Image synthesis from corrupted contrasts increases the diversity of\ndiagnostic information available for many neurological diseases. Recently the\nimage-to-image translation has experienced significant levels of interest\nwithin medical research, beginning with the successful use of the Generative\nAdversarial Network (GAN) to the introduction of cyclic constraint extended to\nmultiple domains. However, in current approaches, there is no guarantee that\nthe mapping between the two image domains would be unique or one-to-one. In\nthis paper, we introduce a novel approach to unpaired image-to-image\ntranslation based on the invertible architecture. The invertible property of\nthe flow-based architecture assures a cycle-consistency of image-to-image\ntranslation without additional loss functions. We utilize the temporal\ninformation between consecutive slices to provide more constraints to the\noptimization for transforming one domain to another in unpaired volumetric\nmedical images. To capture temporal structures in the medical images, we\nexplore the displacement between the consecutive slices using a deformation\nfield. In our approach, the deformation field is used as a guidance to keep the\ntranslated slides realistic and consistent across the translation. The\nexperimental results have shown that the synthesized images using our proposed\napproach are able to archive a competitive performance in terms of mean squared\nerror, peak signal-to-noise ratio, and structural similarity index when\ncompared with the existing deep learning-based methods on three standard\ndatasets, i.e. HCP, MRBrainS13, and Brats2019.\n