Siyuan Wei

Sonic Millip3De: A massively parallel 3D-stacked accelerator for 3D ultrasound

Three-dimensional (3D) ultrasound is becoming common for non-invasive medical imaging because of its high accuracy, safety, and ease of use. Unlike other modalities, ultrasound transducers require little power, which makes hand-held imaging platforms possible, and several low-resolution 2D devices are commercially available today. However, the extreme computational requirements (and associated power requirements) of 3D ultrasound image formation has, to date, precluded hand-held 3D capable devices. We describe the Sonic Millip3De, a new system architecture and accelerator for 3D ultrasound beamformation-the most computationally intensive aspect of image formation. Our three-layer die-stacked design features a custom beamsum accelerator that employs massive data parallelism and a streaming transform-select-reduce pipeline architecture enabled by our new iterative beamsum delay calculation algorithm. Based on RTL-level design and floorplanning for an industrial 45nm process, we show Sonic Millip3De can enable 3D ultrasound with a fully sampled 128×96 transducer array within a 16W full-system power budget (400× less than a conventional DSP solution) and will meet a 5W safe power target by the 11nm node.

High Frame Rate 3-D Ultrasound Imaging Using Separable Beamforming

Recently, there has been great interest in 3-D ultrasound imaging, but power constraints have precluded practical implementation of high-resolution and high-frame-rate 3-D ultrasound in handheld imaging platforms. In this paper, we propose a separable beamforming procedure for both 3-D synthetic aperture and plane wave systems that drastically reduces computational and hence power requirements. Separable beamforming approximates 2-D array beamforming for 3-D images through a series of beamforming operations on 1-D arrays. Our proposed method is based on a separable delay decomposition method that minimizes phase error. We show that the proposed separable synthetic aperture system achieves 19-fold complexity reduction and the proposed plane wave separable system achieves 12-fold complexity reduction compared to the corresponding non-separable beamforming baseline systems. Furthermore, we verify the performance of the fixed-point-precision separable beamforming and iterative delay calculation through Field II simulations. Our results show that both the synthetic aperture system and the plane wave system can produce images with the same quality as images generated by non-separable beamforming. We also briefly describe how the two types of separable beamformer can be implemented on a modified version of Sonic Millip3De, our recently proposed hardware accelerator for the digital front-end of a 3-D ultrasound system.

Separable Beamforming For 3-D Medical Ultrasound Imaging

Three-dimensional ultrasound imaging is a promising medical imaging technology because of its ease of use and improved accuracy in diagnosis. However, its high computational complexity and resulting high power consumption has precluded its use in hand-held applications. In this paper, we present a separable beamforming method that greatly reduces computational complexity. Our method is based on decomposing the delay term in a way that minimizes the root-mean-square error caused by the decomposition. We analyze tradeoffs between the approximation error caused by the decomposition and computational complexity. Then, we present enhancements to the Sonic Millip3De hardware accelerator for ultrasound beamforming to implement separable beamforming. Using hardware synthesis targeting standard cells in 45 nm, we show that the proposed method allows us to boost the Sonic Millip3De frame rate from 1-2 Hz to 32 Hz while maintaining power consumption at 15 W. We validate image quality of our method using cyst phantom simulations in Field II. Our evaluation demonstrates that the proposed separable beamforming method can produce 3-D images with high quality that are comparable to those generated by non-separable beamforming.

Sonic Millip3De with dynamic receive focusing and apodization optimization

3D ultrasound is becoming common for non-invasive medical imaging because of its accuracy, safety, and ease of use. However, the extreme computational requirements (and associated power requirements) of image formation for a large 3D system have, to date, precluded hand-held 3D-capable devices. Sonic Millip3De is a recently proposed hardware design that leverages modern computer architecture techniques, such as 3D die stacking, massive parallelism, and streaming data flow, to enable high-resolution synthetic aperture 3D ultrasound imaging in a single, low-power chip. In this paper, we enhance Sonic Millip3De with a new virtual source firing sequence and dynamic receive focusing scheme to optimize receive apertures in multiple depth focal zones. These enhancements further reduce power requirements while maintaining image quality over a large depth range. We present image quality analysis using Field II simulations of cysts in tissue at varying depths to show that our methods do not degrade CNR relative to an ideal system with no power constraints. Then, using RTL-level design for an industrial 45nm ASIC process, we demonstrate 3D synthetic aperture with 120×88 transducer array within a 15W full-system power budget (400x less than a conventional DSP solution). We project that continued semicondutor scaling will enable a sub-5W power budget in 16nm technology.

FPGA implementation of low-power 3D ultrasound beamformer

3D ultrasound is common for non-invasive medical imaging in cardiology and OB-GYN because of its accuracy, safety, and real-time ease of use. However, high bandwidth requirements and extreme computational complexity have precluded hand-held and low-power 3D systems, limiting 3D applications. In previous work, we presented Sonic Millip3De, a hardware design that can efficiently handle the high computational demand of real-time 3D synthetic aperture beamforming, even in handheld and mobile applications. The design combines a custom, highly parallel hardware system with a novel delay approximation method to quickly produce high quality 3D image data within an estimated 15 W full-system power budget. Prior evaluations of the design relied on software prototypes; this work extends previous evaluations with an FPGA implementation of the beamforming accelerator, validating the results of earlier prototypes. In particular, we carry out image quality analyses of our beamforming architecture using simulated 3D echo data (from Field II) and 2D artificial tissue phantom data acquired using a Verasonics V-1 system and Philips P4-1 probe. We compare results from the FPGA implementation to an ideal software beamformer and prior software prototypes of the Sonic Millip3De design.

