MQ 53448
MQ 53448
MQ 53448
PROPAGATION IN TISSUE
I*m
National Library
of Canada
Bibliothque nationale
du Canada
Acquisitions and
Bibliographie Services
Acquisitions et
services bibliographiques
Canada
Canada
ABSTRACT
A new, computationally efficient approach for rnodelling finite amplitude ultrasound
order operator splitting approach, enabling the effects of diffiaction, nonlinearity, and
absorption to be calculated separately over relatively large incremental distances using a
hctional step marching scheme. A computationally efficient angular spectnim algonthm
has also been developed to calculate the diffkactive propagation fkom non-axisymmetric,
non-separable sources. Results of our model have shown close agreement with published
data. Moreover, our approach may offer computational savings compared with existing
models. Indeed, with our algonthm it should be possible to simulate the nonlinear
propagation of sound beams from realistic medical ultrasound scanners, and perhaps to
investigate ways to improve the design of tissue harmonic imaging systems.
Acknowledgernents
1would like to express my sincere appreciation to Prof Richard S.C. Cobbold, and to Dr.
Jahangir Tavakkoli for their mentorship, fiiendship, and patience. Prof. Cobbold has
always been cheerfully available, and open to discuss matters acadernic or casual. Dr.
encouragement and love. My son Justin has been with us for the last eight months of my
thesis, and it is my farnily that has given me much inspiration and happiness.
Table of Contents
ACKNKOWLEDGEMENTS
....
...............................................................
CHAPTER 1
...............................................
...
111
iv
1
CHAPTER 2
2.1
APPROIYIMATION.........................................................................................
2-2 ATTENUATION.............................................................................................
7
-8
CHAPTER 3
DISTORTION................................................................................................. 18
3.2.3
...........26
CHAPTER 4
EFFECTS OF NONLINEARITY.DIFFRACTION.AND
ABSORPTION...............................................................................................
33
4.4.1 The Choice of a Nonlinear Algorithm........................................... 34
4.4.2 The Inclusion of Absorption ......................................................... 34
4.4.3 The Choice of a DifEaction Algorithm ........................................ 34
A COMPARATIVE EVALUATION OF ALGORITWMS FOR
CALCULATION OF DFFTUCTION ........................................................ -35
4.5.1 The Rayleigh Integral ..................................... .............................. 36
4.5.2 Impulse Response Formalism ....................................................... 36
4.5.3 Angular S p e c t m Method ............................................................ 38
THE ANGULAR SPECTRUM MEWOD: THEORY ................................ -40
4.6.1 The Spatial Frequency Interpretation of a CW Transducer
Excitation ...................................................................................... 41
4.6.2 The Angular Spectrum .................................................................. 42
4.6.3 Difictive Propagation Using the Angular Spectnim .................. 42
CHAPTER 5
SOURCE.........................
.
.
.
.
..
58
CHAPTER 6
CHWTER 7
CHAPTER 8
CONCLUSIONS .........................................................105
8.1 SUMMARY.................................................................................................
105
REFERENCES ............
APPENDLX B
............................................................................. 112
and safety due to use of non-ionizing radiation. Some limitations of ultrasound compared
to other modaiities inciuae infenor resoiution and poor pencation dcpth. B e c r ~ s t
ultrasound attenuation is more severe for higher fkequencies, there is typically an implicit
tradeoff between resolution, and penetration depth. Moreover, artefacts due ro clutter,
beam defocusing due to tissue path inhomogeneities, and multiple reflections can diston
Mallart and Fink, 1994; Zhu and Steinberg, 1993 a & b; Karaman et. al. 1993 ) have had
lllnited clinical success.
Chagter I
Introdzrction
-7
relationship between pressure and density. The compression phase of a wave will travel
faster than the rarefaction phase, and thus, as the wave propagates, it will undergo
distortion, which will be more severe for higher pressure amplitudes. In the Erequency
domain, nonlinearity means that harmonics and sum and difference frequencies will be
generated.
The concept of nonlinear propagation, goes back to the work of Euler in 1755. In the
medical field, although the vast majority of research and developrnent in medical
ultrasonics has assumed linear propagation, it has been well undentood that nonlinear
effects play a non-negligible role - especially for ultrasonic devices which use high
amplitude sources. Therapeutic applications such as Iithotripsy for minimally invasive
kidney stone hgmentation, generate very hi&
as well as sum and difference frequencies. Contrast Harrnonic Imaging is based on ushg
harmonic echoes fiom contrast microbubbles to form an image. When investigating the
potential use of ultrasound contrast agents, researchers found that relatively clear B-mode
ultrasound images could be formed by receiving at hvice the fkequency of the transmitted
Chapter I
Introduction
beam, even when no contrast agents were present. The second harmonic image in the
absence of contrast agents, is fomed from backscatter of harmonics generated in
nonlinear propagation through tissue.
A key paper by Ward et. al. (1997) demonstrated the feasability of applying nonlinear
propagation to B-mode imaging. Averkiou et. al. (1997) m e r demonstrated that
adecpate harmonic signal level could be obtained nom tissue propagation to enable its
use for imaging. Chnstopher (1997, 1998) has demonstrated with modeling and
experimental work that sufficient harmonic signal can be obtained within current
standards, and that harmonic imaging can reduce the degrading effects of phase
aberration. Li et. al. (2000) have recently published a computer mode1 for simulating
realistic tissue harmonic images using an axisymetric source, and have studied the
improvement in resolution and image contrast derived fi-omTH1 compared with
conventional B-mode imaging. Clinically, TH1 has been used in a number of fields,
inciuding cardiology, and has shown promising results. Clinicai images show a marked
improvement in haze and other artefacts cornpared with fundamental imaging, not to
mention higher resolution (Tranquart et. al. 1999). Tissue Harmonic hnaging is already
in widespread clinicd use, and medical ultrasound manufacturen such as ATL, Acuson,
General Electric, Hewlet Packard, Siemens, etc. al1 have machines with TH1 capabilities.
Many investigaton will agree, however, that TH1 is still in its infancy, and has much
room for optirnization. Humphrey, in a recent revew article (Humphrey, 2000)
comrnents on the need for "more efficient algonthms and simpler models to be able to
predict the effects of nonlinearity in a given field." He notes some recent developments,
simulate a technology before manufacturing. Often modeling not only saves money by
dlowing virtual research and developrnent, but it also helps foster an understanding of
Chapter I
Introduction
principles needed for an optimal design. Our work centen on modeling nonlinear
propagation in tissue from a realistic medical ultrasound scanner.
The model we have developed contributes a new perspective in modeling methodology.
Currently, there are a number of approaches to modeling nonlinear propagation. Many
are based on the 'KZK' equation, which is valid for nonlinear propagation in
thermoviscous fluids. for fairly directional sources, and for sources which are not too
focused. Results are not valid in the extreme nearfield, or far off axis as it rnakes the
parabolic approximation. Numencal solutions have been investigated in the time domain
(Lee and Hamilton, 1994), and the fiequency dornain (Aanonsen et. al., 1984). Much of
the work has been done for propagation in water, and for axisyrneaic (piston or focused
disk) transducen. Cahill and Baker (1998)have simulated nodinear propagation in water
from a (non-axisymmetric) phased array transducer.
Christopher and Parker (1991) have developed another method, not based on the KZK
equation, wherein they model nonlinear propagation by breaking up difiaction and
nonlinearity over m a l 1 steps, and solve for diffraction using a fast transfomi - based
angular spectnim method, and solve for nonlinearity using a frequency domain approach.
Their results are valid for propagation in arbitrary media (not just thermoviscous fluids),
and their work can be applied to non-axisyrnmetric sources.
Tavakkoli et. al. (1998) developed a tirne domain model based on a second order operator
splitthg. They used the Rayleigh integral to compute the effects of diffraction, a time
domain algorithm to compute nonlinearity, and a minimal phased FIR filter to calculate
attenuation. Their model has been verified for axisymmetric lithotripter sources.
One of the difficulties in using computational modeling of nonlinear propagation is the
heavy computational burden. Axisymmetric codes dominate the literanire, and even with
this symmetry, computational complexities of curent models are fairly intense, requiring
Chopter 1
Introduction
requirements. The step fiom continuous wave sources to pulsed sources requires yet an
additional degree of computational complexity.
So far, published literature is devoid of papers explorhg the modeling of nonlinear
propagation in tissue fiom a linear phased m a y . The computational burden needed to
accomplish such a task may be one explmation for ihis. Simulating fuiite amplitude
propagation in tissue from a phased array will be addressed in this work.
Simulation of various types of transducen (inciuding linear phased arrays) and arbitrary
waveforms should be possible. Cornputational efficiency is a primary concem, and
we hope that our modeling rnethodology rnay prove to yield some degree of
cornputational savings compared to existing models. Simulation of nonlinear pulse
propagation in tissue fiom non-axisymmetric sources should be realizable on a personal
cornputer with a nin time of a few hours. Verification of our model should be possible, by
cornparison with published experimental and theoretical results.
Ultimately it is hoped that our model can be used to investigate optimal design schemes
and parameters for a tissue harmonic imaging system. With this motivation, peripheral
investigations stemming from this thesis will include using coded excitation and pulse
inversion to improve the signal to noise ratio in harmonic imaging.
Chapter I
Introduction
choose to Mplement will be discussed in the Chapter 4. In Chapter 5 we will discuss the
nonlinear pulsed propagation. The concluding chapter will summarize this work and its
main conibutions, give recornmendations for future work, and outline conclusions.
differential equation. The assumption of linear propagation, however, is valid only for
relatively srnall disturbances.
In considering wave propagation in fluids, one can obtain the linear wave equation from
the Navier-Stokes Equation by making a srnall signal approximation for the density and
pressure, given by
where the subscript 'O' denotes the equilibnum quantity, and the subscnpt '1' indicates
the changes in the quantities, which are small. This small signal approximation leads to
the well known equation
where p~ is the buik viscosity, p is the shear viscosity, and K is the adiabatic
compressibility. This is a linear equation, and can also be expressed in terms of the
normal particle velocity u, or the velocity potentia.14.
V'Q,
+ k-2 0 = O
where
t2
Chapter 2
Theoretical Backpround
In the absence of viscous loss, the time domain equation becomes the familiar
homogeneous wave equation:
and the ftequency domain equation looks the same except that
S=k .
It is useful to consider the 1-D version of this equation for particle velocity - and express
it in the fonn:
which can be factored to obtain two uncoupled wave equations called reduced equations
or evolution equations, one of which is:
which descnbes plane waves propagating in the positive z-direction. Going to retarded
time, r = r - z / co we have:
2.2 Attenuation
The assumption of viscous loss in the equation of state gives nse to a quadraric frequency
dependence for the attenuation of the wave propagation. This assumption is only valid for
thenno-viscous fluids, however, and is not true for tissue. Tissues generally have a more
cornplicated loss model, and the attenuation coefficient is govemed by af dependence
where n is typically in the range 1.1 to 1.5.
are large enough to invalidate the small signal approximation are often referred to as
Chapter 2
Theoretical Rackpround
beginning and end of a pulse propagate with the small signal speed c,, within the pulse,
the propagation speed varies. The variation of propagation speed with initial amplitude is
actually due to two separate effects: convection, and nonlinearity of the medium.
Convection effects cm be thought of as being like an oscillating wind travelling with the
wave. Overall, the oscillation propagates with small signal speed c,, however, the peak of
the oscillation wiIl also have a local particle vefocity ii above and beyond the wave
velocity c..
Effects due to the nonlinearity of the medium can be understood as a dependence of the
speed of sound with temperanire and pressure. The compression phase of a wave will
cause a local increase in pressure and temperanire compared with the rarefaction phase.
Locally, an increase in pressure and temperature causes an increase in the speed of sound.
Thus the compression phase of a wave travels faster than the rarefaction phase. Note that
because the speed of sound is dependent on density, the plane wave impedance relation,
is no longer a linear relation. The slope of a graph of pressure venus density, is thus not a
straight line, but is rather a curve, where the local slope is proportional to the square of
Cha~ter2
Theoretical Backaround
IO
The general propagation speed of sound can thus be written as (Beyer, 1974)
where we will cal1 p the parameter of nonlinearity. We will show in section 2.3.2 that
where fint term (unity) is due to convection, and the second term (Bm) is a parameter
related to the nonlinear relationship between pressure and density. in the case where rc is
very low, the speed of sound reduces to c., the mal1 signal speed. Convective and
nonlinear effects can collectively be refened to as nonlinear effects, as both these effects
will contribute a nonlinear term in the differential equations describing nonlinear
propagation.
C h ater
~ 2
Theoretical Backmound
II
These nonlinear effects can contribute to distortion of a aven initial waveform. The
compression phase of a sinusoid, for example, will have a propagation speed greater than
that of the rarefaction phase. In an ideal dissipationless medium, a sinusoid will thus
distort into somethig which may approach a sawtooth wave, as illustrated in Fig. 1.
Figure 1 Effects of nodinear distortion of a plane sinusoidal wave. (a) Initial waveform
of a 1 MHz wave. (b) Showing the distortion after propagating 0.8 of the shock distance
in a lossless medium where the coefficient of nonlinearity is given by P=3.5.
Noniinear distortion of a waveform in the kequency domain represents generation of
Chapter 2
Theoretical Background
12
E
O
0.04
0.02
* - * .
'
20
10
30
40
--
50
Frequency (MHz)
Figure 2: Frequency spectnim of the waveform in Fig. 1b. Harmonies above IO MHz are
present in the spectnim.
2.3.2 The Parameter B/A
The thermodynamic relationship between pressure and density, is in general not a linear
one. The pressw-density relation cm be written as p=p(p,s), where p=p,+p, and
p=po+p,, and the subscripts denote the equilibriurn and perturbation values respectively.
A Taylor series expansion can be done:
where
3
= POCO
,d'p
B=POg
The first order measure of nonlinearity, HA, is the parameter of nonlinearity (Beyer,
1974). The coefficient o f nonlinearity is defhed as
(13)
B
2A
p=1+-.
Chapter 2
Theoretical Backaround
13
The Eulerian speed of sound under adiabatic conditions c m be written as (e.g. Hamilton
It should be noted that for a linear medium, P=O and NA=-2 (rather than O).
This is often called the equation of nonlinearity, and has been solved in both tirne and
fiequency domain.
2.3.4 The Poisson Solution
Given the initial value problem ii(.r,O)=G(x), or the boundary value problem u(O:t)=F(t), a
solution to the reduced wave equation cm be written intrinsically as:
Essentially the solution represents distortion of the initial waveform, govemed by the
nonlinear compression of the argument.
Chapter 2
Theoretical Background
14
c:
x =Pm.
The Fubini Solution represents a frequency domain solution, and gves the amplitude of
the nth tiarmonic as
-2
'O
b'
0.1 '0.2
0.; 0.4'0.5
0.6 0.7 0.8 0.9
Propagation distance normalized to shock
Figure 3 First five harmonic amplitudes generated fiom nonlinear plane wave
propagation in a lossless medium.
wave travelling in a non-attenuating medium, a shock wave foms when the maximal
slope of the wave becomes infinite. This happens at the shock formation distance X .
