Disentangled Representation Learning For Deep Mr To Ct Synthesis Using Unpaired Data

Many different methods have been proposed for generation of synthetic CT (sCT) from MR images. Most of these methods depend on paired-wise aligned MR and CT training images of the same patient, which are difficult to obtain. In this paper, we propose a novel disentangled representation learning method for MR to CT synthesis using unpaired data. Specifically, we first embed images onto two spaces: a modality- invariant geometry space capturing the shared anatomical information across different imaging domains, and a modality- specific appearance space. From the embedding, a sCT image can be synthesized from a MR image by taking the encoded geometry features from the MR image and an appearance vector sampled from the appearance space of a CT image. To handle the challenging of distinguishing cortical bone from air in MR images, where both of them have low intensity values, we propose a novel Geometry Similarity Module (GSM) to take the context information into consideration. Experimental results demonstrated that our approach achieved better or equivalent results than the state-of-the-art.