基于GPU异构平台的实时CT图像重建系统的研究
|
摘要
针对采用单CPU CT图像重建时间长,采用CPU集群重建成本及能耗高的问题,提出了CPU多线程+GPU的异构重建模型。采用CPU多线程流水线模式,将整个任务分解为若干个处理阶段,相邻的两个阶段之间以循环缓存连接,上一阶段完成一次计算任务后将数据放到循环缓存里,然后继续下一次的计算任务,下一阶段探测到循环缓存里有数据后从缓存里取出数据开始计算。各个任务是并行处理任务的,针对某一耗时瓶颈模块再采用GPU并行加速,充分发挥CPU和GPU的计算资源。CPU多线程+GPU模型相对于CPU多线程模型加速了16. 45倍,相对于串行CT图像重建加速了20. 5倍以上。将CPU多线程+GPU模型重建的图像与CPU串行程序重建的CT图像相比较,数据结果在误差范围内,满足实验设计要求。提出的图像重建模型采用成本较低的GPU显卡就实现了性能大幅提升,大大降低了CT图像重建系统的成本及功耗,而成本及功耗的降低会引起CT医疗诊断费用的降低,最终惠及广大病患。
In order to solve long time consuming problems of single CPU computed tomographic( CT) reconstruction and the problems of high cost and high energy consumption in the reconstruction using CPU cluster,this paper proposed a heterogeneous reconstruction model of CPU multithreads + GPU. This model used the CPU multithread pipelining pattern,it divided the whole task into several stages,and connected each two adjacent phases by a loop buffer. Once the calculation of the current task stage completed,the data would be put into the loop buffer,and then the next computing task would continue to execute.When the data in the loop buffers was detected by the latter stage,the data would be removed from the loop buffers and calculated. In this way,each thread processed tasks in a parallel way. For a time consuming bottleneck module,it adopted the parallel acceleration of GPU to give full play to the computing resources of CPU and GPU. The CPU multithreads + GPU model is16. 45 times faster than the CPU multithreaded model,and accelerates more than 20. 5 times faster than serial CT image reconstruction. By comparing with experimental results of CPU serial program,the result of the CPU multithreads + GPU model can meet the experimental design requirement in the range of error tolerance. The image reconstruction model proposed by using low cost GPU has greatly enhanced performance,greatly reduces the cost and power consumption of the CT system. The reduction in cost and power consumption will lead to a reduction in the cost of CT medical diagnosis,which will eventually benefit patients.
In order to solve long time consuming problems of single CPU computed tomographic( CT) reconstruction and the problems of high cost and high energy consumption in the reconstruction using CPU cluster,this paper proposed a heterogeneous reconstruction model of CPU multithreads + GPU. This model used the CPU multithread pipelining pattern,it divided the whole task into several stages,and connected each two adjacent phases by a loop buffer. Once the calculation of the current task stage completed,the data would be put into the loop buffer,and then the next computing task would continue to execute.When the data in the loop buffers was detected by the latter stage,the data would be removed from the loop buffers and calculated. In this way,each thread processed tasks in a parallel way. For a time consuming bottleneck module,it adopted the parallel acceleration of GPU to give full play to the computing resources of CPU and GPU. The CPU multithreads + GPU model is16. 45 times faster than the CPU multithreaded model,and accelerates more than 20. 5 times faster than serial CT image reconstruction. By comparing with experimental results of CPU serial program,the result of the CPU multithreads + GPU model can meet the experimental design requirement in the range of error tolerance. The image reconstruction model proposed by using low cost GPU has greatly enhanced performance,greatly reduces the cost and power consumption of the CT system. The reduction in cost and power consumption will lead to a reduction in the cost of CT medical diagnosis,which will eventually benefit patients.
引文
[1]杨莹,张学军. X射线在医学诊断领域的应用机理[J].真空电子技术,2015(1):20-23.(Yang Ying,Zhang Xuejun. The mechanism of x-ray imaging in medical diagnosis[J]. Vacuum Electronics,2015(1):20-23.)
