Ultrasonics 51 (2011) 247–252

Contents lists available at ScienceDirect

Ultrasonics
journal homepage: www.elsevier.com/locate/ultras

Short Communication

In vitro microemboli classification using neural network models and RF signals

N. Benoudjit a, K. Ferroudji a, M. Bahaz a, A. Bouakaz b,⇑
a
Laboratoire d’Electronique Avancée, Université de Batna, Algeria
b
UMR Inserm U930 CNRS ERL 3106, Tours, France

a r t i c l e i n f o a b s t r a c t

Article history: Emboli classification is of high clinical importance for selecting appropriate treatment for patients. Sev-
Received 25 April 2010 eral ultrasonic (US) methods using Doppler processing have been used for emboli detection and classifi-
Received in revised form 29 August 2010 cation as solid or gaseous matter. We suggest in this experimental study exploiting the Radio-Frequency
Accepted 5 September 2010
(RF) signal backscattered by the emboli since they contain additional information on the embolus than
Available online 15 September 2010
the Doppler signal. The aim of the study is the analysis of RF signals using Multilayer Perceptron
(MLP) and Radial-Basis Function Network (RBFN) in order to classify emboli.
Keywords:
Anthares scanner with RF access was used with a transmit frequency of 1.82 MHz at two mechanical
Microemboli
Gaseous embolus
indices (MI) 0.2 and 0.6. The mechanical index is given as the peak negative pressure (in MPa) divided
Solid embolus by the square root of the frequency (in MHz). A Doppler flow phantom was used containing a 0.8 mm
RF signals diameter vessel surrounded by a tissue mimicking material. To imitate gas emboli US behaviour, Sonovue
Neural network microbubbles were injected at two different doses (10ll and 5ll) in a nonrecirculating at a constant
flow. The surrounding tissue was assumed to behave as a solid emboli. In order to mimic real clinical
pathological situations, Sonovue concentration was chosen such that the fundamental scattering from
the tissue and from the contrast were identical. The amplitudes and bandwidths of the fundamental
and the 2nd harmonic components were selected as input parameters to the MLP and RBFN models.
Moreover the frequency bandwidths of the fundamental and the 2nd harmonic echoes were approxi-
mated by Gaussian functions and the coefficients were used as a third input parameter to the neural
network models. The results show that the Gaussian coefficients provide the highest rate of classification
in comparison to the amplitudes and the bandwidths of the fundamental and the 2nd harmonic compo-
nents. The classification rates reached 89.28% and 92.85% with MLP and RBFN models respectively.
This short communication demonstrates the opportunity to classify emboli based on a RF signals and
neural network analysis.
Ó 2010 Elsevier B.V. All rights reserved.

1. Introduction
The ultrasound community has experienced dramatic technical
advances over the last decades, such as blood flow measurements
with elaborate Doppler techniques or real time three-dimensional
imaging with 2-D phase array transducers. This was partly ascribed
to the advantages of ultrasound over other diagnostic modalities,
including its low cost, real time character and safety.
Transcranial Doppler (TCD) [1] is a non-invasive ultrasound
method used to assess blood flow velocity in the major basal intra-
cranial arteries on a real time basis. Since the early 60’s, experi-
mental studies have been carried out to evaluate the ability of
TCD to detect, quantify and classify intracranial emboli. One of
the techniques consists in detecting the appearance of abrupt
changes in the Doppler signal with amplitudes that are higher than
⇑ Corresponding author. Address: CHU Bretonneau, B1A, UMR Inserm U930 CNRS
ERL 3106, 2 Boulevard Tonnellé, 37044 Tours Cedex 9, France.
E-mail address: bouakaz@med.univ-tours.fr (A. Bouakaz).
the remaining recorded signal. These changes are known as high
intensity transient signals in the TCD waveform and are used as
indicators of circulating microemboli, in a wide variety of clinical
areas: cardiac and carotid surgery [2,3], following prosthetic heart
valve insertion [4] and cerebral angiography [5]. Many reports
have emphasized the ability of TCD, combined with sophisticated
signal processing techniques, to identify microemboli in the brain
circulation. Currently, TCD based techniques rely on the appear-
ance of high intensity transient signals (HITS) which are used as
indictors of the presence of emoboli. Unfortunately, this technol-
ogy presents today some limitations to first detect the emboli
and then to distinguish between emboli and artefacts [6–8] and
more importantly to determine the embolus composition [9,10].
Especially before performing interventional procedures for neuro-
protection where the nature of the emboli is an important diagnos-
tic criteria. This is the case during carotid endarterectomy and
cardiac surgery where air and solid emboli are strongly intricate.
They can be detected with TCD but it has been very difficult to
distinguish solid from gaseous microemboli. Solid emboli are

