RetinaSynth: Diffusion‑Based Synthetic Imaging

Amina Kurtovic, Amila Akagic

Keywords: Artificial intelligence (AI), Healthcare data, Synthetic data, Generative models, Medical image synthesis

Abstract: Recent advances in artificial intelligence (AI) have unlocked remarkable potential across a wide spectrum of application domains. For AI models to be successfully deployed, they usually need access to vast amounts of high-quality data. In healthcare, however, such data is very constrained due to the sensitive nature of patient information, and strict legal frameworks (e.g., HIPAA in the United States and GDPR in Europe) govern its use. This creates a major obstacle for building robust AI‑driven medical solutions. A promising approach to overcome this barrier is the generation of synthetic data that closely resembles real-world data distributions while protecting individual privacy. Synthetic datasets can therefore accelerate model development and validation while staying fully compliant with privacy regulations. Beyond augmentation, synthetic data unlocks other new opportunities: translating images between modalities, simulating imaging under different acquisition parameters, and producing realistic material for training and education. In this paper, we investigate if generative models (VAEs, diffusion, Pix2Pix) can create realistic retinal images, while also assessing their clinical usefulness, fairness, and regulatory acceptability. We propose a three-stage pipeline, shown in Figure 1., that (i) extracts vascular structures via segmentation, (ii) generates synthetic vessel skeletons using either Variational Autoencoders (VAEs) or diffusion models, and (iii) translates skeletons into retinal images using Pix2Pix in both paired and unpaired settings. To the best of our knowledge, this work is first to apply diffusion models to vessel-segmented fundus images for synthetic retinal image generation.

Submission Number: 140