Diabetic Retinopathy Detection with Deep Learning: A ResNet-CNN Model Enhanced by Attention Mechanism and Ensemble Learning

Diabetic retinopathy (DR), a leading cause of vision impairment worldwide, primarily impacts individuals with diabetes, making early detection vital to prevent irreversible vision loss. Leveraging deep learning, particularly Convolutional Neural Networks (CNNs), has become instrumental in the automated analysis of retinal fundus images for DR detection. This study reviews recent advancements in CNN-based DR detection, focusing on techniques like ensemble learning and attention mechanisms that improve model accuracy and interpretability. Despite significant progress in classifying DR stages, challenges remain around data imbalance, image quality variation, and the need for model transparency in clinical settings. Using the APTOS 2019 Blindness Detection dataset, which includes diverse, labeled retinal images, we train, test, and benchmark deep learning models under standardized conditions, employing Python and TensorFlow for model development. Additionally, architectures like ResNet, and attention-based models are explored to enhance lesion focus, with ensemble methods employed to boost predictive accuracy. Results demonstrate improved model interpretability and robust DR detection, highlighting deep learning’s potential for clinical use and suggesting future directions, such as integration with electronic health records (EHR) and mobile-based applications.