RetinaSim: Synthetic, whole-retina blood vessel networks for training deep neural networks and simulating retinal pathophysiology

Abstract

Purpose : Training of deep neural networks can be limited by access to well-curated, manually-labelled data. We have developed a novel approach for simulating realistic, fully-connected retinal blood vessel networks, and for predicting whole-retinal blood flow using established biophysical models, in healthy retinas and diabetic retinopathy. We used generative deep learning (DL) to create synthetic clinical image data from the synthetic vascular networks, with an aim to reduce bottlenecks to clinical implementation. This physics-informed approach has potential for enhancing the explainability of DL models and reducing reliance on manually-labelled data.

Methods : A novel, hybrid approach was used to generate whole-retina, fully-connected arterial, venous and capillary networks. Our approach combined Lindenmayer systems, constrained constructive optimisation and space colonization. The geometry of arteries, arterioles, venules and veins were optimised using a realistic biophysical model (including Murray’s Law). As networks were fully-connected, blood flow could be estimated using Poiseille’s Law. The resultant networks were compared against manually-segmented retinal vessel clinical images. Generative adversarial networks (GANs), with CycleGAN architecture, were used to perform image-to-image translation between simulated networks (domain A), widefield retinal photographs (domain B), optical coherence tomography angiography (domain C) and fluorescein angiography (domain D). Frechet inception distance was used to evaluate generated image quality.

Results : Synthetic retinal vascular networks compared favourably with real-world data (Figure 1). After 400 training epochs, the GAN produced fake clinical images with clear retinal vasculature and detail resembling the target images at native resolution from domains B-D. The average Frechet Inception Distance (FID) of the generated images was 5.06 (sd 1.95), indicating a small distance between feature vectors for real and fake images.

Conclusions : We demonstrated a physics-informed approach combining biophysical models of retinal vasculature with DL. The framework showed promise in predicting blood flow, synthesising diabetic retinopathy, and generating synthetic clinical images. As simulated retinal networks have a known ground-truth, this bypasses the need for manual labelling.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

RetinaSim: a physics-informed framework