A Two-Stage GAN for High-Resolution Retinal Image Generation and Segmentation

In this paper, we use Generative Adversarial Networks (GANs) to synthesize high-quality retinal images along with the corresponding semantic label-maps, instead of real during training a segmentation network. Different from other previous proposals, employ two-step approach: first, progressively growing GAN is trained generate which describes blood vessel structure (i.e., vasculature); second, an image-to-image translation approach used obtain realistic generated vasculature. The adoption two-stage process simplifies generation task, so that network requires fewer consequent lower memory usage. Moreover, learning effective, and only handful samples, our generates high-resolution images, can be successfully enlarge small available datasets. Comparable results were obtained by employing synthetic in place data training. practical viability proposed was demonstrated on two well-established benchmark sets for segmentation—both containing very number samples—obtaining better performance respect state-of-the-art techniques.