Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5829_Библиотеки_им_академика_М_И_Перельмана.pdf
Activity Index: A Tool to Identify Active Carotid Plaques 175
31. Carr S, Farb A, Pearce WH, et al (1996) Atherosclerotic plaque rupture in symptomatic carotid
artery stenosis. J Vasc Surg 23:755
32. Gertz SD, Roberts WC (1990) Hemodynamic shear force in rupture of coronary arterial
atherosclerotic plaques. Am J Cardiol 66:1368–1372
33. Davies MJ, Richardson PW, Woolf N et al (1993) Risk of thrombosis in human atherosclerotic
plaques: role of extracellular lipid, macrophage and smooth muscle cell content. Br Heart J
69:377–381
34. Bassiouny HS, Sakaguchi Y, Mikucki SA, et al (1997) Juxtalumenal location of plaque necrosis
and neoformation in symptomatic carotid stenosis. J Vasc Surg 26:585
35. Gronholdt ML, Nordestgaard BG, Wiebe BM (1998) Echolucency of computerized ultrasound
images of carotid atherosclerotic plaques are associated with increased levels of triglyceriderich lipoproteins as well as increased plaque lipid content. Circulation 97:34–40
36. Lammie GA, Path MR, Sandercock PA, Dennis MS (1999) Recently occluded intracranial
and extracranial carotid arteries. Relevance of the unstable atherosclerotic plaque. Stroke
30:1319–1325
37. Torvik A, Svindland A, Lindboe CF (1989) Pathogenesis of carotid thrombosis. Stroke
20:1477–1483
38. Gonc¸alves I, Moses J, Dias N, Pedro LM, Fernandes e Fernandes J, Nilsson J, Ares MP (2003)
Changes related to age and cerebrovascular symptoms in the extracellular matrix of human
carotid plaques. Stroke 34:616–622
39. Gonc¸alves I, Moses J, Pedro LM, Dias N, Fernandes e Fernandes J, Nilsson J, Ares MP (2003)
Echolucency of carotid plaques correlates with plaque cellularity. Eur J Vasc Endovasc Surg
26:32–38
40. Gonc¸alves I, Lindholm MW, Pedro LM, Dias N, Fernandes e Fernandes J, Fredrikson GN,
Nilsson J, Moses J, Ares MP (2004) Elastin and calcium rather than collagen or lipid content
determine the echogenicity of human carotid plaques. Stroke 35:2795–2800
41. El-Barghouty GG, Nicolaides AN et al (1995) Computer-assisted carotid plaque characterization. Eur J Vasc Endovasc Surg 9:389
42. Russell DA, Wijeyaratne SM, Gough MJ (2007) Changes in carotid plaque echomorphology
with time since a neurologic event. J Vasc Surg 45(2):367–372
43. Madycki G, Staszkiewicz W (2006) Detailed plaque texture analysis as the alternate method
of ultrasound image analysis in predicting the risk of intraoperative microembolism and
perioperative complications. Vasa 35(2):78–85
Coronary Atherosclerotic Plaque
Characterization By Intravascular Ultrasound
Francesco Ciompi, Oriol Pujol, Josepa Mauri Ferr´e, and Petia Radeva
Abstract The accurate characterization of in vivo atherosclerotic plaques
represents an important task during percutaneous intervention. Intravascular
Ultrasound (IVUS) is a catheter-based imaging technique that provides a detailed
cross-sectional representation of the internal morphology of the vessel, thus
allowing to assess plaque amount and composition. In this chapter the state of the
art methods for automatic plaque characterization in IVUS are analyzed. The main
classification techniques as well as the most discriminative features are illustrated.
Furthermore, a recently presented technique for the fusion of in vivo and in vitro
IVUS data is illustrated.
1 Introduction
Coronary heart disease represents the 21% of mortality cause and accounts for
2 millions of lives in Europe. Atherosclerotic plaque formation results from the
thickening of the intimal-medial segments and an overall thickening of the vessel
wall. The term vulnerable plaque is commonly attributed to an atherosclerotic
plaque that is likely to rupture or fissure, leading to thrombosis, and then to acute
coronary syndrome: myocardial infarction, unstable angina pectoris, or sudden
cardiac death [1–6]. An accurate analysis of in vivo plaque composition is then
an important task in diagnosis and detection of vulnerable atheroma before plaque
rupture.
