous score (i-ROP score) from 1 (normal retinal vasculature) to 9 (severe plus disease) was assigned to each exam. Infants were subdivided using a reference standard diagnosis based on clinical exam and expert grading of clinical images into groupings based on need for treatment. The disease course was then assessed between the two groups longitudinally across multiple exams until the infants were lost to follow-up, the retina fully vascularized, or treatment-requiring ROP developed.

Results : A total of 5255 clinical examinations in 871 babies (1692 eyes) were analyzed. 91 eyes (5.4% of the cohort) progressed to treatment-requiring disease, which occurred at an average i-ROP grade of 5.80 and 37.6 weeks post-menstrual age (PMA). The mean gestational age and weight at birth of those that required treatment was 25.1 weeks and 683 grams, in contrast to 27.3 weeks and 966 grams in those that did not require treatment (p<0.05 for both comparisons). Average i-ROP scores of the two groups were significantly different at all time-points analyzed but became more apparent with advancing PMA. At 36-38 week exams, the average i-ROP score of infants developing treatment-requiring disease was 5.22 compared to 1.20 in those that received no treatment (p<0.05). When assessing the rate of change in i-ROP scores over time, eyes that developed treatment-requiring disease had a faster average increase in i-ROP score at a given time (0.4 – 1.32 i-ROP points/week) than untreated eyes (0.07 – 0.20 i-ROP points/week) (p<0.05).

Conclusions : The i-ROP score, an automated algorithm using deep learning, effectively correlates with observed clinical progression of ROP in this cohort of infants born prematurely. Going forward, automated computer-based image analysis may be considered as a means to monitor disease progression in infants undergoing screening for ROP.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.