Computer Science ›› 2013, Vol. 40 ›› Issue (4): 306-309.

Previous Articles     Next Articles

2D-3D Medical Image Registration Based on GPU

DANG Jian-wu,HANG Li-hua,WANG Yang-ping and DU Xiao-gang   

  • Online:2018-11-16 Published:2018-11-16

Abstract: In 2D-3D medical image registration,the DRR generation and the computation of the similarity measure between the DRR and the x-ray image are the most important and time-consuming registration procedures.Because of the problem of large amount of calculation in registration procedure,this paper combined the pattern intensity with gradient to simplify the calculation,and used GPU multithreading parallel computing to generate the DRR and comput the similarity on the GPU,introduced the gradient descent and multi-resolution strategies in the registration optimization procedure to complete the registration process.Compared with other similarity measures and registration on the CPU,this registration method ensures the registration precision and improves the registration speed.

Key words: 2D-3D medical image registration,Digitally reconstructed radiograph (DRR),Pattern intensity (PI),Graphic processing Unit (GPU),Gradient descent

[1] van de Kraats E B,Penney G P,et al.Standardized evaluation methodology for 2D-3D registration [J].IEEE Transaction on Medical Imaging,2005,24(9):1177-1189
[2] Groher M,Bender F,Hoffmann R T,et al.Segmentation-driven 2D-3D registration for abdominal catheter interventions[J].Medical Image Computing and Computer-Assisted Intervention,2007,4971:527-535
[3] Birkfellner W,Stock M.Stochastic rank correlation:A Robustmerit function for 2D/3D registration of image data obtained at different energies[J].Medical Physics,2009,6(8):3420-3428
[4] Penney G P,Weese J.A Comparison of Similarity Measures for Use in 2D-3D Medical Image Registration [J].Lecture Notes in Computer Science,1998,17(4):594
[5] Zollei L.2D-3D Rigid-Body Registration of X-Ray Fluoroscopy and CT Images[D].Electrical Engineering Computer Science,2001
[6] Kubias A.2D/3D Image Registration on the GPU[J].PatternRecognition and Image Analysis,2008,8(3):382-383
[7] Weese J,Buzug T M.An Approach to 2D/3D Registration of a Vertebra in 2D X-ray Fluoroscopies with 3D CT Images[C]∥Computer Vision Virtual Reality and Robotics in Medicine and Medical Robotics.1997:124
[8] ,假设将I(v,w)=Iv(v,w)代入式(9)中可以得到模式强度的计算公式为: S=1WHπr2∑Wx=1∑d2≤r2σ2σ2+(∑((λiμi)T·x→)-∑((λiμi)T·v→))2(16) 通过对模式强度计算公式的变形,模式强度的计算得到了简化,计算时间得到了缩短。 在计算差值图像的模式强度时,需要遍历整个图像的每个像素,并且以像素为中心,半径r内的像素点也要参与运算,可以看出每一个像素在计算的时候至少会被计算两次,每次遍历计算一个像素模式强度时,这个像素是其周围像素的中心
[9] 。所以当从左到右和从上到下扫描整个图像时,只计算每个像素点右边和下面的像素点,如图2所示,这样可以使计算的复杂度降低一半,缩短计算时间。 图2模式强度计算示意图  3.42D-3D配准的GPU实现 近年来,GPU作为一个强大的计算平台逐渐受到了人们的关注,它提供了灵活的编程环境,并且可以处理大量复杂数据的并行计算。GPU被成功地应用于计算机辅助手术中,同时也被应用于2D-3D医学图像配准中。GPU提供的线程可以与图像的大小相同,线程与图像的像素一一对应,将线程索引作为图像的坐标,就可以并行地计算每个像素的灰度,相比于在CPU上完成2D-3D医学图像配准,可以大大加速配准的速度。 2D-3D图像配准过程主要包含DRR图像的生成与DRR图像与X线图像之间相似度计算两个主要步骤,本文将这个两个步骤在GPU的并行计算平台上完成,整个配准的过程如图3所示。 图32D-3D配准流程图  3.5优化 配准优化的过程是寻找优化参数使DRR图像与X线图像之间的相似度最大,得到最好的配准效果。本文的优化过程是使用梯度下降优化策略来寻找最优参数,在不断的迭代过程中,通过得到的参数指导下一步迭代的图像变换,得到最大的相似性测度。为了能够提高配准的计算速度以及算法的鲁棒性,避免优化出现局部极值,采用了多分辨率策略
[10] 。配准的过程首先在粗糙的低分辨率完成,一旦完成一层的配准,用这层得到的优化的变换参数作为下一层较高分辨率的配准的输入,依次迭代完成整个配准。 4实验结果与分析 本文实验采用的硬件配置为:内存4.0GB,显卡Intel(R) HD Graphics Family,静态处理器为1GB。实验环境为Windows XP下的Visual Studio 2008,实验以盆骨的X线图像以及CT体数据为输入,其中X线图像大小为512×512,CT体数据大小为512×512×360。实验以X线图像为参考图像,DRR图像为浮动图像
[11] ,配准效果图如图4所示。 图4配准效果图  为了进一步验证实验结果,本文同时使用相关系数、互信息以及传统模式强度测度方法对配准过程作了相似性测度,并对结果作了总结,与本文相似性测度方法作了对比,如表1所列。 表1几种相似性测度方法比较相似性 测度方法txtytzrxryrzCPU (s)GPU (s)加速比相关系数15.220.31.021.621.630.40350.210.533.4互信息16.822.51.152.012.120.36380.414.326.6模式强度12.818.30.950.701.500.30400.518.521.6本文方法10.215.40.800.651.400.25360.28.244.0 通过实验表明,模式强度相似性测度在2D-3D医学图像配准中效果最好,而本文通过将模式强度与梯度相结合,对模式强度的计算进行了改进,配准的精确度和速度都有所提高。而基于GPU的配准在保持高的配准精度的同时,相比于基于CPU的配准速度有了明显的提高,很好地加速了配准的过程。 由于配准过程中计算量大,本文在采用梯度下降优化寻找最优参数的同时,引进了多分辨率策略

No related articles found!
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!