Design and evaluation of medical ultrasonic adaptive beamforming algorithm implementation on heterogeneous embedded computing platform
© The Author(s) 2017
Received: 29 July 2016
Accepted: 3 January 2017
Published: 23 January 2017
Medical ultrasonic imaging has been utilized in a variety of clinical diagnoses for many years. Recently, because of the needs of portable and mobile medical ultrasonic diagnoses, the development of real-time medical ultrasonic imaging algorithms on embedded computing platforms is a rising research direction. Typically, delay-and-sum beamforming algorithm is implemented on embedded medical ultrasonic scanners. Such algorithm is the easiest to implement at real-time frame rate, but the image quality of this algorithm is not high enough for complicated diagnostic cases. As a result, minimum-variance adaptive beamforming algorithm for medical ultrasonic imaging is considered in this paper, which shows much higher image quality than that of delay-and-sum beamforming algorithm. However, minimum-variance adaptive beamforming algorithm is a complicated algorithm with O(n 3) computational complexity. Consequently, it is not easy to implement such algorithm on embedded computing platform at real-time frame rate. On the other hand, GPU is a well-known parallel computing platform for image processing. Therefore, embedded GPU computing platform is considered as a potential real-time implementation platform of minimum-variance beamforming algorithm in this paper. By applying the described effective implementation strategies, the GPU implementation of minimum-variance beamforming algorithm performed more than 100 times faster than the ARM implementation on the same heterogeneous embedded platform. Furthermore, platform power consumptions, computation energy efficiency, and platform cost efficiency of the experimental heterogeneous embedded platforms were also evaluated, which demonstrated that the investigated heterogeneous embedded computing platforms were suitable for real-time portable or mobile high-quality medical ultrasonic imaging device constructions.
KeywordsEmbedded GPU implementation Medical ultrasonic adaptive beamforming High-performance computing
There are several useful medical imaging modalities for clinical diagnoses, which are radiography , magnetic resonance imaging , nuclear imaging , ultrasonic imaging , and computed tomography . Among the mentioned medical imaging modalities, medical ultrasonic imaging is a very common imaging technique which has been widely applied to a vast range of clinical diagnoses for many years . Medical ultrasonic imaging has overwhelming advantages over other medical imaging modalities, such as real-time imaging at a smooth video frame rate, high safety without electromagnetic radiation, and relatively low cost as compared to other medical imaging modalities . Therefore, medical ultrasonic imaging is usually utilized to observe heart movements , fetal developments , blood flows , and so on. Recently, as the needs for portable and mobile medical ultrasonic diagnoses increase, the development of medical ultrasonic imaging algorithms on embedded computing platforms is a rising research direction. The key evaluation feature of an embedded medical ultrasonic imaging algorithm implementation is its real-time imaging capability. A medical ultrasonic imaging algorithm can be used for real-world clinical diagnostic applications only if its implementation can run at a real-time video frame rate.
However, these high-quality imaging algorithms are computationally demanding and difficult to implement in real-time, especially on embedded platforms with limited computing resources. The real-time implementation of MV adaptive beamforming is challenging, whose computational complexity is O(n 3) in sequential implementation . Among various embedded computing platforms, field-programmable gate arrays (FPGAs) are good potential development platforms to implement real-time medical ultrasonic imaging algorithms. However, an FPGA platform with sufficient computational resources to implement a regular high-quality medical ultrasonic imaging algorithm at real-time video frame rate is usually of high price. Hence, in order to reduce the implementation cost of a portable or mobile medical ultrasonic imaging device, alternative high-performance embedded computing platform with relatively lower implementation cost is required. As a result, the heterogeneous embedded computing platforms with high-performance graphics processing units (GPUs) and advanced RISC machine (ARM) processors will be investigated in this paper.
