Ultrasound in Med. & Biol., Vol. 27, No. 9, pp.

1231–1238, 2001
Copyright © 2001 World Federation for Ultrasound in Medicine & Biology
Printed in the USA. All rights reserved
0301-5629/01/$–see front matter

PII: S0301-5629(01)00423-9

● Original Contribution

PERFORMANCE OF ULTRASOUND ECHO DECOMPOSITION USING

SINGULAR SPECTRUM ANALYSIS

WAGNER COELHO DE ALBUQUERQUE PEREIRA* and CARLOS DIAS MACIEL†

*Biomedical Engineering Program, COPPE/UFRJ, Rio de Janeiro, RJ, Brazil; and †Electrical Engineering
Department, State University of Londrina, Londrina, PR, Brazil

(Received 5 February 2001; in final form 11 June 2001)

Abstract—Diagnostic ultrasonography has its well-established role in medicine. Nevertheless, the quanti-
tative characterisation of biological tissues by ultrasound (US) is still a main topic of research. Several
parameters have been explored with this purpose, (e.g. attenuation, backscatter coefficient, US speed). More
recently, mean scatterer space (MSS) has been proposed as a characterisation parameter. The objective of
this work was to investigate the potential of the singular spectrum analysis (SSA) to estimate MSS. This
method proposes the reconstruction of the periodic part of the original US signal from where the MSS of the
medium can be estimated. SSA is applied to simulated and real backscattered echoes from a phantom and
a bovine liver sample. Consistent results were obtained from both Monte-Carlo simulation and real data.
They were compared with literature. Presently, precision, accuracy and sensibility of SSA are being
investigated. (E-mail: wagner@peb.ufrj.br) © 2001 World Federation for Ultrasound in Medicine &
Biology.

Key Words: Singular spectrum analysis, Ultrasound, Tissue characterisation, Mean scatterer space.

INTRODUCTION gation medium to quantify tissues. Physical parameters

The conventional ultrasound (US) grey-scale images are
widely used in the medical diagnostic field. This kind of
examination has been growing at important annual rates
(Kadah et al. 1996). These images are made based on the
principle of acoustic impedance mismatch between dif-
ferent biological tissues illuminated by an ultrasound
beam, and only a small part of the information contained
in the acoustical backscattered signal is used for this
representation. This approach limits the identification of
diseases to a stage where the ultrasonographer can visu-
ally detect anatomic differences. During the last decades,
different approaches have been under development to
produce new kinds of images from US signals, with
special attention to the issue of quantitative US charac-
terisation (Bridal et al. 1998). More recently, statistical
signal processing is being used to provide parameters
from US echoes, opening a new branch of studies.
The objective of quantitative US is to extract pa-
rameters related to the physical properties of the propa-
Address correspondence to: Dr. Wagner Coelho de Albuquerque
Pereira, Biomedical Engineering Program-COPPE/UFRJ, PO Box
68510, Zip 21945-970 Rio de Janeiro-RJ, Brazil. E-mail: wagner@
peb.ufrj.br
can be obtained from transmitted and/or scattered US
signals by modelling the interaction of the acoustic wave
within the propagation medium. It is expected that these
quantities may be applied to discriminate between
healthy and diseased tissues.
In particular, in the field of biological tissue iden-
tification, statistical signal processing tools have been
explored, using basically the spectral analysis approach
from radiofrequency (RF) and envelope signals to quan-
tify parameters such as attenuation, propagation speed,
mean particle diameters and mean scatterer spacing
(MSS).
The MSS is a parameter related to spatial organisa-
tion of tissues and has been recognised as an effective
tool for characterisation and discrimination of tissues
that present a semiregular lattice of scatterers (Felling-
ham and Sommer 1984). The liver, for instance, has a
well-organised microstructure and, with different dis-
eases, its spatial characteristics can be changed. Much
work has been done to estimate MSS, using techniques
based on autocorrelation, power spectrum and cepstrum
(Fellingham and Sommer 1984; Landini and Verrazzani
1990; Wear et al. 1993; Narayanan et al. 1997).
Varghese and Donohue (1993, 1994, 1995) intro-

