Thanks to their ability to learn data distributions without requiring paired\ndata, Generative Adversarial Networks (GANs) have become an integral part of\nmany computer vision methods, including those developed for medical image\nsegmentation. These methods jointly train a segmentor and an adversarial mask\ndiscriminator, which provides a data-driven shape prior. At inference, the\ndiscriminator is discarded, and only the segmentor is used to predict label\nmaps on test images. But should we discard the discriminator? Here, we argue\nthat the life cycle of adversarial discriminators should not end after\ntraining. On the contrary, training stable GANs produces powerful shape priors\nthat we can use to correct segmentor mistakes at inference. To achieve this, we\ndevelop stable mask discriminators that do not overfit or catastrophically\nforget. At test time, we fine-tune the segmentor on each individual test\ninstance until it satisfies the learned shape prior. Our method is simple to\nimplement and increases model performance. Moreover, it opens new directions\nfor re-using mask discriminators at inference. We release the code used for the\nexperiments at https://vios-s.github.io/adversarial-test-time-training.\n