激光与光电子学进展, 2017, 54 (10): 102001, 网络出版: 2017-10-09
内镜超声合成孔径成像算法的并行实现 下载: 679次
Parallel Implementation of Synthetic Aperture Imaging Algorithm for Endoscopic Ultrasound
光计算 内镜超声成像 并行实现 合成孔径 统一计算设备架构 optics in computing endoscopic ultrasound imaging parallel implementation synthetic aperture compute unified device architecture
摘要
为了提高合成孔径成像算法在医学超声内镜系统中的计算效率, 提出一种在图形处理器(GPU)上并行实现的合成孔径成像方法。首先介绍了合成孔径算法的基本原理和图像重构过程; 然后对该算法进行并行化处理分析; 最后采用CUDA编程模式单指令多线程(SIMT)的灵活架构, 实现了基于GPU的内镜超声合成孔径成像算法。对多组散射点仿真成像实验进行对比分析, 并采用自行搭建的超声内镜实验系统对铁丝、肿囊假体及猪皮组织进行成像实验验证。实验结果表明, 所提方法在保证成像结果和成像质量不变的前提下, 大幅度提高了计算效率, 在计算数据规模为1.47 GB(5305×581×64×8 byte)时, 获得了50.93倍的最大加速比。
Abstract
In order to improve the computational efficiency of synthetic aperture imaging algorithm in the medical endoscopic ultrasound system, the synthetic aperture imaging approach with parallel implementation on graphics processing unit (GPU) is proposed. Firstly, the basic principle and image reconstruction process of synthetic aperture algorithm are introduced. Then, the algorithm is analyzed in parallel processing. Finally, synthetic aperture imaging algorithm for endoscopic ultrasound based on GPU is implemented by using the flexible architecture of single instruction multiple threads (SIMT) of compute unified device architecture (CUDA) programming mode. Multiple simulation experiments of scattering points imaging are compared and analyzed, and the imaging experiment verifications of iron wire, cyst prosthesis and pigskin tissue are carried out by using a self-built endoscopic ultrasound experimental system. The experimental results show that the proposed method can greatly improve the computational efficiency while keeping the same imaging quality and results. When the calculated data size is 1.47 GB (5305×581×64×8 byte), the maximum speedup ratio reaches 50.93.
李溦, 陈晓冬, 李嘉科, 汪毅, 郁道银. 内镜超声合成孔径成像算法的并行实现[J]. 激光与光电子学进展, 2017, 54(10): 102001. Li Wei, Chen Xiaodong, Li Jiake, Wang Yi, Yu Daoyin. Parallel Implementation of Synthetic Aperture Imaging Algorithm for Endoscopic Ultrasound[J]. Laser & Optoelectronics Progress, 2017, 54(10): 102001.