DeepROP: Automated Stage-Wise Classification of Retinopathy of Prematurity Using Enhanced Retinal Fundus Imaging

DeepROP presents a hybrid deep learning framework designed for automated stage-wise classification of Retinopathy of Prematurity (ROP) using enhanced retinal fundus imaging. ROP is a potentially blinding condition affecting premature infants, and its diagnosis often relies on subjective expert interpretation of retinal features. To address this challenge, DeepROP integrates the strengths of EfficientNetB4 and Swin Transformer models to extract both local fine-grained features and global structural patterns critical for ROP staging. The model is trained on the clinically annotated “ROP Stages Reviewed” dataset, with rigorous preprocessing, data augmentation, and stratified sampling techniques to ensure class balance and generalizability. Our system achieves a classification accuracy of 96.5 %, outperforming existing models such as ResNet-50, VGG-16, and InceptionV3. Visual explanations via Grad-CAM enhance interpretability, making DeepROP a reliable assistive diagnostic tool. This research highlights the potential of AI-driven solutions in improving early ROP detection, reducing diagnostic variability, and supporting timely clinical interventions in resource-constrained environments.