1231
1232 Ultrasound in Medicine and Biology
duced a characterisation of the scatterer distribution of
tissue microstructures with a technique called spectral
autocorrelation (SAC). One of their results is the estima-
tion of MSS from 2-D maps. The main difference be-
tween the classical power spectrum and the spectral
autocorrelation is that the SAC function includes the
phase differences between different spectral components.
Simon et al. (1997) proposed a computationally
efficient algorithm for MSS estimation. They used an
analysis based on the spectral redundancy present in
backscattered echo signals and generated spectral lines
through a quadratic transformation of the RF echo signal.
The main advantage of this approach is the possibility of
real-time implementation and superior results when com-
pared with amplitude-only algorithms. When compared
with amplitude-phase techniques, their results are noted
as comparable.
The basic assumption that supports the concept of
MSS estimation is that the periodic and nonperiodic
structures can be separated in their corresponding US
backscattered echoes. This separation can be repre-
sented, for example, by the discrete (“fence-like” well-
defined frequencies) and continuous part of the signal
spectrum, respectively.
From this point of view, singular spectrum analysis
(SSA) seems to be a powerful tool for time-series anal-
ysis, different from the common approaches and not yet
well explored for biological tissue characterisation. It has
the capability to decompose periodic and aperiodic struc-
tures from the time signal, even in a noisy ambient.
The main objective of this work was to investigate
the potential of SSA to estimate MSS when applied to
simulated and real backscattered echoes from phantoms
and bovine liver. Results from a Monte-Carlo simulation,
from an experimental phantom, as well as from a bovine
liver sample, are presented and discussed.
THEORY
SSA is a tool commonly used in nonlinear dynamics
studies (Broomhead and King 1986; Vautard and Ghil
1989; Fowler and Kember 1998) and in several signal-
processing areas (Therrien 1992). It is based on the
classical principal component analysis (PCA), where a
time-series data vector x with length N-points can be
expanded with respect to a complete orthonormal basis
composed of M vectors (Ek, 1 ⱕ k ⱕ M).
The vectors Ek are called eigenvectors and are ob-
tained from the diagonalisation of the correlation matrix
(Cx) from the original N-data points. This matrix is
constructed by moving a window of M points (M ⬍ N)
along the sequence x. This window is shifted by one
point at a time. The position of the window is indicated
by the index “i”, so, “i” varies from 1 to N ⫺ M ⫹ 1. For
Volume 27, Number 9, 2001
each position of the window, the M data points inside it
can be described by a linear combination of the vectors
Ek, according to the following equation:
冘a E , 1 ⱕ j ⱕ M,
M
x i⫹j ⫽ k
j
k
(1)
k⫽1
where the coefficients akj are called the principal com-
ponents (PCs).
Each data window can be understood as one single
point in a M-dimensional phase space. The singular
vectors (Ek) then represent the principal axes (bases) of
this phase space. Each singular vector (eigenvector)
points toward the directions with most energy (following
an optimisation criterion); thus, the projection of the
original points onto them is the most efficient for that
specific data set.
Two important results from the eigenstructure of
principal component analysis can be summarised:
1. Eigenvectors of the correlation matrix Cx pertaining
to a zero-mean random vector x define unit orthog-
onal vectors representing the principal directions
along which the statistical variance has its extreme
values; and
2. Each associated eigenvalue defines the extreme
value of the variance in that direction. The total
variance of the time series is divided among the M
different basis.
PCA and SSA have the same mathematical descrip-
tion. PCA has been applied to data compression, chang-
ing the original data set to a new one based on the highest
variance information described by the principal axes (the
variance is associated with the eigenvalues.) SSA uses
information from eigenvalues/eigenvectors to describe
particularities present in time series. In our approach, we
looked for the basic periodic behaviour from the data,
and it may not necessarily be described by the highest
variance components.
Vautard and Ghil (1989) showed that the presence
of eigenvalues with values close enough to be identified
as a “pair” could be associated with a periodic compo-
nent of the signal. The frequency of this periodic com-
ponent can be found by taking the average of the peak
value of the amplitude spectrum of the two associated
eigenvectors. After we can identify the eigenvalue pairs
present in the data, it is possible to reconstruct and
characterise separately the periodic and nonperiodic
structures of the data just by selecting the correct eigen-
vectors. In the same way, this decomposition can opti-
mally capture certain parts of signal behaviour, selecting
a correct subset of bases from all bases (Vautard et al.
1992).
Some drawbacks with this technique have to be
 Singular spectrum analysis on US simulated signals ● W. C. A. PEREIRA and C. D. MACIEL 1233

