基于SIMD技术的锥束ART算法快速并行图像重建

Fast parallel image reconstruction with cone-beam ART algorithm based on SIMD technology

ART(algebraic reconstruction technique)算法是一种迭代图像重建方法,适合于大型工业构件的无损检测,其缺点是计算量大、重建时间长。为了提高锥束ART算法的重建速度,本文提出一种快速并行图像重建方法。首先根据锥束CT扫描方式下三维射线的对称性提出一种权因子和体素索引的并行计算方法,通过一次计算可同时得到两条射线的权因子和体素索引;然后采用Intel处理器的单指令多数据(single instruction multiple data,SIMD)技术,一次性加载多个打包数据,利用SSE(streamingSIMD extension)指令实现了投影、计算图像校正和反投影的并行运算。实验结果表明本文提出的方法非常有效,在保证图像重建精度的同时取得了约1.5倍的重建加速比。

Algebraic reconstruction technique（ART） is an iterative image reconstruction method,which is suitable for nondestructive detection of large industrial components.The shortcomings of ART are large amount of computation and extremely time cost.In order to improve the reconstruction speed of cone-beam ART,this paper presents a fast parallel image reconstruction method.Firstly,we propose a parallel method to calculate the weight coefficients and voxel indexes according to the symmetry of 3D ray in the scanning mode of cone-beam CT,which can get the weight coefficients and voxel indexes of two rays at the same time.Then,we use SIMD technology of Intel CPU,which can load a number of packed data at one time and realize parallel implementation of projection,calculation of image correction and back projection using SSE instructions.Experimental results show that our method is very effective and a reconstruction speedup about 1.5 times can be obtained without any loss of image reconstruction precision.