Physics-informed deep generative learning for quantitative assessment of the retina

This paper introduces a novel algorithmic approach using physics-informed generative adversarial networks (PI-GAN) to generate realistic digital models of human retinal blood vessels. The PI-GAN approach, based on biophysical principles like Murray's Law and fluid dynamics, enables accurate segmentation and reconstruction of blood vessel networks without human input, surpassing human labelling performance. Evaluated on DRIVE and STARE datasets, it achieved near state-of-the-art vessel segmentation with training on a small simulated dataset. This method holds significant potential for improving early disease detection and monitoring in retinal diseases.

Emmeline E. Brown, Andrew A. Guy, Natalie A. Holroyd, Paul W. Sweeney, Lucie Gourmet, Hannah Coleman, Claire Walsh, Athina E. Markaki, Rebecca Shipley, Ranjan Rajendram, Simon Walker-Samuel