American Journal of Biomedical Engineering 2013, 3(1): 14-30

DOI: 10.5923/j.ajbe.20130301.03

Digital Signal Processing Methodologies for Conventional

Digital Medical Ultrasound Imaging System
Mawia Ahmed Hassan1,*, Yasser Mostafa Kadah2

1
Biomedical Engineering Department, Sudan University of Science & Technology, Khartoum, Sudan
2
Systems & Biomedical Engineering Department, Cairo University, Giza, Egypt

Abstract Ultrasound imaging is an efficient, noninvasive, method for med ical diagnosis. A commonly used approach
to image acquisition in ultrasound system is digital beamforming. Digital beamforming, as applied to the medical
ultrasound, is defined as phase align ment and summation of signals that are generated fro m a co mmon source, by received
at different times by a mu lti-elements ultrasound transducer. In this paper first: we tested all signal processing
methodologies for digital beamforming which included: the effect of over samp ling techniques, single trans mit focusing
and their limitations, the apodization technique and its effect to reduce the sidelobes, the analytical envelope detection
using digital finite impulse response (FIR) filter appro ximations for the Hilbert transformat ion and how to co mpress the
dynamic range to achieve the desired dynamic range for display (8 bits). Here the image was reconstructed using physical
array elements and virtual array elements for linear and phase array probe. The results shown that virtual array elements
were given well results in linear array image reconstruction than physical array elements, because it provides additional
number of lines. Ho wever, physical array elements shown a good results in linear phase array reconstruction (steering) than
virtual array elements, because the active elements number (Aperture) is less than in physical array elements. We checked
the quality of the image using quantitative entropy. Second: a modular FPGA-based 16 channel digital u ltrasound
beamforming with embedded DSP for ultrasound imaging is presented. The system is imp lemented in Virtex-5 FPGA
(Xilin x, Inc.). The system consists of: t wo 8 channels block, the DSP wh ich co mposed of the FIR Hilbert filter b lock to
obtain the quadrature components, the fractional delay filter block (in-phase filter) to co mpensate the delay when we were
used a high FIR order, and the envelope detection block to compute the envelope of the in-phase and quadrature
components. The Hilbert filter is imp lemented in the form whereby the zero tap coefficients were not computed and
therefore an order L filter used only L/2 mu ltiplications. This reduced the computational time by a half. Fro m the
implementation result the total estimated power consumption equals 4732.87 mW and the device utilization was acceptable.
It is possible for the system to accept other devices for further processing. Also it is possible to build 16-,32-, and
64-channel beamformer. The hardware architecture of the design provided flexib ility for beamforming.
Keywords Medical Ult rasound, Dig ital Beamforming, FIR Hilbert Transform Filter, FPGA, Embedded DSP

the tissues of the body where apportion is reflected, which

1. Introduction used to generate the ultrasound image. Emp loyed ultrasound
waves allow obtaining informat ion about the structure and
Ultrasound is defined as acoustic waves with frequencies nature of tissues and organs of the body[3]. It is also used to
above those which can be detected by the ear, fro m about 20 visualize the heart, and measure the blood flows in arteries
KHz to several hundred MHz. Ultrasound for medical and veins[4].
applications typically uses only the portion of the ultrasound The commonly use arrays are linear, curved, or phase
spectrum fro m 1 MHz to 50 MHz due to the combined needs array. The important distinctions arise from the method of
of good resolution (small wave length) and good penetrating beam steering use with these arrays. For linear and curve
ability (not too high a frequency)[1]. linear, the steering is acco mp lished by selection of a g roup of
They are generated by converting a radio frequency (RF) elements whose location defines the phase center of the
electrical signal into mechanical vibration via a p iezoelectric beam. In contrast to linear and curve linear array, phase array
transducer sensor[2]. The ultrasound waves propagate into transducer required that the beamfo rmer steers the beam with
switched set of array elements[5]. These requirements
* Corresponding author:
[email protected] (Mawia Ahmed Hassan) mention important differences in complexity over the linear
Published online at http://journal.sapub.org/ajbe and curved array. Beamformer has two functions: directivity
Copyright © 2013 Scientific & Academic Publishing. All Rights Reserved to the transducer (enhancing its gain) and defines a focal
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 15

point within the body, from which location of the returning 2.1. Over Sampling Techni que
echo is derived. Over samp ling is used to achieve high delay resolution.
A commonly used approach to image acquisition in However, this increase the data volu me has to acquire. This
ultrasound system is digital beamforming because the analog is usually avoided by sampling just above the Nyquist rate
delay lines impose significant limitations on beamformer and interpolating to achieve the required delay resolution.
performance and more expensive than digital implementatio Radial sampling resolution is a relat ionship between the
ns. Dig ital beamforming, as applied to the medical depth and the number of delay values and it equals to the
ultrasound, is defined as phase align ment and summat ion[2] speed of sound over twice the samp ling frequency. Fro m the
of signals that are generated from a co mmon source, by literature[7], wide band transducer required delay resolution
received at different times by a mu lti-elements ultrasound in order of 1/ 16 the signal period.
transducer[6].
The beamforming process needs a high delay resolution to 2.2. Delay Equati on
avoid the deteriorating effects of the delay quantization lobes

RF signal N
on the image dynamic range and signal to noise ratio

RF signal 3

RF signal 2

RF signal 1
(SNR)[7]. If oversampling is used to achieve this timing
resolution[8], a huge data volu me has to be acquired and
process in real time. This is usually avoided by sampling
just above the Nyquist rate and interpolating to achieve the

