Monitoring response to treatment in severe retinopathy of prematurity using a deep learning based quantitative severity scale

Abstract

Purpose : To evaluate the clinical utility of quantitative image analysis using a deep learning plus disease severity score to monitor response to treatment in patients with type 1 (treatment-requiring) retinopathy of prematurity (ROP).

Methods : Images from clinical exams performed between July 2011 and December 2016 of infants in the multicenter Imaging and Informatics in ROP (i-ROP) study were reviewed to identify babies with type 1 ROP. Using the output of a fully automated deep learning algorithm trained to evaluate plus disease severity, we developed a continuous score (i-ROP score) from 1 (normal retinal vasculature) to 9 (severe plus disease). We examined the pre-treatment and post-treatment i-ROP score for the four weeks preceding and following treatment with either laser or intravitreal injections with anti-vascular endothelial growth factor (anti-VEGF). Eyes were only included if they had at least one clinical exam before and after treatment.

Results : A total of 43 eyes received treatment with either laser (n=34) or anti-VEGF therapy (n=9). The mean i-ROP score 2 weeks prior to treatment was 3.9, compared to 7.1 at time of treatment (p< 0.0001, student's t-test). Two weeks post treatment, the mean score decreased to 4.0 (p < 0.0001, student's t-test).

Conclusions : The i-ROP score, an automated ROP severity score developed using deep learning, correlates with observed clinical progression and regression of disease in ROP. Using this technology, it may be able to objectively monitor ROP response to treatment. Future work is needed to determine whether there may be observed differences in response to treatment between treatment groups, and whether this technology may identify disease recurrence requiring retreatment.

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