F. Ciompi ()•O.Pujol•P.Radeva
Computer Vision Center, Campus UAB, Edifici O, Bellaterra, Spain
University of Barcelona, Gran Via de Les Cortes Catalanes, 585, 08007 Barcelona, Spain
e-mail: fciompi@maia.ub.es; oriol@maia.ub.es; petia@cvc.uab.es
J.M. Ferr´e
University Hospital Germans Trias i Pujol, Carretera de Canyet s/n. 08916 Badalona, Spain
e-mail: mauritri@gmail.com
J.M. Sanches et al. (eds.), Ultrasound Imaging: Advances and Applications,
DOI 10.1007/978-1-4614-1180-2
8, © Springer Science+Business Media, LLC 2012
177
178 F. Ciompi et al.
Fig. 1 (a) In vivo IVUS image in cartesian coordinates, representing the cross-sectional IVUS
image and (b) in polar coordinates
Fig. 2 Histological image (at the microscope) of ex vivo coronary section (a). The three layers are
marked and (b) the detail of the interfaces (internal and external elastic lamina) is shown
Intravascular Ultrasound (IVUS) is a catheter-based imaging modality that
provides an accurate luminal and transmural image of vascular structures [7]. It
allows the visualization of the full circumference of the vessel wall, thus showing
the tissue morphology and composition [8,9](seeFig.1).
The use of IVUS facilitates direct measurements of lumen size, including minimum and maximumdiameter and cross-sectionalarea as well as the characterization
of atheroma size, plaque distribution, and lesion composition [10]. Three main
regions, surrounding the luminal area of the vessel, can be distinguished in an IVUS
image: the intima,themedia,andtheadventitia (see Fig. 2).
Coronary Atherosclerotic Plaque Characterization By Intravascular Ultrasound 179
The intima is normally a thin layer of endothelial cells: this layer substantially
and often unevenly thickens into atherosclerosis. The media consists of multiple
layers of smooth muscle cells arranged helically and circumferentially around the
lumen of the artery. Finally, the adventitia is the external layer, essentially composed
by fibrous tissue, i.e., collagen and elastin [11]. Estimation ofthe vessel area is based
on the measurements of the media-adventitia border, and plaque area is derived by
subtracting lumen area from vessel area.
Several IVUS-based approaches for the automatic assessment of plaque
composition have been proposed in the last ten years, mainly based on image
texture analysis [12–14] as well as on raw Radio Frequency (RF) signals processing
[15–20]. Some approaches use Autoregressive Models (ARM) [15–17]orthe
Fourier Transform [18, 20] to obtain a description of the tissue in the frequency
domain, from whichspectral features are extracted. Approachesusing the Integrated
Backscatter parameter [21–23]ortheWavelet coefficients [24] have been also
proposed.
Most of the proposed methods for the automatic plaque characterization are
based on a model that learns the plaque properties by labeled IVUS examples.
In order to correctly train a model, reliable labeled IVUS data are necessary. The
reliable correspondence between an atherosclerotic plaque and the corresponding
IVUS data can only be obtained by histological analysis of post-mortem coronary
arteries. Histology is in fact the only procedure that allows to know the real
plaque nature in a certain position of the vessel. Unfortunately this methodology
suffers from the complicated procedure of obtaining in vitro data: scarce arteries
availability, frequent tissue spoiling during analysis and the difficulty in finding the
right correspondence between IVUS and histological image.
In addition, in almost all the known plaque characterization methods, the
discriminative power is assessed on in vitro plaques, then implicitly extending
the discriminative property to in vivo data as well. This erroneous assumption is
implicitly based on the hypothesis that differences between in vivo and in vitro
IVUS data are negligible.
In this chapter some of the most important state-of-the-art techniques for
automatic plaque characterization in Intravascular Ultrasound are presented. Furthermore, recent advances in the research topic of fusing in vivo and in vitro IVUS
data for plaque characterization are described, and some results are provided.
2 In Vitro Data Validation Procedure
A possible procedure for the acquisition and validation of post-mortemIVUS data is
now described. The coronary artery (separated from the heart) is first put on a midsoft plane and filled (using a catheter) with physiological saline solution at constant
pressure (around 100 mmHg), simulating blood pressure (see Fig.3). In the panel
the distal and proximal position, together with left and right hand are marked to be
used as a reference during histology. The probe is then introduced and RF data are
180 F. Ciompi et al.
Fig. 3 Set up for IVUS acquisition from post-mortem artery. (a) Coronary artery on the
plane with marked positions clearly shown; (b) plaque segmentation on histological image and
(c) segmentation on corresponding IVUS image
acquired in correspondence of plaques. These positions can be clearly marked on
the externalpart of the artery. The artery is then cut in correspondenceof previously
marked positions and plaque composition is determined by histological analysis.
