Recently, the generative adversarial network (GAN) has been extensively applied to the cross-modality conversion of medical images and has shown outstanding performance than other image conversion algorithms. Hence, we propose a novel GAN-based multi-domain registration method named multiscale diffeomorphic jointed network of registration and synthesis (MDJRS-Net). The deviation of the generator of the GAN-based approach affects the alignment phase, so a joint training strategy is introduced to improve the performance of the generator, which feedbacks the structural loss contained in the deformation field. Meanwhile, the nature of diffeomorphism can enable the network to generate deformation fields with more anatomical properties. The average dice score (Dice) is improved by 1.95% for the computer tomography venous (CTV) to magnetic resonance imaging (MRI) registration task and by 1.92% for the CTV to computer tomography plain (CTP) task compared with the other methods.