Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
I. INTRODUCTION
Manuscript received March 3, 2002; revised September 23, 2002. The work
of O. Michailovich was supported by the Israeli Planning and Budgeting Committee (VATAT) of the Council for Higher Education. The Associate Editor
responsible for coordinating the review of this paper and recommending its publication was M. W. Vannier. Asterisk indicates corresponding author.
*O. Michailovich is with the Department of Bio-Medical Engineering,
Technion-Israel Institute of Technology, 32000 Haifa, Israel (e-mail: bmoleg@
tx.technion.ac.il).
D. Adam is with the Department of Bio-Medical Engineering, TechnionIsrael Institute of Technology, 32000 Haifa, Israel.
Digital Object Identifier 10.1109/TMI.2002.806570
1491
after the separation. The advantage of such a harmonic-separation technique is self-evident: it could enhance the quality
of harmonic imaging, without the usual tradeoff between
harmonic detectability and imaging resolution.
The main contribution of the current paper is the proposal
of a nonlinear harmonic-separation method, which fulfills the
above requirements, and which can be efficiently implemented
by means of convex optimization. The approach is taken from
the field of sparse signal representations [11], which was successfully applied to the problem of blind source separation [12].
The superior ability of the sparse signal representation to separate components of signals, made up of a superposition of a few
disparate phenomena, was reported [13]. The results obtained
here confirm the advantages of the sparse decompositions over
the traditional linear approaches.
Since the harmonic separation procedure is iterative, the
complexity of the problem should be maximally reduced so as
to minimize the computational load. Such a reduction can be
achieved by means of the demodulation technique that is an
inherent part of almost all current commercial imaging systems.
In order to handle more efficiently the harmonic signals, a
slight modification of the standard demodulation scheme is
proposed here, which is explicitly explained in Section IV
below. It will be shown that the demodulated signals can be
efficiently represented using frames, derived from the familiar
Gabor frame by an appropriate modulation and bandwidth
adjustment (see Section V below).
The paper is organized as follows. Section II introduces theoretical preliminaries, which are necessary to form a basis for
the following derivations. Different models of the signals under
consideration and appropriate demodulation strategies are introduced in Sections III and IV, respectively, where also a proof
is provided that these procedures are optimal from the viewpoint of lossless reduction of the data size. Section V provides
a brief introduction to Gabor frames, with extension of its definition also to the discrete case. It also describes modifications
of standard Gabor frames, which are necessary for efficient representing the harmonic signals at hand. The algorithm structure
is summarized in Section VI. Section VII contains examples of
(2.1)
and
be two subsets of , namely
,
Now, let
, and
, where denotes the space
of square summable sequences. Then, the harmonics separation
problem can be formulated as the following optimization task,
searching for:
(2.2)
and index sets
In words, one needs to find the sequence
, , such that the linear combination of the functions
,
,
with the coefficients, respectively given by
,
, will be as close as posthe subsequences
, correspondingly.
sible to the signal components ,
When the optimal solution is found, the desired signal components are estimated as
(2.3)
The in (2.3) denotes an estimate of the appropriate value.
Unfortunately, the minimization problem (2.2) cannot be
solved directly, until we are given an oracle, providing us with
and . Hence, it is useful to find another
the functions
(possibly sub-optimal) optimization procedure, which will
possess the following three important properties.
1) It must be solvable (that implies obtaining the solution in
finite time, without requiring either infeasible additional
information or conditions).
1492
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
(2.4)
(3.1)
Here
denotes the central frequency of the transmitted ultraare the description pasound pulse, and
and
are peaked
rameters. The two-sided spectra of
and
, correspondin the vicinity of the frequencies
ingly. While the spectrum of the fundamental harmonic can be
, namely
assumed to be enclosed within a band of width
Support
(3.2)
that of the second harmonic might have its support, enclosed in
an interval of rather different width.
Generally, there are two cases, which depend strongly on the
transmitted waveform and on a specific kind of the imposed
nonlinearity. In the first case, the bandwidth of the second harmonic component is approximately equal to that of the fundamental harmonic, formally
Support
(3.3a)
Such a situation frequently takes place, when the nonlinear
propagation of the ultrasound wave is of concern [18]. In the
second case, the spectrum of the second harmonics, when
received from disrupting contrast agents, can have even a
doubled width, as compared with the fundamental band [19].