[2] Tang Min,Zhao Jieyi,Tong Ruofeng,et al. GPU accelerated convex hull computation[J]. Computers&Graphics,2012,36(5):498-506.
[3]丁科,谭营. GPU通用计算及其在计算智能领域的应用[J].智能系统学报,2015,10(1):1-11.(Ding Ke,Tan Ying. A review on general purpose computing on GPUs and its applications in computational intelligence[J]. CAAI Trans on Intelligent Systems,2015,10(1):1-11.)
[4] Hsieh J. Computed tomography:principles,design,artifacts,and recent advances[M]. 2nd. Washington DC:SPIE Press,2009:36-41.
[5] Yu Tianshu,Wang Ruisheng. Scene parsing using graph matching on street-view data[J]. Computer Vision and Image Understanding,2016,145(4):70-80.
[6]孙万捷,高瞻,潘海燕,等.动态任务分配CUDA线程束步进体绘制[J].计算机辅助设计与图形学学报,2016,28(10):1630-1638.(Sun Wanjie,Gao Zhan,Pan Haiyan,et al. Volume rendering using dynamic CUDA warp marching[J]. Journal of Computer-Aided Design&Computer Graphics,2016,28(10):1630-1638.)
[7] Dzmitry S. Real-time 3D cone beam reconstruction[C]//Proc of IEEE Symposium Conference Record Nuclear Science. Piscataway,NJ:IEEE Press,2004:3648-3652.
[8] Zhao X. GPU-based 3D cone-beam CT image construction:application to micro CT[C]//Proc of IEEE Symposium Conference Record Nuclear Science. Piscataway,NJ:IEEE Press,2004:3648-3652
[9] KachelrieM. Hyperfast parallel-beam backprojection[C]//Proc of IEEE Nuclear Science Symposium Conference Record. Piscataway,NJ:IEEE Press,2006:3111-3114.
[10]Knaup M,KachelrieM. Real-time cone-beam CT image reconstruction using a mercury’s dual cell-based system(DCBS)and a Sony’s playstation 3(PS3)cluster[C]//Proc of IEEE Nuclear Science Symposium Record. Piscataway,NJ:IEEE Press,2007:3926-3928.
[11]孙涛,韩善清,汪家旺. PET/CT成像原理、优势及临床应用[J].中国医学物理学杂志,2010,27(1):1581-1582,1587.(Sun Tao,Han Shanqing,Wang Jiawang. Imaging theory,predominance and clinical applications of PET/CT[J]. Chinese Journal of Medical Physics,2010,27(1):1581-1582,1587.)
[12]柳有权,刘学慧,吴恩华.基于GPU带有复杂边界的三维实时流体模拟[J].软件学报,2006,17(3):568-576.(Liu Youquan,Liu Xuehui,Wu Enhua. Real-time 3D fluid simulation on GPU with complex obstacles[J]. Journal of Software,2006,17(3):568-576.)
[13] NVIDIA. NVIDIA CUDA compute unified device architecture:programming guide V1. 0[R/OL].(2007-06-23). http://developer.download. nvidia. com/compute/cuda/1. 0/NVIDIA_CUDA_Programming_Guide_1. 0. pdf.
[14]Seherl H,Keek B,Kowarsehik M,et al. Fast GPU-based CT reconstruction using the common unified device architecture(CUDA)[C]//Proc of IEEE Nuclear Science Symposium Conference Record.Piscataway,NJ:IEEE Press,2007:4464-4466.
[15]Pratx G,Chinn G,Olcott P D,et al. Fast,accurate and shift-varying line projections for iterative reconstruction using the GPU[J]. IEEE Trans on Medical Imaging,2009,28(3):435-445.
[16]Smith B D. Image reconstruction from cone-beam projection:necessary and sufficient condition and reconstruction methods[J]. IEEE Trans on Medical Imaging,1985,4(1):14-25.
[17]Muller K,Xu Fang. Practical considerations for GPU-accelerated CT[C]//Proc of IEEE International Symposium on Biomedical Imaging:Macro to Nano. Piscataway,NJ:IEEE Press,2006:1184-1187.
