Fig. 2

Basic diagrams of DL generative modelling architectures. A GAN structure where the Generator is decomposed into Encoder + Decoder. The Generator generates MR images from CT images, and the Discriminator discerns between real MR and CT-generated ones. B VAE structure which feeds the original input into a probabilistic encoder; the encoder learns vectors μ and σ from the data and the reparameterization trick is used to obtain a parametrised latent space z (note ⊙ is used for element-wise multiplication) from which images can be reconstructed. GAN: generative adversarial network; VAE: variational autoencoder