2D/3D多模态医学图像配准算法研究  

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

作　　者：徐密 诸葛斌[1] 袁非牛 尹正虎 董黎刚[1] 蒋献 孙应绮 宋嘉琦 史晓彤 苏雷 周屹博 林诗凡 XU Mi;ZHUGE Bin;YUAN Feiniu;YIN Zhenghu;DONG Ligang;JIANG Xian;SUN Yingqi;SONG Jiaqi;SHI Xiaotong;SU Lei;ZHOU Yibo;LIN Shifan(School of Information and Electronic Engineering,Zhejiang Gongshang University,Hangzhou 310020,China;School of Information and Mechatronics Engineering,Shanghai Normal University,Shanghai 201418,China;Pediatrics Hospital of Fudan University,Shanghai 201102,China)

机构地区：[1]浙江工商大学信息与电子工程学院,浙江杭州310020 [2]上海师范大学信息与机电工程学院,上海201418 [3]复旦大学附属儿科医院,上海201102

出　　处：《电信科学》2024年第3期75-88,共14页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.62272308);浙江省新型网络标准与应用技术重点实验室项目(No.2013E10012);浙江省“尖兵”“领雁”研发攻关计划项目(No.2023C03202);浙江省研究生一般项目(No.1120XJ0622033);国家级大学生创新创业训练计划项目(No.202210353022,No.202310353040,No.202310353069);浙江省大学生科技创新活动计划—新苗人才计划项目(No.1120KZN0223068G)。

摘　　要：2D/3D多模态配准在医学影像导航手术中起着重要作用,主要用于提供术前三维图像和术中二维图像的实时信息,帮助医生精准定位病灶,规划手术路径,从而提高手术的安全性和效率。提出了一种2D/3D多模态医学图像配准算法,首先利用Swin Transformer优秀的特征提取能力,构建了初始姿态估计模型,实现姿态参数的快速预测;接着,为了提升整个配准方法的鲁棒性,引入基于Grangeat关系的粗配准方法;最后设计了基于梯度下降的精配准模块,以提升整个配准过程的精确性,且在该模块将Sobel微分算子与归一化互相关结合,提升了参数优化过程中的灵敏度。实验结果表明,所提配准方法在正位和侧位配准中的误差满足配准要求,配准成功率有显著提升。2D/3D multimodal alignment plays an important role in medical image navigation surgery,which is mainly used to provide realtime 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.

关 键 词：2D/3D图像配准 多模态 初始姿态估计 粗配准 精配准 相似度测量 

分 类 号：P237[天文地球—摄影测量与遥感]