Figure 1.

Overview of our framework. From input images ${\mathrm{I}}_f,{\mathrm{I}}_m$ the semantic features ${\mathrm{f}}_f,{\mathrm{f}}_m$ are extracted by U-Net and passed on to the registration networks (RegNet) after the concatenation. Deformation field ${\mathrm{V}}_1,{\mathrm{V}}_2$ are then estimated by the first and the second RegNet, respectively. As input of the second RegNet, the semantic features of the moving image are transformed using the output of the first RegNet. Finally, two deformation fields are combined to obtain the final deformation field V.