0041-624X/$ - see front matter Ó 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ultras.2010.09.002
248 N. Benoudjit et al. / Ultrasonics 51 (2011) 247–252
potentially far more damaging than gaseous emboli and they imply
different therapeutic strategies [11]. Indeed, gaseous emboli are
considered as a risk indicator and their detection implies to seek
for the cause of the emboli. However, when solid emboli is de-
tected, an anticoagulant or antiagregant treatment will be used.
In this short communication, we propose a new approach to de-
tect and classify microemboli using artificial neural networks and
the backscatter RF (Radio-Frequency) signal instead of the Doppler
signal. The rest of this manuscript is organized in four sections.
Section 2 describes the basics of Multilayer Perceptron (MLP) and
Radial-Basis Function Network (RBFN) models when used as clas-
sifiers. Section 3 presents the material and methods used in this
experimental study. Section 4 provides the experimental results
obtained for two concentrations (10 ll and 5 ll) of microbubbles
at low (0.2) and high (0.6) MI. Section 5 draws the conclusion of
this study.
2. Artificial neural networks
Artificial neural networks (ANN) are widely used in applications
involving classification or function approximation. It has been pro-
ven that several classes of ANN such as Multilayer Perceptron
(MLP) and Radial-Basis Function Network (RBFN) are universal
function approximators [12,13]. In the next two subsections, we
will briefly recall the basis of MLP and RBFN models.
2.1. Multilayer Perceptron (MLP)
The most widely used neural classifier today is the MLP network
which has also been extensively analyzed and for which many
learning algorithms have been developed. MLP networks are non-
linear models consisting of a number of units organized into multi-
ple layers. Given enough hidden units and enough data, it has been
shown that MLPs can approximate virtually any function to any de-
sired accuracy. In other words, MLPs are universal approximators
[12,13].
The MLP neural network model consists of a network of pro-
cessing elements or nodes arranged in layers. Typically it requires
three or more layers of processing nodes: an input layer which ac-
cepts the input variables used in the classification procedure, one
or more hidden layers, and an output layer with one node per class.
For our case, two classes (gaseous and solid embolus) will be used.
The principle of the network is that when data from an input pat-
tern is presented at the input layer, the network nodes perform cal-
culations in the successive layers until an output value is
computed at each of the output nodes. This output signal should
provide the appropriate class for the input data, i.e. we expect to
have a high output value on the correct class node and a low out-
put value on all the rest. The multilayer feed-forward neural net-
work is trained by supervised learning using the iterative back-
propagation algorithm [12]. In the learning phase a set of input
patterns, called the training set, are presented at the input layer
as feature vectors, together with their corresponding desired out-
put pattern which usually represents the classification for the in-
put pattern. Starting with small random weights for each input
pattern, the network is required to adjust the weights attached
to the connections so that the difference between the network’s
output and the desired output for that input pattern is decreased.
The weights between the output layer and the layer below (hidden
layer) are then adjusted by the generalized delta rule [12].
2.2. Radial-Basis Function Network (RBFN)
Radial-Basis Function Network (RBFN) can be used for a wide
range of applications mainly due to the fact that they can approx-
imate any function [13,14] and their training is faster compared to
Multilayer Perceptrons (MLP). This fast learning speed comes from
the fact that RBFN has just two layers (hidden layer and output
layer) of parameters (centers + widths and weights) and each layer
can be determined sequentially. The training of RBF networks can
be split into an unsupervised part and a linear supervised part.
Unsupervised updating techniques are straightforward and rela-
tively fast. The supervised part of the learning consists in solving
a linear problem, which is therefore also fast, with the additional
benefit of avoiding the problem of local minima usually encoun-
tered when using Multilayer Perceptrons (MLP) [13–15].
A RBF network is a two-layered ANN and is defined as:
X
K
 