Delay Stage
Over Sampling Technique
required delay resolution[7].
Beamforming required Apodization weighted to
decreasing the relative excitation near the edges of the
Single Transmit Focusing
radiating surface of the transducer during transmit or
receiving, in order to reduce side lobes. Just as a time side
lobes in a pulse can appear to be false echoes.
After delay and sum the envelope of the signals is detected. The Apodization Technique
The envelope then compressed logarithmically to reduce the
dynamic range because; the maximu m dynamic range of the
human eye is in the order of 30 dB[9]. The actual dynamic Physical Array Virtual Array
range of the received signal depends on the ADC bits, the Elements Elements

Summation Stage
time gain co mpensation (TGC) amp lifier used in the front
end, and the depth of penetration. The signal is compressed
to fit the dynamic range used for display (usually 7 or 8 b its). Linear Array Linear Phase
It is typical to use a log co mpressor to achieve the desired Reconstruction Array
dynamic range for display[10]. Reconstruction
With the growing availability of high-end integrated
analog front-end circu its, distinction between different FIR Hilbert Transform
digital ult rasound imag ing systems is determined almost Filter Design
exclusively by their software co mponent.
We develop a compact low-cost digital ultrasound
imaging system that has almost all of its processing steps Real Quadrature
done on the PC side. In this paper first: we tested all signal Components Components
processing methodologies for dig ital beamforming which
DSP Stage

included (Figure 1): the effect of over sampling techniques,

single transmit focusing and their limitations, the
Envelope Detection
apodization technique and its effect to reduce the sidelobes,
the analytical envelope detection and how to co mpress the
dynamic range to achieve the desired dynamic range for Compressed Dynamic
display (8 bits). Here the image was reconstructed using
Range
physical array elements and virtual array elements for linear
and phase array probe.
Second: a modular FPGA -based 16 channel dig ital Reconstructed Ultrasound
ultrasound beamforming with embedded DSP for u ltrasound Line
imaging is presented. The system is imp lemented in Virtex-5
FPGA (Xilin x, Inc.). Figure 1. The technique block diagram

Figure 2 shows the geometry wh ich is used to determine

2. Methodology the channel and depth-dependent delay of a focused
16 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

transducer array. After a wave-front is trans mit into the summed together after phase shift and some signal
med iu m an echo wave propagates back from the focal points conditioning to produce a single output. This reconstruction
(P) to the transducer. The distance from P to the origin is technique divides the field of view (FOV) into different point
equal to the Euclid ian distance between the spatial point targets (raster points), P(i,j).
(xi ,y i ) ( is the center for the physical element number i) and Each point represented as an image p ixel, which is
the focal point (xf,y f) (is the position of the focal point). The separated laterally and axially by small d istances. Each
original time, t i, , equal the distance from P over the speed of target is considered as a point source that transmits signals to
sound. A point is selected on the whole aperture (AP) as a the aperture elements as in figure 4. The beamfo rming timing
reference (xc,y c) for the imag ing process. The propagation is then calculated for each point based on the distance R
time (tc) for this was calculated as the above, but the distance between the point and the receiving element, and the velocity
here fro m P to the reference (xc ,y c). The delay to use on each of ultrasonic beam in the media. Then the samples
element of the array is the difference between t c and ti . corresponding to the focal point are synchronized and added
to complete the beamforming as the follo wing:
y (axial) PD (i, j) = ∑Nn=1 Xn �K ij �, (1)
(xc ,yc) (xi ,yi ) where PD (i,j) is the signal value at the point whose its
coordinates are (i,j), and Xn (Kij ) is the sample
x (Lateral)
corresponding to the target point in the signal Xn received
by the element number n. The sample number Kij which is
equivalent to the time delay is calculated using the equation
below:
P (x ,y )
f f
R n (i,j )
K ij = . (2)
T∗c
Scan Line Scan Line Here Rn (i, j) is the distance from the center of the element
to the point target, c is the acoustic velocity via the media,
Figure 2. Geometry of a focused transducer array and T is the sampling period of the signal data.
2.3. physical Array Elements y

2.3.1. Linear Array Reconstruction

Selected of a group of elements (aperture) whose location
is defines the phase center of the beam[1]. Electronic R
focusing was applied on receive for each aperture. Received R1
P
at the aperture elements are delayed by focusing delays and
summed to form scan line in the image. After that one NxN (Matrix)
elements shift is applied to the aperture and the process was
repeated till the end of the array elements at the outer side x (Lateral)
processing all image scan lines (Figure 3 where aperture
equal 32 elements). The number of lines equal to the total Figure 4. The raster point technique
number of elements minus the number of the aperture
2.4. Virtual Array Elements
elements plus one.
2.4.1. Linear Array Reconstruction
1 2 3 …. 32 33 34 … 128
In this technique we used central elements and assumed
Line1 there were v irtual elements on the left of the central element.
Line2
For examp le, if we want to use aperture equal 128 elements
we take element nu mber 1 as center element and selected 64
Line3 physical elements on the right (already exist) and 64 virtual
elements on the left (not exist), as in figure 5.
Figure 3. Linear physical array elements To compensate the loss in energy, we mu lt iplied by
factor (Ma) because this line really taken fro m 64 elements
2.3.2. Linear Phase Array Reconstruction
instead of 128 elements. The Ma factor equal the nu mber of
In contrast the linear array, phase array transducer aperture elements divided by the number of the physical
required that the beamformer steered the beam with an elements.
unswitched set of array elements[1]. In this process, the time Image line was obtained from the summation of physical
shifts follow a linear pattern across of array fro m one side to elements mult iplied by Ma factor. There for the Line number
another side. In receive mode, the shifted signals are one equal the summation of the 64 physical elements ×
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 17

