Learning Fixed Points in Generative Adversarial Networks: From Image-to-Image Translation to Disease Detection and Localization | Synapse