Research on 2D/3D multimodal medical image registration algorithm

Abstract

2D/3D multimodal alignment plays an important role in medical image navigation surgery, which is mainly used to provide real-time information of preoperative 3D images and intraoperative 2D images to help doctors accurately locate the lesions and plan the surgical paths, so as to improve the safety and efficiency of surgery. A 2D/3D multimodal medical image alignment algorithm was proposed, which firstly utilized the excellent feature extraction capability of Swin Transformer to construct an initial pose estimation model to realize the fast prediction of pose parameters. Then, in order to improve the robustness of the whole alignment method, a coarse alignment method based on the Grangeat relation was introduced. Finally,a fine alignment module based on gradient descent was designed. A fine alignment module based on gradient descent was designed to improve the accuracy of the whole alignment process, and the Sobel differential operator was combined with normalized correlation in this module to improve the sensitivity of the parameter optimization process. The experimental results show that the proposed alignment method meets the alignment requirements in the orthogonal and lateral alignment errors, and the alignment success rate is significantly improved.

Keywords

• 2D/3D image registration;multimodality;initial pose estimation;coarse registration;fine registration;similarity measurement