An efficient deep learning model for reliable detection and classification of retinopathy of prematurity

Abstract

Retinopathy of prematurity (ROP) is a leading cause of childhood blindness in very low birthweight and premature infants, and its timely diagnosis is hampered by the shortage of specialized clinicians. Recent advances in artificial intelligence have demonstrated expert-level performance in detecting severe ROP in retinal images, motivating the development of automated screening tools. This work introduces ROPDeepX, a hybrid convolution neural network (CNN) architecture that integrates ResNet50 and EfficientNet-B4 architectures to enhance ROP detection and severity classification. To address the scarcity of publicly accessible ROP datasets, three datasets were employed: the newly released FARFUM-RoP dataset, the Retinal Image Dataset of Infants and ROP (RIDIRP), and the HVDROPDB dataset.The FARFUM-RoP dataset served as the primary source for model training, while RIDIRP and HVDROPDB were used to validate cross-dataset robustness. VGG19, ResNet50, EfficientNet-B4, and ROPDeepX were evaluated on these datasets. ROPDeepX outperformed the pretrained models, achieving 96. 9% accuracy on the FARFUM-RoP dataset. ROPDeepX demonstrated superior results in different evaluation metrics, in both multiclass and binary classification tasks.