Although we do not expect shock waves in a tissue medium when diagnostic pressure
levels are used, E will be used as a scaling parameter in the nonlinear algorithm.
Chapter 2
Theoretical Background
16
which is known as Burgers Equation (Beyer, 1974). We may note that in an attenuating
medium, a plane wave will never really shock in the true sense, but raher, the 'shockfiont' will have a finite thickness, rather than an abrupt discontinuity.
The most widely used equation for modeling fmite amplitude sound bearn propagation,
and which accounts for the effects of difiction, nonlinearity, and absorption, is that due
to Khokhlov, Zabolotskaya and Kuznetsov, which is commonly referred to as the KZK
equation. It c m be written as (Kuznetsov, 1971)
where
is the transverse Laplacian operator. Moreover, it can be cast in the form of an evolution
equation by taking the temporal (retarded time) integral of the above equation. We could
also express the KZK equation in terms of the z-component of the particle velocity by
using a linea.plane-wave impedance relation, uz 5 p /(poco), which is valid to the order
of the approximations made in the derivation of the KZK equation.
Chapter 2
leoretical Backaround
17
The nght hand side of equation (22) describes absorption and nonlineuity, and the left
hand side describes spatial evolution and difnaction. The KZK equation describes quasiplane wave propagation, and is based on the so called parabolic approximation, which is
that the angular spectrum is assumed to be sufficiently narrow so that the wave will be
close to planar, and evolving slowly in the z-direction. This approximation is not valid for
strongly focused beams or for beams with strong irregulaxities in the transverse structure,
such as occurs in the nearfield zone. Thus, the mode1 is limited to cases where difiaction
effects are minimal and focusing gains are small. The parabolic approximation is valid
for focused sources provided that
In the next chapter, a discussion of numerical methods for solving the KZK equation, will
also include a bief description of how arbitrary attenuation may be accounted for in a
computational scheme to solve the KZK.
1-D(plane wave) and 3-D nonlinear propagation, including the effects of difiaction,
attenuation and nonlinearity, that are relevant to the contributions of this thesis.
3.1 Computational Models for Nonlinear Plane Wave Distortion
where v&)
This expression involves two quadratic surns, the first representing accretion of the nth
harmonic by a nonlinear combination of other harmonics tliat have a surn fiequency of
n/,. With conjugation, the second quadratic sum may be interpreted as depletion of the
nth h m o n i c harmonics with ciifference fiequemies nf,. Note that to make the
computation redizable, (2) assumes that the solution to Burgers equation can be
represented by a finite number of haxmonics. When there is no attenuation, (2) is valid
o d y up to the shock distance, and the solution becomes equivdent to Fubini's solution
given by (18) in chapter 2. Fenlon (1971) was the fust to derive coupled spectral
equations that are equivalent to (2). He investigated mono and bi-fiequency sources
(including cylindrical and spherical spreading. Korpel(198 0) presented coupled spectral
Chapter 3
19
for arbitrary absorption and dispersion in progressive plane waves. To do this one rnay
make the substitution
(3)
a,f n 2 + a, + jd, ,
in (2), where a, is the attenuation coefficient, and d. is the dispersion coefficient for the
nth hamonic. Attenuation rnay often be modeled by a power law relation, described by
(4)
a, =a,fff,
where a,is a constant,f, is the frequency of the nth harmonie, and b is an arbitrary real
number. For biological tissue, b is typically between 1.1 and 1S. The dispersion
coefficient rnay be expressed as
(5)
d, = n w, ?(ci'
-ci' ) ,
and accounts for the deviation of the phase speed c,, of the 11th hamonic component from
the infinitesimal sound speed of the fundamental c.,
Haan and Cook (1 983) also investigated plane wave nonlinearity using the FDSBE, but
which we were unable to verify.
used a somewhat different f o m from that given by (3,
Christopher and Parker (199 1) also cite a form for the FDSBE that diffen fkom (2). It
should be noted that the form of the FDSBE given by (2) is quoted in Hamilton and
Blackstock (1 998; pp.3 13-3 14). We have venfied that in the limiting case where there is
no attenuation, the numerical results obtained by using (2) reduce to those obtained using
the Fubini solution.
The main disadvantage of the FDSBE is that hundreds or thousands of harmonics rnay be
required to accurately propagate pulses or waves which develop shockfionts, and as a
result, the cornputational burden can be large. A promising algorithmic approach to
reducing the number of harmonics required for propagation of wavefoms with one (and
only one) shockfont is that introduced by Pischkal'nikov et. al (1996). The premise of
their scheme is that a shockwave rnay be thought of as a sum of a perfect sawtooth wave
and smoothly varying (unshocked wave), which is well represented by only a few (for
example 30 or less) harmonics. Because the Fourier transfomi of a sawtooth wave is
exactly known, an infinite nurnber of harmonic~rnay be retained in the spectnun of a
Cha~ter3
20
Equation (9) can be nurnerically irnplemented by sampling the temporal waveform, and
where m is the index of the temporal wavefom and n represents the nth step Az, in r (e.g.
Tavakkoli et. al. 1998). The discrete version of the inequality (7) c m be written as
Chapter 3
21
du 1 p ( u ( t
-=dt 2
+ ~ t- ~) ( t )-)(
At
Al
Further precautions are needed to propagate shockfronts. First, shock segments are
identified by locating consecutive sarnples for which there is a change above some
threshold level. Next, consecutive shock segments are consolidated into shockfronts, and
then propagated as a whole without M e r steepening. T i , Christopher's scheme is
implicity harmonc-limited, and allows considerable computational savings, while
There are few analytic solutions to the equations of noniinear propagation, and such
solutions are for very specialized conditions. In general, numencal methods must be
used, and these will be discussed under the following headings:
Chapter 3
-77
used standard backward difference methods to propagate step by step the effects of
difbction and absorption, and of nonlinearity. Their calculations were restricted to
axisymenic Gaussian, plane and focused sources. Alternative algorithms that solve for
attenuation and d i f i c t i o n in the fkequency domain, and nonlinearity in the time domain
have been presented by McKendree (198 1) and Froysa et. al. (1993).
Aanonsen et. ai. (1984), and is widely referred to as the Bergen code. They substituted
expansion (1) into the KZK equation to get a system of equations, which they then
integrated numerically using a simple implicit backward difference scheme. Their work
was primarily for nearfield studies. Baker et. al. (1988) verified the accuracy of this
numerical Ka( solution by comparing with experiment in the case of a CW plane disk
radiator in a water medium. Hamilton et. al. (1985) introduced a transformation to
Chavter 3
23
improve the efficiency of the aigorithm in the far-field. For focused beams, a
modification of the coordinate transformation was introduced by Hart and Hamilton
(1988). Other frequency domain studies have investigated radiation from bifrequency
(Kamakura et. al. 1989; Naze Tjotta et. al. 1990; Naze Tjotta et. al. 1991), pulsed (Baker
and Humphrey, 1992), and rectangular (Kamakura et. al., 1992; Baker et. al. 1995)
sources. Averkiou et. al. (1995) used the spectral code developed by Naze Tjotta et. al.
(1991) to compare theoretical mode1 predictions with experiments for finite amplitude
Cahill and Baker (19974b; 1998) used a non-axisyrnetnc version of the Bergen code, to
simulate the acoustic field of a phased array medical scanner for water propagation. They
demonstrated that nonlinearity can interact with difbction in such a way as to move the
peak region of intensity off-mis, and io cause the focal region to shift towards the
transducer.
JP
-=h
j(Vip)dr.LDbp
2 -
(Di fiaction)
Chapter 3
24
To first order in terms of the propagation distance Az, one may show that these equations
are independent of each other, so that the total change in pressure can be approximated by
the sum of these contributions, Le.,
As illustrated in Fig. 1, the pnnciple underlying the fractional step algorithm is to first
solve the diffraction equation for a plane to plane propagation, then apply absorption, and
findly to solve the nonlinearity equation over the distance Az.
z+Az
Plane 1
Plane 2
thus propagating the sound beam in a . incremental march over maIl distances Az.
Chapter 3
25
The tirne domain algorithm is well suited to pulsed propagations, and to very large
amplitude continuous wave propagation. Of particular note is the work of Averkiou et. al.
propagation in water, Little work has been done to venfy the accuracy of the KZK for
other media such as tissue. Computational requirements can become an issue for
nonlinear modeling. Baker (1998) cornrnents that the continuous wave field response
from an axisyrnmetric source at moderate drive levels can be run on a persona1 computer
in a matter of minutes. When the drive Ievel is increased, more hatmonics are needed,
thus incurring more memory and computational tirne requirements. Pulsed waveforms
are also more demanding since a wide spectnim of fiequency components must be
included. A non-axisymmetric source geometry can cause an order of magnitude increase
in the computational burden. The results of Cahill and Baker (199%) required about
For pulsed propagation, Tavakkoli's method has the advantage of using fairly big Asteps between propagation planes, but suffers fiom long computation times for the
Chapter 3
26
Figure 2 The second order operator splitting method of Tavakkoli et. al. (1 998). In this
scheme the combined effects of diffraction, attenuation, and nonlinearity are
approximated to a second order in the incremental distance &.
Chanter 3
27
has been applied to the modeling of highly focused lithotripter sources (1994), and for
simulation of nonlinear-based irnaging (1 997).
coefficient with a quadratic frequency dependence, and which does not describe the nearlinear fiequency relationship of most tissues. Many of the numencal methods based on
the KZK equation were restricted to thennoviscous fluids, and thus are not appropriate
for modeling finite amplitude propagation in tissue. Although it is possible to model
finite amplitude propagation in a tissue medium using the frequency domain solution to
the KZK equation, the computational burden cm be heavy, especially for pulsed
propagation, and non-axisymmetric sources. The methods of Chnstopher and Parker
(1 99 l), and Tavakkoli et. al. (1998) have enabled accurate modeling of nonlinear
propagation in tissue. In the approach used by Tavakkoli et. al. the Rayleigh integral is
used to compute the effecrs of diffraction, and is thus cornputationally burdensome.
Chnstopher and Parker use a fast transform approach to difictive propagation.
Although it fiords some computational savings, their approach, like others, is
computationally intense for propagation of pulsed excitations, and for non-axisymmteric
sources. In this chapter, a new method of modeling is introduced, which not only builds
on the strengths of previous algorithms, but which also aims to gain some degree of
computational savings compared with these schemes.
one. We start with a model evolution equation, valid for quasi-plane wave sound beams,
Our hypothesis is that operators Lo, LA,and LN exist such that the evolution equation (1)
accurately describes finite amplitude propagation in tissue. It is our belief that such a
hypothesis is well founded. Partial justification of using (1) is that the KZK equation can
be written in a form equivalent to (1) by integrating with respect to retarded time, to
obtain (3.22), where the operators are defhed by (3.21). We feel that propagation in
tissue is likely sirnilar to propagation in thermoviscous fluids, with only attenuation
mechanisrns being sipificantly different. Experirnental confirmation of this hypothesis
stems ffom the efforts by Tavakkoli et. al. (1998). They used equation (1) as a basis for
their model and obtained reasonable agreement with experiments of finite amplitude
propagation in a tissue mimicking fluid. Moreover, it can be shown that the approach of
Christopher and Parker is equivalent to a first order operator splitting of an equation
equivalent to (1). Thus although (1) is not derived from elementary pnnciples of physics
as is the KZK equation, there is ample evidence that it can accurately model finite
amplitude propagation in arbitrary media, including biological tissue.
Our approach to solving (1) will employ an operator splitting scheme so as to solve for
the equations of dif'ftaction, attenuation, and nonlinearity separately over small steps.
To intmduce the operator splitting techniques of this thesis, some notation will fint be
developed. Consider fint the process of difiaction. Given the normal particle velocity
field profile v(xry,zl;t) across a plane z = zl, we may represent the field profile across
plane s = 1 2 assuming difictive propagation only (no nonlineety or attenuation) by
introducing an operator D.&,
such that v(x,y,z2;t) = ~g.h-v(.r,y,zl;t).The N O planes z=z,
absoqtion respectively. There is a relationship between the operators r and the operators
29
30
L of (l),which will be explained in Appendix A. For now it will suffice to Say that
v(x,y,z2;t)=
TA,d_~(~,y,~r;t)
is a solution to the absorption sub-equation given the initial condition v =
v(.ry,,zi;t) on plane z = z,.
As our approach is similar to that of Chnstopher and Parker (1991), it is important to note
that thek mode& methodology is equivalent to a fint order operator splitting, which
where ~+I+N.h-~(.r.y,~l,t)
represents a solution at z = q+d= to (l), and &+.4.il-~(.r,y.~l,t)
represents a solution at r = :l+&
to the sub-equation
given the initial contition v = v(.r,y.z,,t).In words, (3) says that the combined effects of
diffraction, absorption, and nonlinearity c m be approximated by one incremental step of
combined difhction and attenuation, followed by one step of nonlinearity. In Appendix
A, we will prove the fint order nature of this scheme.
v(~,y,~~;t)~r~+~+~.~~(~,}',~~;f)=r~&~~~~+~
or thousands of fiactional steps out to the focal region of a transducer. Sixnilarly, the
finite difference techniques used to solve the IUK equation required relatively fine axial
meshing.
31
v(x,y,~Z;~)~rD,A+N,kv(x,y,~,;t)=r~r~.~,2~;I-r~+d.h-,r~(~y,~I;t
which, iike the Tavakkoli (1998) operator splitting, is second order in the small ordering
parameter k . This operator splitting is illustrated in figure 2.
-.-
+ --
Plane z=zl
Figure 2 illustration of the proposed second order operator splitting theorem.
Proof of (6) is provided in Appendix A. It should be noted that both second order
operator splitting methods can be used within the context of the KZK equation. Proof of a
second order operator splitting aigorithm for the KZK equation is also given in
Appendix A.
It is hoped that by taking advantage of the larger axial steps which these second order
operator splitting schemes allow, and by using faster algorithrns for difiaction,
considerable computational savings cm be obtained.
The second order operator splitting presented in section 4.2 can be applied to solve for
the effects of diffraction, nonlinearity, and absorption over these small steps. Thus, we
propose a fractional step marching scheme, which progresses fiom plane-to-plane dong
the direction of propagation.
In solving the difiction sub-equation (2), or the sub-equation combining diffraction and
attenuation (4), we may guess that a solution to either of these equations must also be a
solution to the homogeneous wave equation, as this is the equation which govems linear
propagation phenornena. When the parabolic approximation is made, as is the case for the
KZK equation, the solution to the sub-equations are only approximations to the true field
which is a solution to the homogeneous wave equation. The approaches of Tavakkoli and
Christopher do not make the parabolic approximation, and dimction is computed
exact ly.
