磁共振扩散张量成像数据分析中基于统一计算设备架构的高速行处理求解超定线性方程组方法  

Row action method with compute unified device architecture in the process of solving over-determined equations for magnetic resonance diffusion tensor imaging

作　　者：王飞[1,2] 高嵩[2] 

机构地区：[1]北京大学基础医学院生物物理系,北京100191 [2]北京大学医学部医学影像物理与工程实验室,北京100191

出　　处：《中国医学影像技术》2012年第6期1226-1229,共4页Chinese Journal of Medical Imaging Technology

基　　金：国家自然科学基金(81171130);北京市自然科学基金(7102102)

摘　　要：目的提出一种运行于普通个人电脑平台上的并行方法,用于求解MR DTI中的超定线性方程组。方法利用统一计算设备架构(CUDA)使中央处理器(CPU)与图形处理器(GPU)协同求解超定线性方程组。CPU用于数据准备与生成扩散矩阵,GPU中的大量流处理器并行用于迭代计算。结果 CUDA模式下行处理运算速度远快于CPU串行计算,图像矩阵增大时这一优势更加明显。结论与CPU串行模式相比,CUDA模式可显著提高DTI数据处理速度。Objective To propose a parallel mode running on conventional personal computer to accelerate the process of solving over-determined equations of magnetic resonance DTI.Methods Compute unified device architecture（CUDA） was used to combine the central processing unit（CPU） and graphic processing unit（GPU） to solve over-determined equations.All data preparation and diffusion matrix generation were achieved by CPU.Then,the iterative row action method was parallel performed by stream processors in GPU.Results The computing speed of the proposed CUDA method was much faster than that of the CPU based method.The advantage of CUDA was more pronounced when the imaging matrix was increased.Conclusion CUDA approach can produce significant performance gain compared to conventional CPU method in the data analysis process of DTI.

关 键 词：扩散磁共振成像 图像处理器 统一计算设备架构 

分 类 号：R445.2[医药卫生—影像医学与核医学] R311[医药卫生—诊断学]