Hybrid Deep Learning Framework for Diabetic Retinopathy Detection using Generative Adversarial Networks and Transfer Learning

Main Article Content

Dr. T. V. Hyma Lakshmi, P. Hema Sree, Hasti Venkata Subbaiah, K Punnam Chandar, Gangu Rama Naidu , Yogendra Narayan

Abstract

                        Diabetic retinopathy (DR) is a leading cause of vision loss worldwide, making early detection critical for preventing severe visual impairment. Traditional methods for DR diagnosis are limited by the availability of specialists and the time-consuming nature of manual screenings. In this study, we propose a hybrid deep learning framework that combines Generative Adversarial Networks (GANs) for data augmentation with transfer learning from pre-trained convolutional neural networks (CNNs) to improve the accuracy and efficiency of DR detection. The dataset, consisting of retinal fundus images categorized into five DR stages, suffers from class imbalance, particularly in severe and proliferative stages. GANs were employed to generate synthetic images to address this imbalance, while transfer learning with models such as ResNet-50, VGG-19, and Inception-v3 enabled effective feature extraction from the images. The results demonstrate significant improvements in classification performance, with the ResNet-50 model achieving the highest accuracy of 93.5% and an AUC-ROC of 0.96. The GAN-augmented models notably enhanced the detection of minority classes, improving the F1-scores for severe and proliferative DR by 15% compared to traditional augmentation techniques. The use of early stopping ensured stable training, while the confusion matrix showed minimal misclassifications between adjacent DR stages. These findings suggest that the proposed framework can significantly improve the accuracy and robustness of DR detection, especially for underrepresented disease stages. The proposed hybrid framework offers a scalable and efficient solution for automated DR screening, with potential for integration into clinical workflows to assist in early diagnosis and intervention. Future work includes real-world validation on larger datasets and exploring advanced architectures for further performance enhancement.

Article Details

Section
Articles