Figure 5.

Overview of involved convolutional sub-networks (blue) in the two-phase training according to section 3.2 in the text. In the first phase (A) the shape space is trained with E₁ and D by using loss L_Shape (red) to reconstruct R_c and R_cp. In the second phase (B) E₂ is trained to map segmentations from U into the shape space via the second loss L_Prediction (green). The latent code ŷ_i interpolated between ŷ_c and ŷ_cp is decoded with the previously trained D of phase (A) for predicting the desired follow-up lesion Ŝ_l.