128/64. Then one elements shift was applied to the virtual 2 sin 2 (π (n −α) ⁄2)
,n ≠ α
and physical aperture and this process is repeated till the h[ n] = �π n −α , (4)
factor equal 128/128. 0 ,n = α
64 virtual elements 64 physical elements α = (N − 1)/2 . We chose the filter length equal
16,20,24,28 and 32-tap with a Hamming window used to
reduce the sidelobe effects. According to the normalized
1 2 … 63 64 1 2 … 63 64 … 128 root mean square error (RMSE) between the designed FIR
Hilbert filter and ideal Hilbert transform filter. The values
of the RMSE fo r the five FIR filters are shown in table 1.
Center element # 1 for image line #1
We selected 24-tap FIR filter because it provided a good
result for the qadrature components compared to ideal
Figure 5. Linear virtual array elements
Hilbert transform filter.
2.4.2. Linear Phase Array Reconstruction
Table 1. The Normalized RMSR for FIR Hilbert filters
We used the same techniques as in section (2.3.2).
FIR Hilbert filter order Normalized RMSE value
2.5. Apodizati on
16 0.0109
Apodization is amplitude weighting of normal velocity 20 0.0096
across the aperture[9][11], one of the main reasons for
24 0.0092
apodization is to lo wer the side lobes on either side of the
28 0.0091
main beam[12]. Just as a time side lobes in a pulse can
appear to be false echoes[9]. Aperture function needed to 32 0.0090
have rounded edges that taper toward zero at the ends of the
aperture to create low side lobes levels. We used windowing 2.6.2. Co mpressed the Dynamic Range
functions (hamming, Blackman, and Kaiser (β=4)) as The envelope then compressed logarithmically to ach ieve
apodization functions to reduce the side lobes. There is the desired dynamic range for d isplay (8 bits). It is typical to
trade-off in selecting these functions: the main lobe of the use a log compressor to achieve the desired dynamic range
beam broadens as the side lobes lower[9]. for d isplay. Log transformation co mpressed the dynamic
range with a large variation in pixels values[18].
2.6. Envel ope Detection and Compressed The Dynamic
Range 2.7. Implementation Steps
2.6.1. FIR Hilbert Transform Filter Design
After delay and sum, the analytic envelope of the signal is
calculated as the square root of the sum of the squares of the
real and quadrature components[13]. The most accurate way
of obtaining the quadrature components was to pass the echo
signal through a Hilbert transform[14], because it provides
90-degree phase shift at all frequencies[15].
The Hilbert transformation filter acts like an ideal filter
that removes all the negative frequencies and leaves all
positive frequencies untouched. A number of authors
suggested the use of digital FIR filter appro ximations to
implement the Hilbert transformat ion. For linear time
invariant (LTI) a FIR filter can be described in this form[16]:
𝑦𝑦 [ n] = 𝑏𝑏0 𝑥𝑥[ 𝑛𝑛] + 𝑏𝑏1 𝑥𝑥[ 𝑛𝑛 − 1] + ⋯ + 𝑏𝑏𝑀𝑀 𝑥𝑥[ 𝑛𝑛 − 𝑀𝑀]
𝑀𝑀
= �𝑖𝑖=0 𝑏𝑏𝑖𝑖 𝑥𝑥[𝑛𝑛 − 𝑖𝑖] . (3)
Where 𝑥𝑥 is the input signal, 𝑦𝑦 is the output signal, and the
constants 𝑏𝑏𝑖𝑖 , 𝑖𝑖 = 0,1,2,…,M, are the coefficients. The
designed FIR Hilbert filter can be used to generate the
Hilbert transformed data of the received echo signal. The
impulse response of the Hilbert filter with length N (odd Figure 6. Architecture implementation of the modular FPGA-Based 16-
channel digital ultrasound receive beamformer blocks
number) is defined as[17]:
18 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

(a)

(b)
Figure 7. The inside contents of the implementation blocks. (a) The 8 channel block, (b) the reconstructed line block

A typical architecture imp lementation of the modular Xilin x block to read the one d imension RF data fro m
FPGA -based, 16-channel digital u ltrasound receive workspace.
beamformer with embedded DSP fo r ult rasound imaging 2. The RF data then convert the double precision data type
was shown in figure 6. The system consist of: Two 8 to fixed po int numeric precision for hard ware efficiency.
channels block and reconstructed line b lock. The 3. Verified the fixed-point Model by co mpared the
beamfo mer is done by using Xilin x system generator fixed-point results to the floating-point results and determine
(Xilin x, Inc.) and MATLAB simu lin k (MathWorks, Inc.). if the quantizat ion error is acceptable.
The system is imp lemented in Virtex-5 FPGA. The 4. A fter verified the model, each channel data saved in
implementation steps are (Figure 7): memo ry b lock. The memory word size is determined by the
1. The RF data saved in MATLA B workspace and used bit width of the channel data. The memo ry controlled by wire
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 19

