Temporal Progression Analysis of Diabetic Retinopathy Using Recurrent-CNN Architectures

Abstract

Diabetic Retinopathy (DR) is a progressive eye disease that requires timely and accurate detection to prevent vision impairment. While convolutional neural networks (CNNs) have shown high efficacy in detecting DR from retinal images, they often fall short in capturing temporal changes across longitudinal patient data. This research proposes a hybrid deep learning framework integrating Recurrent Neural Networks (RNNs) with CNNs—termed Recurrent-CNN (R-CNN)—to analyze the temporal progression of DR. The model leverages sequential retinal images and clinical metadata to model disease evolution over time, enabling more granular stage prediction. We train and validate our approach on publicly available and proprietary longitudinal DR datasets, achieving notable improvements in progression prediction accuracy and temporal consistency compared to baseline CNN models. Our findings suggest that incorporating temporal dynamics significantly enhances the interpretability and clinical relevance of DR grading systems, providing a robust tool for ophthalmologists in proactive patient management.