^f ðxÞ ¼ kj /j kxi  cj k ; ð1Þ
j¼1
where
 2 !
1 kxi  cj k
/j ðkxi  cj kÞ ¼ exp  ; ð2Þ
2 rj
where ^f ðxÞ is an unknown function, uj is a set of multi-dimensional
Gaussian kernels and kj are weight factors. According to Eq. (1), the
model is characterized by three types of parameters. The cj are
called centres (centroids). They are chosen into dense regions of
the input space, via a vector quantization technique [15,16]. The
number K represents the number of centres in the hidden layer
while rj is the width of each centre. The choice of the rj is selected
according to a previous study [15]. The cj and rj are chosen regard-
less of the outputs f(x) = 1 (gaseous embolus) or f(x) = 0 (solid
embolus), based only on the properties of the distribution of the in-
puts space x. Finally, the kj are found by linear regression analysis.
Before building a model, the samples are often subdivided into
learning (training and, validation) and test sets. Our experimental
data consist of 102 samples. In order to calculate the predictive
ability of a model, we need to set aside some data which is not used
in the learning phase. This set is known as the test set. Approxi-
mately one-third (1/3) of the data are to be randomly assigned to
the test set (26). The rest will belong to the learning set, i.e., the
set of data used for building the models. The learning set itself will
be divided into a training set (approximately two-third) and a val-
idation set (approximately one-third). The first set is used to obtain
the network parameters (for MLP and RBFN) (50); the second set is
used to choose the optimal number of neurons in the hidden layer
of the MLP and RBFN networks (26).
3. Material and methods
The experimental set-up consists of a Doppler flow phantom
containing three tubes of 0.2, 0.4 and 0.8 mm in diameter. The tube
of 0.8 mm was chosen since its size approximates the size of a hu-
man vessel (Fig. 1). A continuous flow carries the Sonovue micro-
bubbles through the insonified vessel. Sonovue was used since it
is the only commercially available contrast agent in Europe. The
Fig. 1. Experimental set-up.
 N. Benoudjit et al. / Ultrasonics 51 (2011) 247–252 249

