基于深度学习的2D/3D医学图像配准研究  

2D/3D Medical Image Registration Using Convolutional Neural Network

作　　者：陈向前[1] 郭小青 周钢[1] 樊瑜波[1] 王豫[1] Chen Xiangqian;Guo Xiaoqing;Zhou Gang;Fan Yubo;Wang Yu(School of Biological Science and Medical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]北京航空航天大学生物与医学工程学院,北京100191

出　　处：《中国生物医学工程学报》2020年第4期394-403,共10页Chinese Journal of Biomedical Engineering

摘　　要：2D/3D配准在临床诊断和手术导航规划中有着广泛的应用,可解决医学图像领域中不同维度图像存在信息缺失的问题,能辅助医生在术中精准定位患者的病灶。常规的2D/3D配准方法主要依赖于图像的灰度进行配准,但非常耗时,不利于临床实时性的需求,并且配准过程中容易陷入局部最优值。提出用深度学习的方法来解决2D/3D医学图像配准问题。采用一个基于深度学习的卷积神经网络,通过网络对数字影像重建技术(DRR)进行训练并自动学习图像特征,预测X光图像所对应的参数,从而实现配准。以人体骨盆的模型骨为实验对象,根据骨盆的CT数据生成36000张DRR图像作为训练集,同时通过C臂采集模型骨的50张X光图像作为验证。结果显示,深度学习算法在相关系数、归一化互信息、欧式距离3个精度评价指标上的测试值分别为0.82±0.07、0.32±0.03、61.56±10.91,而常规2D/3D算法对应的测试值分别为0.79±0.07、0.29±0.03、37.92±7.24,说明深度学习算法的配准精度优于常规2D/3D算法的配准精度,且不存在陷入局部最优值的问题。同时,深度学习的配准时间约为0.03 s,远低于常规2D/3D配准的时间,可满足临床对于实时配准的需求,未来将进一步开展临床数据的2D/3D配准研究。2D/3D registration is widely used in clinical diagnosis and surgical navigation planning,which can solve the problem of missing information in different dimensions of medical images and assist doctors to accurately locate patients'lesions during surgery.The conventional 2D/3D registration method mainly relies on the gray level of the image for registration,but the registration process is very time consuming,which is not conducive to the clinical real-time requirements,and the registration process is easy to fall into the local optimum.This study proposed a deep learning approach to solve 2D/3D medical image registration problems.The method used a deep learning-based convolutional neural network to train the DRR and automatically learned image features to predict the parameters corresponding to the X-ray image to achieve registration.In the study,the human pelvis model bone was used as the experimental object.A total of 36,000 DRR images were generated as training sets,and 50 X-ray images of the model bone were collected by C arm for verification.Results showed that the test values for the three precision evaluation indicators of the correlation coefficient,normalized mutual information and Euclidean distance were 0.82±0.07,0.32±0.03,61.56±10.91 and the corresponding test values of the conventional 2D/3D algorithm were 0.79±0.07,0.29±0.03,37.92±7.24.These results meant the registration accuracy of deep learning algorithm was better than the conventional 2D/3D algorithm and there was no local optimal value for deep learning algorithm.Meanwhile,the registration time of deep learning was about 0.03 s,which was much lower than the time of conventional 2D/3D registration,which can satisfy the clinical demand for real-time registration.In the future,2D/3D deep learning registration research of clinical data will be further carried out.

关 键 词：2D/3D配准 深度学习 卷积神经网络 

分 类 号：R318[医药卫生—生物医学工程]