Large Deformation Registration with A Confidence-Guided Network

Xiaolong Wu, Bo Hu, Ruikang Xu, Shenglong Zhou, Ruisheng Gao, Zhiwei Xiong

Abstract: Large deformation is a challenging optimization problem in medical image registration. To deal with it, recent deep learning-based registration methods employ multi-step network structures to break the large deformation into smaller ones and register them separately. However, these methods have two problems. First, they cannot effectively discriminate between hard-to-optimize large deformations and easy-to-optimize small deformations. This indistinctive registration manner may result in inaccurate registration results. Second, since the deformation range to register is wide but local, existing convolution-based and transformer-based methods cannot have a proper receptive field. In this paper, we propose a novel unsupervised registration network, with a confidence-guided part to focus on different scale deformations with different degrees. Furthermore, we propose a recurrent large kernel attention registration part to register the image with the proper receptive field. Comprehensive experimental results on the brain and liver datasets show that our proposed confidence-guided network significantly improves registration accuracy over existing methods for large deformation registration. Code is available at https://github.com/aicaomei/CGRNet.

External IDs:dblp:conf/wbir/WuHXZGX24