flow and the concentration of microbubbles are controlled by the
operator. The concentration of contrast agent microbubbles is cho-
sen such as its scattering amplitude at the fundamental frequency
is comparable to the scattering of the surrounding tissue at the
same frequency. This choice is dictated by the clinical situation
where the scattering amplitude of emboli approaches that of blood.
In the study, two concentrations of the contrast agent were used:
5 ll and 10 ll. The microbubbles were administered into a
200 ml volume of Isoton.
Several studies have shown that the acoustic behaviour of gas-
eous emboli and microbubbles are similar, mainly the nonlinear
behaviour [17]. For that reason, contrast microbubbles have been
chosen to mimic the ultrasonic behaviour of gaseous emboli.
However solid emboli have been shown to exhibit acoustic
properties similar to those of biological tissue. Therefore the scat-
tering of solid emboli is purely linear and no harmonic components
can be seen in their scattered signal. This ultrasonic behaviour is
similar to that of biological tissues and thus the tissue mimicking
material surrounding the vessel was chosen to mimic the behav-
iour of solid emboli.
The ultrasound waves were generated by a VF13-5 probe con-
nected to a Siemens Antares scanner (Anthares, Siemens, MV,
CA). The acquisitions were performed at 1.82 MHz transmit fre-
quency in Tissue Harmonic Imaging (THI) mode at a frame rate
of 14 fps. The acoustic focus was set at 2 cm which is the depth
at which the flowing contrast microbubbles are located. The User
Research Interface was used to grab the unfiltered RF signals to a
personal computer for further analysis.
The MLP and RBFN were developed with MATLAB neural net-
work toolbox (Mathworks, Massachusetts, USA). The following
parameters were selected as input to the models:
 The bandwidths of the scattered RF signals at the fundamental
and at the second harmonic frequencies.
 For every RF signal scattered from the microbubbles (gas
emboli) or surrounding tissue (solid emboli), its FFT is calcu-
lated. The amplitudes at the fundamental frequency and at
the second harmonic frequency are selected. These values are
used as an input parameter.
 The frequency spectra of the fundamental and the second har-
monic were approximated by a Gaussian shape function using
the following equation:
 2 !
ðx  b1 Þ
gðxÞ ¼ a1 exp  ð3Þ
c1
The Gaussian coefficients a1, b1 and c1 were used as a third input
parameter.
Fig. 2 shows the general bloc diagram of the neural network
classification models used in our study. For each input parameter,
a vector containing the values of the parameters (amplitude or
bandwidth or the Gaussian coefficients) is constructed. This vector
is used as an input to the MLP or RBFN model which provides in its
output a value of 1 or 0 for gaseous or solid emboli respectively.
4. Results and discussion
Fig. 3 shows an example of a typical grey scale images obtained
at an MI of 0.2 (panel A) and an MI of 0.6 (panel B) for a concentra-
tion of microbubbles equal to 5 ll. The regions of interest corre-
sponding to gaseous and a solid emboli are shown on each of the
images.
Fig. 4 displays two examples of RF signals extracted from the re-
gions of interest at the two different mechanical indices. The
acoustic pressures are given through the mechanical index as dis-
played on the scanner monitor. The mechanical index is defined as
the peak negative pressure (in MPa) divided by the square root of
the frequency (in MHz). Fig. 4A displays an RF signal backscattered
by solid and gaseous emboli at MI of 0.2. The frequency spectra of
both signals include only a component at the fundamental fre-
quency. The acoustic pressure is not sufficiently high to generate
non linear microbubbles oscillations characterized by the

1: Gaseous embolus
Neural Network
Classification model
(RBFN or MLP) 0: Solid embolus
Input parameter output

Fig. 2. General bloc diagram of the neural network classification models.

