In this article a new hybrid deep learning system is proposed which incorporates both a Generative Adversarial Network (GAN) and a Convolutional Neural Network (CNN) to automatically detect and classify the severity of Diabetic Retinopathy (DR) from fundus images. This technique aims to facilitate large-scale clinical screening, by solving problems related to class imbalance and scarcity of medical data. To address these challenges, a Conditional GAN (cGAN) is used to generate realistic fundus images of all five DR severity grades, thus expanding the training set. This expanded dataset is then used to train a ResNet-50 CNN model, which conducts end-to-end multi-class classification and automatically learns discriminative features, avoiding the need for extraction of lesions. The proposed system is trained and evaluated on benchmark fundus image datasets and exhibits good generalization. The experimental results show the model achieves a classification accuracy of 95.2% with a cross-entropy loss of 0.48 and surpasses the state-of-the-art by 15%. The system also exhibits higher sensitivity to under-sampled severe stages of DR and is computationally efficient. In conclusion, the hybrid model provides a robust, scalable, and efficient solution for automatic DR diagnosis and has the potential to be deployed for clinical use and when compared with existing classifier’s like SVM and DT in which the proposed classifier gave the best accuracy.