Can deep machine learning accurately diagnose retinopathy of prematurity?

A novel deep machine learning algorithm for telemedicine screening was able to accurately detect zone and different stages of retinopathy of prematurity (ROP), according to a study.

Researcher prospectively obtained 1400 consecutive wide-field retinal images of 111 prematurely-born infants, of which 69% (n = 966) had some stage of ROP>

Novel computer-aided approaches were verified for traceability and clinical accuracy using a confusion matrix analysis. Scores for vessel detection were 98.5% for accuracy, 92.9% for sensitivity, 98.6% for specificity, and 74% for F1. ROP stage detection included global accuracy of 98.3%, sensitivity of 99.3%, precision of 98.9%, and F1-score of 99.1%.

Zone detection and detection of plus and pre-plus disease was done with image processing resulting in 95% accuracy.

Reference
Punjabi OS, et al. Retinopathy of prematurity screening using a novel method of advanced image processing and deep machine learning. Presented at: 2020 ASRS Virtual Meeting.