[2] Tang Min,Zhao Jieyi,Tong Ruofeng,et al. GPU accelerated convex hull computation[J]. Computers&Graphics,2012,36(5):498-506.
[3]丁科,谭营. GPU通用计算及其在计算智能领域的应用[J].智能系统学报,2015,10(1):1-11.(Ding Ke,Tan Ying. A review on general purpose computing on GPUs and its applications in computational intelligence[J]. CAAI Trans on Intelligent Systems,2015,10(1):1-11.)
[4] Hsieh J. Computed tomography:principles,design,artifacts,and recent advances[M]. 2nd. Washington DC:SPIE Press,2009:36-41.
[5] Yu Tianshu,Wang Ruisheng. Scene parsing using graph matching on street-view data[J]. Computer Vision and Image Understanding,2016,145(4):70-80.
[6]孙万捷,高瞻,潘海燕,等.动态任务分配CUDA线程束步进体绘制[J].计算机辅助设计与图形学学报,2016,28(10):1630-1638.(Sun Wanjie,Gao Zhan,Pan Haiyan,et al. Volume rendering using dynamic CUDA warp marching[J]. Journal of Computer-Aided Design&Computer Graphics,2016,28(10):1630-1638.)
[7] Dzmitry S. Real-time 3D cone beam reconstruction[C]//Proc of IEEE Symposium Conference Record Nuclear Science. Piscataway,NJ:IEEE Press,2004:3648-3652.
[8] Zhao X. GPU-based 3D cone-beam CT image construction:application to micro CT[C]//Proc of IEEE Symposium Conference Record Nuclear Science. Piscataway,NJ:IEEE Press,2004:3648-3652
[9] KachelrieM. Hyperfast parallel-beam backprojection[C]//Proc of IEEE Nuclear Science Symposium Conference Record. Piscataway,NJ:IEEE Press,2006:3111-3114.
[10]Knaup M,KachelrieM. Real-time cone-beam CT image reconstruction using a mercury’s dual cell-based system(DCBS)and a Sony’s playstation 3(PS3)cluster[C]//Proc of IEEE Nuclear Science Symposium Record. Piscataway,NJ:IEEE Press,2007:3926-3928.
[11]孙涛,韩善清,汪家旺. PET/CT成像原理、优势及临床应用[J].中国医学物理学杂志,2010,27(1):1581-1582,1587.(Sun Tao,Han Shanqing,Wang Jiawang. Imaging theory,predominance and clinical applications of PET/CT[J]. Chinese Journal of Medical Physics,2010,27(1):1581-1582,1587.)
[12]柳有权,刘学慧,吴恩华.基于GPU带有复杂边界的三维实时流体模拟[J].软件学报,2006,17(3):568-576.(Liu Youquan,Liu Xuehui,Wu Enhua. Real-time 3D fluid simulation on GPU with complex obstacles[J]. Journal of Software,2006,17(3):568-576.)
[13] NVIDIA. NVIDIA CUDA compute unified device architecture:programming guide V1. 0[R/OL].(2007-06-23). http://developer.download. nvidia. com/compute/cuda/1. 0/NVIDIA_CUDA_Programming_Guide_1. 0. pdf.
[14]Seherl H,Keek B,Kowarsehik M,et al. Fast GPU-based CT reconstruction using the common unified device architecture(CUDA)[C]//Proc of IEEE Nuclear Science Symposium Conference Record.Piscataway,NJ:IEEE Press,2007:4464-4466.
[15]Pratx G,Chinn G,Olcott P D,et al. Fast,accurate and shift-varying line projections for iterative reconstruction using the GPU[J]. IEEE Trans on Medical Imaging,2009,28(3):435-445.
[16]Smith B D. Image reconstruction from cone-beam projection:necessary and sufficient condition and reconstruction methods[J]. IEEE Trans on Medical Imaging,1985,4(1):14-25.
[17]Muller K,Xu Fang. Practical considerations for GPU-accelerated CT[C]//Proc of IEEE International Symposium on Biomedical Imaging:Macro to Nano. Piscataway,NJ:IEEE Press,2006:1184-1187.