Like the KZK equation, out model equation (1) of this chapter is only valid for quasiplane wave sound beams. The quasi-plane approximation assumes that the sound beam is
directional and that the angular spectmm is nmow. Thus the nonlinear substep used in a
hctional step marching scheme, based on an operator splitting of our model equation is
lacking multi-directionality of propagation. It is assumed that the nonlinear propagation
32
linear way. Moreover, the parabolic approximation need not be made and diffraction may
be calculated exactiy. Because of these reasons, a fractional step rnarching scheme based
on (6) may prove to be more accurate in the near-field than the W( approach. This
prospect will not, however, be investigated, as the very nearfield is of little importance to
simulating medical ultrasound systems.
One disadvantage of using a fiactional step marching scheme to solve for nonlinear
propagation is that modeling errors may accumulate over many fiactional steps, and the
scheme rnay become inaccurate at best and unstable at wont. Algonthmic burden, on the
other hand is problematic. OAen there is an inherent tradeoff between accuracy and
computational efficiency, and good judgernent must be used in reaching a suitable
compromise between the two. The choice of an accurate yet efficient way of computing
the effects of diffraction, attenuation, and nonlinearity is thus key to developing a
simulation tool for harmonic imaging.
33
(FDSBE)is a good choice for a nonlinear (plane-wave) algorithm. We use the FDSBE
presented in chapter 3, when moderate drive levels are simulated. For very high
amplitude excitations, a number of authors comment on the benefits of calculating
nonlinearity in the time domain.
The thne domain nonlinear algorithm presented by Christopher (1994), and described in
section 3.1.2, will be used for pulsed excitation modeling, since it offers great stability, is
implicitly harmonie-limited (and thus efficient), and is reasonably accurate.
4.4.2 The Inclusion of Absorption
Absorption can be caiculated together with either the diffraction or the nonlinear
algorithms. If nonlinearity is modeled in the tirne domain, we will couple absorption with
the diffiction algorithm. If nonlinearity is modeled in the frequency domain, absorption
may be included with the nonlinear algorithm (for stability), or with the diffraction
algorithm.
4.4.3 The Choice of a Diffraction Algorithm
The choice of an efficient difhction algorithm was a pnmary focus of this thesis. Its
importance is illustrated by the observation that for the algorithms of Tavakkoli et. al.,
which used the Rayleigh integral for the difhction computation, 90% of the CPU time
was being used on the diffraction substeps.
following requirements:
34
Computationally efficient
Plane to plane
Suitable for a linear phased array, and other transducers with and without apodization.
Should include the possibility of beam steenng.
in terms of N, Nsand N,. We will, in particular, investigate the foIlowing three methods:
(1) The Rayleigh approach (2) The impulse response fomalism, and (3) the angular
speceum method.
35
v,(f-Rlc,)
4(r,O=-IJ
dS0 ,
sa
where $I is the velocity potential at the observation point, vn is the normal particle velocity
acmss the s o u c t , R is the distance between the source element and S, is the surface
across the source. Let us consider the cornputational complexity of the Rayleigh integral.
interest, we would evaluate a double integral over the surface of the transducer/ource
plane. Thus for al1 times of interest, this would require Nl(NxNy)calculations. To do this
for dl points in the plane n
i
,
would require N,(N~N~)'calculations.
If there is axial syrnmetry, the Rayleigh double integral reduces to a single integral, and
only the radial component of the plane profile needs to be evaluated to represent the
entire plane z=z,. Thus the computational complexity for the Rayleigh integral (axial
cornputational burden due to the Rayleigh integral can be large if planar sampling is
dense. Otherwise, the Rayleigh integral satisfies the other requirements given in section
4.4.3.
The impulse response of a transducer is, as the name suggests the field response of a
perfect impulse excitation:
36
Here 5 is an apodization factor across the source, and the other variables are the sarne as
those given in the Rayleigh integral. The transient response due to a given excitation may
be obtained by convolving the impulse response with the source excitation:
@ ( r , t=
) vJt)
* h(r,r).
Note that this method assumes that the source velocity v(x,y,t)is separable, Le. can be
written as v,(t)~(x,y)
where 5 is an apodization factor. One of the requirements for a
suitable diffraction aigonthm in the pioposed aonlinex proprg~titiooAgorithm ic tha! the
approach be valid for non-separable distributions. To see why this is so, consider a point
source which emits a delta-function excitation. The impulse response on a plane a
distance away fiom the transducer will consist of an impulse aniving first at the center of
the plane (the point directly normal to point source), followed by impulses arriving later
in isochronal rings off a i s . The impulse response is therefore given by
which is clearly no t separable. Since we are concemed with plane to plane diffractive
propagation, it is essential that the method be capable of propagating non-separable
distributions. The impulse response formalism is thus discounted fiom our list of
potential candidates for a difiction algorithm. Nevertheless, we will consider its
computational complexity and compare its efficiency with other techniques. If it is
computationally efficient, it may be used for the fint step propagation from a transducer
surface where excitation is separable.
In this case, to calculate the velocity potential at a point on the plane P=z,, requires a
convolution of the impulse response with the surface velocity of the transducer/source
plane. This would require on the order of 3Ntlologz(N,) + N,computations if the FFT
method of convolution is employed. Thus to evaluate the velocity potential at each point
37
single integral (NINrcornputations). Once we compute the impulse response for a point in
the plane z=z, for al1 times, we then need to convolve the impulse response with the
surface velocity to get the velocity potential at a point in the observation plane. Thus for
this computation for each point in the plane r=zl, we require N,Nr(NINXNs+3N,log2(~)
+
which we shall cal1 the Tjotta impulse response method (Tjotta and Tjotta, 1982). By this
method, one does a coordinate transformation so as to integrate over isochronal ring
elements (points equidistant fiom the observation point). In this case, suppose that the
angle to integrate over a single annular isochronal ring is P, and that there are Ng discrete
points to surn over. Then for each observation point we require Np computations to find
the impulse response at a single tirne, and NgN, computations for K times. Once the
impulse response has been found for a aven observation point, we do a convolution over
+ Nt computations. We
the source velocity v. This convolution would require 3Ntlo~(Ni)
then would need to do the identical calculation for al1 points in the observation plane.
Overall, we would thus need N ' ( N B N , + 3N,logt(N,) + Nt)computations, which is
approximately NXNyNflt.
The theory of the angular spectnim method d l be given later on in the chapter.
38
It will suffice to mention that for each frequency component of the source excitation,
there exists a spatial fiequency representation of the source plane. Propagation fkom one
plane to another may be accomplished through use of a transfer function H.
Let us suppose that the surface velocity of the source plane is v(x.y.z,.t). First we need to
perform a FFT to ind the firequency domain representation of the velocity s(x.y,zo,a)for
each point (.Y,)))in the source plane. In general, this will reuuire NrNJVtl~g2~Nt)
computations. Once s(x,y,z,,o) is found for al1 points (x,y) in the transducer plane, and for
al1 fiequency components a,the following steps are needed:
For each frequency component a,
(1) Perform a 2D-FFT on s(x,y.z,. o)to get S(k, k,; r=z,). (Nx~,log2[NXNJ2]
computations).
(2) Multiply S by the transfer function H(knky; zolzl)defined in section 4.6. (IV&',
computations).
(3) Perform a 2D-IFFT to obtain the fkequency domain velocity s(x,y,zl,o).
(NxN,.log2[N'N,J2] computations).
computations for al1 the frequency
Thus we require N,[UV,N,log2(NxNJ2)+NXNY]
components. (These computations are made once the values for v(x,y,O,t) are computed).
Finally, once s(x,y,zl,a)has been computed for al1 points (x,y) on the plane z=zi,and for
al1 fkequencies o, we do an IFFT on s(x,yIrl,& enabling the time domain velocity
waveform v(x,y,zi,t) to be obtained. This should be done for al1 points (x.y) in the plane
z=z, (Le. NxN@'logs(Nt) computations). Thus in total, we require
2N,Nfltlog2(Nt)+Nt[2NXh~logz(N,N,12)+NxNy]
computations.
n i e anpular spectnim method is very computationally efficient compared to the Rayleigh
and impulse response methods, given a non-uniform, non-axisymmetric surface velocity
on the transducer/source plane.
Below is an exarnple using acnial numben to illustrate this point. Suppose we want to
compute the particle velocity for a11 points in a plane a distance Az fiom a
39
40
Method
# Computations
Rayleigh
3.5 184x10"
1 Impulse response *
1 3.5188~10'~
1 Angular spectnim
1 7.1135~10'
Besides being computationally efficient, the angular spectmm method is valid for planetoplane propagations - even when the source plane is non-separable. It can be very
accurate in the nearfield as will be demonstrated in subsequent chapters. Attenuation,
refiaction, and reflection can be easily incorporated, and sarnpling theorems will be
developed to ensure accurate modeling results. in short, the angular spectrum method is
(IgSO), and to compute the field fiom a plane piston transducer (1982). Stephanishen and
Benjamin used the angular spectnun for forward and backward propagations (1982), and
Waag et. al. explored windowing and other issues in discrete implementation of the
angular spectrum technique (1 985). Shafer (1W),and Christopher and Parker (199 1)
have applied the angular spectnim method to transducer characterization using
backpropagation, and nonlinear propagation.
4.6.1 The Spatial Frequency lnterpretation of a CW Transducer Excitation
the acoustic spatial frequency specrnim as the linear sum of al1 these point source
contributions. Mathematically speaking, for a flat transducer, we can consider the
aperature surface s(x,y,zO)to have a particular spatial fkequency spectrum, S(k, k,) given
The distribution S tells us that a particular wave vector k*k,k,k) has weighting S(k,k,).
Note that for a particular frequencyf, a particular wave vector (kl = 2nf/c, may be
specified by referring only to the (k, k,) components since t is given by the relation kL=
k2 (k:+k;).
*
41
For a disk transducer, the spatial fiequency spectnim corresponds to ajinc function, and
for a rectangular transducer, the spatial fiequency spectnim corresponds to a 2-D sinc
function.
4.6.2 The Angular Spectrum
We can also write the spatial kequency spectnim as an angular spectrum. Given the
direction cosines
=-ka.
a C COS^ =LY
P=cosw=
k.f=-kfi
y = (l-(u'+~2))'"= -@k
-02
For fume note, we will often refer to S(kfiky),given in (7) as the angular spectrum, even
though it is technically expressed in terms of spatial fiequency vecton, and not direction
cosines as is the case for (8).
4.6.3 Diffractive Propagation Using the Angular Spectrum
On a plane a distance dz away from a harmonically excited transducer, difhctive
propagation will alter the spatial frequency spectxum on the transducer surface S(k,k,.; z
=z,)
to something new, S(k,k,; z = r,) related once again to the spatial field profile at the
42
into the homogeneous Helmholtz equation, and solving the resultant second order
homogeneous differential equation in S,given the initial condition S = S ( Q P : ~ O ) .
The transfer function H is thus given by
(9)
H(k,yk,,:h.)=
for
(ki + k ; ) > k 2
Thus, given the angular specmim S(k,,k,;zJ on plane,, we can know the angular
spectnim on a plane z,, a distance k from plane, by the following transfer function
relation:
(10)
S ( k x y k , ; = ,=) H ( k x y k , ; bS(k,,k,;-,)
)
The field profile on the plane 21 can then be found by taking the inverse 2-D Fourier
transform of the above result. The sequence of steps used to compute the angular
spectrum is illustrated in Fig. 2 below.
The transfer function H can be thought of as a complex phase weighting, that governs
Huygen's principle of wave superposition. The ultrasonic field from a source may be
thought of as a collection of constituent waves - each with its own directionality and
44
where
Thus, given the field profile ~(xJ,,) on a plane,, we c m know the angular spectnim on
a plane rl a distance Az away from zo by the following 2-Dconvolution relation:
We can think of the point spread function h(x.y;k), as the field profile or response on
plane z=zo given a one point source at (xo.y,.~cO)is harmonically excited with frequency
45
where LI denoted the 2D-Fourier Transform of u, and H is the transfer function. Similar
expressions may be written for pressure field to pressure field d i f i c t i v e propagation.
If we want to know the pressure, or velocity potential at a plane a distance dz away fiom
the transducer, given the normal particle velocity on the source plane (transducer
For example, suppose we want to compute the pressure over a plane z, a distance away
f?om the transducer:
p(x,y.z,)=u(x,y,s=O)**hu,,(x,y:&)
The point spread function h..,, is in this case given by
R = Jx'+y'+&
It cm also be shown that the transfer h c t i o n Hu.>,
is given by
where G is the Fourier transform of the Greens function. Using Hu.>pcan be numencally
troublesome because of the singularity for &=O (Maynard and Williams, 1982), and the
use of
46
small incremental steps, and the field profile is marched along in a plane-to-plane fashion
using a fiactional srep marching scheme. The choice of'an efficient aiid accate
difictive propagation algorithm suitable for such a modeling scheme is important
considering previous models suffer from computational bottlenecks in the diffraction
portion of the propagation algorithm. The angular spectrum method was s h o w to be a
good candidate for such an algorithm. Discrete hplementation of the angular speccnim
scherne involves sampling and windowing issues which are nontivial. In this chapter, we
consider difiactive propagation from non-axisyrnmetric sources. The reader is refened to
Appendix B for the slightly different analysis needed for radially symmetric sources.
with some general insights into the use of the angular spectnim approach for cases with
radial symrntery, is discussed in Appendix B. They argue that the pnncipIes behind their
algorithm can easily be extended to non-axisymetric sources by using the 2DF'FT in place
of the DHT.In fact, Christopher (1999) implernented a nonaxisymmetric version of his
FSC algorithm, using the ZD-FFT. However, he used axial increments of 20 per cm,
which is a subwavelength step size. Our intention is to exploit the second order operator
splitting introduced in Chapter 4 to enable the use of much larger step sizes, while
concurrently saving computational cost. Christopher's use of subwavelength distances for
the angular s p e c t m likely elicited mal: wraparound errors. For large-step diffractive
propagation, however, there are some unforiunate problems which will require careful
analysis. First, however, we will fomalIy introduce the notation, theorerns, and
Chapter 5
48
5.3.1 Notation
There are two approaches one may take to irnplement the angular spectnim method
numencally. One approach, which we shall cal1 the kequency sampled convolution
plane difiactive propagation, and the source plane is actually some distance fiom the
transducer, then the field profile across the source will be of infinite extent, and must be
truncated and windowed. Windowing issues will be dealt with in section 5.7. We shall
denote the desired spatial extent of the source in the r and y directions by t X, and f Y,.
The spatial extent of the source plane should be chosen so that the buik of the field
energy across the plane is contained within these boundaries. Now we must sample s(x,y)
over the interval [ IX, ,t Y, 1, wi th sarnpling intervals Ax and Ay. Let Nx,be the number
of samples over OUrcX,, and Nv, be the number of samples over WFY,.The sarnpling
should adequately capture the plana variations in the source distribution, as determined
by the 2D-Nyquist Theorem.
Next assume that we wish to find the d i h c t e d field profile across a plane z =
distance dz away fiom the source. Again, the d i h c t e d field profile on the observation
Cha~ter5
49
plane may be of infinite extent, however, we must choose a finite window in which we
desire correct results. Ca11 the x and y spatial extents of the observation plane
+ Y,
+ Xoand
respectively.