enable port with 1 indicates that the value of the channel data channels was 128 channels, and the A/D sampling rate was
should be written to the memory address pointed to by 13.8889 MSPS. Linear shape transducer was used to acquire
step-up counter. the data with center frequency of 3.5 M Hz, and element
5. The delay process is based on sampled delay focusing spacing of 0.22mm. Each u ltrasonic A-scan was saved in a
(SDF). The delays calculated using the same method in record consisted of 2048 RF samp les per line each
section 2.2. Then the delays converted to number of samp les represented in 2 bytes for the phantom data and 4 byte for the
by divided the delays by the sampling time. SDF consist of real data, and the signal averages was 8. The speed of the
addressable shift register (ASR) to delay the sampled signals, ultrasound was 1480 m/sec. The data were acquired for
and M-Code Xilin x block that contain the calculated delays. phantom within 6 p ins at different positions. The data was
The samples are delayed by the value in the address input of used to simulate the N-channel beamformer on receive as
the ASR. discussed in methodologies. The radio frequency (RF)
6. After delay ing each RF channel samples, the signals, A-scan, were recorded fro m every possible
summation is applied using M-Code block to summate the 8 combinations of transmitter and receiver for all elements in
channel signals. the 128 elements.
7. The summat ion of the two 8 channels is connected to
adder to reconstructed the final focus line. 3.2. Over Sampling Techni que
8. Modify the bit of the signal to 16 bit using the bit Table 2 was shown the effect of oversampling to the delay
modifier b lock. resolution. As can be shown when sampling ratio increased
9. The Reg ister block (data presented at the input will the delay resolution decreased according to the radial
appear at the output after one sample period). sampling resolution. Fro m the table the Delay resolution
10. We used the FIR Hilbert filter (with 24-tap as in equal to (1/(2×sampling ratio )).
section 2.6.1) b lock for applying the quadrature Fro m the table when the sampling ratio (F) equal 8 as an
components. interpolation factor, it gave better delay resolution (1/16)
11. The Fractional delay filter (in-phase filter) to the signal period. However, this increase the data volume
compensate the delay when we are being used a high FIR has to acquire.
order.
12. Then we mod ified the bit of the signals from step 10 3.3. Physical Elements Array
and 11 to 16 bit again using the bit modifier blocks.
13. The Envelope detection block wh ich was co mputed 3.1.1. Linear Array Image Reconstruction
the envelope of the two signals coming fro m step 10 and 11. In this situation we reconstructed an image using
14. In order to obtain performance and logic utilizat ion N-channel beamformer on receive where N= 4,8,16,32 and
figures for the suggestion architecture, it was implemented 64. Figure 8 was shown images reconstructed for six p in
in the hardware description language (VHDL) and phantom by linear technique (figure3). In figure 8 (b,c,d and
synthesized with Virtex-5 FPGA . e) the images was reconstructed with focusing using AP
equal 4,8,16,32and 64. In figure 8 (a) image was
reconstructed without focusing and AP=4 elements. When
3. Results and Discussions the AP size increased the lateral resolution was improved.
However, the FOV was reduced when the aperture size
3.1. The Ultrasound Data increased. Figure 9 shown the signal to noise ratio (SNR)
We used correct data obtained from the Bio med ical when aperture equal 4,8,16,32,and 64. As can be shown
Ultrasound Laboratory[19], University of Michigan; the when the size of A P increased that improved the SNR. A lso
phantom data set that was used to generate the results here is figure 10 described the effect of oversamp ling technique.
under "Acusonl7" and real data (cyst4_a100). The
parameters for this data set are as follo ws: the number of
Table 2. The effect of the oversampling to the delay resolution
Sampling Radial sampling resolution(=Ts/2×c)
Delay resolution
Sampling fre quency(Fs) period Delta-distance (Dd)
(1/(2×sampling ratio))
(Ts=1/Fs) c=1.480mm/µsec
Fs=13.8889
Ts=72 nsec Dd=0.0533 mm/samples 1/2
Sampling Ratio=1
Fs=2×13.8889=27.7778
Ts=36 nsec Dd=0.0267 mm/samples 1/4
Sampling Ratio=2
Fs=4×13.8889=15.5556
Ts=18 nsec Dd=0.0133 mm/samples 1/8
Sampling Ratio=4
Fs=8×13.8889=111.1112
Ts=9 nsec Dd=0.0067 mm/samples 1/16
Sampling Ratio=8
20 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

Here we used image with AP equal to 64 with F equal to

1,2,4,and 8 then we calcu lated the SNR for each images. As
can be shown The SNR was imp roved when the F increase.

Figure 9. SNR for AP equal 4,8,16,32,and 64 using physical linear array

(a)

(b)
Figure 10. The effect of oversampling technique to the SNR. Here we
used image with AP equal to 64 with F equal to 1,2,4,and 8

3.3.2. Linear Phase Array Reconstruction

Due to the small FOV of the linear reconstruction and its
limited lateral resolution, we used linear phase
reconstruction to reconstructed image of six p ins phantom
(c) fro m the data set. Figure 11 shown images reconstructed
using raster point technique. Element number 128 was the
transmitter and received with all 128 elements. In figure 11
(a and b) image was reconstructed using F equal 4 and 8. As
we saw when F was increased it improved the image quality
because from table 1 radial resolution equals 0.0067 fo r (F=8)
compared to 0.0133 for (F= 4). Figure12 shown image
reconstructed by the same technique using all elements for
transmitted and received. The six p in of the phantom are
(d) clearly described with a moderate lateral resolution.

(e)
(a)

(f)
Figure 8. Physical Linear array image reconstruction, (a) Using AP=4 (b)
without focusing, (b) Using AP= 4 with focusing, (c) Using AP= 8with Figure 11. Physical Linear phase array image reconstruction, element
focusing, (d) Using AP= 16 with focusing, (e) Using AP= 32 with number 128 for transmitted and received with all 128 elements, (a) Using
focusing, (f) Using AP=64 with focusing F=4, (b) Using F=8
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 21

shown when the size of AP increased the SNR was

improved.

