Enhancing the performance of diabetic retinopathy grading models using synthetic images

To determine whether strictly curated synthetic images can improve grading performance in diabetic retinopathy (DR) diagnosis within limited data regimes, and to compare the utility of representative Generative Adversarial Network (GAN) and diffusion architectures.

StyleGAN3 and Medfusion models were pre-trained on an auxiliary dataset. To prevent mode collapse and severity downshift, synthetic images were curated via a domain-matched Oracle network using Latent Space Rejection Sampling and Class-Conditioned SDEdit Escalation. Efficacy was evaluated on an isolated downstream DR classification task (IDRiD dataset) across three architectures (ConvNeXt, ResNet50, VGG16) using five random seeds.

Curated StyleGAN3 dataset achieved better Fréchet Inception Distance (24.25) and Padded SSIM (0.9037) compared to Medfusion (58.70 and 0.8852), though Medfusion yielded a higher Inception Score (2.35). In the isolated downstream classification task, baseline ConvNeXt models trained solely on real IDRiD data achieved a Quadratic Weighted Kappa (QWK) of 0.6350 ± 0.0135. Augmentation with curated StyleGAN3 improved the QWK to 0.7185 ± 0.0309, while Medfusion augmentation achieved 0.7695 ± 0.0254 (an absolute improvement of 0.1345). Crucially, Medfusion augmentation successfully doubled the sensitivity (Recall) for Proliferative DR from 0.2615 ± 0.1595 to 0.5385 ± 0.0942.

Data augmentation using strictly curated synthetic images successfully addresses data imbalance and scarcity, significantly improving the ordinal accuracy and minority-class sensitivity of DR diagnosis models. Furthermore, standard global image quality metrics like FID may not fully capture localized pathological utility; diffusion models with higher FID scores ultimately provided superior structural diversity and downstream clinical performance.