In this paper, the implementation of MV adaptive beamforming algorithm on high-performance embedded GPU computing platform will be discussed. The implementation strategies for a high-performance GPU on a heterogeneous embedded computing platform will be described, and the performance of the GPU implementation and its ARM processor counterpart on the same embedded computing platform will be evaluated. The performance features evaluated in this paper will include the algorithm computing speed, the computation energy efficiency, and the platform cost efficiency. The following section will interpret the detailed sequential calculation steps of the medical ultrasonic MV adaptive beamforming algorithm. Then, the implementation design of the MV adaptive beamforming algorithm on the embedded GPU computing platform will be described in Section 3. Section 4 will illustrate the experimental setup and the performance evaluations of the embedded implementations. Finally, the paper concludes in Section 5.
2 Medical ultrasonic adaptive beamforming sequential calculation
The MV adaptive beamforming algorithm starts with the input matrices which contain unaligned ultrasonic echo signals from a series of ultrasonic echo channels. The number of the ultrasonic echo channels is denoted as M, which is also called the receive aperture of the ultrasonic echo signals. In order to obtain the aligned (M×1) echo signal vector e c h o(p s ) for a specific image pixel p s , the algorithm calculates the delay information of all the ultrasonic echo channels for the specific image pixel.
where a is a steering vector with all ones, which is because the ultrasonic echo signals used for covariance matrix calculation are already delayed and aligned.
The pixel amplitude estimate value v(p s ) is the output of the MV adaptive beamforming algorithm. The calculation steps repeat until all the pixel amplitude estimate values of the whole image are obtained.
3 Implementation design on embedded GPU computing platform
3.1 Heterogeneous embedded computing platform
Although MV adaptive beamforming algorithm outputs high-quality medical ultrasonic images, it is computationally demanding. The high computational complexity of MV adaptive beamforming algorithm hinders MV algorithm being implemented in real-time on conventional embedded computing platforms, such as conventional ARM processors. Therefore, the implementation of MV adaptive beamforming algorithm on heterogeneous embedded computing platform with high-performance GPU is explored in this paper, so as to validate the real-time imaging capability of MV adaptive beamforming algorithm on embedded platforms.
3.2 Implementation strategies for embedded GPU
In order to effectively utilize the high-performance embedded GPU in the heterogeneous embedded computing platform to implement the MV adaptive beamforming algorithm efficiently, the following implementation strategies are applied.
3.2.1 Allocation of GPU computing resources
The first implementation strategy is about the allocation of the GPU computing resources. According to the GPU compute unified device architecture (CUDA) programming principles, the GPU CUDA programming model consists of three programming hierarchy levels, i.e., GPU compute grid, GPU compute block, and GPU compute thread. At the top level of the programming hierarchy, all the algorithm computations are executed within one GPU compute grid. Meanwhile, at the second level of the programming hierarchy, the program tasks are allocated into a set of GPU compute blocks. The computation in different GPU compute blocks can be executed in parallel. Besides, at the third level of the programming hierarchy, the computational workloads are assigned to a series of GPU compute threads. The programs in different GPU compute threads are executed simultaneously, while the program instructions within one GPU compute thread are executed sequentially.
3.2.2 Overall design overview
In the design, the MV beamforming implementation took M receive channels as its inputs to generate an amplitude estimation of one image pixel. The input data from the ultrasonic echo receive channels was also known as pre-beamform data before the beamforming process, and the output pixel value was also known as post-beamform data after the beamforming process.
Each receive channel streamed input echo samples to the delay calculation block, forming an M×1 vector of delayed echo samples as its output. The purpose of the delay calculation block was to align the input ultrasonic echoes based on the delay information among the receive channels. The delayed echo vector must subsequently be multiplied by the adaptive apodization weights. Hence, while the adaptive weight calculation block was calculating the adaptive weights, the echo vector was stored in the data storage buffer for later usage, which was built with GPU shared memory. The purpose of data storage buffer was to let the delayed echo vector wait until the adaptive weight calculation block finished its workload.
Finally, the pixel amplitude estimation block outputs the final pixel value. It first multiplied (M−L+1) segmented delayed echo vectors e c h o subk(p s )(k=0,1,…,M−L) and their adaptive weights. The results of these (M−L+1) pixel value estimates were then averaged to obtain the final pixel amplitude value output.
