Fig. 3

CMRGenNet structure. A StyleGAN2 with Adaptive Discriminator Augmentation (ADA) (orange blocks) generates realistic CMR images while the label generator branch (light blue blocks) predicts a semantic segmentation and also provides a segmentation uncertainty estimate as a byproduct. Note that the two branches are trained independently. In particular, there is no gradient back-propagation from the semantic label into the StyleGAN2 architecture. In addition, the mapping of class labels to colors is provided on the right