considered. Both periodic signals and noise can produce

eigenvalue pairs and, if the pairs are miselected, all the
following reconstruction will be corrupted and a correct
MSS estimate will not be possible. Therefore, the effi-
cient choice of eigenvectors/eigevalues is crucial to the
method.
A schematic illustration of this situation is in Fig. 1,
where there are two circles and, in both, it is possible to
observe two main eigenvalue data sets that represent
signal and noise subspaces. The first one, on top (Fig.
1a), is ideally what we would like to have: the periodic
behaviour present only in the signal subspace. Actually,
the periodic behaviour is also present in the noise sub-
space (Fig. 1b), and so there are eigenvalue pairs related
to them. So, an eigenvalue pair may be related to the
periodic activity in signal subspace, as well as the ape-
riodic activity in the noise subspace. In the analysis here
proposed, the eigenvalue pair identification is made spe-
cifically to find the bases of the periodic signals, leaving
apart all other bases that are related to aperiodic or noise
subspace signals.
More specifically, in this work, we intended to
decompose backscattered envelope US signals in two
parts: 1. periodic signal and 2. aperiodic signal and noise.
The original signal is projected on the Ek basis. Then,
individual eigenvalue pair analysis is made. Mean scat-
terer space may be estimated from the association of the
periodic behaviour, indicated by eigenvalue pairs, and
the power (energy) of the corresponding eigenvectors.
Each individual eigenvector spectrum of interest
concentrates power (Vautard and Ghil 1989; Vautard et
al. 1992), where there are sharp peaks (“fence-like”
well-defined frequencies) in the original spectrum. By
selecting the correct eigenvectors/eigenvalues, it is pos-
sible to “filter” the data window sequence and to find its
periodic component. This kind of filtering process is
cited in literature as being a nonlinear filtering (Fowler
and Kember 1998). It has a good performance, because it
is not necessary to set previously the exact frequencies,
as is needed in the regular filtering techniques.
The main difficulty of analysing periodic behaviour Fig. 1. Schematic representation of an eigenvalue data set for a
correlation matrix estimated from a time-series sequence. This
using SSA is to choose the “correct” eigenpairs. Some
drawing represents the eigenvalues’ separation in two main
authors (e.g., Allen and Smith 1996) pointed out that parts: noise and signal subspaces. (a) The hypothetical case
modification of M (the size of correlation matrix) within where the periodic components can be described only by the
a range of values may confirm a stable eigenpair. As an signal subspace (b) the real situation where eigenpairs may be
example of this situation, in Fig. 2 are represented the found in signal as well as in noise subspaces.
eigenvalues from the correlation matrix generated by the
following simulated signal:
value of M is changing from 10 to 75 and, in the y-axis,
there is the representation of the corresponding eigenval-
x n ⫽ sin ␻ 1t n ⫹ sin ␻ 2t n ⫹ r n (2)
ues (represented by ⫹). At each M value, there is exactly
the same number of eigenvalues in the vertical lines.
where ␻1 ⫽ 11 rad/s, ␻2 ⫽ 1 rad/s and rn is white noise. Arrows 1 show the two main eigenvalue pairs related to
The data window has 150 samples. In the x-axis, the frequencies ␻1 and ␻2. Arrow 2 shows a spurious result;
1234 Ultrasound in Medicine and Biology
Fig. 2. Eigenvalues for the example with two sinusoidal signals
as a function of the size of the covariance matrix (10 ⬍ M ⬍
75). For each dimension M in the x-axis, there is the same
quantity of eigenvalues in the y-axis. It can be observed that
there are two stable pairs (corresponding to the two sinusoidal
frequencies of the signal) clearly away from the cluster at the