3.2.3 Adaptive weight calculation parallelization
In order to implement a highly parallelized MV algorithm, probability theories and linear algebra theories were used to optimize the detailed implementation and reduce the computation operations. The integration of the mathematical theories into the implementation will be described in the following three parts.
The adaptive weight calculation block in Fig. 6 performs the core computation of the MV adaptive beamforming algorithm. It consists of three major units: covariance matrix calculation, linear equation solver, and final weight calculation. Here, the inner working principles of these blocks will be elaborated.
18.104.22.168 Covariance matrix calculation
As a result, only the diagonal elements and the lower (L) or upper (U) triangular matrix elements of the covariance matrix need to be calculated. Therefore, L×(L+1)/2 calculations are needed in stead of L×L calculations. Taking the advantage of the symmetry makes the covariance matrix implementation nearly twice faster.
22.214.171.124 Linear equation solver
Therefore, using a system solver reduces extra operation time and storage resources. The covariance matrix has positive-semidefinite and symmetric properties , hence, Cholesky decomposition  is applicable to this weight solver in regular MV algorithm. Cholesky decomposition is derived from Gaussian elimination, but halves the decomposition operations and is more stable than L U decomposition which is the matrix form of Gaussian elimination. L D L T decomposition form of the Cholesky decomposition was adopted.
Since the weight solver was iterative, the iterations cannot be parallelized. But inside the iterations, parallelization was achievable. For example, the L matrix was formed column by column and only one column in one iteration, but the element calculations within each column of L could be parallelized.
126.96.36.199 Final weight calculation
As a result, the final weight calculation step can be parallelized.
3.2.4 Assignment of GPU memory access
The memory access strategy of GPU implementation has an important impact on the overall GPU computational speed. There are basically three types of memory modules inside the GPU, i.e., global memory, shared memory, and register files. The lifetime of the data in the three memory types is associated to the GPU computing resources respectively. The lifetime of the data in the global memory is associated to the GPU compute grid, the lifetime of the data in the shared memory is associated to the GPU compute block, and the lifetime of the data in the register files is associated to the GPU compute thread.
These three types of memory modules are fabricated according to different architectural hierarchies. The register files reside right next to the GPU processing cores, the shared memory locates in a place with a farther distance to the GPU processing cores, and the global memory locates farthest to the GPU cores. The distance to the GPU cores determines the access speed of the specific memory type. As a result, the access speed of the register files is the fastest, the access speed of the shared memory is slower than that of the register files, and the access speed of the global memory is the slowest among the three types of memory modules. However, on the other hand, the memory size of a specific memory type is inversely proportional to its memory access speed. Thus, the size of the global memory is the largest, the size of the register files is the smallest, and the size of the shared memory is between the size of the global memory and the size of the register files. Therefore, in a GPU program memory assignment, small-size and frequently-used variables can be stored in the register files, while large amount of data should be stored in the global memory. Furthermore, the shared memory is usually used for the shared data space utilized in the computation among the GPU compute threads within a specific GPU compute block.
4 Implementation evaluations and discussions
4.1 Experimental setup and simulated scenario
The experimental heterogeneous embedded computing platforms used in our experiments were Jetson TX1 and TK1 evaluation embedded platforms from Nvidia Corporation. Both evaluation boards utilized similar heterogeneous embedded computing platform architecture modules as illustrated in Fig. 3. The major differences between the two evaluation boards were the different GPU module and the different ARM processor module. TX1 platform was using a GPU module of Maxwell GPU architecture and a quad-core Cortex-A57 ARM processor , while TK1 platform was using a GPU of Kepler GPU architecture and a quad-core Cortex-A15 ARM processor . The computational capabilities and the power consumption requirements of the two heterogeneous embedded computing platforms were different, which will be shown in Section 4.2.
