Generative Adversarial Network with Patch Selection for Deformable Registration of Medical Images

Image registration is a fundamental topic in medical image research. Traditional registration methods have a large computational cost and longtime consumption in the optimization process. Recently, learning-based registration frameworks have not only improved computational speed but also greatly enhanced registration performance. However, local registration accuracy is often difficult to maintain during registration for images with complex deformation. Therefore, we proposed a generative adversarial network with an image patch selection strategy and attention block. The discriminator applies additional regularization to image patches with complex deformations, while improving overall registration accuracy. We also introduced the anti-folding and smoothness loss in the registration network to generate reasonable deformation fields. The experimental results demonstrate that compared with some current studies, the proposed method achieves better registration performance, especially in regions with complex deformations.

I. Introduction

Medical image registration refers to the process of aligning medical images obtained at different times, from different imaging devices, or from different anatomical positions. This is an important research direction in the field of medical imaging, which is widely used in image-guided medical image analysis, radiotherapy planning, image guidance for surgery, motion tracking, segmentation, image reconstruction, and other fields [1],[2]. Medical image registration methods can be roughly divided into two categories: those using traditional methods and those using deep learning-based methods.