bottom. For dimension M around 60, a spurious pair starts to be
detached from the cluster.
i.e., a third pair that belongs to the noise space is being
detached from its group.
The periodic behaviour suggested from the two first
eigenpairs as M grows may not be a sufficient condition
to get all the periodic information from the signal. Ther-
rien (1992) mentions the work of Wax and Kailath,
akaike information criterion (AIC) and minimum de-
scription length (MDL) to separate the eigenvalues of
signal and noise subspaces. These criteria were applied
in systems where noise has white properties. In those
cases, the plot of the singular spectrum (eigenvalues
organised in descending order and in logarithmic scale)
shows an evident eigenvalue “plateau” that can be used
as a criterion to separate noise from signal subspace.
When applied to backscattered US signals, it produces a
poor result because the noise backscattered from the
medium has coloured properties, due to the transducer
frequency bandwidth. In these cases, there is no plateau
present in the singular spectrum plot (Palus and Dvorák
1992; Allen and Robertson 1996).
If the original sequence is dominated by oscillation,
it is reflected directly on the variance and then its singu-
lar spectrum will have eigenpairs with corresponding
eigenvectors delayed from each other by a factor of ␲/2,
and these eigenpairs will tend to have higher values than
the others.
As mentioned before, a nonoscillatory process may
also generate eigenpairs, so, in this work, an heuristic
criterion was developed to identify eigenpairs. It is di-
vided into two steps: 1. choose the first or the two first
pairs if they represent at least 65% of the accumulated
Volume 27, Number 9, 2001
Fig. 3. Singular spectrum from the envelope signal of Fig. 5.
(original US signal from the parallel-nylon-line phantom im-
mersed in water). Arrows point to the two chosen eigenvalue
pairs. The correlation matrix dimension (M) was 100. Contin-
uous line represents the accumulated variance of the data.
variance in that frequency, and 2. the frequency associ-
ated to their eigenvectors are spaced at no more than
2.5% of each other. A reason for this criterion is based on
the fact that, if a periodic behaviour is present, it will
tend to project their energy only in two bases (forming a
pair), but aperiodic signals will have the energy distrib-
uted among many bases.
The implementation of this criterion is made as
follows. The first step is to plot the singular spectrum and
to identify the eigenvalue pairs of higher amplitudes (as
in Fig. 3, explained in the Results section). The second
step is to plot 2-D graphics with the power spectrum of
each eigenvector previously multiplied by its eigenvalue,
following the same sequence of the singular spectrum (as
in Fig. 4, explained in the Results section). The more is
the variance, the more important is the corresponding
eigenvector. Nevertheless, there is still a possibility that
this procedure could be excluding some important (if it
exists) information with low variance.
MATERIALS AND METHODS
All the software implementation was made in
MATLAB (MathWorks Inc., USA). Results will be
shown of SSA applied to three different kinds of US
backscattered signals. The first one consists of a real
signal obtained from a phantom containing a set of
parallel nylon lines (spaced 3.0 mm from each other)
immersed in water. The inspection pulse was generated
by a 10-mm diameter pz36 (Ferroperm, Denmark) ce-
ramic with low quality factor Q and centre frequency 5.9
MHz (pulse duration 0.5 s). The received signal was
sampled in a oscilloscope (TDS 420-Tektronix, USA)
 Singular spectrum analysis on US simulated signals ● W. C. A. PEREIRA and C. D. MACIEL
Fig. 4. Power spectra of the 16 first eigenvectors, previously
multiplied by each corresponding eigenvalue following the
