Deep Learning for ROP

Researchers at University College London and Moorfields Eye Hospital have developed new deep learning (DL) models for the accurate detection of retinopathy of prematurity (ROP) (1). ROP is increasing in prevalence as global survival rates for premature babies increase, but the study authors are optimistic that these models could lead to worldwide improvements in detecting the disease, making ROP – currently a leading cause of childhood blindness – “a thing of the past.”

Specifically, two DL models – one bespoke, one code-free deep learning (CFDL) – were developed by the UCL-Moorfields team to differentiate between “healthy,” “pre-plus,” and “plus disease” – the latter a hallmark of severe ROP characterized by abnormal posterior retinal vessel dilatation and tortuosity. The team used the DL tools to retrospectively screen a sample of 7,414 images taken from 1,370 newborns previously admitted to Homerton Hospital, London, between 2008 and 2018. The hospital is known to serve an ethnically and socioeconomically diverse community – an important side note, when considering that ROP is said to be influenced by both ethnicity and socioeconomic factors (2).

The AI results were then compared with results taken from assessments by ophthalmologists (each image was graded by two junior ophthalmologists with all disagreements adjudicated by a senior pediatric ophthalmologist). Comparisons indicated that both types of deep learning models performed at a similar diagnostic level to that of ophthalmologists when discriminating between healthy retinal images and those with either pre-plus or plus disease features. The team did note, however, that the CFDL model did not perform as well as the bespoke model when it came to pre-plus disease in minority classes.

Despite this downside, the CFDL model can be easily optimized by individuals without prior coding experience. In either case, the UCL-Moorfields work indicates that DL models show promise as an alternative to manual review of scans, which could free up ophthalmologists’ time by allowing trained nurses to initiate the diagnostic process.