The point spread function h must also be sampled with adequate sarnpling out to some
f i t e extent. The point-spread function h is of infinite extent, and must be tnincated. The
required convolution is only nurnericaiiy redizabie irboth s and ii iue of finie exteni.
Note that h must be sampled out to at least [ X ,+ X,, Y, + Y,] to be able to get accurate
for -Nr,<m< +NT,-1, and -NT,<nc+NT,- 1. Correct results are obtained in the region
-Nxo<mc +NY.1, -Ny,<nc +Nb-1, with aliasing erron dominating the region exterior to
this.
{X,
v
A
(spanning Nu
-points):ischosenso-as
to contain the bulk of
the energyin the
source plane :-- - -
Y, (spanning Nys
points) is chosen so as
to contain the bulk of
the energy in the
( source plane
)
U, (spanning Nu,,
points)
is the extent to wh&h
correcttesu'itsare-desirecl in the .
:
observation.pTane---
Figure 6 The notation and variables used in the 2D-F'FT based angular spectrum
algorithm for (a) the source plane and (b) the observation plane.
Chapter 5
51
2D-FFT as follows:
Having chosen [T&Ty],
and [dr.Ay],and sampling the sequences s and h as outlined
above,
1. Zero pad s[m,n] out to [NTuNn].Ifs is of infinite extent, finr use a tapered window o
taper the edge discontinuities fiom tmcation to zero. This will prevent edge artefacts
in the convolution. Call the zero padded (and tapered) window s,[ni.n].
2. Take the 2D FFT of both hd and s, and cal1 the results f i and S,, respectively.
3. Take the 2D IFFT of the product Hdx
S'.
4. The correct result of the convolution can be found by extracting the central [Nx,,Nyo]
core results of the inverse transfomi. The results outside this core have wraparound
emor.
transfer function H in the frequency domain with discrete spatial fiequency sampling
defined above. Cal1 this sampled version Hd.
3. Zero pad s[m,n]
out to [NnNy].Ifs is of infnite extent, first use a tapered whdow to
taper the edge discontinuities fiom truncation to zero. This will prevent edge artefacts
Chapter 5
52
Thus, before using most 'black-box' 2D FFT algorithms, we must fmt shift the sequences
so their bonom edge coincides with the orign. M e r using the 2D FFT algorithrn, the
sequences must be shifted back so that they are centered about the origin. This shifting
process can be accomplished by a function (called ntshifE in MATLAB),which swaps
5.2 Accurate Sampling of the Point Spread Function h for the Case of
Single Step CW Diffiractive Propagation
The Nyquist theorem demands that the maximal frequency content of the point spread
Function h ultimately dictates the sampling we should use in the spatial dornain. Thus, by
iooking at H,the fkequency dornain representation of h, we can get a pichire of what kind
of sampling dx and Ay we need in the spatial domain.
Inside the radiation circle,
(6)
kj+k:=(hrflc)?,
the transfer function H(k,,k,) has magnitude 1, and tapers off to zero exponentially
beyond it. Thus if we choose a threshold for the maximal frequency content [k-k,]
of h as being slightly beyond the radiation circle, (Le. ,k
= 2#/c
+ E, where c is a mal1
amount), the point spread function will be well represented, and the 2D FFT of h will
have negligible wraparound error. This can be understood by considering that the
Chopter 5
53
wraparound errors associated with Hd will be negligible. Thus the SSC algorithm using
the 2D FFI' will be accurate. Note that the sampling defined by the Nyquist threshold
case, one might consider a much derser sampling, or? altematively. sarnpling the transfer
fnction H in the spatial fiequency domain,
5.3 Accurate Sampling of the Transfer Function H
Sampiing the transfer function H in the spatial frequency domain in 2-D somewhat more
subtle than Christopher's axisyrneteric (1-D)
treatment. The crux of the problems are
associated with what one might cal1 circle-square issues.
Coiisider the following three cases.
1. The radiation circle is completely within the spatial frequency (transform) domain
2. The radiation circle is only partially within the spatial frequency (transform) domain
and very dense oscillations, and cannot be adequately sampled. Sampling the point spread
function h in these cases is the appropnate route. For case (l), adequate sampling of h is
guaranteed, since the radiation circle is contained within the transfonn domain, and
wraparound error Erom the evanescent tail of H is negligible. For case (2), h must be
sampled at U2 or better, even though this may not be the sampling scheme for the source.
The 2D-FFTof h may,however be muicated to the extent of the angular spectrum of the
source,enabling the computation to be done.
Chapter 5
54
In case (3), where the radiation' circle is completely beyond the finite-extent spatial
fiequency domain, the oscillations of H can be adequately sampled, &en the Nyquist
rate of the most rapidly varying portion of the 2-D transfer function H (within the finite
spatial frequency domain).
contained within the transform domain. The highest kequency oscillations will be at the
corner of the transform domain, since we know the oscillations become more dense dose
to the radiation circle. We can use the 2-D Taylor expansion of the argument of H in the
epsilon neighborhood of [k,,k,,]
transfonn domain.
Given that
H(k,,k,;Az)
=e
j J k~Z - ( k : + k t )
is the transfer function for the fi-equencyf (and k = t@), we may define a fimction g,
such that
to fint order
Chapter 5
55
and a similar result c m be obtained for P,. Thus the sampling rate Ak, is given by
(7)
Ak,S-
2P
=
.c
n Jk'
-(kt,
+ kt,
)
3
2 A Z r ~m u
and a similar result is obtained for bk,. Al1 of the above analysis assumes that the
transform domain is completely within the radiation circle.
Essentially, we can use such a scheme where the extent of the angular spectmm of the
source is well contained within the extent of the domain, or at least the bulk of the energy
is contained within the transform domain. This will be the case for field distributions
which are relatively smooth, and slowly varying in the spatial sense. Thus the nearfield of
a source would be a poor candidate for using this scheme, and sampling h would likeiy be
a better choice.
5.4 The Limitations of the Ray Theory Truncation for 2-D Diffractive
Propagation
The idea of the ray theory truncation, described in Appendix B, is to limit the maximum
spatial fiequency extent based on the maximal angle between the source and the edge of
the plane where correct results are desired. Tmcation in the spatial fkequency domain
can reduce the computational requirements for the SSC and FSC a l g o n h s . The ray
theory truncation may be used to uncate the product H x S, and thus decrease the size of
the matrix on which to perform the inverse 2D-DFT. However, rnay be unprofitable if the
2D-FFT is used, which requires that the matrix be of size 2 x 2m points (where n and rn
are integen). The ray theory tmcation may not be used to decrease the spatial sampling
of s or h directly, as this may incur senous aliasing errors in the convolution.
Chapter 5
56
transducer surface. A course sampling of the aperature results in poor nearfield accuracy
of not only the angular spectnun method, but also the Rayleigh method, and the impulse
response method. The farther the observation plane from the source, the more accurate
the computed field profile (b-sed or? the $ven gridding scheme) will be. Reasonably
accurate results may be obtained for observation planes which are in the farfield of each
sarnpling element. Nearfield inaccuracies may be a source of instability for a plane-toplane fiactional step nonlinear propagation algorithm, which we intend to develop.
Nearfield erron may propagate incrementally in the rnarching scheme, and so it is
important, if at al1 possible to secure reasonably accurate nearfield results. One way of
ensuring nearfield accuracy is to use an extremely dense gridding scheme. This however
cornes at the expense of greater computational burden.
For phased arrays, an optimal design for eliminating grating lobes involves using
elements whose inter-element spacing is less than h/2. Should we sample the m s d u c e r
in the spatial domain, very fine mesh would be needed to define the smdl detail present.
Aitematively, sampling the transducer in the spatial fiequency domain, has an additional
advantage of using moderate sampling rates for representing the angular spectnim of a
source.
The angular spectrum only needs to be defined out to a spatial fkequency km, of slightly
better than k, = 2 S c , since fter a step of difhctive propagation, the transfer fnction H
Chapter 5
57
will filter dl spatial kquencies beyond the radiation circle. In fact, we have found that
sampling the aperature in the spatial fkequency domain, can give very accurate results even in the nearfield.
where A,, is an apodization factor for the nth element, e'" is a phase delay for element n,
and 5 represents the complex phase delay and apodization of a lens applied to the
elevation plane of the entire array and used for additional focusing. Here the function
E(x,y)=rect(xAV)rectO,/..
a lens is used, the y-component of the Fourier Transform may be difficult to evaluate
for diffictive propagation nom array transducers is to use a non angular spectnun
technique for the first step, such as the impulse response method of Ullate and San
Emeterio (1992), which is appropriate for calculating the transient near-field of phased
array transducers. The angular spectrurn method could then be applied in a plane to plane
Chapter 5
58
transducer was compared with the lateral profile computed with the Rayleigh integral.
Our approach shows good agreement in the farfield, at a h . Figs. 7(a) and (b) show the
cornparison at the axial distance a%.
-5
-3
-2
-1
Chaater 5
59
Chapter 5
60
domain sampling outlined in section 5.6.2 was simulated. The array had 16 elements, and
was of height H=2cm,width W=h/2, and had inter-element spacing s=;V4. It used no
focusing or apodization, and was excited atf = 1 MHz, with a normal particle velocity of
lcmk The lateral field profile was computed at a distance z=x(H/~)'/A.away fiom the
transducer. The results are show in Fig. 8(a).
To assess the accuracy of the technique, results were compared with a calculation based
on the Fresnel approximation. The expression used was (Crombie et. al., 1997)
wbere
and where o =
4:-
Results of this calculation are shown in Fig. 8(b). Note the close agreement between the
two methods.
-5
-3
-2
-1
Chapter 5
'
0
-5
-3
-2
-1
61
profile with the known analytic curve. As shown in Fig. 9(a), the agreement between our
algorithm and the analytic solution are in excellent agreement. Fig. 9@) shows the field
Chrinter 5
62
disiribution of the transducer over a plane dehed by the lateral and axial coordinate
axes. The dark region near the edges ofthe bearn profile are due to windowing.
10
15
20
Chapter 5
63
-2
xin an
Figure 10 (a) The normal velocity field profiles from a linear phased array (a) in the
N = 64
elevation plane () in the azimutha1 plane. The parameters are: f i 2
elements, F = 10 cm, Po= 347kPa,height 1cm, and length = 3.7cm.Note that this is a
finite amplitude signal level. Propagation was in tissue, with a = 0.3 dB/cm and b = 1.1.
(1992) be used to calculate the first-step linear field profile across a plane normal to the
Chmter 5
64
direction of propagation. Subsequent to this fist step, the angular spectrum approach
could be used for the incremental plane-to-plane propagations.
5. I O Pulse Propagation
So far in this chapter, we have discussed the numerical implernentation of the angular
spectrum method for monofkequency sources. In considering pulsed sources, the source
waveform must be broken up into frequency components. Each frequency component
will have its own angular specmun, and may be propagattd saparatdy using a propagator
function appropriate for the distance and frequency in question. When choosing the sire
of the mesh over the source and observation planes, sampling at slightly better than the
Nyquist rate of the highest Erequency present in the spectrum should be used. This
corresponds to 7J2 sampling, where X is the wavelength of the peak fiequency
component. Pulsed propagation will be considered more fully in Chapter 7.
the propagator function. The choice of sampling intervals for the source determines
the extent of the spatial kequency domah.
If the extent of the spatial fkequency domain is completely contained within the
radiation circle, the FSC algorithm mey be used. The theorem in section 5.4 tells us
what samphg intervals are needed within the spatial fiequency domain to adequately
Chupter 5
65
sample H.These spatial frequency domain sampling intervals, in tum determine the
spatial extent to which the source should be zero padded.
If the extent of the spatial kequency domain completely contains the radiation circle,
spatial sampling of h is better. Because the radiation circle is contained within the
transfomi domain, adequate sampling of h in the spatial domain is guaranteed.
if the radiation circle is only partially contained within the transform domain, h must
be sampled at a rate slightly better than the source. the sarnpling rate being slightly
better than J2 (so as to increase the spatial frequency extent of h to completely
encompass the radiation circle). Upon taking the Fourier Transform of h, the resulting
angular spectrum rnay be muicated to match the spatial frequency extent of the
source.
The ray theory tmcation rnay be used to truncate the product H x S, and thus
decrease the size of the matrix on which to perfom the inverse ZD-DFT.This
however rnay be unprofitable if the ZD-FFTis used, which requires that the matrix be
of size 2" x 2mpoints (where n and m are integers). h e ray theory tmcation rnay not
be used to decrease the spatial sampling of s or h directly, as this rnay incur serious
aliasing errors in the convolution.
Frequency Domain sampling of transducer surfaces rnay prove to be more accurate in
near-field computations, and rnay offer computational savings.
With the above approach, accurate results for continuous wave propagation were
obtained, as verified by comparison with other methods.
This distortion gives rise to h m o n i c s not present in the initial spectnun, md thus in
modeling nonlinear propagation, each of these harmonic components must be accounted
for. The ger,ad ilpproach we propose to modeling nonlinear progression has been
outlined in chapter 4. This chapter will give greater attention to the details of o u
proposed scheme when the excitation signal is narrowband. In particula., we will
comment on our choice of operator splitting approaches, the choice of nonlinear
algorithm, and we will investigate modifications in the angular spectrum algorithm
needed to propagate harmonics generated in nonlinear propagation. Approaches used to
validate our algorithm will be explained, and results of our modelling will be compared
The general idea behuid the proposed modelling rnethodology is to solve for the effects
of diffiction, attenuation and nonlinearity separately over mal1 steps. As explained
earlier, this can be accomplished by using an operator splitting approach with a fiactional
step rnarching scheme. In selecting an operator splining technique, two second order
schemes were presented in Chapter 4 which are suitable for simulation of finite amplitude
propagation in tissue for CW sources. One of these schemes was originally presented by
Tavakkoli et. al. (1998),and consists of an incremental halfstep of difiction, followed
by a whole step of nonlinearity and attenuation, followed by an additional half step of
difnaction. The other approach was to fkst do a halfstep of combined diffraction and
67
attenuation, followed by a full step of nonlinearity, and finishing with a half step of
combined difhction and attenuation.
The choice of these operator splitting schemes may be based on whether it is desirable to
combine difiction and attenuation together, or whether combined nonlinearity and
attenuation is preferred.
of nonlinearity in the fkequency domain, since this approach would bypass the need to
transfomi the spectra back into the time domain. Because of this reason, and because of
the computational efficiency afforded by needing only a few hannonics, the FDSBE is an
attractive algonthm for CW nonlinear propagation. The FDSBE,allows the numencaI
evaluation of the combined effects of nonlinearity and attenuation. The inclusion of
attenuation in the nonlinear algorithm is attractive for the sake of stability of the
nonlinear substep, as it will tend to dampen higher harmonies. This may be particularly
crucial when large substeps, (made feasible due to the second order nature of the operator
splitting scheme) are chosen. If the FDSBE is used, the operator splitting scheme
presented by Tavakkoli et. al. (1998) would be the natural choice to use in the fractional
step marching scheme. Such a scheme has been implemented by us with moderate
success. It should however be noted that the other second order operator splitting scheme
for the nonlinear substep should be not only accurate, but also stable. Ofien times in
numencd modeling, instability is an undesirable product of nonlinearity. We consider
here validation of our nonlinear algorithm which will be implemented in the Eractiond
step NLP algorithm.