（a）

(a)

（b）
Figure 12. Physical Linear phase array image reconstruction, transmitted
and received with all 128 elements (a) Using F=4, (b) Using F=8
(b)
3.4. Virtual Array Elements

3.4.1. Linear Array Image Reconstruction

Table 3 shown a co mparison between linear array
reconstruction with physical and virtual elements, as can be
seen virtual elements provide additional scan lines leaded to
improved the FOV and lateral resolution than using physical
array elements. Also virtual array elements provided a
possibility of using AP equal 128 elements with 65 lines and
that could not be acceptable in physical array elements,
(c)
because we had only one line.
Figure 13. Virtual linear array image reconstruction, (a) Using AP=32, (b)
Table 3. Number of lines in linear array image reconstruction with Using AP=64, (c) using AP=128
physical and virtual elements
Physical elements Virtual elements
AP=32 97 lines 113 lines
AP=64 65 lines 97 lines
AP=128 1 lines 65 lines

Figure 13 showed 32, 64, and 128 elements channel

beamformers. Figure 13 (a and b) shown 32 and 64 channel
respectively, as can be shown, when increased the AP the
lateral resolution was improved and also FOV been better
than physical array elements in figure 8 (e and f). In figure 13
(a)
(c) the AP equal 128 wh ich it's unacceptable in physical
array elements.

3.4.2. Linear Phase Array Reconstruction

Figure 14 (a, b, and c) shown images reconstructed using
raster point technique for v irtual elements when AP equal 32,
64, and 128 elements respectively. As we saw virtual
elements array in steering reduced the FOV but the lateral
resolution was good compared to the physical array elements
which gave a good FOV and lateral resolution. Figure 15
shown the SNR when AP equal 16,32,and 64. As can be (b)
22 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

The frequency response of apply 16-, 20-, 24-, 28-, and

32-tap FIR Hilbert filter shown in Figure 17. Figure 18 had
shown the different between the FIR Hilbert filters and the
ideal Hilbert filter. The normalized root mean square error
(RM SE) between the FIR hilbert filter and ideal Hilbert
transform filter for the five FIR filters are shown in table 1.
Fro m the result 24-tap provided a good result because it gave
the mediu m RMSE (0.0092) between the five selected FIR
filters that it gave a good result because the different between
（c） it and the ideal Hilbert filter look like rando m signal.
Figure 14. Virtual Linear phase array image reconstruction, (a) Using
AP=32, (b) Using AP=64,(c) AP=128

(a)

Figure 15. SNR for AP equal 16,32,and 64 using raster point technique for
virtual elements

3.5. Effect of Apodizati on

Table 4 shown the effect of Hamming, Blackman, and
Kaiser (β =4) apodizat ion windows compared to rectangular
window. Because the aperture is rectangular unfortunately,
the far field beam pattern is a sinc function with near in-side
lobes only -13 dB down fro m the maximu m on axis value.
Fro m table 4 there is trade-off in selecting these functions: (b)
the main lobe of the beam broadens as the side lobes lower.
Also the table described the SNR between these windows.
As can be shown the Blackman apodization function was
given a better SNR, better reduction in the sidelope and good
main lobe width compared to the other windows.
Table 4. Hamming, Blackman, and Kaiser (β=4) apodization functions
compared to rectangular window (without Apodization)
Approximate
Type of Peak Side Lope SNR
Width of
Window Amplitude(Relative) (Relative)
Main Lobe (c)
Rectangular -13.3 0.0137 32.6680
Hamming -42.6 0.0195 37.6472
Blackman -58.1 0.0252 38.4245
Kaiser (β=4) -30.0 0.01758 37.0413

Figure 16 was shown a comparison between images

reconstructed without apodization using physical array
elements (transmitted and received by all elements) to
images reconstructed with apodization. As can be shown in
figure 16 (b, c, and d) there is trade-off in selecting these
functions: the main lobe of the beam broadens as the side
lobes lower co mpared to figure 16(a) for rectangular (d)
aperture. Figure 16. Comparison between images reconstructed with and without
apodization. (a) Image without apodization, (b) Image apodized with
3.6. Envel ope Detection and Compressed Dynamic Hamming,(c) Image apodized with Blackman,(d) image apodized with
Range kaiser (β=4)
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 23

Fig.19 was shown comparisons between envelope images

used physical and v irtual linear array elements. In figure 19(a
and b) image reconstructed using aperture equal 32 and 64
with physical array and fig.19(c, d, and e) shown images
reconstructed using aperture equal 32, 64, and 128
respectively with virtual elements. As can be shown virtual
elements array provide best FOV and lateral resolution than
physical array elements.