In this case
Support
(3.3b)
Although the case (3.3a) is included in (3.3b), both these cases
will be treated separately throughout the paper, because they
lead to two distinct algorithms, which have considerably different complexity.
It is important to note, that basically the ultrasound pulse is
produced by a resonance system (the transducer piezoelectric
element), where its motion can be described by a second-order
differential equation. It implies an ultrasound pulse of an oscillatory nature and, hence, its fundamental spectrum should
. As a result,
converge to zero in the neighborhood of
. Regarding the overall signal
one can conclude, that
, one concludes that Support
for
for the case
the case (3.3a) and Support
(3.3b). It implies, that these signals are to be digitized at the samand
, respectively. Because the
pling rates
sampling rate is proportional to the bandwidth of the bandpass
signal, it is often convenient to represent it in terms of a low-pass
signal, whenever the latter can be sampled at a lower sampling
rate. It is performed through a cancellation of a high-frequency
component of the bandpass signals, known as demodulation. Its
modified versions, specially adapted to the harmonic separation
problem, are introduced in the next section.
IV. AUGMENTED DEMODULATION
Since the received RF sequences are real, their Fourier
transforms exhibit conjugate symmetry. It implies a possibility
to describe these signals in the Fourier domain only over the
. Consequently, the
positive semi-axis, i.e., for
signals
and
, when treated separately, can be completely characterized by the complex signals
and
, respectively, conditioned that
is known
in advance [17, Ch. 7]. It was found to be very useful to
execute the harmonics separation procedure in the domain
of low-pass signals. The reason for this is twofold. First, it
leads to efficient localization of the signal spectrum, and,
consequently, to optimal reduction of the problem size. Second,
this representation facilitates construction of a frame (as shown
in the section below), which is useful in the context of the
harmonics-separation problem.
in the low-pass domain, but
In order to express the signal
in an equivalent form, a technique similar to the familiar demodulation can be used. It differs from the standard procedure by the
requirement that the information, characterizing both harmonic
components, must be saved in the low-pass domain without loss
or aliasing. Slightly different demodulation schemes are used for
the two cases, when the bandwidth of the second harmonic is similar to that of the fundamental component, and when it is broader.
A. Equal Harmonic Bandwidths
In this case the demodulation is performed by the following
is multiplied by a complex expotwo simple steps. First,
and
, i.e.,
nent, whose frequency is the average of
. Second, the resultant (complex) signal is subjected
to a low-pass filter, whose passband is defined by the interval
. According to the assumption (3.3a), the signal,
so obtained, is given (accurate up to the error introduced by the
filtering) by
(4.1)
From the context of the problem it is obvious, that the complex
and
represent activities of the first and second
signals
harmonic components, correspondingly. Moreover all the original information is preserved by these signals, up to the error,
induced by the filtering. It should be noted that the frequency
is smaller by a factor of two, in
band, spanned by the signal
. The components of
have their
comparison to that of
frequency support included within the intervals, given by
Support
Support
(4.2)
1493
(4.3)
to belong
It is possible now to consider the sequence
, which is a space of -seto the signal subspace
quences, whose Fourier series are supported within
. Analogously, one can consider the sequence
as being in the subspace
, which is comprised of the -sequences, having its Fourier series supported
.
within
B. Nonequal Harmonic Bandwidths
From purely theoretical point of view, the bandwidth of
the second harmonic component depends on the waveform
of the transmitted pulse and specific kind of the nonlinearity,
imposed by the insonified medium. To be as general as possible
under such circumstances, one can assume that the bandwidth
of the second harmonic component can be at most twice the
fundamental bandwidth. Consequently, in order to preserve
this component maximally undistorted in the low-pass domain,
and the
the frequency shift (demodulation) is set to be
subsequent low-pass filter has its passband defined by the
. In this case the resultant complex signal
interval
can be represented as
(4.4)
As in the previous case the resultant bandwidth is half the origand
are peaked around
inal, and the complex signals
and 0, respectively, and have their frequency bands supported within
Support
Support
(4.5)
(4.6)
Analogously to the previous case, one can define the signal suband
to which the sequences
spaces
and
are supposed to belong. The next
section addresses the construction of a frame that will be repreand .
sentative for the signal components included in
1In order to simplify the notation below, the same regular case characters will
denote both the sampled signals and the corresponding analog signals.