68
propagation in a dissipationless medium, and compared the harmonic trends with that of
the (analytic) Fubini solution, presented in Chapter 3. This cornparison (shown in Figs.
l(a) and ()) is, of corne, not valid beyond the shock distance. Fig. l(c) shows typical
:
....................................................................................................
..*...L................................................................ ........................................... !
I
0.1
0.2
O .3
0-4
propagation distance nomalized to shock distance
0-5
69
Figure 1 (a) The propagation curves of the first 5 harmonics (f= M H z ... 5MHz) using
the FDSBE in a dissipationless medium. (b) The Fubini solution. (c) Nonlinear plane
wave propagation in tissue for the first three harmonics, assuming an attenuaton
coefficient of 0.3dB/(cm MHz), (Le. a linear frequency dependence of the attenuation)
using the FDSBE.The source amplitude for a11 three plots was O. l d s , and the nonlinear
parameter p was 3.5.
Step size is an important factor in contributhg to the stability of the algorithm. If the
steps used in the FDSBE algorithm are too big, the algorithm incorrectly extrapolates the
nonlinear distortion, causing instability problems. Christopher and Parker (199 1) found
that limiting the step increments to a certain size resulted in stable cornputations. In
particular, their step sizes were no larger than the distance over which the highest
harmonie might be attenuated by a factor of 0.7. We will calI this the 0.7 rule. Mthough
70
it seems counter-intuitive to consider step sizes which limit attenuation, the issue is more
with regards to the nonlinearity. The FDSBE cannot handle large changes robustly. Our
work confimis the general utility of the 0.7 nile for the plane wave FDSBE algorithrn.
We did, however, notice that the d e did not always apply when diffraction present.
Some instabilities were observed for moderate amplitude wave propagation in water.
Fewer problems were encountered for tissue propagation.
Each nonlinear substep rnay be broken down into smaller substeps to allow the 0.7 rule to
be implemented. Larger step sizes, however, were still reiained for the diffraction
substeps, and thus some of the instabilities we noticed may have been for cases where the
large step size of difiaction perturbed the subsequent application of the nonlinear
algorithrn. Further investigation of stability for Our approach is recommended for Future
work.
For shock wave propagations, hundreds or thousands of harmonics may be needed,
greatly increasing the computation time of the FDSBE algorithrn. Specifically,
,
N is the number of harmonics
computational complexity on the order of N ~where
retahed. Christopher and Parker used a harmonic limiting scheme whereby artificially
high attenuation was applied to higher harmonics, thereby allowing shockfiont
propagations to be carried out reasonably accurately with 50-100 harmonics. Because we
are prirnarily concerned with nonlinear propagation in tissue, attenuation may work
naturally to our advantage in this regard, and so artificial ramping of the attenuation for
higher harmonics was not considered in this work. The scheme proposed by Pischai'nilov
et. al. (1996) may also be useful for shock wave propagations, but was not considered
here, as shock waves are not anticipated in tissue for amplitudes used in diagnostic
medical ultrasound.
71
be applied to nonlinear propagation within the context of the second order operator
splitting scheme. In particular, choosing a planar sampling grid may be a non-trivial task.
For each harmonic profile, it must M e r be decided whether to use the FSC or SSC
approach to calculate diffraction.
For simplicity, let us presume that each harmonic profile has an identical planar sampling
scheme. Moreover, although it is possible to use an adaptive gridding scheme, we found
it simpler to use the identical gridding scheme for each plane dong the direction of
propagation.
Intuitively, one might presume that the sampling intervals Ar and Ay which are needed to
represent planar distributions in propagation of harmonics would be at most kN/2,where
AN is the wavelength of the highest harmonic. Although this is the rate needed to
adequately sarnple the point spread function h for step sizes k > A . the harmonic profile
itself may be much smoother than this Nyquist rate requires.
In the work of Christopher and Parker (199 l), it was found that reasonable modeling
results could be obtained by sarnpling at rates far less than the Nyquist rate of the highest
harmonic. In fact, they observed that for focused field propagations involving up to 50
harmonics, a radial sampling rate of 4 times the Nyquist rate of the fundamental has
given very good results. For unfocused fields, 1 or 2 times the Nyquist rate of the
fundamental worked very well. Like them, we can contirm of the adequacy of the
sampling rate by examinhg the radial profiles of the highest harmonics. If the radial
sampling rate is insufficient, then these profiles will loose coherence.
Selection of a lateral sampling rate for the harmonic profiles of the planes, essentially
reduces to an educated guess. If increasing the sampling rate across the lateral profiles of
the h m o n i c s does not significantly produce different results, one may surmise that a
convergent solution has been reached and that sampling is adequate.
72
Once lateral sampling rates have been chosen, each harmonic profile must be propagated
an incrementd distance using either the FSC or the SSC algonthms. Based on the
analysis in the previous chapter, there is a reliable way of choosing whether to sample h
or to sample H. Ifthe radiation circle of the nth harmonic is completely contained within
the discrete transform domain, sampling of h should be done at the rate determined by the
gridding of the harmonic distribution itself. If the radiation circle is partially contained
within the transform domain, once again h should be sampled, however, sampling should
be done at the Nyquist rate of the nth harmonic (hd2) or better - even if it requires
sampling at a different rate that the source field. Sampling H rnay be done when the
radiation circle is exterior to the extent of the spatial kequency domain and when al1 the
oscillations of H contained within this domain rnay be adequately sampled by the spatial
fiequency sampling scheme of the source.
Suppose that the source plane harmonic profiles are sampled at a rate y tirnes the Nyquist
rate of the fundamental, where y is a real number, typically between 1 and 3. For
harmonics n > y, the discrete spatial frequency transform domain will be completely
inside the radiation circles k,'+k,'=k,,'.
possible to sample H at the sarne rate as the angular specum of the source. For
harmonics n W l and below, sampling H is not advisable since there rnay be convolutional
wraparound errors which rnay give erroneous results of the FSC algorithm. Exceptions to
this hypothesis rnay be in situations where either the propagation distance is small
compared to the harmonic wavelena&, or where the angular spectnun of the source plane
is very narrow. The minimum harmonic n~ for which sampling H is reliable can be
obtained from the theorem given in section 5.4. Equation (7) in this section, rnay be rewritten as
where k, = nk = 2nny7c is the magnitude of the wave vector for the nth harmonic.
73
where
4). Thus restricting the spatial frequency extent over the observation plane by a ray
theory truncation may lead to inaccuncies. For this reason, we chose not to implement
fl
Assume that sampling the hamionic source planes at y times the Nyquist rate will
adequately represent the angular spectra of the highest harmonic. This determines the
and,,k
as well as the spatial sampling intervals Ax and dy.For simplicity, we take equal x
and y coordinate s a m p h g Iengths.
74
Choose the maximal extent [X,,Y,] of the observation plane which one desires correct
results. Again, for simplicity, let X, = Y.. Moreover, consider for simplicity that the
sampling extent [X,YS]ofthe source plane is the same size as the region of interest in
the observation plane. Thus, X,= X, and Y, = Y,.
Choose the maximal extent of the spatial dornain transform space: T, 5 X,c X,,
and
T' 2 Y, + Y,.
extent to whicti the source plane must be zero padded. It also esrabiisiies the spatiai
fiequency domain gridding
*,
and Ak,.
Note that the radiation circle of the harmonics n, >y are contained within the spatial
fiequency transform domain.
Determine the desired step size Az between planes. Note that a constant step size
algorithm will be the most computationally efficient since the propagator functions
would only need to be cornputed cnce.
For harmonics n~ (given by Eq. (2)) and above, sample H in the spatial fiequency
domain with the sarnpling prescribed by dk,= f l - d
Ak,
= dc..
For each of the harmonics 11 = 1, ..., (nH-1),sample h. For harmoaics 1 ... n, the
sampling of h should be done with the intervals Ax and A4v used to sample the source
plane. For harmonics n, ... n r l , sample h with intervals determined by the Nyquist
rate of the nth harmonic (i,R)or better.
The product Sd x Hd c m be truncated in the spatial frequency domain to yk, which is
the spatial fiequency extent detennined by the sampling AqAy of the hamonic source
p lanes.
NLP algorithm is important because the high fiequency hannonics generated in nonlinear
propagation need to be propagated accurately. Such a test is difficult to devise because
there is no known gold standard to test whether a paaicular harmonic generated after a
nonllliear substep is being propagated properly in the subsequent difhctive substep.
75
Our approach to testing the diffraction algorithm is to quantitatively assess the accuracy
of linear propagation of several fiequencies from a large piston source - where hamonics
are not generated due to nonlinear propagation, but rather originate frorn the source
directly. The astute reader will be cautious of such a test. Having selected a sarnpling rate
y at between 1 and 4 times the Nyquist rate of the fundamental, it may be that the angular
spectxum for harmonics n q extend beyond the extent of the spatial fiequency domain -
thus incuning wraparound error. Indeed. even after difhctive propagation, where the
angular spectnim S is tapered by a transfer function H,the extent of the angular spectnim
rnay still extend out to the radiation circle of the nth harmonie, which is exterior ?O the
spatial fiequency dornain for n?. For large sources and focused sources, however, the
sound beam will be fairly directional, and the angular spectrum may be narrow - such
that the bulk of the energy will be contained within the extent of the spatial frequency
domain. This is the motivation for using a large disk.
Results of our multistep difiaction algorithm for a large disk radiating several harmonics
was compared with results of another diffkaction technique of hown accuracy, where the
fiequencies were propagated one by one, and not al1 together. Harmonic profiles of the
two techniques are compared in Fig. 2.
76
.......,. .....................;
$ o.lZ\---.--;
.....-,.--....
C
za 01 ...........
!t ,
-.-----r-.--8
LI
.....-. ........................
1
'
c 0,mc.........................;..............................
O
L.........,...............
02'-
.............. I.............
............
0mC............9...*
...*.+--.;
1
OQC
*'
=.-.
.........
4
-
Figure 2 The figures on the left are latenl profiles of a disk of radius 1.9 cm using our nonlinear
propagation algorithm, with P=O, and harmonics of 1,2, and 3 MHz being transrniaed fiom the source. The
figures on the ri@ are lateral profiles of fiequenciesf = 1, 2, and 3 MHz, respectively using a 'Gold
Standard' one-step angular spectrum method which bas been cross verified with the (slower)Rayleigh
integral method. This cornparison is to verify that the proposed sampling scheme for the diffraction
algorithm c m accurately propagate multiple frequency harmonics. The parameters for the resuls on the
right are: #planes = IO, pressure amplitude of each fiequency coniponent = 100 kPa, axial distance at which
the pronles were computed = a'/&, where A. is the wavelength of the 1MHz signal. Sampling was done at
1.1 times the Nyquist rate of the fundamental.
77
were used to avoid loops in the code, as MATLAB is an interpretive language, and
handles loops very slowly. However, array operations are a forte of MATLAB, and most
algorithms could be vectorized to take advantage of MATLAB's optirnized anay
processing capabilities. Sohvare exists to compile MATLAB code into C/C++, however,
this route was not taken since run times were reasonable enough in the MATLAE!
envuoment.
6.6 Verification of the NLP Algorithm
The source pressure was only 100 Wa. Baker (1988) obtained experimental and
theoretical fields which we use to compare with our results in Fig. 3 and 4. Also shown in
Fig. 5 is a cornparison with results reported by Christopherts (199 1).
Figure 4 Laterd profile at z = 275 mm ofthe first three harmonics produced fiom a 1.9
cm plane piston transducer, radiating at 2.25 MHz. The colored lines are the results from
our modei, the black solid Iine is experiment (Baker, 1988), and the black dashed iine is
results of KZK modeling (Baker, 1988).
79
Figure 5 Cornparison with the modeling results of Christopher (1991) for the same
paramters as given in Fig. 2.
Our simulations for Figs. 3-5 used only 20 propagation planes. In contrast, Baker used a
smaller step sue in his finite difference code corresponding to over 1O00 planes.
Christopher, too uses small steps (often 20 per cm)thus requiring hundreds or thousands
of increments. The second order operator splitting scheme of our appmach appean to
in Figs 3-5 that the lateral profile of the findamental drops off at around 35mm on either
side of the mainlobe. This is due to the windowing we applied to the propagation planes.
A M e r observation is that the 3" harmonic is slightiy lower than the expermental and
KZK results in Fig. 4, but in very good agreement with Christophers results s h o w in Fig.
5 . This rnay be due to the use of the plane impedance relation we used to relate normal
particle velocity with pressure. Both our method and Christopher's technique used normal
80
particle velocity in the NLP algorithm. The plane irnpedance relation is less accurate in
the nearfield, and may give rise to a source of error.
It is interesting that that in Figs. 4 and 5, the on axis fundamental level is actually below
the levels of the harmonics. Moreover, the levels of the harmonics are fairly high overall.
In Fig. 3, the second harmonic is less than lOdB down f'rom the fundamental, even out to
about 20 mm off auis. This is encouraging with regards to harmonic imaghg, since
extracthg the harmonic signals may be very challenging if the signal level is too low.
One might wony more about signal levels in tissue propagation, since tissue is typically a
strongly attenuating medium. Ln the next section we show some encouraging modeling
(1997) have simulated nonlinear propagation kom a focused disk in a liver-like medium
using a KZK modeling approach. We use our mode1 to compare with his results. For
these simulations, the source, operating at 2 MHz, had a radius of 1 cm and a focal length
of 10cm. In our modeling approach, we did not sample the aperature in the spatial
and 8, only 20 propagation planes were used with a lateral sampling of 2.1 times the
Nyquist rate of the fundamental. For Fig. 6, more propagation planes were used so as to
capture the axial variations of the harmonics. The deep nulls which are seen in Averkiou's
results, but not in o u n may be due to higher axial sampling on their part, and less axial
sampling on our part.
81
82
Figure 8 Lateral profile of the first three harmonics at 0.7 of the focal distance. (a) From
Averkiou et. al. (1997). @) Resuits from Our model. Parameters were the same as for
Fig. 6.
Note that in the axial profile (Fig. 6) the harmonic build-up due to nonlinearity occurs as
the beam propagates, and that the signal is not well formed in the extreme nearfield. This
can be advantageous for imaging, where surface inhomogeneities would othenvise
perturb the coherence of the Fundamental signal. Because harmonics are pnrnarily formed
after these surface regions, their utility for imaging deeper structures is more robust
compared with using the fundamental.
Note also that there is a prefocal shift of the 1st maximum of the fundamental curve.
This occurs because energy is being depleted fYom the fundamental tu feed the generation
of higher haxmonics. Such a shift may be accounted for in designing a tissue hannonic
imaging system.
