Does an ensemble of GANs lead to better performance when training segmentation networks with synthetic images?