proposed criterion. They were obtained from the SSA applied
to the US signal from the parallel-nylon-line phantom (Fig. 5).
In x-axis is the eigenvector number in the y-axis is the fre-
quency and, in grey scale, is the amplitude from each eigen-
vector. Notice that the same two first eigenvalue pairs (arrows)
of Fig. 3 have the strongest peak and have very close frequency
values. That, according to the adopted criterion, is an indication
of the presence of periodicity.
with sampling frequency of 25 MHz. Figure 5a shows
the RF echo backscattered from the phantom. The two
vertical dashed lines show the signal window used for
the analysis. Figure 5b shows a zoom of this window
and, Fig. 5c, its envelope signal. Time and space are
related by the equation d ⫽ 2ct, where d is the distance
travelled by the US wave during time t with average
Fig. 5. Ultrasound RF backscattered signal obtained from the
parallel-nylon-line phantom immersed in water. (a) Vertical
lines around 120 ␮s and 140 ␮s show the signal window to be
analysed by SSA; (b) zoom of the window; (c) envelope signal
of the window.
1235
speed c. Multiplication by 2 accounts for the fact that the
wave propagates twice inside the medium in pulse-echo
mode.
The second kind of data are simulated backscttered
echoes. A Monte-Carlo analysis was applied to these
signals to see how SSA behaves in different scenarios of
jitter and SNR. The simulation model adopted is a 1-D
RF echo from tissue, using a convolution of tissue-
scattering amplitude and a RF pulse inspection, as de-
scribed in Simon et al. (1997). The simulation parame-
ters are: US pulse centre frequency of 3.5 MHz, MSS of
1.25 mm, signal-to-noise ratio (SNR) in the range 12 to
⫺5 dB (12 dB, 5 dB, 0 dB, ⫺3 dB and ⫺5 dB) and jitter
of the regular structure of 1%, 5%, 10%, 20% and 30%.
The concept of a jitter was introduced to take into ac-
count possible variations of the target positions along the
US beam axis. It is expressed as a percentage of the
nominal target separation. SNR (in dB) is estimated from
the amplitude relation between the echoes from the reg-
ular structures to the echoes from diffuse ones. It is
important to note that both signals represent information
from biological tissue. For each combination of jitter and
SNR, 500 signals were simulated, which makes a total of
12,500 signals.
The third kind of data are US backscattered signals
from a bovine liver sample. The regular backscattered
signal from liver is the result of a regular distribution of
its lobular structure (Pereira et al. 2000). This distribu-
tion has fluctuations in its periodicity. In addition to that,
there are many randomly distributed particles that act as
lower amplitude US scatterers. In the SSA analysis, 500
backscattered signals from bovine liver collected by
Pereira et al. (2000) were used. Their basic experimental
setup is composed of two transducers. The pulse from the
transmitter transducer has a centre frequency of 3.3 MHz
and duration of 3.5 ␮s. This transmitter was made with a
low mechanical factor PZT ceramic from Ferroperm威
whose diameter is 12.5 mm. The reception transducer is
a PZT ceramic hydrophone, of 0.8 mm diameter and
centre frequency of 15 MHz. The sample frequency was
50 MHz. For the MSS estimation, the collected signals
were decimated by a factor of two.
To do the SSA analysis on the bovine liver US
signals, the correlation matrix (CD) of dimension 150 ⫻
150 was calculated as a Toeplitz matrix from the auto-
correlation function. To make the SSA signal reconstruc-
tion, only the first 4 eigenvalues were considered, be-
cause this subset has more than 90% of all variance
present in the signals.
RESULTS
Figure 3 shows the singular spectrum with the eig-
envalues (circles) in a descending order, for the signal
1236 Ultrasound in Medicine and Biology Volume 27, Number 9, 2001