In the lateral profiles, as for water propagation, the harmonic sidelobes fa11 off more
quickly than the fundamental sidelobes. This is a major factor in the attraction of using
83
harrnonics for imaging. Lower sidelobes may mean an hprovement in the signal to
clutter ratio. One reason that haxmonic sidelobes are lower may be that harmonic
generation is amplitude dependent. Where the fundamental level is highest (in the
mainfobe region) the hamonic level will also be hi&, but when the fundamental level is
lower (inthe sidelobe regions), the harmonic levels &op off rapidly. Propagation in
tissue may actudly work to our advantage in this regard. In the mainlobe region,
nonlinearity will dominate over attenuation if the signal is strong enough, however, in the
sidelobe region, attenuation may dominate over nonlinearity, and the harmonic sidelobes
will fa11 off very rapidly.
The lateral profile of the fundamental is still very similar to the profile one would see in
pure linear propagation. Note the classic Bessel directivity of the fundamental at the
focus. Sidelobe oscillations called f i n g n may be seen in the harmoinic profiles and are
characteristic of nonlinear propagation. They have been noted by a number of authon,
inciuding Averkiou ( 1997).
Finally it should be noted that the second h m o n i c level is still quite high (only 9dB
down fi-om the huidarnental at the focus), even for low amplitude excitation, and for
propagation in tissue. As noted, generation of sufficient harmonic levels is important for
realizing the feasibility of signal detection for harmonic imaging.
plane, We have used the spatial fiequency domain sampling approach outlined in section
5.7.2. The array considered had 64 elements of width N2, and spaced h/4 apart, where
the operating frequency was Z M H z . Elements were of height lm, and the array was
phased such that the focal distance was IOcm. The length of the transducer was 3.7 cm,
84
and the continuous wave excitation was of amplitude 347 kPa. Propagation was in a
liver-like medium with B=5 and a,= 0.3dBlcm at 1MHz. Shown in Fig. 10 are the
azimutha1 (.Y-z) and elevation (y-z) planar distributions. As expected, the harmonic
mainlobes are much nmower than the fundamental. Moreover, as rnay be seen in Fig. 9,
the h m o n i c sidelobes &op oDFextremely fast compared with the fundamental. In the
elevation plane, even when no lens is used, the harmonic profiles are narrower. In phased
seems to be slightly closer to the transducer than the 10 cm mark which we would expect.
radial ranQe(m)
Figure 9 Lateral profile of first three harmonies of a linear phased array.
Chapter 6.Mode
(a) fundamental
Wme Sources
85
(c}third harmonic
(d) fundamental
86
(f)third hmonic
Figure 10 (a)-@) Azimuthal plane and (d)-(e) elevation plane harmonic profiles from a 2
MHz linear phased array propagating in a liver-like medium. The array had 64 elements
of hight lm, and was 3.7cm in total length. The elements were Al2 across, and spaced
7J4 apart, where h = df The focal length was IOcm.The tissue was assumed to be
characterized by P=5, a,= 3 &/(cm MHz), and b=l. 1. No lens was used to focus the
transducer in the elevation plane.
lateral extent and sanipling of propagation planes used. Because Our mode1 may handle
non-axisymmetric sources, a 2D grid of sample points must be used to represent each
propagation plane. Increasing the number of samples from 5 12 x 5 12 to 1024 x1024, for
example, produces a large increase in computation t h e . For a given aperature size,
87
sampling.
The other major factor affecting the speed of the algorithm is the number of harmonics
included. Luckily, for tissue, we can attain reasonable results with only a few hannonics.
For water. there is less attenuation. and. in general, more harmonics are needed.
the initial spectrum, and thus both the nonlinear substeps and the difhctive substeps are
burdened by a computational load much greater than in the narrowband case. Efficient
modeling of nonlinear propagation of wideband signals requires a slightly different
approach than that used for CW excitation as outlined in the previous chapter. In this
chapter we intend to develop a methodology appropriate for pulsed excitation.
used. For pulsed sources hundreds or even thousands of fiequency components may be
needed to accurately sarnple the temporal variations of a wideband signal. Artificial
absorption of higher hannonics has been used in the continuous wave case to limit the
harmorks created due to nonlinearity. However, such as scheme when applied to pulsed
waveforms, may cause unnatural distortion. Most investigaton agree that using a thne
domain nonlinear algorithm is a much more attractive option. For difiaction, our
intention is to use the angular spectnim method, as this approach has shown great
computational savings over the Rayleigh method. The disadvantage, then, of using a time
domain nonlinear operator is that we will need to transform back and forth between tirne
and fiequency domains for each substep in the NLP algorithm. Not only will this process
take additional time, but such numencal book-keeping will likely incur some errors.
However, the computational savings to be won by using a time domain nonlinear
algorithm, may outweigh any potential t h e Iosses or inaccuracies due to the transform
algonthms.
Of the two time domain algonthms for nonluiear plane wave propagation presented in
Chapter 3, we selected the algorithm developed by Christopher, as it is very stable, is
cornputationally efficient, and is reasonably accurate. To ver@ the accuracy of our code,
Chapter 7
89
Fubini solution in the pre-shock region for a sinusoidal excitation. The two methods
compared well (Figs. l@) and '(c)), although discrepancies may be observed when the
step size Az is too large (Fig. 1 (a)). The TDSNE also captured relatively well the
dynamics of plane wave nonlinearity of pulsed sources (Fig. 2). Both step size and
sampling fiequency of the waveform play an importait role in the accuracy of the
nonlinear substep. When this algorithm is integrated into the fractional step NLP
dgoritha, multiple nonlinear substeps may be needed for each large diffractive substep.
Chapter 7
90
0.2
0.6
0.4
0.8
."
0.1
0.2
03
0.4
0.5
0.6
0.7
0.8
0.9
Figure 1 The harmonic profiles from the time domah nonlinear algorithm for plane wave
propagation with (a) 20 harmonics and 20 Az steps. @) with 20 harmonics and 100 &
steps compared with (c) the Fubini solution out to the shock distance. Note that there are
some discrepancies in (a) due to insufficient step size.
Cha~ter7
8
time (ps)
time (ps)
12
91
14
Chapter 7
92
frequeuncy (MHz)
.O6
.osI
i
.O4
1
i
.O3
1
1
j
1
i
1
1
l
I
frequency (MHz)
Figure 2 Plane wave pulse propagation. (a) Initial waveform @Distortedwaveform after
propagating 0.7 of the shock distance (c) Initial spectnim of the waveform, and (d)
spectnim of the distorted waveform conesponding to 0.7 of the shock distance.
Chapter 7
93
(1998) to implement the fiactional step marching scheme of the NLP algorithm. In large
part this choice stemmed fiom the hope that by combining absorption with the nonlinear
substep the FDSBE would be more stable. For pulsed sources, we have opted to evaluate
nonlinearity in the time domain, and there is no obvious way to include attenuation in the
TDSNE algorithm. One option would be to evaluate nonlinearity and attenuation step by
step over very mal1 intervals (smaller than the difeactive substeps which may be used in
the NLP aIgorithm). Another alternative is to calculate the attenuation with the difiactive
substep. The TDSNE is fairly stable, so the absence of attenuation in the nonlinear
substep will not greatly affect the stability of the NLP algorithm. Moreover integration of
attenuation in the diffraction algorithm is relatively easy. By modifying the propagator
fnctions h and H to include absorptive effects,
r
'i
and
attenuation may be accounted for in a very computationally efficient way. In light of our
choice to combine the effects of attenuation and difhction, the operator splitting scheme
we choose to implement is not that presented by Tavakkoli et. al., but rather the new
scheme proposed in section 4.2, and illustrated in Fig. 2 of chapter 4. In this scheme, the
combined effects of diffraction and attenuation are propagated over an incremental half
step, followed by a full step of nonlinearity, and finishing with a half step of attenuation
and difiaction.
Chapter 7
94
FSC or SSC algorithm for each discrete spectral frequency, selection of temporal
sampling and detemiining an adequate temporal extent to the signal. We will deal with
the later two issues first.
Nyquist rate, which is defined to be twice the rate of the highest frequency component in
the spectrum. If a pulse has center fiequencyh and bandwidth AL then the sampling rate
should be at leastfN,,,,=2(f&Af12).
For a finite amplitude pulse, hannonics will be generated during nonlinear propagation
which are not present in the initial spectrum. Because it is difficult to anticipate the
complex interaction between difhction, attenuation, and nonlinearity, selection of a
sampling rate becomes nontrivial.
For noniinear propagation in tissue, it is likely that attenuation will filter high fiequencies
generated due to nonlinearity. Based on our experience fiom CW propagation in tissue, it
is expected that no more than 5-10 harmonies of the center fiequency rnay be required for
the kinds of amplitudes allowed in diagnostic imaging. Thus, we should choose the
sampling fkequency to be 5-1 0 times the Nyquist rate fmUist=2&+Afl2) of the initial
spectrum.
scheme, a sufficiently long temporal window, or pulse repetition intervai (PRI) must be
used in the initial signal. To choose an adequate temporal window, b o t . linear and
nonlinear effects must be considered.
Chopter 7
95
Linear Analysis
Consider a piston transducer of radius a, and consider that we wish to send a pulse f?om
the transducer where the pulse has duration cr. Now consider that we observe the
waveform at a point (x,y) on a plane a distance z away from the source.
The time marking the beginning of the observed waveform will be given by the shortest
transit time betsveen the tmnsducer and the observation point. The time marking the end
of the observed waveform is given by the Iongest transit time between the transducer and
the point plus the length of the pulse.
Thus, the absolute longest duration the pulse could be is given by:
If (x,y) is not in the shadow of the transducer, the length of the pulse is given by
which is maximal for any given r when z=O, and in which case, once again we have:
Should we want to ensure that a pulse is adequately represented in the fkquency domain,
we would then need sampling given by the Nyquist theorem:
If an insufncient temporal extent is used for pulsed propagations, the waveform at the
observation point may be aliased. The closer the observation plane to the source, the
Chapter 7
96
greater the extent of a temporal window is needed to prevent aliasing errors. A plot of the
required temporal window venus axial distance fiom source is shown in Fig. 3.
Nonlinear Analysis
A peculiar phenomenon occurs in noniinear propagation, where difference frequencies
accumulate in a spectral region below the fundamental band of the signal. This
phenornemon has been coined "self demodulation" by Averkiou (1993) and others. In
highly absorptive media, attenuation may moreover filter out a good deal of the spectral
fiequencies above this demodulated band, and thus produce a signal which is of lower
fkequency than initially transmined. This self-demodulation phenomenon can sometimes
be accompanied by signal elongation. If the temporal extent of the pulse is shorter than
the resultant pulse Iengthening, wraparound error may occur. It is our experience,
however, that the temporal window required for the linear analysis of the previous section
Chapter 7
97
is more than adequate to accommodate the small degree of pulse lengthening and selfdemodulation which occurs in nonlinear propagation through tissue.
(thus dr
A,&
sources, it may well be that such dense spatial sampiing is not needed farther away frorn
the transducer, and a ray theory truncation rnay enable reduction of this sampling rate.
For finite amplitude propagation, harmonic profiles rnay Vary more spatially than the
fundamental profiles, and greater sampling rnay be needed. Borrowing insight from the
<
=f,+A\f/2. This
such that for frequency components/, up to nWl, the point spread function should be
sampled. For n a r r , sampling of the transfer nicntion H may be employed. For n <y, the
sampling of h should be done with the intervals Ax, Ay used to sample the source plane.
Chapter 7
98
For each fiequency componentJ, such that y<n<nirl, the point spread function h should
be sampled with at least the Nyquist rate off, (i.e. & = A&).
mcthod. Specificdly, a Icm disk transmining a pu!se with a peak amplinide 1cm/s, a
center fiequency of MHz, and 50% bandwidth (figure 4(a)) was simulated. Excellent
agreement was obtained, as s h o w in figures 4(b) and (c). The observation point for these
figures was on axis at a distance of a'/&, where h, is the wavelength of the center
-0.02
time (ps)
-0.02
time (ps)
Fig. 4 Cornparison of small signal difhction calculations (a) Initial pulse wavefom
calculated using (b) our angular spectnun approach, (c) the impulse response technique.
For a disk transducer at a ? ~ .
Chapter 7
100
dgorithm, venfication of the full nonlinear propagation (NP) algorithm was performed
with a few key tests:
the results of Averkiou et. al. (1997), with our own modeling technique for a focused disk
propagating 100 cycles of a 2 MHz sine wave in a liver-like media. We used 25
propagation planes, where each plane was divided up into 5 substeps for the nonlinear
algorithm. In this cornparison, shown in Fig. 5, there are some discrepancies. A cornpiete
explanation for these discrepancies is not understood, however, it is hypothesized that the
step size may be too large to account for the complex interaction between difiaction and
nonlinearity.
Fig. 5 . Cornparison of the new NLP algorithm for pulsed sources (a) in the quasi-CW
limit with the NLP algorithm (b) of the previous chapter. Shown are the lateral harmonic
profiles of a 2MHz fundamental signal at the focal distance (z=lOcm) of a concave disk.
Both results may be compared with Averkiou et. al. ( 1997).
The result of our algorithm, shown in figure 6 is in excellent agreement with the result
obtained by the impulse response method.
Chapter 7
-0.0%
4
time (ps)
102
Fig. 6. The observed waveform computed using our pulsed NLP algonthm in the limit of
vanishing P. This result may be compared with Fig. 4(c).
linear analysis. The amplitude asymmetry shown in Fig. 7 may suggest a more carefil
Chapter 7
103
analysis of standards, to account for nonlinear effects. Averkiou and Hamilton (1 997)
have shown that the negative pressure is ofien maximized before the focal region of the
transducer, and thus particular attention should be directed to monitoring possible adverse
cavitation phenornena in the prefocal region. Their simulations, however, were for water
propagation. It is hoped that future investigations using this mode1 for tissue may shed
additional light on the important area of standards.
time (p)
"O
4
5
Frequency (MHz)
Figure 7 (a) Waveform at the focus of a Gaussian apodized, focused disk transducer after
propagation through a tissue medium. The disk radius was 1.5 cm, the focal length was 6
cm, and the Gaussian amplitude shading of the source was such that the half-amplitude
radial distance was 0.84 cm. The pulse used was a sinusoid modulated by the square root
of a Gaussian function with 50% bandwidth, and amplitude 257 kPa. The tissue had
parameters P=5, ~ = dBIcm
3
at 1MHz2and &L3.
-
management was required. Several Gigabytes of fiee hard-disk space were used as a
swap-space to temporarily store data. Fortunately, even with the time consurning
processes used to manage the deluge of data, the code could be run within a few hours.
Qualification should be given, however, to clarify that increasing the nurnber of lateral
samples in the propagation planes above 256 x 256 or 5 12 x 5 12 points may incur a huge
Chapter 8: Conclusions
8.7 Summary
The primary contribution of this thesis has been the development of an efficient computer
mode1 for sirnulaihg nonlinear ultrasound propagation in tissue. This model shall enable
accurate modeling of finite amplitude ef5ects in modem clinical ultrasound scanners, and
in particular, may be useful as a tool for optimising the design of tissue harmonic imaging
systems.
