医学图像配准的深度学习方法综述  

Survey on Deep Learning in Medical Image Registration

作　　者：莫晓盈 杨锋[1,2] 尹梦晓 石华榜[1] MO Xiao-ying;YANG Feng;YIN Meng-xiao;SHI Hua-bang(School of Computer,Electronics and Information,Guangxi University,Nanning 530004,China;Guangxi Key Laboratory of Multimedia Communications Network Technology,Guangxi University,Nanning 530004,China)

机构地区：[1]广西大学计算机与电子信息学院,南宁530004 [2]广西多媒体通信与网络技术重点实验室,南宁530004

出　　处：《小型微型计算机系统》2021年第8期1706-1714,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61861004,61762007)资助;广西自然科学基金项目(2017GXNSFAA198267,2017GXNSFAA198269)资助。

摘　　要：图像配准指的是寻找两个或多个图像之间的映射关系,医学图像配准在疾病诊断、手术引导和疾病治疗跟踪等方面具有重要应用价值,如何精确、高效地配准医学图像已成为一个急需解决的课题.近些年来,基于深度学习的医学图像配准方法逐渐崭露头角,一定程度上克服了传统的配准方法上适用范围窄、计算速度不够快等瓶颈.本文将深入地介绍基于深度学习的医学图像配准现状和现存的配准方法技术.本文首先介绍3类基于深度学习的图像配准方法,包括监督变换估计、无监督变换估计和使用生成对抗网络的配准方法;然后在两个主流数据集上对一些常见的配准方法进行配准效果分析比较;最后对基于深度学习的医学图像配准发展趋势进行讨论.Image registration is the process of establishing spatial correspondences between different image acquisitions,medical image registration has important application value in disease diagnosis,surgical guidance and disease treatment tracking.How to register medical image accurately and efficiently has become an urgent problem.In recent years,the development of medical image registration method based on the deep learning gradually perform well.To some extent,it overcome the bottleneck of traditional registration method,such as narrow application scope and insufficient computing speed.This paper first introduces three types of image registration methods based on deep learning,including supervised transform estimation,unsupervised transform estimation and registration method using generated adversarial network.Then analyze and compare the effect of some common registration methods in two main datasets.Finally,discuss the development trend of medical image registration based on deep learning.

关 键 词：图像配准 监督学习 无监督学习 生成对抗网络 变换估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]