Fig. 7. Monte-Carlo analysis for the US–simulated backscat-

Fig. 6. Final result of the signal analysis from the parallel-
nylon-line phantom. (a) Reconstructed periodic signal using the
four eigenvectors corresponding to the two chosen pairs of
eigenvalues. (b) Power spectrum of the reconstructed periodic
signal, from which the position of the highest peak is taken as
the estimation of the MSS. (c) Aperiodic signal, obtained from
the subtraction of the reconstructed signal from the original
one. (d) Power spectrum of the aperiodic signal.
tered signals. Results obtained for the same SNR (around 0 dB)
and four different jitters. Each item of the figure has a cluster-
ing plot (top) and the respective histogram (bottom). The clus-
tering is composed of the MSS estimates (y-axis axis is the
SNR level and x-axis the MSS frequency). For the histograms,
the x-axis is the same MSS and the y-axis the number of events
(estimates). The size of the window (M) is 256 samples. Sim-
ulated MSS of 1.25 mm (corresponding to a MSS frequency of
616 kHz). (a) 5% of jitter; (b) 10% of jitter; (c) 20% of jitter;
and (d) 30% of jitter.

presented in Fig. 5 (parallel nylon-line phantom). It is

observed that it has two eigenvalue pairs that may repro- (which is equal to c/(2*MSS), with the parameter c as the
duce a quasiperiodic behaviour (arrows). The continuous US propagation speed) and the y-axis is the number of
curve in this figure represents the total accumulated events (estimates). In Fig. 7a and b, it is possible to
variance from the progressive summation of the eigen- observe an artefact of the analysis, which is that there are
values. From a total of 100 eigenvalues, the first 20 have some estimates locked in the harmonics of the expected
near 90% of all variance that is distributed among all response.
bases by the process. In Fig. 4 is presented the frequency
response of the respective eigenvectors in the same order
of the eigenvalues of Fig. 3. In the x-axis is the eigen-
vector number; the y-axis is the frequency and, in grey-
scale, is the linear amplitude response from each eigen-
vector. Notice that the same two first eigenvalue pairs
(arrows) have the strongest picks and their frequency
values are very close to each other (for each pair).
According to the criteria adopted here, it is an indication
of the presence of periodicity.
In Fig. 6a, the reconstructed signal from the paral-
lel-nylon-line phantom can be observed, using these two
pairs of eigenvectors and, in Fig. 6b, its power spectrum.
The aperiodic signal (Fig. 6c) is obtained by subtracting
the periodic signal from the original one. In Fig. 6d is the
power spectrum of the aperiodic signal.
The result for the Monte-Carlo analysis applied to
the US-simulated signals can be seen in Fig. 7, where, for Fig. 8. SSA performance curves as a function of the jitter,
each individual item (a, b, c, and d), there are two obtained from the Monte-Carlo analysis for the US-simulated
graphics. The upper one is the plotting of the clustering backscattered signals at different SNR levels (12 dB, 5 dB, 0
dB, ⫺3 dB and ⫺5 dB). Window data length with 256 samples.
made by the estimates (y-axis is the SNR level and x-axis Vertical axis is the percentage of correct MSS estimates. An
is the MSS frequency) and below is the corresponding assumed correct estimate is within the range of 5% of the exact
histogram, where the x-axis is the same MSS frequency value.
 Singular spectrum analysis on US simulated signals ● W. C. A. PEREIRA and C. D. MACIEL
Fig. 9. SSA performance curves as a function of jitter, obtained
from Monte-Carlo analysis for the US–simulated backscattered
signals at different SNR levels (12 dB, 5 dB, 0 dB, ⫺3 dB and
⫺5 dB). Window data length with 256 samples. Vertical axis is
the percentage of correct MSS estimates. An assumed correct
estimate is within the range of 25% of the exact value.
The figure of merit adopted for this Monte-Carlo
analysis is depicted in Figs. 8 and 9 and is the same cited
in Simon et al. (1997). It was chosen as such to ease the
comparison of the results. There, one can find the per-
formance curves for five different levels of SNR, as a
function of jitter. In Fig. 8, a correct estimate is defined
as the one that falls within the range of 5% of the correct
value. In Fig. 9, this range is set to 25% of the correct
value. The interpretation of them is in the next section.
The US bovine liver samples were processed with
