Figure 3:

Alternative model architectures for learning the effect of registration hyperparameters. In these approaches, the hyperparameters Λ are provided as input to an auxiliary convolutional network $d_{{\theta }_d}$ (blue), which is integrated directly with the primary registration network $g_{{\theta }_g}$ (grey). The output of $d_{{\theta }_d}$ is either added to the output channels of the registration U-Net (full-integrative model) or provided as an additional input to the first layer (pre-integrative model) or last upsampling layer (post-integrative model).