1494
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
Fig. 2. Demodulation for the case of equal bandwidths of the first and second harmonics. (A) Power spectra of the signals f (t) and f (t). The spectrum of
f (t) has lower amplitude for convenience of visualization. Note that this sketch is merely schematic, intended to clarify the spectra supports. (B) Power spectra
of p (t) and p (t) (after demodulation). Note the symmetry around the origin.
1495
Fig. 3. Demodulation for the case of nonequal bandwidths of the first and second harmonics. (A) Power spectra of the signals f (t) and f (t). The spectrum
of f (t) has lower amplitude for convenience of visualization. (B) Power spectra of p (t) and p (t) (after demodulation).
Due
to
periodicity
of
the
Fourier
series
, there are only
significant shifts,
in (5.2). Now it is possible
implying
to define the following collection of vectors, obtained by
applying a composition of the shift and modulation operators
:
to
(5.3)
It was explicitly investigated in [22], [23] that the collection,
with good timegiven by (5.3), can constitute a frame in
frequency localization properties, on condition that
(5.4)
must span all of
in order to be
The vector set
.
a frame in this space, implying that
.
This results in the requirement
matrix, whose columns are formed by
Let be a -by, as given by (5.3). Now,
,
all vectors
one has [24]
(5.5)
is known as a frame opwhere the symmetric matrix
erator, and and are minimal and maximal eigenvalues of ,
respectively. Consequently it can be concluded that every matrix with column rank uniquely represents a frame, where
1496
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
Fig. 4. Frame design for the case of equal bandwidths of the first and second harmonics. (A) Supports of the Fourier series of the first and second harmonic
[n] and g
[n], as given by (5.10).
components of the signal p[n], as given by (4.3). (B) Supports of the Fourier series of the vectors g
Support
in its exact
The requirement for compact support of
sense may be too limiting, resulting in noncompactness of
(which is highly undesired). Thus, the rethe sequence
quirement is hereby relaxed, with the phrase supported in the
interval interpreted as having most of its energy concentrated
in the interval.
an operator of modulation by , one can
Denoting by
as follows:
obtain another vector set in
(5.8)
, this set also constitutes a frame
Due to the unitarity of
, and, moreover, it is characterized by the same frame
for
bounds. The new frame can also be considered as a union of
three subsets given by
where
(5.9)
It can be seen that for the specific choice
(5.11)
where
the window
Support
(5.12)
Support
Support
(5.10)
(5.13)
by
Support
Denoting
Support
Support
Support
it is obvious that
, and, moreover,
Support
Support
(5.14)
1497
Fig. 5. Frame design for the case of nonequal bandwidths of the first and second harmonics. (A) Supports of the Fourier series of the first and second harmonic
components of the signal p[n], as given by (4.6). (B) Supports of the Fourier series of the vectors g [n], k = 1; 2; 3 and g [nj ], as given by (5.13). Note
that the parameter controls the bandwidths of g^ (!), and it should be set adaptively to fit optimally the support of p^ (! ).
current study, because it is simple and accurate. After the discrete window is defined, all the frame vectors are computed and
stored as the columns of a -bymatrix , which will be
referred to as a frame matrix.
VI. SOLUTION BY CONVEX OPTIMIZATION
where
(5.15)
differs
Comparing (5.11) with (5.15) one can see that the set
from only by a subset, comprised of nonmodulated vectors
. Moreover the supports of their Fourier series
(which are continuously controlled by ) are larger or equal to,
, but never exceeding
as compared with those of
Support
. The advantage of the frame, as given by
(5.15), is illustrated in Fig. 5. On one hand, the coincidence of
and
is obvious, implying that
the supports of
the first harmonic component can be efficiently represented
. On the other hand, appropriately
in
choosing the parameter , one can achieve a coincidence
and
, implying that the
of the supports of
second harmonic component can be efficiently represented
. An optimal value of
might be
in
obtained by a calibration procedure, which may be needed for
each given case.