C= 5 µl
(a) LOW MI (b) HIGH MI

Solid Emboli
Solid Emboli

Gaseous Emboli
Gaseous Emboli

A B

Fig. 3. Examples of grey scale images acquired: (A) at low mechanical index (MI = 0.2) and (B) high mechanical index (MI = 0.6) for a concentration of microbubbles equal to
C = 5 ll. The regions of interest corresponding to a solid embolus and gaseous embolus were extracted, as shown on each image.
250 N. Benoudjit et al. / Ultrasonics 51 (2011) 247–252
Fig. 4. Examples Radio-Frequencies (RF) signals and their corresponding frequency spectrum represented with the Gaussian approximation (dashed line): (A) at low
mechanical indice (MI = 0.2); (B) high mechanical index (MI = 0.6).
formation of a second harmonic component. Thus the solid embo-
lus responds linearly to the ultrasound excitation in a similar way
as the gaseous embolus. At this mechanical index and frequency
(0.2 and 1.8 MHz respectively), the peak negative pressure is
260 kPa. The frequency spectra of the scattered signals from the re-
gion of the microbubbles do not show any harmonic components
and thus we assume that at this acoustic pressure, the microbub-
bles scatter only linearly.
Fig. 4B shows the scattered RF signals of both solid and gaseous
emboli at a higher MI (0.6). We appreciate for this excitation pres-
sure the generation of nonlinear components at the second har-
monic frequency by both gaseous and solid particles. For the case
of a gaseous embolus, this component is produced by the nonlinear
oscillations of the microbubbles and therefore is considered as a
classification parameter since solid embolus scatters only linearly.
Nevertheless, at high acoustic pressure (or MI’s), which is the case
at MI of 0.6; the propagation of an ultrasound wave becomes non-
linear, meaning that harmonic components (2nd, 3rd and higher)
are generated in the propagation path. As a consequence, a solid
embolus that is located at a distance from the transducer will be
hit not only by the main (or fundamental) component but also
by the harmonics that were generated during the propagation
path. Since a solid embolus scatters linearly, it will scatter all the
impinging components including the fundamental and the 2nd
harmonic. Thus and at this applied MI, a generated second har-
monic component does not necessarily indicate that a gas bubble
is present since it can be generated during the nonlinear propaga-
tion process. Therefore the frequency spectrum of the signal back-
scattered by a solid embolus will also include some nonlinear
components. Since these nonlinear propagation effect will contam-
inate the scattering nonlinearity of the gas microbubbles, the har-
monic generation cannot be used as a discrimination factor
between solid and gaseous matter.
Tables 1 and 2 summarize the percentage of correct classifica-
tion of microemboli using the MLP and RBFN analysis as a function
of the different input parameters and the mechanical index for the
two microbubble concentrations (10 ll and 5 ll).
The bandwidths of the linear and non linear components for the
two concentrations do not provide a correct average rate of classi-
fication neither at the low MI (0.2) and nor at the high MI (0.6).
Only 50% of classification rate is obtained for both models (RBFN
and MLP) using the bandwidths as input parameter. When
amplitudes of the fundamental and the second harmonic compo-
nents are introduced as input parameters into the neural networks
models, the correct average rate of classification of microemboli at
high MI (0.6) reached 87.5% for the RBFN classifier and 82.14% for
the MLP classifier for the higher concentration. At the low micro-
bubble concentration (5 ll), the correct average rates of classifica-
tion for microemboli at high MI (0.6) were 83.13% and 78.56% for
RBFN and MLP models respectively. Here we talk about RF signal
scattered from the microbubbles (gas emboli) or surrounding tis-
sue (solid emboli), its FFT is calculated. The amplitudes at the fun-
damental frequency and at the second harmonic frequency are
selected. These values are used as an input parameter.
When the Gaussian coefficients of the spectral envelopes are
used as input parameters in both neural network models (RBFN
and MLP), the highest rate of classification reached a value of
92.85% for the RBFN model at the high MI (0.6). Using the MLP
 N. Benoudjit et al. / Ultrasonics 51 (2011) 247–252 251

Table 1
Classification rates of gaseous and solid emboli with concentration of microbubbles (10 ll) at low MI (0.2) and high MI (0.6) for three different input
parameters: the bandwidths and the amplitudes of the fundamental and the second harmonic and the Gaussian parameters issued form Eq. (3).

Bandwidths Amplitudes Gaussian coefficients

Low MI High MI Low MI High MI Low MI High MI
Results with MLP (C = 10 ll)
Gaseous emboli 0% 0% 100% 100% 85.71% 78.57%
Solid emboli 100% 100% 64.28% 64.28% 85.71% 100%
Average rate 50% 50% 82.14% 82.14% 85.71% 89.28%
Results with RBF (C = 10 ll)
Gaseous emboli 0% 0% 78.57% 85.71% 92.85% 85.71%
Solid emboli 100% 100% 71.42% 85.71% 78.57% 100%
Average rate 50% 50% 74.99% 85.71% 85.71% 92.85%

Table 2
Classification rates of gaseous and solid emboli with concentration of microbubbles (5 ll) at low MI (0.2) and high MI (0.6) for three different input
parameters: the bandwidths and the amplitudes of the fundamental and the second harmonic and the Gaussian parameters issued form Eq. (3).