（a）

Figure 17. The frequency response of apply 16-, 20-, 24-, 28-, and 32-tap
FIR Hilbert filter
(b)

Figure 18. The different between the FIR Hilbert filters and the ideal
Hilbert filte (c)

In figure 20 we shown the envelope of images used raster

point technique (There were 128 elements for t ransmits and
receives), apodized by Hamming, Blackman, and Kaiser
(β=4) respectively. As can be viewed the results were better
than figure 16 (b,c, and d), because the envelope provided
more signal strength as can be shown in figure 21.The same
was also said for figure 19 (a and b) co mpared to figure 8 (e
and f) and figure 19 (c,d, and e) co mpared to figure 13
(a,b,and c).
Figure 22 (a,b, and c) shown the images in figure 20 (a, b, (d)
and c), which co mpressed the dynamic range using
logarith mic comp ression to achieve the desired dynamic
range for display (8 b its).
Figure 23 and figure 24 had shown an image line with and
without compress the dynamic range. As can be shown the
maximu m dynamic range without compress equal to 105dB
and 13d B after co mpress the dynamic range.
Figure 25 (a) shown pin three in the co mpressed image as
a sub-image without apodizat ion. Figure 25 (b, c, and d)
shown the same pin apodized by Hamming, Blackman, and （e）
Kaiser (β=4) respectively. Figure 25 (e, f, and g) was a lateral Figure 19. Comparisons between envelope images used physical and
profiles shown the effect of apodization to the side lobes and virtual linear array elements (a) Image with physical array elements using
main lobe as function in the frequency domain. As can be AP=32, (b) Image with physical array elements using AP=64, (c) Image
shown there is trade-off in selecting these functions: the with virtual elements using AP=32, (d) Image with virtual elements using
AP=64, (e) Image with virtual elements using AP=128
main lobe of the beam broadens as the side lobes lower.
24 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

Fig.26 shown images reconstructed using raster point

technique for real data (Cyst). The over sampling factor (F)
equal 8, apply the envelope detection to over sampling data
and compressed the dynamic range used Log compression.

（a）

(a)

(b)

(b)

(c)
Figure 22. The compressed envelope images used log compression
(transmitted and received with all elements) (a) Compressed envelope
(c) image apodized with Hamming , (b) Compressed envelope image apodize
Figure 20. The envelope images used raster point technique (There were d with Blackman , (c) Compressed envelope image apodized with Kaiser
128 elements for transmits and receives) (a) Envelope image apodized with (β=4)
Hamming , (b) Envelope image apodized with Blackman , (c) Envelope
image apodized with Kaiser (β=4)

Figure 21. The analytical envelope Figure 23. The dynamic range without compression
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 25

(e)

Figure 24. The dynamic range with compression

(f)

(a)

(g)
Figure 25. Pin three in the phantom as a sub-image. (a) Without
apodization,(b) Apodized with Hamming , (c) Apodized with with
Blackman , (d) Apodized with Kaiser (β=4), (e) Frequency spectrum of
fig.24 (b), (f) Frequency spectrum of fig.24 (c), (e) Frequency spectrum of
fig.24 (d)

(b)

Figure 26. Image reconstructed using raster point technique for real data
(c) (Cyst)

3.8. Image Quality

We selected pin number two and pin number three as
sub-images (64X64) and (100X100) respectively, to show
the quality of the images using an entropy function. Table 5
was shown a co mparison of images quality (entropy) without
apodization (figure 16(a)) and image with apodization
(figure 16 (b)), Apodization was done using Hamming
window. Table 6 was shown entropy of images without
apodization (figure 16 (a)) and image apodized with
（d） Blachman window (figure 16 (c)). Table 7 was shown an
26 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

entropy of image without apodization (figure 16(a)) and 3.9.2. Delays

image with apodization (figure 16 (d)). Apodization was Figure 28 (a) and Figure 28 (b) illustrated 100 samples
done using Kaiser (β=4) window. fro m channel 8 and channel 9 data respectively before and
Table 5. Entropy of sub-images with apodization (Hamming) and without after synchronization. As can be shown in Figure 28(a) the
apodization time of arrival is different for the two channel compared to
Compress without Compress with Figure 28(b) with the same arrival times.
apodization apodization
Pin3 3.3756 2.9286
Pin4 3.2017 2.8832

Table 6. Entropy of sub-images with apodization (Blackman) and without

apodization
Compress without Compress with
apodization apodization
Pin3 3.0857
Pin4 3.0026 (a)

Table 7. Entropy of sub-images with apodization (Kaiser(β=4)) and

without apodization
Compress without Compress with
apodization apodization
Pin3 3.2180
Pin4 3.1205

Fro m the tables entropy of image is minimized when the

image was uniform. Image with apodizat ion was given
(b)
entropy value less than image without apodization, indicated Figure 28. Alignment for RF signals.(a) 100 samples of channel 8 and 9
for suppressed side lopes. Table 8 was shown an entropy before synchronization,(b) 100 samples of channel 8 and 9 after
ratio between co mpressed sub-image without apodization synchronization
and with apodizat ion. As can be seen Blackman g iven better
results than Hamming and Kaiser in co mpressed image with 3.9.3. The Reconstructed Line
apodization. Figure 29 shown the co mparison between imp lemented
Table 8. An entropy ratio between compressed sub-image without
the first line (after the pipe line adder in Figure 7 (b)) in
apodization and with apodization image reconstruction and the simulated one. As we
Kaiser
mentioned in the methodologies the normalized RMSE
Hamming Blackman
(β=4) between the designed filter and ideal Hilbert transform filter
Pin2 1.0939 1.1526 1.0490 of 24 order FIR Hilbert filter is the mediu m one (0.0092) ,
Pin3 1.0663 1.1105 1.0260 so we use this filter for the simu lation and implementation
of the ultrasound data. Apply this designed Hilbert filter to
3.9. Implementation the received echo line after delay and sum shown in Figure
30(a) co mpared to the ideal Hilbert filter in Figure 30 (b).
3.9.1. Verify the Fixed-point Model Figure 30 (c) had shown the different between the
Figure 27 shown double precision RF data type read fro m reconstructed and ideal Hilbert. Figure 30 (d) described the
MATLAB work space compared to fixed point RF data frequency spectrum for ultrasound data. As can be shown
type for hardware efficiency. We verified the fixed-point the negative frequency was eliminated co mpared to the
Model by subtract the fixed -point RF signal fro m the ideal Hilbert filter frequency spectrum (Figure 30(e)).
floating-point RF signal and the result equal zero. This Figure 31 had shown the imp lemented of the linear phase
means zero quantization error. array image reconstruction to reconstruct image of six p ins
phantom fro m the data set, and the point spread function
(PSF) of the first pin as indicated by white arrow in Figure
31 (b) and Figure 31 (d). Figure 31 (a) described the image
reconstruction after delay and sum and Figure 31(c) shown
the image reconstructed after applying FIR Hilbert filter and
envelope detection. As can be shown Figure 31(c) provided
best FOV and lateral resolution than Figure 31 (a), because
the envelope provided more signal strength. The PSF
presented a quantitative measure of the beamforming
Figure 27. Comparison between floating and fixed point RF signal quality.
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 27