In order to mimic the real-world medical ultrasonic imaging applications, the Field II simulation tool simulated a 128-element medical ultrasonic transducer with 0.3048-mm element pitch. The pulse repetition rate was set as 5 KHz and the sampling rate was 40 MHz. These parameter settings referred to the real-world medical ultrasonic transducer characteristics. During the simulation of the medical ultrasonic imaging process, the simulation program first simulated the emission process of the medical ultrasonic imaging and then simulated the reception process of the medical ultrasonic imaging. The output of the simulation program was the expected unaligned medical ultrasonic echo signal matrices which were adopted as the input for the MV adaptive beamforming algorithm implementation.
4.2 Evaluations and discussions
4.2.1 Algorithm computing speed
The computing speed evaluation of the MV adaptive beamforming algorithm implementation comprised (a) the giga floating-point operations per second (GFLOPS), (b) the output image frame rate of the algorithm implementation, and (c) the computational speedup of the embedded GPU implementation over its ARM processor counterpart. The algorithm computing speed evaluation was conducted on Jetson TX1 heterogeneous embedded development board. The evaluation test cases included various M and L combinations. While doing the evaluation tests, the value of M was first determined, which was chosen as 16, 32, 64, and 128, respectively. Then, the value of L was chosen from 1 to M/2. Therefore, there were 8 M and L combinations when M is 16, 16 M and L combinations when M is 32, 32 M and L combinations when M is 64, and 64 M and L combinations when M is 128. In the experiments, as the number of image pixels, 127×1000 pixels for one image, did not change during the tests, the number of the GPU compute blocks were not changed in the experiments. But the number of the GPU compute threads, a.k.a. GPU compute thread block size, was varying during the experiments. The algorithm computing speed of various GPU compute thread block size was tested so as to find the value of the GPU compute thread block size in the best practices. The GPU compute thread block size test cases covered the numbers from 32 to 1024 which were multiples of 32.
where Timeframe was the algorithm computational time to obtain one frame of image, Timepixel was the algorithm computational time to obtain one image pixel amplitude estimate value, and Numberpixel was the number of the pixels in one image frame.
In the experiments,the shortest algorithm computational time was 0.017 s in the case that M was 16, L was 2, and the GPU compute thread block size was 32. Thus, the highest output image frame rate of the MV adaptive beamforming algorithm implementation was 1/0.017= 58.82 frame/s (fps). Since the number of image pixels was 127×1000=127,000, the fastest pixel computation time for the MV adaptive beamforming algorithm was 0.017/127,000=1.34×10−7 s. Therefore, when the number of the pixels Numberpixel changed, the output image frame rate changed, which was expressed as the following: F r a m e r a t e=1/(1.34×10−7×Numberpixel). The output image frame rate was higher than those in the experiments when Numberpixel was smaller than 127,000, but the output image frame rate was lower than those in the experiments when Numberpixel was larger than 127,000. Therefore, the real-time imaging capability of the MV adaptive beamforming algorithm implementation was not only associated with Timepixel but also associated with Numberpixel. In the experiments with Numberpixel=127,000, many test cases for various M and L combinations achieved real-time video imaging frame rate for medical ultrasonic imaging, which was higher than 20 fps.
4.2.2 Computation energy efficiency
where C o m p u t a t i o n a l p e r f o r m a n c e could be any of the algorithm computing performance evaluation features discussed in Section 4.2.1, but GFLOPS and image output frame rate of the MV adaptive beamforming algorithm were usually adopted. Besides, C o n s u m e d p o w e r was defined as the average power consumption while the embedded computing platform was fully-loaded for the computation.