the SSA algorithm proposed, and the final result is pre-
sented on a histogram in Fig. 10. Each histogram class
has a width of 20 kHz. This figure shows a maximum of
the histogram in 0.20 MHz (4.0 mm) and a range be-
tween 0.19 MHz (4.8 mm) and 0.22 MHz (3.1 mm). The
MSS value, as well as its range, were estimated from the
relation MSS ⫽ c/(2*MSS_Frequency), where c is the
US speed in the medium which, for this experiment, was
estimated as 1630 m/s at 20°C (Pereira et al. 2000). The
parameter MSS_frequency is obtained as the correspond-
ing frequency of the amplitude peak of the FFT (fast
Fourier transform) performed on the SSA reconstructed
US signal.
DISCUSSION
For the results with the phantom signals, it is im-
portant to notice that the peak in Fig. 6b is in 0.254 MHz,
which corresponds to a periodicity of 2.95 mm (the real
spacing is 3 mm). This error seems to be acceptable,
considering the US speed in water was not estimated but
1237
Fig. 10. Histogram of the MSS estimate for the 500 bovine liver
backscattered signals collected by Pereira et al. (2000). The
main peak is near 0.20 MHz (which corresponds to a MSS of
4.0 mm) and ranging from 0.17 MHz to 0.22 MHz (correspond-
ing to a MSS range of 4.8 mm to 3.1 mm).
assumed c ⫽ 1500 m/s at 26°C. The second peak is in the
first harmonic of the previous peak (0.508 MHz).
Figure 7 shows different levels of jitter (a ⫽ 5%;
b ⫽ 10%; c ⫽ 20%; d ⫽ 30%) to the same SNR (near 0
dB). Two results are important in these graphics. The
first one is that some estimates are grouped around the
first harmonic of the simulated original frequency. The
second result is that, as jitter becomes higher, the esti-
mates start to spread around the original clusters. Each
item of the figure has a clustering plot (top) and the
respective histogram (bottom). The clustering is com-
posed of the MSS estimates (y-axis is the SNR level and
x-axis is the MSS frequency). For the histograms, the
x-axis is the same (MSS axis) and the y-axis the number
of events (estimates). For this analysis, the size of the
window (M) is 256 samples.
In our work, we made Monte-Carlo analysis with
different levels of SNR and jitter, as summarised in Figs.
8 and 9. It is possible to observe the consistency of the
method: lower jitter and low SNR generate a high per-
centage of correct estimates. In Fig. 9, there is an inter-
esting aspect; for almost all the curves, the best perfor-
mance is not at jitter equal to zero, but only when it is 5%
or 10%. One possible explanation for this would be that,
for low jitter, some estimates are locked in the harmonics
and, as jitter increases, the cluster spreads and some
estimates may have returned to be around the correct
value. This behaviour can also be observed in Fig. 8
(lines for ⫺3 dB and 5 dB). In this study, we have made
an analysis to quantify the repeatability of the process. If
we assume 50% of correct estimates as a sign of consis-
tency, we can see that, for the curve with SNR of 12 dB,
1238 Ultrasound in Medicine and Biology
jitter goes as far as 17%, approximately. Even for a SNR
as low as 0 dB, which is a severe adverse condition, the
process tends to generate consistent estimates up to
around 14% of jitter (Fig. 8). The performance of the
process continues to decrease as jitter goes up. It is
important to have in mind the question of until what level
of jitter a periodic structure can be understood as such.
Simon et al. (1997) presented a comparison of dif-
ferent types of MSS algorithms. The results shown in
Figs. 8 and 9 are similar to those in their article (Fig. 11c
and d). To generate the Monte-Carlo simulation, they
adopted a parameter Ad, defined as the amplitude rate
between the diffuse and regular scatterers, which is
equivalent to a SNR value of 20 dB in our definition. In
a first approach, consider the line of 12 dB of Figs. 8 and
9 to compare with the results of Simon et al. (1997).
Their results at 15% of jitter indicate a performance
inferior to 50%, but our results are near 60%. From 5%
to 15% of jitter, our results are near 10% higher than
theirs. In the range of 25%, the results are similar.
The simulation model used assumes that all the
scatterers are in one dimension (1-D). When using real
US signals, all the scatterers that are within the US beam
contribute to the signal. In other words, the scatterers are
in a 3-D distribution. So, for the signals from bovine
liver, we believe that the fact that SSA could deal with
