Class balanced diabetic retinopathy image synthesis using a latent diffusion framework
TL;DR
This study establishes the Diabetic Retinopathy Latent Diffusion Synthesizer (DR-LDS) as a highly resource-efficient solution to the medical data bottleneck by leveraging a fine-tuned Variational Autoencoder for domain-adapted latent space compression, alongside an optimized U-Net.
Abstract
Diabetic Retinopathy (DR) is one of the major causes of preventable blindness globally. Automated screening for DR is critical but is severely hindered by data scarcity and class imbalance. Real-world datasets, such as APTOS 2019, exhibit extreme class imbalance, where sight-threatening classes are statistically rare. Traditional augmentation fails to capture complex pathological features, while Generative Adversarial Networks (GANs) often suffer from mode collapse. Existing diffusion approaches typically operate in pixel space, limiting image resolution and requiring extreme computational resources. To address this, we introduce the Diabetic Retinopathy Latent Diffusion Synthesizer (DR-LDS), a class-conditional framework that synthesizes high-fidelity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$512 \times 512$$\end{document} fundus images natively deployable on consumer-grade hardware. By leveraging a fine-tuned Variational Autoencoder (VAE) for domain-adapted latent space compression, alongside an optimized U-Net, our method achieves superior anatomical realism with convergence in just 150 epochs (at 17 min 22 s per epoch). Extensive benchmarking demonstrates that DR-LDS outperforms state-of-the-art baselines, achieving a Fréchet Inception Distance (FID) of 8.05. When used to augment minority classes to a uniform target for the APTOS 2019 dataset across 11 deep learning architectures, our synthetic data significantly improved diagnostic accuracy by up to 15.25% (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.001$$\end{document}), enabling a lightweight SqueezeNet model to reach 95.60% validation accuracy and an F1 Score of 0.96. Moreover, rigorous data leakage analyses and Explainable AI (XAI) verify that DR-LDS learns genuine pathological biomarkers. This study establishes DR-LDS as a highly resource-efficient solution to the medical data bottleneck. The source code is available at: https://github.com/Touhid-Alam/DR-LDS
