Click to edit Master title style

Jesper Mosegaard, PhD Head of Computer Graphics Lab Alexandra Institute Denmark

Mobile Wireless Ultrasound with GPU Beamforming

Click to edit Master title style

•  Private not-for-profit company within IT –  Technology transfer from University research through GTS institutes

(Danish model)

•  Application oriented research •  Consultancy for companies

The Alexandra Institute

Click to edit Master title style FutureSonic

•  ” A new platform and business model for on-demand diagnostic ultrasound imaging“

•  2013-2018 •  Budget: 23 mio US$

19/03/15 3

Click to edit Master title style

•  Center for Fast Ultrasound Imaging, Technical University of Denmark

–  Martin Christian Hemmsen –  Borislav G. Tomov –  Jørgen Arendt Jensen

•  BK Medical –  Carsten Kjær

•  Computer Graphics Lab, Alexandra Institute –  Thomas Kim Kjeldsen –  Lee Lassen –  Jesper Mosegaard

Joint work

19/03/15 Page 4

Click to edit Master title style Presentation based on publications

19/03/15 Page 5

M. C. Hemmsen, T. Kjeldsen, L. Larsen, C. Kjaer, B. G. Tomov, J. Mosegaard, and J. A. Jensen, “Implementation of synthetic aperture imaging on a hand-held device,” in Proceedings of 2014 ieee international ultrasonics symposium, 2014, pp. 2177-2180.

T. Kjeldsen, L. Lassen, M. C. Hemmsen, C. Kjaer, B. G. Tomov, J. Mosegaard, and J. A. Jensen, “Synthetic aperture sequential beamforming implemented on multi-core platforms,” in Proceedings of 2014 ieee international ultrasonics symposium, 2014, pp. 2181-2184.

Click to edit Master title style Medical Ultrasound

19/03/15 6

http://www.bkmed.com/products_en.htm

Click to edit Master title style Mobile transducer

Click to edit Master title style Leveraging disruptive technology

19/03/15 8

Click to edit Master title style Ultrasound

•  From acoustic (pressure) waves to images

19/03/15 9 http://www.sensorwiki.org/doku.php/sensors/ultrasound

Medium Velocity (m/sec)

Fat 1450 Water 1480 Soft tissue 1540 Kidney 1560 Blood 1570 Muscle 1580 Bone 4080

Click to edit Master title style Beamforming to reconstruct images

19/03/15 10

Click to edit Master title style Shooting for a number of scanlines

19/03/15 Page 11

Click to edit Master title style Dynamic Receive beamforming

19/03/15 Page 12 19/03/15 Page 12

Scanline Dynamic focus point

Multiple transducer elements

Click to edit Master title style

•  Traditional beamforming requires a high data bandwidth.

•  A typical system could have 128 channels and use a 12-bit 40 MHz sampling system.

•  This generates 128 × 40 × 106 Hz × 2B = 9.54 GB/s

Beamforming

19/03/15 Page 13

Click to edit Master title style

•  Simple first stage –  Single focal point for both transmit and receive

•  Advanced second stage –  combining information from multiple first stage focused scan lines

•  Reduction in data-transfer requirement –  Reduced by a factor of receive elements (192)

SASB – dual beamforming

19/03/15 Page 14

M. C. Hemmsen, J. M. Hansen, and J. A. Jensen, “Synthetic Aperture Sequential Beamformation applied to medical imaging using a multi element convex array transducer,” in EUSAR , Apr. 2012, pp. 34–37.

