Deep Learning for Classifying Stages of Retinopathy of Prematurity

This work presents a deep learning based approach to classify stages of retinopathy of prematurity (ROP). In particular, the approach focuses on delineating stages 1, 2 and 3 that are relatively crucial for treatment. Training and validation of the models is carried out on multiple subsets of data from 82,295 images collected by Narayana Nethralaya, Bengaluru, using 3nethra fundus cameras. A pre-trained EfficientNetV2S model is used for feature extraction and the datasets are used to train classification models to predict stage, plus disease status and decision for treatment. Experimental results demonstrate the efficacy of training models only on images capturing the temporal region of the retina and manually filtering out images that show no visible signs of ROP, with the best performing model distinguishing stage 2 and stage 3 with an accuracy of 93.44
^%
, AUC of 98.71% and sensitivity of 93.30%.