Average power consumptions of idle and fully loaded embedded platforms
Idle power (Watt)
Fully loaded power (Watt)
Referring to (11) and the results discussed in Section 4.2.1, the highest performance-energy efficiency of Jetson TX1 heterogeneous embedded computing platform could be calculated as 88.98/12.2= 7.29 GFLOPS/Watt or 58.82/12.2= 4.80 fps/Watt. Besides, the highest operation calculation throughput of the MV adaptive beamforming algorithm on Jetson TK1 heterogeneous embedded computing platform was 37.52 GFLOPS in the case that M was 128, L was 32, and the GPU compute thread block size was 128, and the highest output image frame rate of the MV adaptive beamforming algorithm on Jetson TK1 heterogeneous embedded computing platform was 17.24 fps in the case that M was 16, L was 1, and the GPU compute thread block size was 64. As a result, the highest performance-energy efficiency of Jetson TK1 heterogeneous embedded computing platform was calculated as 37.52/9.3= 4.03 GFLOPS/Watt or 17.24/9.3= 1.85 fps/Watt. Therefore, the performance-energy efficiency results of the two heterogeneous embedded computing platform demonstrated that the embedded computing platform consumed more power did not mean that it had a lower performance-energy efficiency. In the experiments, the TX1 embedded computing platform with higher average fully-loaded power consumption exhibited higher performance-energy efficiency.
4.2.3 Platform cost efficiency
where C o m p u t a t i o n a l p e r f o r m a n c e could be GFLOPS throughput or image output frame rate of the MV adaptive beamforming algorithm. Furthermore, the cost of the heterogeneous embedded computing platform P l a t f o r m c o s t was measured in US dollars.
As stated in the official website of Nvidia Corporation, Jetson TX1 heterogeneous embedded computing platform cost $600 and Jetson TK1 heterogeneous embedded computing platform cost $200. When considering the highest computational performance of the heterogeneous embedded computing platforms, the highest cost efficiency of the Jetson TX1 heterogeneous embedded computing platform was expressed as 88.98/600= 0.148 GFLOPS/dollar or 58.82/600= 0.098 fps/dollar. Similarly, the highest cost efficiency of the Jetson TK1 heterogeneous embedded computing platform was expressed as 37.52/200= 0.188 GFLOPS/dollar or 17.24/200= 0.086 fps/dollar. Hence, the Jetson TX1 embedded computing platform has lower platform cost efficiency as compared to Jetson TK1 embedded computing platform in terms of GFLOPS throughput, and slightly higher platform cost efficiency as compared to Jetson TK1 embedded computing platform in terms of image output frame rate performance.
As a result, the highest performance-power-cost efficiency of Jetson TX1 heterogeneous embedded computing platform was 7.29/600= 0.01215 GFLOPS/Watt/dollar or 4.80/600= 0.00800 fps/Watt/dollar. Similarly, the highest performance-power-cost efficiency of the Jetson TK1 heterogeneous embedded computing platform was 4.03/200= 0.02015 GFLOPS/Watt/dollar or 1.85/200= 0.00925 fps/Watt/dollar. Therefore, the Jetson TK1 embedded computing platform has higher performance-power-cost efficiency over Jetson TX1 embedded computing platform in both computing performance measurement means.
In this paper, the MV adaptive beamforming algorithm implementation on heterogeneous high-performance embedded computing platforms was investigated. According to the effective high-performance GPU implementation strategies, the MV adaptive beamforming algorithm implementation in high-performance embedded GPU on Jetson TX1 platform can fulfil the real-time imaging requirements for medical ultrasonic imaging in many test cases, and can achieve 102.7 times speedup over its ARM processor counterpart. Besides, the power consumptions of the two experimental heterogeneous high-performance embedded platforms illustrated that the heterogeneous embedded computing platforms can be used as portable or mobile medical ultrasonic devices. Furthermore, the computation energy efficiency and platform cost efficiency of the heterogeneous embedded computing platforms demonstrated that the heterogeneous embedded computing platforms had relatively good implementation efficiency. As a result, the heterogeneous embedded computing platforms investigated in this paper were suitable to construct real-time portable or mobile high-quality medical ultrasonic imaging devices, especially the Jetson TX1 platform.
This work is supported by “Guangdong Natural Science Foundation” (No. 2016A030310412), “Guangzhou Science and Technology Program” (Key Laboratory Project, No. 15180007), and “Guangzhou Science and Technology Program” (No. 201605130108484).