Figure 29. Comparison between implemented the first line (after the pipe
line adder in Fig.6(b)) in image reconstruction and the simulated one

(a)

（a）

(b)

(b)

(c)

(d)

(e)
(d)
Figure 30. Hilbert filter apply to ultrasound line data. (a) Reconstructed
FIR Hilbert, (b) Ideal Hilbert, (c) the different between the reconstructed Figure 31. Image reconstruction and the PSF of the first pin. (a) Image
and ideal Hilbert, , (d) frequency spectrum of reconstructed FIR Hilbert,(e) after delay and sum, (b) PSF of the first pin in (a), (c) Image after applying
The ideal Hilbert filter frequency spectrum FIR Hilbert filter and envelope detection, (d) PSF of the first pin in (c)
28 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

simu lated FIR Hilbert filter and implemented envelope take

after apply the 24-tap imp lemented FIR Hilbert for all the
ultrasound line data. The result was good if it co mpared to
the simulat ion results.

3.9.5. Timing and Clock System

Table 9 and 10 were shown the 2x system clock (Discrete
Pulse Generator) and continuous source (Discrete Pulse
Figure 32. Comparison between envelope take after apply simulated ideal Generator).
Hilbert filter ,the 24-tap simulated FIR Hilbert filter and implemented
envelope take after apply the 24-tap implemented FIR Hilbert for all the 3.9.6. Power Consumption
ultrasound line data
Table 11 was shown the summary of the power
Table 9. 2x system clock (discrete pulse generator) consumption in the whole imp lementation. The total
estimated power consumption equal 4732.87 mW .
Parameter Value
Pulse type Sample based 3.9.7. Device Ut ilization
Time (t) Use simulation time
Table 12 was shown the device utilizat ion summary for
Amplitude 1
the whole implementation, the used devices, available in the
Period (secs) 2
port, and the utilization in percentage using Virtex-5
Pulse Width (% of period) 1
FPGA.
Phase delay (secs) 0
Sample time simulink_period

Table 10. 2x system clock (discrete pulse generator) 4. Discussions

Parameter Value Fro m the results when increasing the oversampling ratio
Pulse type Time based this improved the delay resolution but increased the
Time (t) Use simulation time hardware volu me in the implementation. A further problem
Amplitude 1 in conventional imaging is the single transmit focus, so that
Period (secs) 2*simulink_period the imaging is only optimally focused at one depth. This can
Pulse Width (% of period) 50 be overcome by making co mpound imaging using a number
Phase delay (secs) 0 of transmit foci, but the frame rate is then correspondingly
Sample time 1 decreased. One alternative is to use synthetic aperture
imaging. The main reason for apodizat ion is to lo wer the side
Table 11. Summary of the power consumption in the whole lobes on either side of the main beam. Just as a time side
implementation
lobes in a pulse can appear to be false echoes but there is
Power summary I(mA) P(mW) trade-off in selecting these functions: the main lobe of the
Total Vccint 1.00V 2980 2980 beam broadens as the side lobes lower. Here the image was
Total Vccaux 2.50V 362.82 907.06 reconstructed using physical array elements and virtual array
Total Vcco25 2.50V 338.33 845.81
elements for linear and phase array probe. The results shown
that virtual array elements were given well results in linear
Clocks - 97.75
array image reconstruction than physical array elements,
DSP - 5.35 because it provides additional number of lines. However,
IO - 873.33 physical array elements shown a good results in linear phase
Logic - 30.13 array reconstruction (steering) than virtual array elements,
Signals - 34.69 because the active elements number (Aperture) is less than in
Quiescent Vccint 1.00V 2799.09 2799.09
physical array elements.
Fro m the implementation results we shown that the
Quiescent Vccaux 2.50V 345.00 862.50
fixed-point Model is the same as the floating point mode and
Quiescent Vcco25 2.50V 12.00 30.00 this is an important for hardware efficiency. Future, the
Total estimated power delays applied as SDF gave a synchronous in the time of
4732.87 arrival. Furthermore, The Hilbert filter is implemented in the
consumption
form whereby the zero tap coefficients were not co mputed
and therefore an order L filter used only L/ 2 mu ltip licat ions.
3.9.4. Envelope Detection
This reduced the co mputational time by a half. The total
Figure 32 described the comparison between envelope estimated power consumption equal to 4732.87 mW and the
take after apply simu lated ideal Hilbert filter ,the 24-tap device utilization was acceptable.
 American Journal of Biomedical Engineering 2013, 3(1): 14-30 29

Table 12. Device utilization summary

Slice Logic Utilization Used Available Utilization Slice Logic Utilization Used Available Utilization

