基于GPGPU的锥束螺旋CT DBPF重建算法的实现与加速方法研究  

Implementation techniques and acceleration of DBPF reconstruction algorithm based on GPGPU for helical cone beam CT

作　　者：沈乐[1] 邢宇翔[1,2] 

机构地区：[1]清华大学工程物理系,北京100084 [2]清华大学粒子技术与辐射成像教育部重点实验室,北京100084

出　　处：《核技术》2010年第11期857-862,共6页Nuclear Techniques

基　　金：国家自然科学基金(60772051)资助

摘　　要：基于PI线的微分反投影滤波(DBPF)算法是新型的锥束螺旋X射线CT精确重建算法,然而,由于计算量大,在CPU上重建速度缓慢。通用图形处理器(GPGPU)是一种SIMD并行硬件架构,具有强大的浮点运算能力。本文介绍一种新的PI线选取和采样方法,利用美国NVIDIA公司的计算统一设备架构(CUDA)为DBPF算法的PI线重建提供加速优化策略。使用模拟数据的重建实验进行方法的性能测试,得到了满意的图像质量,重建速度也显著提升(318倍加速比)。还讨论了重建参数对重建图像质量和计算速度的影响,建议了参数选择方式。The derivative back-projection filtered algorithm for a helical cone-beam CT is a newly developed exact reconstruction method. Due to its large computational complexity, the reconstruction is rather slow for practical use. General purpose graphic processing unit （GPGPU） is an SIMD paralleled hardware architecture with powerful float-point operation capacity. In this paper, we propose a new method for PI-line choice and sampling grid, and a paralleled PI-line reconstruction algorithm implemented on NVIDIA＇s Compute Unified Device Architecatre （CUDA） Numerical simulation studies are carried out to validate our method. Compared with conventional CPU implementation, the CUDA accelerated method provides images of the same quality with a speedup factor of 318. Optimization strategies for the GPU acceleration are presented. Finally, influence of the parameters of the PI-line samples on the reconstruction speed and image quality is discussed.

关 键 词：X射线光学 CT重建 反投影滤波 PI线 通用图形处理器 CUDA 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]