Diabetic retinopathy (DR) is a diabetes-induced retinal disease that damages blood vessels, leading to blurry vision and, in severe cases, blindness. Lesions such as microaneurysms, hemorrhages, and exudates are key indicators of DR progression, and their early detection is essential to prevent irreversible vision loss. This study presents a Computer-Aided Diagnostic System for Diabetic Retinopathy (CAD-DR) that integrates a U-shaped convolutional neural network (U-Net) for lesion segmentation with pre-trained Visual Geometry Group (VGG) models for classification. The system uses the APTOS 2019 Blindness Detection dataset of retinal fundus images, with preprocessing enhanced by the Frangi vesselness filter to improve vessel and lesion visibility. Generative Adversarial Network (GAN)-based augmentation was applied to address class imbalance. Experimental results show that VGG19 with GAN augmentation achieved the highest classification accuracy of 89%, with a precision of 87%, recall of 90%, and F1-score of 88.5%, outperforming models without augmentation. The results demonstrate the effectiveness of combining segmentation, vessel enhancement, and GAN augmentation for reliable DR screening.