Fig. 2.

Generation of input label maps. Smooth 3D noise images p_j (j ∈ {1, 2, …, J}) are sampled from a standard distribution, then warped by random deformations ϕ_j to cover a range of scales and shapes. We synthesize a label map s from the warped images ${\tilde{p}}_j=p_j\circ {\varphi }_j$: for each voxel k of s, we assign label j corresponding to image ${\tilde{p}}_j$ where k has the highest intensity j, i.e. $s_k=\text{arg }{\text{max}}_j\left({\left[{\tilde{p}}_j\right]}_k\right)$. We use J = 26.