Attention U-Net and CNN-Based Deep Learning Framework for Diabetic Retinopathy Severity Prediction and Retinal Vessel Segmentation

Diabetic Retinopathy (DR), affecting approximately 40% of diabetic patients globally, manifests through progressive deterioration of retinal blood vessels, leading to vision loss. Although early detection through automated fundus image analysis is crucial for therapeutic intervention, conventional methodologies face significant challenges in lesion identification and segmentation accuracy. This study proposes a deep learning framework integrating an Attention U-Net architecture for vessel segmentation with a Convolutional Neural Network (CNN) for DR classification. This study employed the DRIVE dataset to train the Attention U-Net model, with the resultant segmentation outputs serving as inputs for the CNN-based classification system. The classification model training was conducted using the LVPEI dataset, encompassing 426 fundus images stratified across five DR severity grades: No DR, Mild, Moderate, Severe, and Proliferative DR. The proposed CNN model exhibited robust performance metrics, achieving an accuracy of 90%, an F1-score of 88%, a precision of 87%, and a recall of 89%, surpassing previous benchmark studies in DR classification. The experimental results demonstrate the effectiveness of our methodology in achieving precise retinal structure segmentation and DR severity classification. This integrated approach demonstrates significant potential as a robust tool for automated DR screening and longitudinal disease progression monitoring in clinical settings.