high performance computing on graphics processing units: hgpu.org

This work explored the design considerations on the real-time medical ultrasound adaptive beamformer implementations using different computing platforms: CPU, GPU and FPGA. Adaptive beamforming has been well considered as an advanced solution for improving the image quality of medical ultrasound imaging machines. Although it provides promising improvements in lateral resolution, image contrast and imaging penetration depth, the use of adaptive beamforming is substantially more computationally demanding than conventional delay-and-sum beamformers. In order not to compromise the real-time performance of medical ultrasound systems, an accelerated solution is desirable. In this work, CPU implementation was used as a baseline implementation, based on which the intrinsic characteristics of the algorithm were analyzed. After the analysis of a particular adaptive beamforming algorithm, minimum-variance adaptive beamforming, two design parameters M and L were found to affect the implementation performance in two aspects: computational demand and image quality. The trends of the two aspects were contradictory with respect to the increment of M and L values. In our experiments, when M and L increased, the computational demand increased in a cubic curve; meanwhile, the image quality did not have much improvement when the increased values of M and L entered certain ranges. Since we targeted at a real-time solution without sacrificing the good image quality that adaptive beamforming proposed, a tradeoff was made on the selection of M and L values to balance the two contradictory requirements. Built upon the theoretical algorithmic analysis of the real-time adaptive beamformer realization, the implementations were developed with FPGA and GPU. While a dedicated hardware solution might be able to address the computational demand of the particular design, the need for an efficient algorithm exploration framework demanded a reprogrammable platform solution that was high-performance and easily reconfigurable. Besides, although a simple processor could provide convenient algorithm exploration via software development environment, real-time performance was usually not achievable. As a result, a reprogrammable medical ultrasound research platform for investigating advanced imaging algorithms was constructed in our project. The use of FPGA and GPU for implementing the real-time adaptive beamformer on our platform was explored. In our test cases, both FPGA- and GPUbased solutions achieved real-time throughput exceeding 80 frames-per-second, and over 38x improvement when compared to our baseline CPU implementation. Moreover, the implementations were also evaluated in terms of portability, data accuracy, programmability, and system integration. Due to its high power consumption, high-performance GPU solution is best suited for bedside applications, while FPGAs are more suitable for portable and hand-held medical ultrasound machines. Besides, while the development time on GPU platform remains much lower than its FPGA counterpart, the FPGA solution is effective in providing the necessary I/O bandwidth to enable an end-to-end real-time reconfigurable medical ultrasound image formation system.