Fig. 2.

Overview of a training step. At each mini-batch, we randomly select a 3D label map from a training set $\left\{S_n\right\}$ and sample a pair $\{I,T\}$ from the generative model. The obtained image is then run through the network, and its prediction $Y$ is used to compute the average soft Dice loss, that is backpropagated to update the weights of the network.