Because current modeling techniques involve great computational burden, a pnmary
focus of this work was to develop an algonthm which is as computationally efficient as
possible. Towards this end, a second order operator splitting approach similar to that of
Tavakkoli et. al. (1998) has been devised, so as to enable the computation of difaction,
attenuation and nonlinearity separately over small steps. The second order nature of this
scheme allows the use of relatively large axial Ulcrements in a Eactional step marching
scheme. Our model has accurately reproduced published expenmental and theoretical
data with as few as 20 axial increments. This may be contrasted with other techniques
which typically require hundreds or thousands of axial steps. Our computer model has
been able to evaluate the nonlinear field response of continuous wave sources in as M e
as a few minutes, and wideband transient responses in a few houn. In cornparison,
numerical modeling of nonaxisymrnetric pulsed sources using current methods has taken
based difhction algonthm which they used in their nonlinear propagation model took
90% of the computation t h e , even for axisymrnenic sources. We investigated several
Chapter 8
Conclusions
106
The angular spectnim approach was found to be the most cornputationalIy efficient of the
difbction algo&hs
spectnun method is vdid for non-separable field distributions. Because of these reasons,
we chose the angular spectnun method to calculate the effects of difiaction in our
nonhear propagation algorithm.
Christopher and Parker (1991) have used the angular s p e c t m approach for acoustic field
propagations, however, the techniques they descnbe are for axisymrnetric sources.
Although Christopher has also simulated non-axisymmetric sources (e.g, Christopher,
1997), the step sizes he used were very srnaIl (ofien less than a wavelength), and thus his
approach does not well descnbe the large step approach we wish to take. In this thesis, a
new numencal scheme for implementing the angular specaim technique, applicable for
large axial steps, has been devised for sources which do not possess radial symmtery.
For continuous wave signals, attenuation and nonlinearity were combined together, and
the nonlinear substep was computed using the kequency domain solution to Burgen
equation. For CW sources propagating in tissue, only a few harmonic were nonnegligible, and so evaluation of the nonlinear substep in the Frequency domain was
relatively efficient. It was also seen that the attenuation of tissue afforded additional
stability to the FDSBE algorithm.
For wideband signals, a tirne domain algorithm was used to compute the nonlinear
substep, as the frequency domain approach has an undesirable computational load for
large numben of eequency components. Attenuation was combined with the diffraction
substep in this case. Pulsed simulations were a good deal more t h e consurning than the
continuous wave scenario, and often required extensive memory management. Several
gigabytes of Bee hard disk space were sometimes needed- even with 5 12 Mbytes of
system RAM.
Chapter 8
Conclusions
107
The continuous wave algorithm was verified with the linear algorithm in the case of
vanishing nonlinear parameter P. Moreover, our model compared well with published
results for nonlinear propagation in both water and tissue.
The algorithm for nonlinear pulsed propagtions was verified in the quasi-continuous
wave case, and in the linear limit.
8.2 Recommendatons and Future Work
The most obvious recornrnendation for extension of this research is the experirnental
verification of the pulsed nonlinear model. Once the model is verified experimentally,
various investigations cari be done using the model.
So far, only one way (transmit) propagation has been considered. A pulse echo imaging
system may very well be simulated by adding code for scattering of objects, and for
propagation of the echoes back to the transducer. Since tissue scattenng is typically fairly
weak, linear propagation is sufficient for simulation of the reception portion of the
algorithm. The impulse response method of Ullate and San Emeteno (1988) would be a
good candidate for simulating the received response.
Because Tissue Harmonic Imaging is a relatively new modality, there is still much to be
undentood about optimizing system parameten. Design of an imaging systern based on
nonlinearly generated harmonies should have some distinct differences as compared with
design of a typical pulse-echo B-mode imaging system. To this end, our model may be
Chapter 8
Conclusions
108
extracting the nonlinear portion of the signal, and rejecting the fundamental band. Use
of our model to better understand pulse inversion and its role in harrnonic imaging
would be a valuable for tissue harmonic irnaging system design.
The dynamic range of the A-D converter in an imaging system must be large enough
to not only detect the second harmonic, but to differentiate the second harmonic from
the noise floor. The small signal amplitude is a significant concem in tissue harmonic
imaging. This means that one typically will want to employ either large drive levels
fiom the source, or use aggressive focusing. Aggressive focusing, however must be
traded for depth of field. The investigation of drive levels, focusing and depth of field
using Our model would be very worthwhile.
0
Given a certain availab le bandwidth for the transducer, it must be decided in what
band the transmitted pulse may be sent at, and what band the second harmonic signal
should be received at. We must also defme what we mean by the second harmonic
signal. We rnay use as our second harmonic signal the absolute magnitude of the
pressure, or the second harmonic level relative to the fundamental. If we select the
absolute defuiition of second harmonic signal, we wish to transmit in a band such that
the fundamental p, is maximized AND the received signal pz is maximized. Suppose
that the amplitude of the paisband for the fundamental is LI,and that the amplitude of
the passband for the second harmonic is LI.Ifwe make a cmde approximation that
the amplitude of the fundamentai signal is approximately the square of the second
Chapter 8
Conclusions
109
Should we choose as our definition, however, the second harmonic level relative to
the fundamental, and we approximate this level as the fundamental squared, we want
to rnaximize L,' x ( L ~ I L ~ or
' ) ,in other words, maxirnize L2.This means receiving at
the center frequencyJ of the transducer and transmitting at K/, as shown in Fig. 2.
Ushg our mode1 to investigate which of these schemes is optima1 may enable a gain
of a few decibels of signal level, or some amount of improvement in system
performance.
cha~ter8
Conchisions
110
For hannonic imaging, one must use caution when trying to transmit very short
pulses. Short pulses will have a wide spectral envelope, which, may interfere with the
second harmonic signal in the receive band, complicating the filtering procedure, as is
shown in Fig. 3. Thus, sometirnes axial resolution must be sacrificed for contrast
resolution. The pulse inversion scheme mentioned above may help in this regard,
allowing wider bandwidth pulses to be sent without sacnficing contrast resolution.
Simulation may play a valuable role in understanding how to choose the transmit
waveform, and how pulse inversion may be used to overcome the problem of
harmonic band interference,
The signal to noise ratio (SNR) is of critical importance in harmonic imaging. For
linear B-mode imaging, several investigaton (O'Donnell, 1993; Welsh, 1998;
Misaridis, 2000) have explored the use of coded excitation and pulse compression to
improve the SNR of the received tissue echo. The premise of coded excitation is to
send a long coded waveform, and cornpress the received echo - often by a matched
filtering technique. By using a long pulse, one may increase the received energy
without increasing the signal level above standards. Compression of the signal thus
enables the signal energy to be temporally localised, while the noise level is
unaffected by the compression. Li (1999) has recently explored the use of coded
excitation in tissue h m o n i c imaging. One of the problems of the technique of Li is
the inability to completely reject the fundamental band using filtering. It may be
possible to use pulse inversion to reject the fundamental signal. Our mode1 could be
Chapter 8
Conclusions
lil
8.3 Conclusion
We have developed a computer mode1 of nonlinear ultrasound propagation in tissue,
which can simulate the acoustic field from non-axisymrnetric transducrrs. Our mode1
appears to offer some amount of compuiational savings compared with existing
techniques. The continuous wave response of disk transducers has been simulated and
compared with published results for nonlinear propagation in both water and tissue.
Transient excitation simulations have also been done, and have been verified in the quasiCW limi, and in the lin& where nonlinearity is negligible. The mode1 we have developed
References
Aanosen, S.I., Barkve, T., Naze Tjotta, J., Tjona, S., Distortion and harmonic generation
in the nearfleld of a finite amplitude sound beam, J. Acoust. Soc. Am., 75, 749-768, 1984.
Atkinson, K.E., An introduction to numerical analysis, Wiley, New York, 1978.
Averkiou, M.A., Lee, Y .S., Hamilton, M.F., Self-demodulation of amplitude- and
Sequcncy- nodulated pulses in 3 thermoviscous h i c ? , 1 dcoiar. Soc. Am., 94, 28762883, 1993.
Averkiou, M.A., Hamilton, M.F., Measurements of harmonic generation in a focused
finite-amplitude sound beam, J. Acom. Soc. Am., 98,3439-3442, 1 995.
Averkiou, M.A., Roundhill, D.N.,Powen, LE., A new imaging technique based on the
nonlinear properties of tissues, lEEE Ultrasortics Symposium, 1561- 1566, 1997
Baker, A C , Anastasiadis, K., Humphrey, V.F., The nonlinear pressure field of a plane
circular piston: Theory and expenrnent, J. Acoust. Soc. Am., 84, 1483- 1487, 1988.
Baker, A.C., Humphrey, V.F., 1992, Distortion and high frequency generation due to
nonlinear propagation of short ultrasonic pulses from a plane circular piston, J. Acoiisf.
Soc. Am., 92, 1699-1705, 1992.
Baker, A.C., Berg, A.M., Sahin, A., Naze Tjotta, J., The nonlinear pressure field of plane,
rectangular aperatures: Experimental and theoretical results, J. Acoust. Soc. Am., 97,
35 10-35 17, 1992.
Burns, P.N., Powes, LE.,Simpson, D.H., Uhlendorf, V., Fntzsh. T., Harmonic imaging:
Principles and preliminary results, Angiology, S63-S74, 1996
Cahill, MD., Baker, A.C., Increased O ff-axis energy deposition due to diffraction and
nonlinear propagation from rectangular sources, J. Acoust. Soc. Am., 102 , 199-203.
1997a.
Cahill, MD.,Baker, A.C., Numerical simulation of the acoustic field of a phased array
medical ultrasound scanner, J. Acoust. Soc. Am., 1O4 , 1274- 1283, 199%.
Cahill, M.D., Baker, A.C., Numerical simulation of the acoustic field of a phased-array
medical ultrasound scanner, J. Acousl. Soc. Am., 104, Pt. 1, 1274-1284, 1998.
Christopher, P.T., Parker, K.J., New approaches to the linear propagation of acoustic
fields, J. Acoust. Soc. Am., 90, 507-521, 1991.
Chnstopher, P.Te,Parker, K.J. New approaches to nonlinear diffiactive field propagation,
J. Acoust. Soc. Am., 90,488-499, 1991.
Christopher, P.T., A nonlinear plane-wave algorithm for difhctive propagation
involving shock waves, J. Comp. Acoustics, 1,371-393, 1993.
Christopher, T., Modeling the Dornier HM3 lithotripter, J. Acoust. Soc. Am., 95, Pt.1,
3088-3095, 1994.
Christopher, T., Finite amplitude distortion-based inhomogeneous pulse echo ultrasonic
imaging, IEEE Trans. Utrason Ferroelect. Freq. ConP., 44, 1 25- 139, 1997.
References
II4
Froysa, K.E., Naze Tjotta, J., and Bemtsen, J., Finite amplitude effects in sound bearns.
Pure tone and pulse excitation. In Advances in NonINiear Acousrics, H . Hobaek, ed.,
World Scientific, Singapore, pp.233-238, 1993.
Fubini, E., Anomalie nella propagazione di ande acustiche de grande arnpieua. Afro
Frequenza, 4,530-581, 1935.
Gaskill, J.D.,Linear Systems. Fourier Transforms, and Optics, Wiley, New York, 1978.
Goodman, J.A., Introduction to Fourier Optics, McGraw-Hill, New York, 1968.
Godunov, S.K., Difference method for the numencal calculation of discontinuous
solutions of hydrodynarnical equations, Mat. Sb.. 47,271-306 (in Russian), 1959.
Hamilton, M.F.,
Naze Tjoaa, I., Tjona, S., Nonlinear effects in the farfield of a directive
sound source, J. Acoust. Soc. Am., 78,202-2 16, 1985.
Hamilton, M.F., and Blackstock, D.T.(Eds.), Nonlinear Acoustics, Academic Press, New
York, 1998.
Hari, T.S., Hamilton, M.F., Nonlinear effects in focused sound beams,J. Acoust. Soc.
Am., 84, 1488-1496, 1988.
Haran, M.E., Cook, B.D., Distortion of finite amplitude ultrasound in lossy media, J.
Acoust. Soc.Am., 73,774-779, 1983.
References
il5
Hinkleman, LM., Liu, D.L., Metlay, L.A., Waag, R.C., Measurements of ultrasonic pulse
arrival tirne and energy level variations produced by propagation through breast wall, J.
Acumt. Soc. Am., 95,530-541, 1994.
Humphrey, V.F., Nonlinear propagation in ultrasonic fields: measurements, modelling
and harmonic imaging, UItrasonics 38,267-272,2000.
Kamakura, T., Hamada, N., Aoki, K., Kumamoto, Y .,Nonlinearly generated spectral
components in the nearfield of a directive sound source, J. Acoust. Soc. Am., 85,233 12337,1989.
Kamakura, T., Tani, M., Kumamoto, Y., Ueda, K., 1992, Harmonic generation in finite
amplitude sound beams fiom a rectangular aperature source, ,J. Acoust. Soc. Am., 91,
3144-3151, 1992.
Karaman, M., Atalar, A., Koyrnen, H., O'Domeii, M., A phase aberration correction
rnethod for ultrasound imaging,IEEE Trans. Ultrason. Ferroelect. Freq. Con.., 40, 275288, 1993.
Kino, G.S., Acoustic Waves:Devices, Imaging, and Analog Signal Processing,
Englewood Cliffs, NJ: Prentice-Hall, 1987.
Korpel, A., Frequency approach to noniinear dispersive waves, J. Acourr. Soc. Am., 67,
1954-1958, 1980.
Kumetsov, V.P.,Equations of nonlinear acoustics, Sov. Phys. - Acoust., 16,467-470,
1971.
Hamilton, M.F., Time domain modeling of pulsed finite-amplitude sound
Lee, Y.S.,
beams, J. Acoust. Soc. Am., 97,906-917, 1995.
Li, P.C.,Pulse compression for finite amplitude distortion based harmonic imaging using
coded waveforms, UZtrasonic Imagng, 21, 1- 16,2000.
Li, Y., Zagzebski, J.A., Cornputer rnodel for harmonic ultrasound imaging, IEEE Trans.
Ultrason. Ferroelect. Freq. Contr., 47, 1000- 1013,2000
Liu, D.L.,Waag, R.C., Time shift compensation of ultrasonic pulse focus degradation
using least mean-square estimates of mival tirne, J. Acoust. Soc. Am., 95, 542-555, 1994.
Liu, D.L.,Waag, R.C.,Correction of ultrasonic wavefiont distortion using
backpropagation and a reference waveforrn method for the-shifk compensation, J.
Acoust. Soc. Am., 96,649-660, 1994.
Referen ces
116
Macovski, A., Medicd Imaging System. Englewood Cliffs, NJ: Prentice-Hall, 1983.
Mallart, R., Fmk, M., Adaptive focusing in scattering media through sound speed
inhomogeneities: The van Cittert Zernicke approach and focusing cntenon, J. Acoust.