A low complexity scheme for accurate 3D velocity estimation in ultrasound systems

Vector flow imaging is a critical component in the clinical diagnosis of cardiovascular diseases; however, most current methods are too computationally expensive to scale well to 3D. Less complex techniques, such as Doppler-based imaging (which cannot provide lateral flow measurements) and basic speckle tracking algorithms (which have poor lateral accuracy), are incapable of producing high quality 3D measurements. In this paper, we first extend a technique designed to improve lateral flow accuracy for 2D velocity vector estimation, the synthetic lateral phase method, to 3D (SLP-3D). We then show that a straightforward implementation of this algorithm is too computationally complex for modern systems. Instead, we propose a two-tiered method that uses low complexity sum-of-absolute differences (SAD) for coarse-grained search and an optimized version of SLP-3D to fine tune the search for sub-pixel accuracy. We show that the proposed method (SAD+SLP-3Dopt) achieves a 9× reduction in computational complexity compared to the naive SLP-3D. Field II simulations for plug and parabolic flow using our method show a fairly high degree of accuracy in both the axial and the lateral components. Finally, we show our technique can support accurate flow imaging with up to 130 velocity estimations/sec within the power constraints of a handheld device.

Sonic Millip3De: An Architecture for Handheld 3D Ultrasound

3D ultrasound is becoming common for noninvasive medical imaging because of its high accuracy, safety, and ease of use. Unlike other modalities, ultrasound transducers require little power, which makes handheld imaging platforms possible, and several low-resolution 2D devices are commercially available today. However, the extreme computational requirements (and associated power requirements) of 3D ultrasound image formation have, to date, precluded handheld 3D-capable devices. The authors describe the Sonic Millip3De, a new system architecture and accelerator for 3D ultrasound beamforming--the most computationally intensive aspect of image formation. Their three-layer die-stacked design combines a new approach to the ultrasound imaging algorithm better suited to hardware with a custom beamforming accelerator that employs massive data parallelism and a streaming pipeline architecture to achieve high-quality 3D ultrasound imaging within a full-system power of 15 W in 45-nm semiconductor technology (400× less than a conventional DSP solution). Under anticipated scaling trends, the authors project that Sonic Millip3De will achieve the target 5-W power budget by the 16-nm technology node.

Low Complexity 3D Ultrasound Imaging Using Synthetic Aperture Sequential Beamforming

Synthetic aperture sequential beamforming (SASB) is a technique to achieve range-independent resolution in 2D images with lower computational complexity compared to synthetic aperture ultrasound (SAU). It is a two stage process, wherein the first stage performs fixed-focus beamforming followed by dynamic-focus beamforming in the second stage. In this work, we extend SASB to 3D imaging and propose two schemes to reduce its complexity:(1) reducing the number of elements in both transmit and receive and (2) implementing separable beamforming in the second stage. Our Field-II simulations demonstrate that reducing transmit and receive apertures to 32×32 and 16×16 elements, respectively, and using separable beamforming reduces 3D SASB computational complexity by 15× compared to the 64×64 aperture case with almost no loss in image quality. We also describe a hardware architecture for 3D SASB that performs first-stage beamforming in the scan head, reducing the amount of data that must be transferred for offchip processing in the second stage beamformer by up to 256×. We describe an implementation approach for the second stage that performs an optimized in-place update for both steps of separable beamforming and is well suited for GPU.

Low cost clutter filter for 3D ultrasonic flow estimation

3D blood velocity estimation in medical ultrasound systems is revolutionizing the diagnosis of vascular diseases. However, the accuracy of blood velocity estimation is greatly affected by clutter signals from the vessel wall and the tissues surrounding the vessel. Filters used today to remove clutter are computationally expensive, limiting their practicality in portable 3D systems. In this paper, we present clutter filters for arterial flow that reduce computational complexity by orders of magnitude while maintaining the clutter removal performance of existing techniques. We achieve this goal by combining the existing Hankel-SVD clutter filter with the power iteration method to eliminate unnecessary SVD calculations. For the filters which use power iteration exclusively, we achieve excellent filtering performance with only 14.2% computational overhead to our previous flow estimation system. With these filtering methods, our pipelined architecture can compute velocity fields at a rate of 85 frames per second.

High volume rate, high resolution 3D plane wave imaging

3D plane-wave imaging systems can support the high volume acquisition rates that are essential for 3D vector flow imaging and sonoelastography but suffer from low resolution and low SNR. Coherent compounding is a technique to improve the image quality of plane-wave systems at the expense of significant increase in beamforming computational complexity. In this paper, we propose a new separable beamforming method for 3D plane-wave imaging with coherent compounding that has computational complexity comparable to that of a non-separable non-compounding baseline system. The new method with 9-fire-angle compounding helps improve average CNR from 1.6 to 2.2 and achieve a SNR increase of 9.0 dB compared to the baseline system. We also propose several enhancements to our beamforming accelerator, Sonic Millip3De, including additional SRAM arrays, configurable interconnect, and embedded DRAM. Overall, our system is capable of generating high resolution images at 1000 volumes per second.