Enhanced Detection of Diabetic Retinopathy using GAI-Net: Deep Learning Model based on GAN with Autoencoder

Abstract

Diabetic Retinopathy (DR) is a serious illness associated with diabetes and is one of the leading causes of universal visual impairment. Early and proper diagnosis is crucial to prevent permanent eye damage. The proposed paper introduces a new hybrid deep learning framework, GAI-Net, that uses Generative Adversarial Networks (GANs), Autoencoders, and InceptionV3 to detect DR in five levels of severity. The model addressed multiple issues in DR detection, including small datasets and class imbalance, through quality synthetic images and discriminative hierarchical features. Generative component increases data variation, and the discriminative modules increases the representation of the features and accuracy in recognition. To compare the performance of the proposed GAI-Net, a comparative analysis was undertaken with advanced models, including traditional GANs, DCGAN, Attention-GAN, EfficientNet GAN, and InceptionV3 through the use of Kaggle dataset images are included. The obtained results indicate that GAI-Net outperforms all other comparative models, achieving a detection accuracy of 0.99 and an F1-score of 0.98, which is a high level of effectiveness. The methodology shows the strengths of combining the generative and discriminative deep learning methods in medical image analysis. GAI-Net is able to diagnose diabetic retinopathy with a scalable, accurate, and entirely automatic model by constructing balanced datasets and improving the encoding of features. The current research paper explains the potential use of hybrid architectures in computer-aided diagnosis, as it helps strengthen the development of the next-generation tools to find diabetic retinopathy with extensive medical uses.