Soc. Am., 96,3721-3732,1994
McKendree, F.S.,A numerical solution to the second-order nonlinear acoutic wave
equation in one and three dimensions, Ph.D., dissertation, Pennsylvania State University,
1981.
Misaridis, T.X., Gammelmark, K., Jorgensen, C.H.,Lindberg, N., Thornsen, A.H.,
Pedersen, M.H., Jensen, J.A., Potential of coded excitation in medical ultrasound
irnaging, Ultrasonics, 38, 183-189,2000.
Naze Tjotta, J., Tjotta, S., V e h g , E.H.,Effects of focusing on the nonlinear interaction
between two collinear finite amplitude sound beams, J. Acoust. Soc. Am., 89, 10 17-1027,
1991.
Nock, L., Trahey, C.E.,Phase aberration correction in medical ultrasound using speckle
brightness as a qudity factor, J. A c o w Soc. Am., 85 1819-1833, 1989.
O'Donnell, M., Coded excitation system for improving the penetration of real-time
phased m y irnaging systems, IEEE T r a m Ultrason. Ferroelect. Freq. Contr., 39,34 1351,1992.
Pishchal'nikov, Y.A., Sapozhnikov, O.A., Khokhlova, V.A., A modification of the
spectral description of nonlinear acoustic waves with discontinuities,Acoustical Physics,
42,362-367, 2996.
Ratciiffe, J.A., Some aspects of difiction theory and their applications to the ionsphere,
Rep. Prog. Pb.,19, 188-263,1956.
Schafer, M .E.,Transducer characterization in inhomogeneous media using the angular
spectruxn method, Ph-D. Thesis, Drexel University, 1988.
Simpson, D.H, Chien T.C.,Burns, P.N., Pulse inversion doppler: a new method for
detecthg nonlinear echoes from rnicrobubbie contrast agents, IEEE Tram Utrason.
Ferroelect. Freq. Contr., 46,372-3 8 1, 1999.
Tranquart, F., Grenier, N., Eder, V., Pourcelot, L., Clinical use of ultrasound tissue
harmonic imaging, Ultrasound Med. Biol., 25,889-894, 1999.
Tnvett, D.H.,Van Buren, A.L., Propagation of plane, cylindrical, and spherical finite
amplitude waves, J. Acoust. Soc. Am., 69, 943-949, 1991.
Ullate, L.G.,San Emeterio, J.L., A new algorithm to calculate the transient near-field of
ultrasonic phased amys, IEEE Trans. Ultrason. Ferroelect. Freq. Contr., 39,745-753,
1992.
Waag, R.C.,Campbell, LA., Ridder, J., Mesdag, PX., Cross-sectional measurements and
extrapolations of ultrasonic fields, lEEE Trans. Sonics UZ~ason.,SU-32,26-35,1985.
Ward, B., Baker, AC., and Humphrey, V.F., Nonlinear propagation applied to the
improvement of resolution in diagnostic rnedical imaging, J. Acoust. Soc. Am., 101, 143254, 1997.
Welsh, L.R., Fox, M.D.,Practical Spread Spectnun Pulse Compression for Ultrasonic
Tissue Imaging, IEEEE Trans. Ultrason. Ferroelect. Freq. Contr., 45,349-355, 1998.
Williams, E.G., and Maynard, J.D..Holographie imaging without the wavelength
resolution limit, Phys. Rev. Lett., 45,544557, 1980.
Williams, E.G.,and Maynard, J.D, Numerical evaluation of the Rayleigh integral for
planar radiators using the FFT, J. Acoust Soc. Am., 72,2020-2030, 1982.
Wu, P., Stepinski, T., Extention of the angular spectmm approach to c w e d radiaton, J.
Acoust. Soc. Am., 105, 26 18-2627, 1999.
Zhu, Q., Steinberg, B.D., Wavefkont amplitude distortion and image sidelobe levels: Part
1- Theory and cornputer simulations, IEEE Trans. U~lnason.
Fenoelect. Freq. Conrr.,
40,747-753, 1993
Zhu,Q., Steinberg, B.D., Wave~ontamplitude distortion and image sidelobe levels: Part
II - In vivo experiments, IEEE T m . Ultrason. Femoelect. Freq. Contr., 40,747-753,
1993
Our starting point for the proof of the second order operator splitting, presented in
chapter 4 is the mode1 equation:
Operators
cd.&,
rNdr>
rDcAmk,
rD+A+N,dr
etc. can be defined in a similar way.
We want to show that
(3)
rD+A,&O<i2rN,krD+Av~~2v(x9~<z~9t)=rD+A+NJ-v(x,~>z~7t)Co(d-3)**
To do this we fmt rnust show that the operators L commute with the differential d l
a.
We will then be in a position to apply the Taylor theorem to show the relation (3).
Lemrna 1:
Proof:
Lemrna 2:
Amendix A
119
a, =
-=I~(z,s-rt)-v(z,rt)ds'=~,-v,
as --
(5)
Lemrna 3:
Although the nonlinear operator does not commute with d I & ,we c m define an operator
at plane Z + ~ / Z , given the initial condition v.=v(x.y,z,t) over plane z rnay be written as
(z)
Step 2:
rD+AIL.,2
(=)
We may propagate the result fiom step 1, presuming this result is located at plane r using
Now collecting tems with common factors &, A& etc, we obtain
v , ( z ) = v 0 ( z ) + ~{
i y o +
-
+A ,,i
~ , ~ ~ L: A }v O+ d,v lO 'D+A
{ O~+ ~ V ~ +
, ~0 V( k~ 3 )
Step 3:
We may propagate the effects of difhction and attenuation fiom the result fiom step 2,
presuming this result is located at plane z + A d 2 A solution to the evolution equation (7)
for diffraction and absorption given the initial condition defined in step 2 may be written
as
v3 ( 5 )
rDc.4&a-,2
( ~ v , J ~ + A ~) =
i 2~ D~+ A0 & ! ~ V Z
Appendix A
121
+ ~ S & , + ~ V +~ ,~V.I}
We are now in a position to compare the result from step 3 with the field profile obtained
Erorn the combined eflects of difk~ctior.~
nonlineaxity, uid tknuation. Thus we wish to
a distance k away fiom plane z, given the initial profile v(.r.y.z.t) accross plane z. We
have:
&
Comparing this expression with the expression for v3, we find that
Thus we have shown that this operator splitting scheme is second order.
Endofproof.
Corollary 1:A second order operator splitting for the KZK eauation
The retarded time integrai of the KZK equation may be expressed as an evolution
equation (described in equation 3-14 through 3-17) which satisfies the second order
operator splitting theorem proved above.
Proof:
The attenuation operator defined by (3-16),and the diffraction operator (3-14) clearly
commute with the operator a / & . The forrn of nonlinear operator may be redefined as
Appendix A
122
which satisfy an evolution equation (1) and comrnute with 8 /& are subject to the
second order operator splitting theorems of Tavakkoli et. al. (1998) and the scheme
provided above. Thus these second order operator splitting techniques may be extended
A solution to the evolution equation (7) for difiaction and attenuation, at plane z + k ,
rD+A.k~o
(z)
Ste~
2: We may propagate the result fiom step 1, presuming this result is located at plane
z using a nonlinear propagator. A solution to the nonlinear evolution equation (8) an
Amendix A
123
We are now in a position to compare with the field profile obtained fkom the combined
Comparing this with the expression for vz, we see that the two expressions are valid to
k t order, but not to second order. S
8.2 Notation
Consider a disk transducer of radius a, transmitting at a single frequencyJ Let us denote
the source field, and point spread functions as s(r) and h(r), respectively, where r is the
radial coordinate. Their analytic transfoms will be denoted by S(kJ and H(kJ
The discrete implementation of the angular speclnim method uses sampled versions of s
and h, which may be defined as sd[m]=s(r,), and hd[m]=h(rm),where r,= mAr is the mth
radial point, and Ar is the sampling interval. The distance between samples, Ar, can be
125
extent Tof the radial spatial domain, should be chosen such that T is greater than r. + a,
where ro is the extent to which correct results are desired. This choice of T prevents
aliasing in the region of interest O+cr,.
We shall denote the discrete Hankel transform (DHT) of
sd
hat ' N and a subscript '8will be used to denote sampling done in the spatial fiequency
domain. Frequency domain sampling of H, thus gives a sequence
fi
[il = H ( k i ) .
id= IDHT
{&}.
Note that d l these discrete sequences defined above can be thought of as implicitly
periodic.
hd. W e desire the discrete convolution of sequences sd and hd to approximate the analytic
convolution of the fictions s and h. The field profile represented by the analytic
convolution of s and h will of necessity be of infinite extent. Beyond a few source radii in
the lateral direction. the field may very well be nearly zero, however, it will not be
exactly zero. Thus, in using a discrete implementation of the angular spectrum method,
only a finite extent of the field in the observation plane may be calculated. Suppose that
across an observation plane a distance Az away from the disk transducer, we wish to
compute the field profile out to a lateral extent ro.The normal velocity field profile of the
source will have a finite extent of radius a. The analytic point spread function h, however,
will have an infinite extent, but must be truncated at some point. By ensuring that h is
sampled out to a lateral extent of T 2 r, + a , the h e a r discrete convolution sd*hd Will
accurately represent the desired field profile out to ro. In the region between ro and T,the
result of the linear convolution may not accurately reflect the tnie values for the field
profile due to tnincation of the infinite-extent point spread function h.
126
It is well known that the circular convolution can be used to implement the linear
convolution of two sequences. The point spread function h may be sampled out to extent
r., then zero padding may be used to pad sd and hd out to extent T 2 r,
sd
+ a . The DHT of
and hd are computed, then their transforms are multiplied together. The inverse DHT of
this product is the circular convolution of sd and hd. The circular convolution is
quivalent to the dimete lincar comolution of sd and hi t!~ough the whok extint of the
result (out to T), and represents the analytic convolution (the true field profile) in the
region of lateral extent out to r,. The use of zero padding prevents wraparound errors.
Alternatively, h may be sampled out to T 2 r, + a without zero padding (although sd
should be zero padded), and DHT-based circular convolution may be performed on the
resulting sequences. The result of the circular convolution will be equivalent to that of
linear discrete convolution out to r., while the region r, < r < Twill contain a region with
wraparound error. Thus, in either scheme the correct field profile is represented out io r,,
as long as both the source and the point spread fnctions are adequately sampled.
8, = IDHT { H , , ) .
n i e FSC Algorithm zero pads sd out to some spatial extent T 2 r, + a ,while sampling
the transfer f i c t i o n H across the entire frequency domain. Although the region r, < r <
Twill contain "wraparound errors", the resulting field given by (1) represents the correct
field profile out to r,, as long as both the source and the transfer h c t i o n s are adequately
127
sampled. The criterion of adequate sarnpling is, however, a nontrivial issue as will be
discussed in the following sections.
krlk
128
129
Figure 1: The real part of the transfer function H forf = IMHz and (a) z=lOh (b) z =
lOOh , and (c) z = U4.
When using the FSC algorithm for large propagation distances, adequate sampling of H
off more slowly, providing a less abrupt change. Othenvise, sampling of the transfer
function H may be a precarious matter.
Although the FSC algorithm affords the computational shortcut of not having to
transform the point spread fnction into the spatial fkequency domain, it involves perils
associated w ith diasing problems.
It was noted earlier that the FSC approach is equivalent to discrete convolution of the
aperatue sd with
the
correspo~dto
130
6,
convolution. Such aliasing problems are more severe for larger propagation distances and
higher fiequencies, as illustrated in Figs. 2(b) and (d).
131
132
2(a) and (c)), can be done at moderate sampling rates, while the trmsform Hd has
virtually no wraparound enor. I& corresponds to the inhnite extent H wrapping around
itself to produce a finite sequence Hd. Ifsamphg is done at greater than the Nyquist rate
of 2fc, the evanescent tail of Hd will be well represented, and because the evanescent tail
exponentially falls off to zero, there will be very little wraparound error in Hd.
A p p e n d &B
~ The Anmlar Spectrum Ariproach for Radiallv Svmmetric Sources
133
used to rnodify their FSC algorithm. It is based on the premise that at distances greater
than h, the field of a point source c m be well descnbed by a ray theory (which neglects
any evanescent waves). Consider, as depicted in Fig. 3, a plane a distance z away from a
point source.
According to a ray theory interpretation, the radial spatial fiequency, k , at point P in the
observation plane, is given, by
(3)
where 8, is the angIe between the observation point and the point source, as depicted in
Fig. 3.
We cm use this ray theory relation with Huygens principle, by applying it to a collection
of point sources representing a fuiite source. For a finite extent source of radial extent a.
the angfe subtended by the edge of the rransducer and the point P may be labeled O.,
as
134
section
I g n o ~ evanescent
g
components, we may Say that km, = k sin(&,,) is the maximum
spatial fkequency in the angular spectrum of the field across plane P. We can modify
H(z,kr)in the spatial frequency domain by doing a tmcation, so that H is represented
= ksin(O,,,,).
also the cusp at k, = 27$/c, and the dense oscillations in the neighbourhood of this cusp.
Simul~~osy,
smpting requirments fer b t h thel?SCan SSC algmithms wouibbe
reduced.
135
If (r,+a)h = i l f i ,
8, = d3, and the
spatial fiequency
may be m c a t e d ar
r
v 3 / 2 ofk=2nj7c.
Figure 5: B y using a ray theory, H may be truncated at a point where sampling of the
oscillations of H is manageable.
spectral component. The transfer function H, however, bas no finite Nyquist fiequency
limit, and thus infinitely dense sampling would be required to accurately represent H.
Fortunately, the Ray Theory introduced in the previous section allows an elegant and
means we can define a h i t e Nyquist rate for the tmcated propagator function H.
To M e r show how the ray theory truncation surpresses aliasing problems associated
with the FSC algorithm, we prove here a theorem, which we will use as a basis for
sampling in more complicated scenarios in subsequent sections and chapters. Assuming
136
that we have no prior knowledge of the Ray Theory Truncation, suppose that we wish to
truncate H at a particular spatial frequency, which we shall cal1 km,. We will show by
means of a Taylor series expansion that the optimal choice for km, so as to suppress
aliasing errors, is the km, given by the Ray Theory Truncation.
-j & d k * - ( k ;
+k;)
we have
w here
& d m +
r
jp
p
and
k- -,k
P=d-
k' -,;k
Since,,k
13 7
p=L7tan6=r,+a
This is just another way of saying that if we require correct results (without wraparound
errors) out to r,, we can use the ray theory truncation with the FSC approach.
B.1 iConclusions Regarding the Radially Symmetric FSC and SSC Algorithms
Christopher and Parker concluded that for plane-to-plane difiactive propagation, the
SSC algorithm was to be preferred over the FSC algorithm when no ray theory truncation
was used. They also concluded that modifying the FSC to include a Ray theory tmcation
(section B.7) can greatly improve the accuracy of the FSC method, while simultaneously
reducing the computational burden.