Bandwidths Amplitudes Gaussian coefficients

Low MI High MI Low MI High MI Low MI High MI
Results with MLP (C = 5 ll)
Gaseous emboli 0% 0% 71.42% 71.42% 92.85% 78.57%
Solid emboli 100% 100% 85.71% 78.57% 85.71% 100%
Average rate 50% 50% 78.56% 78.56% 89,28% 89.28%
Results with RBF (C = 5 ll)
Gaseous emboli 0% 0% 78.57% 71.42% 78.57% 92.85%
Solid emboli 100% 100% 92.85% 92.85% 78.57% 85.71%
Average rate 50% 50% 85.71% 83.13% 78.57% 89.28%

A 1.5 1.5
Predicted Gaseous embolus Predicted Gaseous embolus
Predicted Solid Embolus Predicted Solid Embolus

1 1

0.5 0.5
Threshold
Threshold

0 0

-0.5 -0.5
0 5 10 15 0 5 10 15
RBFN MLP

1.5 1.5
B Predicted Gaseous embolus Predicted Gaseous embolus
Predicted Solid Embolus Predicted Solid Embolus

1 1

0.5 0.5
Th h ld
Threshold Threshold

0 0

-0.5 -0.5
0 5 10 15 0 5 10 15
RBFN MLP
Fig. 5. Embolus classification using Radial-Basis Functions Network (RBFN) and Multilayer Perceptron (MLP) network with concentration of microbubbles equal to 5 ll (A) at
low mechanical index (MI = 0.2); (B) at high mechanical index (MI = 0.6).
252 N. Benoudjit et al. / Ultrasonics 51 (2011) 247–252
model, the Gaussian coefficients provided a classification rate of
89.28% for at the high MI (0.6) for both microbubble concentra-
tions. These high classification rates might be ascribed to the fact
that the coefficients of the spectral envelopes contain additional
information about the bandwidths and the amplitudes of the linear
and non linear components of the backscattered signals from both
solid and gaseous emboli.
Fig. 5 shows the best model obtained for embolus classification
using MLP and RBFN with concentration of microbubbles equal to
5 ll for respectively low (0.2 upper pannel) and high MI (0.6 lower
panel). For our data, the threshold was set to 0.5 to have an initial
probability of 0.5 (half) for both solid and gaseous embolus. In a
clinical situation, the threshold can be set to a different value
depending on the application.
An output value of 1 corresponds to the gaseous embolus repre-
sented by circles and an output value of 0 corresponds to the solid
embolus represented by stars. For example, Fig. 5A, corresponding
to low MI (0.2) shows that the RBFN classification model used in
this experiment succeeded in classifying 11 gaseous embolus out
of 14 and 11 solid embolus out of 14 also. On the other hand the
MLP classification model used for the same experiment succeeded
in classifying 13 gaseous embolus out of 14 and 12 solid embolus
out of 14. The processing time is an important issue if we have
to implement this technique in a clinical machine. However the
main objective of our study is only to evaluate the performances
of the proposed approach in classifying the emboli. Therefore we
did not focus on the time of processing the data. Obviously, MAT-
LAB will not be used as a processing tool since other software pack-
ages are more appropriate. Nevertheless, it should be noted that
the training phase for the MLP network lasts for a few minutes
while the training phase for RBF network takes less than one min-
ute. The processing time for the test phase (classification) is less
1 s. The current emboli classification techniques are performed in
real time using TCD machines since this differentiation is observed
in the Doppler signal.
5. Conclusions
This short communication demonstrates the opportunity to ex-
ploit RF signal processing for a better classification of microemboli
as solid or particulate matter. A neural network (MLP or RBFN)
analysis using the fundamental and the second harmonic compo-
nents information contained in the RF signal backscattered by an