Click to edit Master title style Algorithmic engineering

19/03/15 15

Click to edit Master title style 1. stage: Fixed focus transmit and receive

19/03/15 Page 16

First stage line Fixed focus point

Receiver r receives at time a+2*b +c

a

b

c

Click to edit Master title style 2. stage: reconstruct focus

19/03/15 Page 17

1. Find index scanline entries 2. Add contribution

Click to edit Master title style The math behind it

19/03/15 Page 18

Click to edit Master title style

for all image samples p with polar coordinates (i,j) for all first stage scanlines k a = calculateWeight(i,k) d = calculateDelay(i,k) s = getScanline(d,k) l(i,j) += a*s

Pseudo code

19/03/15 Page 19

Click to edit Master title style Three implementations

19/03/15 Page 20

AVX Multithreaded

Click to edit Master title style Benchmark, simple implementation

19/03/15 Page 21

600,000  

2950  

699  

3.05  

0  

1  

10  

100  

1,000  

10,000  

100,000  

1,000,000  

Matlab   C  (1  thread)   C  (8  threads)   OpenCL  

ms  p

er  fram

e  

Simple  

Simple  

Intel Core i7 2600 Nvidia GTX 680 GPU

Click to edit Master title style

•  Sampling the beamforming directly in the scan line sample locations

Algorithmic optimization

19/03/15 Page 22

Click to edit Master title style Weights and delays precalculated

19/03/15 Page 23

r

k

Apodization

Delays

r

k

Click to edit Master title style

l = 0 for all image samples p with polar coordinates (i,j)

for all first stage scanlines k - up to N(ri) a = getWeight(i,k) if a=0 then break d = getDelay(i,k) s = getScanline(d,j+k) l(i,j) += a*s if (k>0) then s = getScanline(d,j-k) l(i,j) += a*s

Pseudo code

19/03/15 Page 24

Click to edit Master title style Benchmark, Optimization

19/03/15 Page 25

600,000  

2950  

699  

3.05  

20.9  

5.4  

0.49  0  

1  

10  

100  

1,000  

10,000  

100,000  

1,000,000  

Matlab   C  (1  thread)   C  (8  threads)   OpenCL  

ms  p

er  fram

e  

Simple  

OpAmizaAon  

Intel Core i7 2600 Nvidia GTX 680 GPU

Click to edit Master title style Resulting image quality

19/03/15 Page 26

Matlab SIMD/Multicore OpenGL OpenCL RMSE

0.0044 0.0042 0.0040

PSNR

47.23dB 47.79dB 48.01dB

Click to edit Master title style Going mobile (OpenGL ES 3.0)

19/03/15 Page 27

Click to edit Master title style Mobile hardware

19/03/15 Page 28

LG G2 Samsung Galaxy Tab

Samsung Nexus 10

Nvidia Jetson TK1

HTC Nexus 9

SoC Snadragon 800 Exynos 5 Exynos 5220 Tegra K1 Tegra K1

GPU Adreno 300 Mali T628 Mali T604 Kepler Kepler

Screen 1920x1080 2560x1600 2560x1600 1920x1080 2048x1536

OS Android Android Android Linux4Tegra Android

Click to edit Master title style

•  Need 25.3 MB/s à IEEE 802.11ac

Mobile WIFI capabilities

19/03/15 Page 29

28.8  

18.5  

11.2  

43.4  

35  

0  

5  

10  

15  

20  

25  

30  

35  

40  

45  

50  

LG  G2   Galaxy  Tab  Pro   Nexus  10     Jetson  TK1  +  Intel  7260HMW  

Nexus  9  

MB/s  

WIFI  throughput  

Click to edit Master title style Mobile performance

19/03/15 Page 30

0  

10  

20  

30  

40  

50  

60  

70  

LG  G2  (Adreno  330)   Galaxy  Tab  Pro  (Mali  T628)  

Nexus  10  (Mali  T604)   Jetson  TK1  (Tegra  K1)   Nexus  9  (Tegra  K1)  

ms  

Timings  

Scanconversion  

BeamformaAon  

IQ  demodulaAon  

Datatransfer  

Total  Aming  

Click to edit Master title style Going mobile, Nexus 9

19/03/15 Page 31

Click to edit Master title style Digital or wireless?

19/03/15 32

Click to edit Master title style

Mail: jesper.mosegaard@alexandra.dk Twitter: @mosegaard LinkedIn: linkedin.com/in/mosegaard Please complete the Presenter Evaluation sent to you by email or through the GTC Mobile App. Your feedback is important!

Thank you for your attention