D. Choice of a Window
The last question to be addressed is the choice of the discrete
,
. One option is the envelope of
window
the ultrasound pulse, or its symmetric counterpart, which can
be more conveniently implemented. An important result of this
choice of the window is that the time-support of the pulse
envelope is equal to that of the pulse itself. It means that the
frame functions, derived from it, are ideally suited for the analysis of transient phenomena in the received RF lines, because
their time scale is similar to that of the pulse duration. The pulse
envelope is usually unknown, but can be estimated in a nonparametric manner as shown, for example, in [25]. A different, rather
simpler way to obtain the pulse envelope might be to measure
it directly before the scanning. This approach was used in the
A more concise notation of vectors in the appropriate Euclidean spaces, instead of sequences with indices, will be used
,
, and
be the acbelow. Let
quired RF sequence, the frame matrix [constructed according to
either (5.9) or (5.15)] and a vector of the representation coeffiimplies, that
cients, respectively. being a frame of
(6.1)
Due to the linear dependence of the vectors that comprise the
frame, the definition of the coefficients in (6.1) is not unique.
In the context of the problem of harmonic separation, the representation of by as few frame vectors as possible must be
found. Moreover, in most practical cases, the acquired signals
are contaminated by measurement noises. Since only analysis
of a noise-free part of the signals is of interest, it is tempting
to find a procedure, which can obtain the sparse representation
of the original (i.e., noise-free) signals, ignoring the noise. This
task can be performed by solving the following minimization
problem [13]:
(6.2)
The first term of the functional (6.2) penalizes the misadjustment of the model (6.1), while the second term forces the
solution to be as sparse as possible. Note that the tradeoff
parameter controls the influence of the sparseness penalty.
More about possible choices of , and its connection to the
Lagrange multiplier of the equivalent constrained problem
,
, where
is an
estimated standard deviation of the noise, can be found in
[26]. Note that the noise can be either measured before the
scanning or estimated from the data (see, for example [27]).
However, in this study a different approach to the selection
of was undertaken. It was shown in [11] and [13], that the
solution of (6.2) is closely related to the problem of de-noising
1498
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
1499
Fig. 6. Examples of nonlinear pulses, used in the computer study for simulating the response of a nonlinear medium, e.g., contrast agent. (A) and (C) Pulses
in the time domain. (B) and (D) Spectra of the pulses (normalized to have peak values of 0 dB). Note that the pulses possess different degrees of overlap between
the first and the second harmonic bands.
Fig. 7. Estimates of the second harmonic component of the simulated RF lines, synthesized using the example of a nonlinear pulse, shown in Fig. 6(A) and
(B). (A1) Estimate obtained by MHF. (A2) Estimate obtained by MOF. (A3) Estimate obtained by sparse representation (BP). (B1)(B3) Appropriate estimates in
the time-frequency plane.
1500
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
Fig. 8. Estimates of the second harmonic component of the simulated RF lines, synthesized using the example of a nonlinear pulse, shown in Fig. 6 (C) and
(D). (A1) Estimate obtained by MHF. (A2) Estimate obtained by MOF. (A3) Estimate obtained by sparse representation (BP). (B1)(B3) Appropriate estimates in
the time-frequency plane.
1501
the images was not optimized and, therefore, their quality may
be potentially improved.
Two other pairs of images were produced so as to compare
the MHF and MOF algorithms. In the MHF method two
matching filters were employed, where the first one was
adjusted to transfer the first harmonic frequencies, while the
second one transfers the second harmonic component. Note
that the matching harmonic filtering allows keeping the axial
resolution as high as possible. It is of importance, since the new
approach presented here has the advantage of retaining the axial
resolution, and, thus, the performance of all techniques should
be compared at the same resolution. The images obtained using
the MHF and MOF algorithms were subjected to the standard
I/Q demodulation, and visualized in 8-bit resolution, so that
all the images presented in this paper have the same dynamic
range.
The first harmonic images, obtained by all the algorithms
tested here, are very similar, thus, only one is shown, that is
obtained using MHF. This image is shown in Fig. 10(A), where
the effect of the contrast agent is demonstrated by an increase
of brightness inside the tubes. Fig. 10(B) shows a second-harmonic image, obtained by the MHF algorithm, while Fig. 10 (C)
shows this image, as obtained by the MOF. Note that in both
these cases the vessels are significantly masked by the portion
of the fundamental energy that leaks into the second harmonic
band. The second harmonic image, obtained by BP (sparse solution), is viewed in Fig. 10(D), where the only visible object is the
contrast agent, flowing in the two vessels. There is hardly any
masking by background reflections or strong reflections from
the tube-tissue interfaces.