Number of Slice
1,875 207,360 1% Number of route-thrus 29 - -
Registers
Number used as Flip Number using O6 output
1.875 - - 22 - -
Flops only
Number using O5 output
Number of Slice LUT s 2,283 207,360 1% 7 - -
only
Number used as logic 627 207,360 1% Number of occupied Slices 819 51,840 1%
Number using O6 Number of LUT Flip Flop
885 - - 2,473 - -
output only pairs used
Number using O5 Number with an unused Flip
19 - - 598 2,473 24%
output only Flop
Number using O5 and Number with an unused
23 - - 190 2,473 7%
O6 LUT
Number used as Number of fully used
1,653 54,720 1% 1,685 2,473 68%
Memory LUT-FF pairs

Number used as Shift Number of unique control

1,653 - - 66 - -
Register sets

Number of slice register

Number using O6
1,587 - - sites lost 4 207,360 1%
output only
to control set restrictions
Number using O6
66 Number of bonded IOBs 1,041 1,200 86%
output only
Number used as -Average Fanout of
3 - - 1.86 - -
exclusive route-thru Non-Clock Nets

5. Conclusions Wiley & Sons, New York, (1988).

In this work we applied all signal processing [2] J. A Zagzebski , Essentials of ultrasound physics, St Louis,
M o: M osby, (1996).
methodologies for tradit ional d igital u ltrasound
beamforming. We used two types to reconstruct the [3] W. N.M cDicken, Diagnostic Ultrasonics: Principals and Use
ultrasound images; virtual array elements and physical array of Instruments, JohnWiley & Sons, (1976).
elements for linear and phase array probe, then we studied [4] S. Hughes, M edical ultrasound imaging, Electronic Journals
the effect of each one. A modular FPGA-based 16 channel of Physics Education. 36 ,468–475(2001).
digital ult rasound beamforming with embedded DSP for
ultrasound imaging is presented. The system is [5] K. E. Thomenius, “Evaluation of Ultrasound Beamformers,”
in Proc. IEEE Ultrason. Symp., pp.1615-1621, 1996.
implemented in Virtex-5 FPGA (Xilin x, Inc.). The
computational time was reduced by a half because we [6] R. Reeder, C. Petersen, The AD9271-A Revolutionary
implemented the Hilbert filter in the form whereby the zero Solution for Portable Ultrasound, Analog Dialogue,Analog
tap coefficients were not co mputed and therefore an order L Devices. (2007).
filter used only L/2 mu ltip licat ions. From the [7] C. Fritsch, M . Parrilla, T. Sanchez, O. M artinez,
implementation result the total estimated power “Beamforming with a reduced sampling rate,” Ultrasonics,
consumption and the device utilization were acceptable. It vol. 40, pp. 599–604, 2002.
is possible for the system to accept other devices for further [8] B. D. Steinberg, “Digital beamforming in ultrasound,” IEEE
processing. Also it is possible to build 16-,32-, and Transactions on Ultrasonics, Ferroelectrics and Frequency
64-channel beamformer. The hardware architecture o f the Control, vol. 39, pp. 716–721, 1992.
design provided flexib ility for beamforming.
[9] Szabo, T. L., Diagnostic Ultrasound Imaging: Inside Out,
Elsevier Academic Press: Hartford, Connecticut, 2004.
[10] M . Ali, D. M agee and U. Dasgupta, “Signal Processing
Overview of Ultrasound Systems for M edical Imaging,”
REFERENCES SPRAB12, Texas Instruments, Texas, November 2008.

[1] D. A .Christensen, Ultrasonic Bioinstrumentation, Jonh [11] J. A. Jensen, “Ultrasound imaging and its M odeling”,
30 M awia Ahmed Hassan et al.: Digital Signal Processing M ethodologies for
Conventional Digital M edical Ultrasound Imaging System

Department of Information Technology, Technical University [16] J. O. Smith, M athematics of the Discrete Fourier Transform
of Denmark, Denmark, 2000. (DFT), Center for Computer Research in M usic and
Acoustics (CCRM A), Department of M usic, Stanford
[12] J. Synnevag, S. Holm and A. Austeng, “A low Complexity University, Stanford, California, (2002).
Delta-Dependent Beamformer,” in Proc. IEEE Ultrason.
Symp., pp.1084-1087,. 2008. [17] S. Sukittanon, S. G. Dame, FIR Filtering in PSoC™ with
Application to Fast Hilbert Transform, Cypress
[13] J. O. Smith, M athematics of the Discrete Fourier Transform Semiconductor Corp., Cypress Perform. (2005).
(DFT), Center for Computer Research in M usic and
Acoustics (CCRM A), Department of M usic, Stanford [18] R. C. Gonzalez, R. E. Woods, Digital Image Processing,
University, Stanford, California, 2002. Pearson Prentice Hall, Upper Saddle River, New Jersey,
2008.
[14] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal
Processing. NJ: Prentice-Hall, Englewood Cliffs, 1989. [19] M . O’Donnell and S.W. Flax, “Phase-aberration correction
using signals from point reflectors and diffuse scatterers:
[15] B.G. Tomov and J.A. Jensen, “Compact FPGA-Based measurements,” IEEE Trans. Ultrason., Ferroelect., and Freq.
Beamformer Using Oversampled 1-bit A/D Converters,” Contr. 35, no. 6, pp. 768-774, 1988.
IEEE Transactions on Ultrasonics, Ferroelectrics and
Frequency Control, vol. 52, no. 5, pp. 870-880, M ay 2005.