Fig. 3.
Intermediate steps of the generative model: (a) we randomly select an input label map from the training set, which we (b) spatially augment in 3D. (c) A first synthetic image is obtained by sampling a GMM at HR with randomised parameters. (d) The result is then corrupted with a bias field and further intensity augmentation. (e) Slice spacing and thickness are simulated by successively blurring and downsampling at random LR. (f) The training inputs are obtained by resampling the image to HR, and removing the labels we do not wish to segment (e.g., extra-cerebral regions).