GAN-Driven Retinal Augmentation: A Study of Synthetic Augmentation Strategies and Their Effect on Multilabel Retinal Disease Classification

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness that affects millions of people with diabetes across all income settings. Automated grading of fundus photographs holds real promise for expanding screening access in resource-limited settings. A persistent barrier, however, is class imbalance, where early and intermediate disease grades are under-represented in clinical datasets, degrading per-class sensitivity exactly where early detection matters most. This paper investigates whether synthetic image augmentation using generative adversarial networks (GANs) can address this gap, using the Brazilian Retinal Image Dataset, a patient-wise split fundus dataset with five DR severity grades. Two GAN architectures, DCGAN and FastGAN, were trained on three underrepresented classes (Mild NPDR, Moderate NPDR, and Proliferative DR) to generate 500 synthetic images each. The synthetic images were then used under five augmentation strategies with two classification backbones, DenseNet121 and ResNet50, tested on a held-out set using macro-averaged area under the curve and F1 score. FastGAN consistently achieved lower Fréchet Inception Distance values than DCGAN across all three classes (168.03, 129.17, and 125.86 versus 171.06, 138.78, and 132.81), but GAN augmentation did not improve downstream classification. DenseNet121 with no augmentation yielded the highest macro F1 (0.615), and every GAN-based strategy underperformed relative to this baseline. These findings suggest that at 128-pixel resolution and with only hundreds of real training images per class, GAN-generated retinal images do not yet carry enough diagnostic signal to benefit a classifier and may introduce noise that misleads training.