Unsupervised Medical Image Registration Based on Multi-scale Cascade Network

Abstract

AbstractAs a core technique of medical image analysis task, image registration is the process of finding the non-linear spatial correspondence among the input images. Comparing with supervised learning methods, unsupervised learning methods can ease the burden of manual annotation by exploiting unlabeled data without supervision. For high-resolution medical images, existing methods would maintain the consistency of the global structure and have limited matching accuracy in local details. Moreover, the intensity distribution of the warped image tends to the fixed image, which is quite different from the moving image. To solve the above problems, in this paper, we propose a multi-scale cascade network based on unsupervised end-to-end network registration method. We cascade the registration subnetworks with multi-scale strategy, which can extract local features more comprehensively and realize the registration from coarse to fine. The cascading procedure can further maintain the consistency of both global structure and local details for high-resolution medical images. The cyclic consistency loss is introduced to ensure the content consistency between the moving image and the warped image, and the structural consistency loss is used to ensure the structural consistency between the fixed image and the warped image. Experiments on three datasets demonstrate that our algorithm achieves state-of-the-art performance in medical image registration. The source code would be available online.KeywordsMedical image registrationCascade networkMulti-resolution registration