基于GPU并行计算的超声波束合成方法

Ultrasound Beamforming Based on GPU Parallel Computation

逐列扫描式的聚焦成像方法会限制超声成像帧频的提高。采用平面波发射模式只需要一次发射和接收即可获得一幅完整的超声图像,平面波成像是可以大幅度地提高成像帧频,从而实现超声超高速成像的方法之一。但是现有的超声波束合成器无法达到超声超高速成像对于计算能力的要求。对基于平面波成像的超声延时-叠加波束合成算法进行并行性分析,在此基础上设计并实现两种基于图形处理器(GPU)并行计算的超声平面波成像波束合成方法——基于2个Kernel和基于1个Kernel的并行波束合成方法。两种方法的主要区别在于波束合成中对延时值的计算和存储策略的不同处理,仿体实验证明两种方法的计算帧频分别达到2 178和2 453帧/s。相比于普通的方法,这两种基于GPU的并行波束合成方法的计算帧频分别提速99倍和111倍。实验结果表明,GPU波束合成器相比于传统方法,可以大幅度提高计算能力。

The conventional line-by-line focused mode restricts the improvement of frame rate in ultrasound imaging. In plane wave imaging, each image is obtained by only one transmission and reception, and thus uhrafast imaging can be achieved. However, the current beamformers cannot meet the demand of uhrafast ultrasound imaging for massive computation. In this paper, the feasibility analysis of parallel computation in delay-and-sum beamforming algorithm was performed and two beamforming methods based on graphics processing unit （GPU） parallel computation including the 2 Kernel-based and 1 Kernel-based parallel beamformers were designed and implemented. The main differences between these methods were the calculations and data storage strategies of time delays in beamforming. Phantom experiments demonstrate that the computational frame rates of each method was 2,178 frames per second and 2,453 frames per second, respectively. Each of the methods obtained a speed up ratio of 99 and 111 compared to the normal method, which demonstrated that the GPU-based beamformer could significantly improve the calculating capability.