GAN-based data augmentation with recurrent residual U - Net for exudate segmentation in diabetic retinopathy

Diabetic Retinopathy (DR) is a leading cause of vision impairment worldwide, often characterized by the presence of retinal lesions such as exudates. Hard exudates (HE), which appear as bright lesions in retinal fundus images, are an early and critical indicator of DR, making their accurate detection vital for timely diagnosis and intervention. Despite improvements in medical imaging, creating reliable deep learning models for HE segmentation is largely affected by the scarcity of annotated datasets. Recent studies on classifying the segmentation of sub-lesions in retinal images face a significant challenge due to the limited diversity of datasets. In this study, we propose an approach that combines Generative Adversarial Networks (GANs) for data augmentation with a Recurrent Residual U-Net (R2U-Net) architecture for segmentation. Through the use of GANs, this study addresses the dataset scarcity issue caused by the time-intensive process of manual annotation. The datasets - IDRiD and E-Ophtha - were used to assess our model after it was trained on the IDRiD dataset. Performance comparisons were conducted between models trained on (i) original dataset with no augmentation (ii) original dataset with traditional/offline augmentation, (iii) synthetic dataset made with augmented orig-inal dataset and GAN-generated images. The GAN-augmented dataset achieved the best metrics: IoU 0.9319, DSC 0.9435, accu-racy 0.9926, sensitivity 0.9554, and specificity 0.9930, surpassing recent literature.