IX. DISCUSSION AND CONCLUSION
In this paper a novel approach is presented to the problem
of separation of the harmonic components within an acquired
RF image. The proposed method differs from the conventional
approaches, which employ linear filtration. Linear filtration,
when applied to ultrasound harmonic processing, has several
shortcomings, among which is the inverse relation between
the transmitted bandwidth and the ability to detect the higher
harmonics. In such a case, the filter designed to extract the
second harmonics, allows significant energy related to the first
harmonics, to penetrate into the resultant signal. Since the
first harmonic component is typically stronger in comparison
with the second, this leakage tends to mask the second
harmonic component. In contrast imaging, where the higher
harmonics induced by the contrast agent are measured, it is
generally accepted that the second harmonics are masked by
native tissue harmonics produced by the nonlinear propagation
of the acoustic waves. However, when using linear filtration
during contrast agent imaging, the effect of the leakage of
the fundamental component into the frequency band of the
second harmonic component is much more destructive. It
considerably masks the information related to the nonlinearity
of the interrogated tissue or contrast agent. Consequently, in
order to obtain true harmonic images, this leakage must be
suppressed. It was reported [1], [31] that the higher harmonics
induced by contrast agents are often larger in amplitude than
1502
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 21, NO. 12, DECEMBER 2002
Fig. 10. (A) First harmonic image of the phantom, as obtained by MHF. (B) Second harmonic image of the phantom, as obtained by MHF. (C) Second harmonic
image of the phantom, as obtained by MOF. (D) Second harmonic image of the phantom, as obtained by sparse decomposition (BP).
ACKNOWLEDGMENT
The authors would like to thank Prof. A. Ron from the University of Wisconsin-Madison for a helpful consultation. They
would like also to thank all the anonymous reviewers, whose
useful comments significantly improved the paper.
REFERENCES
[1] X. Verbeek, J. Willigers, P. Brands, L. Ledoux, and A. Hoeks, Measurement of the contrast agent intrinsic and native harmonic response with
single transducer pulse waved ultrasound system, Ann. Biomed. Eng.,
vol. 27, pp. 670681, 1999.
[2] P. Frinking, A. Bouakaz, J. Kirkhorn, F. Ten Gate, and N. de Jong, Ultrasound contrast imaging: Current and new potential methods, Ultrasound Med. Biol., vol. 26, no. 6, pp. 965975, 2000.
[3] D. H. Simpson, C. T. Chin, and P. Burns, Pulse inversion Doppler: A
new method for detecting nonlinear echoes from microbubble contrast
agents, IEEE Trans. Ultrason., Ferroelect., Freq. Contr., vol. 46, pp.
372382, Mar. 1999.
[4] J. Kirkhorn, P. Frinking, N. de Jong, and H. Torp, Three-stage approach
to ultrasound contrast detection, IEEE Trans. Ultrason., Ferroelect.,
Freq. Contr., vol. 48, pp. 10131022, July 2001.
[5] K. E. Morgan, J. S. Allen, P. A. Dayton, J. E. Chomas, A. L. Klibanov,
and K. W. Ferrara, Experimental and theoretical evaluation of microbubble behavior: Effect of transmitted phase and bubble size, IEEE
Trans. Ultrason., Ferroelect., Freq. Contr., vol. 47, pp. 14941509,
Nov. 2000.
[6] P. Frinking, E. I. Cespedes, J. Kirkhorn, and H. Torp, A new ultrasound contrast imaging approach based on the combination of multiple
imaging pulses and a separate release burst, IEEE Trans. Ultrason., Ferroelect., Freq. Contr., vol. 48, pp. 643651, May 2001.
[7] Y. Li and J. A. Zagzebski, Computer model for harmonic ultrasound
imaging, IEEE Trans. Ultrason., Ferroelect., Freq. Contr., vol. 47, pp.
12591272, Sept. 2000.
[8] N. de Jong, P. Frinking, A. Bouakaz, and F. Ten Gate, Detection procedures of ultrasound contrast agents, Ultrasonics, vol. 38, pp. 8792,
2000.
1503