Fig. 6

Illustration of macJNet for CT-MRI registration. Image labels L comprise two subsets: manual annotation label subset L^gt as ground-truth in segmentation network, prediction label subset L^seg is generated by Seg-SubNets. L_M = {Lgt M, Lseg M}, L_F = {Lseg F, Lseg F}. For each iteration, Reg-SubNet takes I_M, I_F and their labels as input, outputs the deformation field ϕ, which provides the cross-modality consistency constrain for Seg-SubNets by mapping L_M to L_F. Seg-SubNets take I_M and I_F as input, and output Lseg M and Lseg F to provide semantic labels as anatomical prior knowledge for registration