The authors declare that they have no competing interests.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- E Quaia, E Baratella, G Poillucci, AG Gennari, MA Cova, Diagnostic impact of digital tomosynthesis in oncologic patients with suspected pulmonary lesions on chest radiography. European Radiology. 26(8), 2837–2844 (2016).View ArticleGoogle Scholar
- S Tang, Q Du, T Liu, L Tan, M Niu, L Gao, Z Huang, C Fu, T Ma, X Meng, H Shao, In vivo magnetic resonance imaging and microwave thermotherapy of cancer using novel chitosan microcapsules. Nanoscale Res. Lett.11(334), 1–8 (2016).Google Scholar
- C Love, CJ Palestro, Nuclear medicine imaging of bone infections. Clin. Radiol.71(7), 632–646 (2016).View ArticleGoogle Scholar
- BK Hoffmeister, MR Smathers, CJ Miller, JA McPherson, CR Thurston, PL Spinolo, S-R Lee, Backscatter-difference measurements of cancellous bone using an ultrasonic imaging system. Ultrason. Imaging. 38(4), 285–297 (2016).View ArticleGoogle Scholar
- D Karimi, P Deman, R Ward, N Ford, A sinogram denoising algorithm for low-dose computed tomography. BMC Med. Imaging. 16(11), 1–14 (2016).Google Scholar
- JF Havlice, JC Taenzer, Medical ultrasonic imaging: an overview of principles and instrumentation. Proc. IEEE. 67(4), 620–641 (1979).View ArticleGoogle Scholar
- RSC Cobbold, Foundations of biomedical ultrasound (Oxford University Press, New York, USA, 2007).Google Scholar
- B Heyde, M Alessandrini, J Hermans, D Barbosa, P Claus, J D’Hooge, Anatomical image registration using volume conservation to assess cardiac deformation from 3D ultrasound recordings. IEEE Trans. Med. Imaging. 35(2), 501–511 (2016).View ArticleGoogle Scholar
- L Zheng, L Gong, F-C Guo, H Chang, G Liu, Application research on three-dimensional ultrasonic skeletal imaging mode in detecting fetal upper jaw bone. Int. J. Clin. Exp. Med.8(8), 12219–12225 (2015).Google Scholar
- A Bar-Zion, M Yin, D Adam, FS Foster, Functional flow patterns and static blood pooling in tumors revealed by combined contrast-enhanced ultrasound and photoacoustic imaging. Cancer Res.76(15), 4320–4331 (2016).View ArticleGoogle Scholar
- J-F Synnevag, A Austeng, S Holm, Benefits of minimum-variance beamforming in medical ultrasound imaging. IEEE Trans. Ultrason. Ferroelectr. Freq. Control. 56(9), 1868–1879 (2009).View ArticleGoogle Scholar
- Y Li, JA Jensen, Synthetic aperture flow imaging using dual stage beamforming: simulations and experiments. J. Acoust. Soc. Am.133(4), 2014–2024 (2013).View ArticleGoogle Scholar
- JA Jensen, NB Svendsen, Calculation of pressure fields from arbitrarily shaped, apodized, and excited ultrasound transducers. IEEE Trans. Ultrason. Ferroelectr. Freq. Control.39(2), 262–267 (1992).View ArticleGoogle Scholar
- J-F Synnevag, A Austeng, S Holm, Adaptive beamforming applied to medical ultrasound imaging. IEEE Trans. Ultrason. Ferroelectr. Freq. Control.54(8), 1606–1613 (2007).View ArticleGoogle Scholar
- A Papoulis, SU Pillai, Probability, random variables, and stochastic processes, 4th edn. (McGraw-Hill, Boston, 2002).Google Scholar
- GH Golub, CF Van Loan, Matrix computations, 3rd edn. (Johns Hopkins University Press, Baltimore, 1996).MATHGoogle Scholar
- Nvidia Jetson TX1 developer kit carrier board specification (Nvidia CorporationSanta Clara, 2016).
- Nvidia Tegra K1 series processors with Kepler Mobile GPU for embedded applications data sheet (Nvidia CorporationSanta Clara, 2015).