高分辨锥束CT并行重建算法在基于NVDIA GPU显卡计算平台上的实现

A Practice on Parallel Reconstruction Algorithm of High Resolution Cone Beam Micro-CT Based on NVDIA GPU Graphic Card

目的:探讨高分辨率锥束显微CT断层重建中引入并行计算的必要性及其加速效果。方法:在具有并行计算功能的GPU显卡(NVIDIA QUADRO K5000显卡,显存4G)中为投影图像和重建体数据分配显存空间,每一个像素分配一个线程进行投影图像的各种校正和滤波,再给每个体素分配一线程进行反投影重建,在显存中实现全部断层重建。程序使用C++面向对象方法实现,内核函数用CUDA实现。结果:重建体数据大小是2 048×2 048×128,每个体素用32位浮点数记录,实验采集1 800张投影,每张投影图像大小为2 048×1 536,重建时间小于9 min,是图像采集时间的1/3,是基于CPU重建耗时的2%。将GPU并行重建得到的图像和CPU单线程重建图像结果进行对比,数据结果一致,满足实验设计的要求。结论:并行计算引入高分辨锥束CT重建可大大提高重建速度,并且能实现采集与重建同步进行。

Objective： To explore the feasibility of parallel computing applying in high-resolution cone beam micro-CT reconstruction and its impact on reconstruction speed. Method： Allocating video memory to projection pictures and reconstruction voxels in GPU graphic card （NVIDIA QUADRO K5 000, Video Memory 4G） with parallel computing, and allocating thread to each pixel for adjusting and filtering projection picture, and then allocating thread to each voxel for Back Projection, thus, all section reconstruction is implemented in graphic card. Result： Less than 9 minutes spent for 2 048 × 2 048 × 128 pixel matrix reconstruction, which is equal to 1/3 of data gathering tirne and 2% of CPU based reconstruction, under the condition that one voxel is recorded by 32 float, and each projection picture size is 2 048 × 1 536, and 1 800 projections are gained in one scanning. Conclusion： Parallel computing applied in cone beam CT reconstruction can greatly increase reconstruction speed and data gathering can simultaneously operate with reconstruction.