An Unsupervised Registration Method for Brain Images Based on Contour Guidance

For medical image registration with large deformation, the huge differences in shape, position and texture make this task challenging. At present, unsupervised registration methods usually use registration software or traditional registration methods for preprocessing, thus only focusing on nonlinear transformation. This paper designs an unsupervised affine registration network to achieve end-to-end registration. The algorithm consists of two parts: (1) Use the designed affine network to predict affine parameters. (2) Use the affine transformer to perform affine transformation on the image. To make the network converge faster, we design a parameter regularization loss for affine parameter constraint. We tested our method on two datasets, including a private dataset and a public dataset, BraTS2018. We use normalized cross-correlation (NCC), dice similarity coefficient (DSC) and mean square error (MSE) as evaluation metrics. Compared with some commonly used unsupervised methods, our method achieves the best results in these three evaluation metrics. Through qualitative analysis, our method can register large deformation image well.

I. Introduction

Medical image registration has important theoretical research and clinical application value. It uses one image as a fixed image and the other image as a moving image for registration. Common medical images include computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), functional magnetic resonance imaging (fMRI), etc. Registration of CT and other anatomical images with PET and other functional images is helpful to obtain both anatomical and functional information, which is of great significance for computer-aided diagnosis [1]. For medical images with large deformation, there are great differences in shape, position and texture between fixed images and moving images, so registration with large deformation is challenging.