embolus allows the classification with sensitivity and a specificity
of 92.85%.
In this in vitro study, a first proof of concept of emboli classifi-
cation based on the combination of neural network processing and
RF information is demonstrated. Although the clinical utility still
needs verification in practical diagnosis, the main message of this
study is to validate this strategy in a simple and a controlled exper-
imental environment before further pre-clinical and clinical valida-
tions are undertaken.
Acknowledgment
The authors would like to acknowledge the support of INSERM
(France)/DPGRF (Algeria) 2007/2008.
References
[1] R. Aslid, T.M. Markwalder, H. Nornes, Non invasive transcranial Doppler
ultrasound recording of flow velocity in basal cerebral arteries, J. Neurosurg. 57
(6) (1982) 769–774.
[2] W. Pugsley, L. Klinger, C. Paschlis, T. Treasure, M. Harrison, S. New-man, The
impact of microemboli during cardiopulmonary bypass on neuropsychological
functioning, Stroke 25 (7) (1994) 1393–1399.
[3] R.G.A. Ackerstaff, C. Jansen, F.L. Moll, Carotid endarterectomy and
intraoperative emboli detection, Echocardiography 13 (5) (1996) 543–550.
[4] D. Georgiadis, D. Grosset, A. Kelman, A. Faichney, K.R. Lees, Prevalence and
characteristics of intracranial microemboli signals in patients with different
types of prosthetic cardiac valves, Stroke 25 (11) (1994) 587–592.
[5] H.S. Markus, A. Loh, D. Israel, T. Buckenham, A. Clifton, M.M. Brown,
Microscopic air embolism during cerebral angiography and strategies for its
avoidance, The Lancet 341 (1993) 784–787.
[6] M. Cullinane, G. Reid, R. Dittrich, Z. Kaposzta, R.G.A. Ackerstaff, V. Babikian,
D.W. Droste, D. GRosset, M. Siebler, L. Valton, H.S. Markus, Evaluation of new
online automated embolic signal detection algorithm including comparison
with panel of international experts, Stroke 31 (6) (2000) 1335–1341.
[7] L. Fan, D. Evans, A. Naylor, Automated embolus identification using a rule-
based expert system, Ultrasound Med. Biol. 27 (8) (2001) 1065–1077.
[8] N. Aydin, S. Padayache, H.S. Markus, The use of the wavelet transform to
describe embolic signals, Ultrasound Med. Biol. 6 (1999) 953–958.
[9] D. Russell, R. Brucher, Online automatic discrimination between solid and
gaseous cerebral microemboli with the first multifrequency transcranial
Doppler, Stroke 33 (8) (2002) 1975–1980.
[10] J.L. Smith, D.H. Evans, P.R.F. Bell, A.R. Naylor, A comparison of four methods for
distinguishing Doppler signals from gaseous and particulate emboli, Stroke 29
(7) (1998) 1133–1138.
[11] G.A. Darbellay, R. Duff, J.-M. Vesin, P.-A. Despland, D.W. Droste, C. Molina, J.
Serena, R. Sztajzel, P. Ruchat, T. Karapanayiotides, A. Kalangos, J.
Bogousslavsky, Solid or gaseous circulating brain emboli: are they separable
by transcranial ultrasound?, J Cerebr. Blood Flow Metab. 24 (2004) 860–868.
[12] S. Haykin, Neural Networks: A Comprehensive Foundation, second ed.,
Prentice-Hall, 1999, pp. 156–256 (Chapter 4 and 5).
[13] Robert J. Howlett, Lakhmi C. Jain, Radial Basis Function Networks 1: Recent
Developments in Theory and Applications, Physica-Verlag, Heidelberg, 2001
(Printed in Germany, Chapter 10).
[14] J. Park, I. Sandberg, Approximation and radial basis function networks, Neural
Comput. 5 (1993) 305–316.
[15] N. Benoudjit, M. Verleysen, On the kernel widths in radial-basis function
networks, Neural Process. Lett. 18 (2003) 139–151.
[16] R. Gray, Vector quantization, IEEE Mag. 1 (1984) 4–29.
[17] P. Palanchon, A. Bouakaz, J. Klein, N. de Jong, Multifrequency transducer for
microemboli classification and sizing, IEEE Trans. Biomed. Eng. 52 (12) (2005)
2087–2092.