As an alternative, a specially designed box containing the artery can be used,
allowing the automatic cut of the vessel at a fixed number of positions by
superimposing a dedicated set of knives [18]. With this procedure, a set of equally-
spaced artery cuts of the artery are obtained.
For each cut, an image of the tissue at the microscope is obtained. Given the
reference points in the paneland in the IVUS image orientation itis possibleto put in
correspondence the plaques detected by histology with their respective areas in the
IVUS image. The different conditions and modalities in which the two images are
acquired are profoundly different. The mechanical consistence given to the artery
while acquiring IVUS data is lost when cutting the vessel. Phenomena of tissue
spoiling and a certain error in finding the exact correspondence between the IVUS
and histological image make hard to get a good registration and, consequently, a
reliable automatic labeling. Hence, the labeling process can be performed manually
by joint cooperation of expertsand pathologist or by using a dedicated software able
to register the two images obtained by different modalities and different approaches
[16,17]. The plaques in IVUS data are thus validated and a groundtruth is obtained.
3 Plaque Characterization by Intravascular Ultrasounds
Different tissue types exhibit different acoustic properties and, consequently,
different intensity and shape of the reflected ultrasonic wave; furthermore, the
contribution of the scattered component is also different for each tissue. As a
result, in the IVUS image, areas corresponding to different tissues exhibit different
grey-level intensity and textures.
Most of the proposed plaque characterization methods are formulated as pattern
recognition problems. For this reason, two main steps can be identified in each
method: (1) the definition and extraction of discriminative features, describing the
properties of different tissues; (2) the training of a classifier, used to characterize
Coronary Atherosclerotic Plaque Characterization By Intravascular Ultrasound 181
RF
Signals
Fig. 4 Schematic example of IVUS image formation from RF signals
TGC BP EV Log
IVUS
image
each position of the IVUS frame as belonging to one of the known plaque types. In
this section the features used in the state-of-the-art methods are first presented, then
the most common classification techniques used in the plaque characterization field
are described.
3.1 Features Extraction
In the state-of-the-art plaque characterization methods, two main groups can be
recognized: the methods based on image analysis and the methods based on radio
frequency analysis. In this section, the features used in the two approaches are
presented.
3.1.1 Image-Based Features
The first plaque characterization method based on the analysis of the texture in
the IVUS image has been presented in [12]. The features used in this approach
are: (1) radial profile,(2)long-run emphasis,and(3)fractal dimension. Despite
of the high accuracy achieved, the main problem of using the IVUS images
obtained with the IVUS equipment is the lack of normalization in data they provide.
In order to improve the tissue visualization, physicians are used to change the
imaging parameters of the IVUS equipment, thus making the set of images and,
consequently, image-based features, not comparable. In order to avoid this problem,
the raw radio frequency signals captured by the IVUS equipment can be used to
reproduce the image formation process [14]. A common approach to IVUS image
formation from RF data foresees, at least, the following steps (see Fig. 4):
• Time Gain Compensation (TGC): as the US propagating in the tissue is af-
fected by an attenuation due to depth, it is necessary to compensate it by a
TGC function. One possible TGC function could be: T (r)=1 −e
α
β
= ln10
f /20
,αis the attenuation factorof the tissue measured in dB/MHz·cm,
−βr
,where
f is the frequency of the transducer in MHz,andr is the radial distance from the
catheter in cm.
• Band Pass filtering: data can be then filtered by a band-pass filter in order to
reduce the noise effect and spurious harmonic components outside the band of
interest. For this purpose, a Butterworth filter is suitable, given that its frequency
response is as flat as mathematically possible in the passband.
182 F. Ciompi et al.
• Envelope: after filtering, the envelope of the signal has to be recovered, in order
to change from bipolar to unipolar signal and to achieve the final conversion
between 0 and 255. This can be done for example by taking the absolute value of
the Hilbert transform of the signal.
• Logarithmic compression: This transformationmaps a narrow range of lowgray-
level values in the input image into a wider range of output levels; the following
log(1+(et−1)RF
formula can be used [25]: RF
log
=
t
RF data after the envelope computation, and RF
)
env
,whereRF
the output of the logarithmic
log
represents the
env
compression.
In [14] the reconstruction process is used to create a set of IVUS images of in
vivo cases. Textural features are then extracted by using: (1) co-occurrence matrix,
(2) local binary patterns,and(3)gabor filters.
