Quantitative analysis of aggressive posterior retinopathy of prematurity using deep learning

Abstract

Purpose : Aggressive posterior retinopathy of prematurity (APROP) is a severe form of ROP with a higher rate of progression to retinal detachment and permanent vision loss. The purpose of this study was to quantitatively characterize APROP using a deep learning (DL) based ROP severity score.

Methods : We conducted a retrospective analysis of the Imaging and Informatics in ROP (i-ROP) cohort from 7 North American centers, consisting of 1029 total patients and 7264 clinical eye exams. A reference standard diagnosis (RSD) was generated for each eye exam using methods previously published combining three independent image-based and one ophthalmoscopic gradings. Eyes without a history of ROP treatment were then categorized by RSD into five ROP severity groups: None, Mild, Type 2 or Pre-Plus, Treatment-Requiring (TR) without APROP, TR with APROP. All images were analyzed using a previously-published deep learning system (i-ROP DL) and assigned a vascular severity score from 1-9. Demographic data, the presence of systemic comorbidities, and the post-menstrual age at peak disease severity were evaluated for each category.

Results : The Table displays the results of the demographics of the 5 ROP severity categories. Infants with TR ROP who developed APROP tended to be more premature by birth weight (617 g vs 736 g, p>0.05) and gestational age (24.3 wks vs 25.5 wks, p<0.01) than those with only TR ROP without APROP. Additionally, these infants reached peak severity at earlier mean postmenstrual age (PMA) than those with only TR ROP without APROP (34.7 wks vs 39.0 wks, p<0.001; Table). The mean i-ROP vascular severity score correlated with the RSD-based ROP severity categories (Figure).

Conclusions : Premature infants in North America with APROP are born younger and develop disease earlier than infants with other categories of ROP, including other infants with treatment-requiring disease. Disease severity is quantifiable with i-ROP score, which correlates with clinically identified categories of disease including APROP. DL based ROP screening may identify these infants developing APROP prior to peak disease severity potentially allowing for earlier treatment in the future.

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

 

Table. Measures of prematurity show difference between TR without APROP and TR with APROP

 

Figure. Vascular severity score correlates with disease severity (median i-ROP with interquartile ranges)