Deep Learning Framework for Classification and Explanation of Early ROP Stages from HVDROPDB Dataset

Early detection and treatment of Retinopathy of prematurity (ROP) in its initial stages (stage 1/2/3) can prevent infant vision loss. The ROP stage is identified from the characteristics of a demarcation line in the retina. This paper introduces an innovative approach to classify fundus images into Normal, Stage 1, 2 or 3 categories. The classification is achieved by segmenting the demarcation line using an attention gate U-Net. Histograms extract the features of the demarcation line and the Random Forest (RF) algorithm selects the important features to train a deep neural network (DNN) for the classification. The stage is explained by locating the demarcation line on the fundus images. The results are at par when compared to the classification results of InceptionV3, Xception, and VGG19 networks using HVDROPDB datasets.