this 3-D distribution and make the estimates converge to
a limited range with a clear peak is, in itself, an indica-
tion of the potential of the SSA analysis. The range in
which the estimates were 3.1 to 4.8 mm is also compat-
ible with the liver microstructure, and depends on how
the sample was cut and exposed to the US beam (i.e., the
orientation of the lobes).
CONCLUSION
A study is presented on the application of the SSA
to US backscattered signals from a simple phantom and
a bovine liver sample to estimate the mean scatterer
spacing. In both cases, it was possible to obtain consis-
tent estimates of the MSS. A Monte-Carlo analysis of the
estimates of MSS based on the SSA was performed on
simulated signals from a simple 1-D model from the
literature, which has shown the behaviour of SSA in the
presence of different levels of jitter and SNR.
SSA was shown to be a powerful way to estimate
MSS, even compared to the algorithm published in the
literature. In addition, SSA opens the possibility of
studying the decomposition of US signals (periodic and
aperiodic subspaces) for a possible characterisation of
biomedical US data.
Volume 27, Number 9, 2001
For this work, the next step is to devise a method-
ology to optimise the separation between periodic and
aperiodic subspaces, so that one can extract an unique
periodic signature of the medium. The precision, accu-
racy and sensibility of SSA are also being investigated
with a more significant amount of experimental data.
After that is reached, we can try to use SSA automati-
cally to localise intermittent oscillations.
Acknowledgement—To the Brazilian agencies CNPq and CAPES.
REFERENCES
Allen MR, Robertson AW. Distinguishing modulated oscillations from
coloured noise multivariate datasets. Climate Dynam 1996;12:775–
784.
Allen MR, Smith L. Monte Carlo SSA: Detecting irregular oscillations
in the presence of coloured noise. J Climate 1996;9:3373–3404.
Bridal SL, Saı̈ed A, Chérin E, Lefèbvre F, Laugier P, Berger G.
High-resolution quantitative echography with tissue parameters. J
Echograph Méd Ultrasons/JEMU. 1998;19(2/3):204 –211.
Broomhead DS, King G. Extracting qualitative dynamics from exper-
imental data. Physica D 1986;20:217–236.
Fellingham L, Sommer F. Ultrasonic characterization of tissue struc-
ture in the in vivo human liver and spleen. IEEE Trans Sonics
Ultrason 1984;SU-31:418 – 428.
Fowler AC, Kember G. Singular systems analysis as a moving-window
spectral method. Eur J Appl Math 1998;8:55–79.
Kadah YM, Farag AA, Zurada JM, Badaw AM, Yossef AM. Classifi-
cation algorithms for quantitative tissue characterization of diffuse
liver disease from ultrasound images. IEEE Trans Med Imaging
1996;15(4):466 – 478.
Landini L, Verrazzani L. Spectral characterization of tissue microstruc-
ture by ultrasound: A stochastic approach. IEEE Trans UFFC
1990;37:448 – 456.
Narayanan VM, Molthen RC, Shankar PM, Vergara L, Reid JM.
Studies on ultrasonic scattering from quasi-periodic structures.
IEEE Trans UFFC 1997;44:114 –124.
Palus M, Dvorak I. Singular value decomposition in attractor recon-
struction: Pitfalls and precautions. Physica D 1992;58:221–234.
Pereira FR, Pereira WCA, Machado JC. Ultrasonic wave speed mea-
surement using the time delay profile of RF backscattered signals.
Ultrasonics 2000;38:708 –710.
Simon C, Shen J, Seip R, Ebbini, ES. A robust and computationally
efficient algorithm for mean scatterer spacing estimation. IEEE
Trans UFFC 1997;44(4):882– 894.
Therrien C. Discrete random signals and statistical signal processing.
Englewood Cliffs, NJ. Prentice Hall Processing Series, 1992.
Varghese T, Donohue KD. Characterization of tissue microstructure
scatterer distribution with spectral correlation. Ultrason Imaging
1993;15:238 –254.
Varghese T, Donohue KD. Mean scatterer spacing estimate with spec-
tral correlation. J Acoust Soc Am 1994;96:3504 –3515.
Varghese T, Donohue KD. Estimating mean scatterer spacing with the
frequency smoothed spectral autocorrelation function. IEEE Trans
UFFC 1995;42:451– 463.
Vautard R, Ghil M. Singular-spectrum analysis in nonlinear dynamics
with applications to paleoclimatic time series. Physica D 1989;35:
395– 424.
Vautard R, Yoiu P, Ghil M. Singular-spectrum analysis: A toolkit for
short, noisy chaotic signals. Physica D 1992;58:95–126.
Wear K, Wagner R, Insana M, Hall T. Application of autoregressive
spectral analysis to cepstral estimation of mean scatterer. IEEE
Trans UFFC 1993;40:50 –58.