Fig. 1

Summary of options to obtain an automatic CMRA segmentation. The border in each image is orange for inputs to the deep learning networks, and green for outputs. Dashed borders indicate that the outputs can be generated with or without corresponding ground truth label maps. Option A is a fully supervised approach (UNet). It only relies on a segmentation module, which takes as input MR images and corresponding label maps to produce as output the segmentations MR-seg. Option B and C are generative modelling approaches; B is a GAN approach which uses the Domain Transfer module plus a Segmentation module. CT and MR images are used as inputs to generate gMR and gCT respectively. gMR and MR are fed into the segmentation module which learns how to segment images from this domain, whether they are real or generated, utilizing the CT label maps. If available, MR label maps can also be used as a supervised segmentation loss, in any quantity; C is a VAE approach which uses the domain decoupling module to generate gCT and gMR images with no domain-specific features. As before, the Segmentation module can be trained with or without ground truth MR-seg, but it requires CT label maps. CMRA: cardiovascular magnetic resonance angiography; GAN: generative adversarial network; VAE: variational autoencoder