Fusion-Based Multimodal Medical Image Registration Combining Inter-Modality Metric and Disentanglement

Multimodal medical image registration is a challenging problem. Inter-modal metric and unifying images into the same domain are two feasible methods, but each has its emphasis. We proposed a fusion-based registration method to combine the advantages of the two ways. In the first stage, two subnetworks register images individually. The code sub-network adopts the disentanglement content code method, and the image sub-network adopts the inter-modality metric method. In the second stage, based on the preliminary information provided by sub-networks, a fusion network is built to perform further registration in a more comprehensive perspective. We compare the typical registration methods quantitatively and qualitatively, including inter-modality, image-to-image translation, and fusion methods. Experiments demonstrate our method overcomes the limit of single methods and is comparable with the state-of-art iterative and learning-based approaches.