Retinal image synthesis through the least action principle

Abstract

Eye fundus image analysis is a fundamental approach in medical diagnosis and follow-up ophthalmic diagnostics. Manual annotation by experts needs hard work, thus only a small set of annotated vessel structures is available. Examples such as DRIVE and STARE include small sets for training images of fundus image benchmarks. Moreover, there is no vessel structure annotation for a number of fundus image datasets. Synthetic images have been generated by using appropriate parameters for the modeling of vascular networks or by methods developing deep learning techniques and supported by performance hardware. Our methodology aims to produce high-resolution synthetic fundus images alternative to the increasing use of generative adversarial networks, to overcome the problems that arise in producing slightly modified versions of the same real images, to simulate pathologies and for the prediction of eye-related diseases. Our approach is based on the principle of the least action to place vessels on the simulated eye fundus.