Hybrid Network Model for the Prediction of Retinopathy of Prematurity from Neonatal fundus images

Retinopathy of Prematurity (ROP) is a vasoproliferative disorder influencing the retinal health of premature neonates with low birth weight. Delay in diagnosis of ROP can lead to permanent blindness of the preterm. This study aims to evaluate the efficiency of the features extracted from three pre-trained networks in the prediction of ROP. Examining the characteristics extracted from different layers of a CNN has the potential to provide deeper insights into the prediction of ROP. Three pre-trained networks such as ConvNeXt, VGG-16, and Inception V3 Convolutional Neural Networks (CNNs), are used to obtain the high-level features from the preterm fundus images and are classified using two methods. The first CNN based classifier, and the second is the Support Vector Machine (SVM). It is observed that the ConvNeXt pre-trained network with SVM classifier outperforms other networks with an 91.6% accuracy, 86.66% sensitivity, 96.66% specificity, F1 measure of 78.95%, and 0.92 as AUC. This proposed automated method for detecting ROP could provide auxiliary support to pediatric ophthalmologists, aiding them in making early treatment decisions to ensure optimal patient care.