3.1.2 Radio Frequency-Based Features
RF data from the unprocessed backscattered ultrasound signal provide an alternative to greyscale image analysis. Theoretically, the analysis of the IVUS-RF
data provides a more accurate and reproducible technique for measuring tissue
properties because it is not subject to machine-dependent processing, subsampling,
interpolation, quantization, and even operator-dependent settings [26].
The most important approaches based on RF data processing are now analyzed
by focusing on the analysis of the used features. Most of the methods process the
RF data frame, formed by N A-lines and M samples.
Acoustic Impedance (Z)
It is known that the US wave propagates with different velocity in different tissues;
this comes from the equation of the wave propagating through a tissue with density
ρ
and with acoustic impedanceζ. Then, the information on the relative acoustic
impedance can beused as a parameterto classify plaque types, since they are considered as different tissues. The method presented in [19] is based on this phenomenon.
It assumes that ultrasound pulse is well approximated by using the Plane Wave Born
Approximation (PWBA) deconvolved inverse scattering technique [27].
Integrated Backscatter (IB)
The first methodbased on IB wasproposed in 1989 and wasoriented to the detection
of acute myocardial infarction and re-perfusionvia M-mode echocardiography [28].
Coronary Atherosclerotic Plaque Characterization By Intravascular Ultrasound 183
Tabl e 1 Integrated
backscatter values obtained
by histological analysis of
post-mortem tissues. The
values are provided as mean
value ±1 dB. Data extracted
from [22]
Histology IB [dB]
Calcified tissue −30 < IB ≤−23
Mixed lesion −55 < IB ≤−30
Fibrotic tissue −63 < IB ≤−55
Lipidic core −73 < IB ≤−63
Thrombus −88 < IB ≤−80
The IB is an intrinsic parameter of the electrical US signal and can be computed
as [21]:
IB = 20log
1
T
1
T
T
2
V
∑
0
, (1)
T
2
V
∑
0
0
where V is the signal voltage from a selected Region of Interest (ROI) in the RF
data frame, V
is the smallest signal voltage that the system can detect, and T is the
0
integration interval.
It has been shown that the IB parameter combined with two-dimensional echo
can differentiatethe tissue characteristics in both in vivo and ex vivo studies [21,22].
In Table 1 the computed value of the IB parameter when characterizing tissues in in
vitro cases are presented when discriminating calcifications, mixed lesions, fibrous
tissue, lipidic core, and thrombus. Since no overlaps in the obtained values are
encountered, the discriminative power of this parameter is confirmed (see Table 1).
A plaque characterization system based only on IB parameter is now distributed
only in Japan (YD Co. Ltd, Tokyo).
Power Spectrum
Based on the hypothesis that different tissues behave differently in the frequency
domain, several approaches based on the analysis of power spectrum have been
proposed. Mainly, given the RF representation of an IVUS frame, for each ROI of
N
samples and Mrcontiguous A-lines, a power spectrum is associated to each point
r
by averaging the power spectra computed in the lines belonging to the ROI. The
power spectrum can be computed by means of the Fast Fourier Transform (FFT)
[15,18, 20] or by using the AutoRegressive Model (ARM) [14,16,17, 29].
Given the power spectrum, a set of spectral features can be extracted. Some
features have been successfully used in plaque characterization techniques, as
the mean power (dB), maximum power (dB), spectral slope (dB/MHz) (a leastsquares linear regression over the given bandwidth), y-axis intercept of spectral
slope (dB) (intercept of the straight line with the y-axis at 0 Hz) [15]. In addition,
the value of the frequencies corresponding to maximum and minimum power can
be used [16, 17]. In [16, 17] these features are extracted in a frequency range of
184 F. Ciompi et al.
17–42MHz, corresponding to the −20 dB bandwidth of the system. This set of
features, commonly known in literature as the seven-features approach, together
with the Integrated Backscatter (IB) represent the basis of a commercial product
called Virtual Histology (VH), implemented in the Volcano (Rancho Cordova,
CA) IVUS clinical scanners that offer near-realtime tissue characterization in
vivo [30].
Alternatively to the presented spectral features, a full-spectrum approach can
be considered, where the Fourier Transform is used [18, 20]. The use of a fullspectrum analysis demonstrated to produce more accurate results, compared with
the seven-features approach, classifying fibrolipidic, lipidic, fibrotic, and calcified
tissue: when using a 40MHz catheter in fact, a lot of variations can be found in the
main bandwidth of the signal (20–60 MHz), thus justifying a full-spectrum analysis
[18]. In the full-spectrum approach the power spectrum, with frequency range from
0 to 100MHz, is discretized using a certain number of bins and then used as feature
vector for each point (the center of a ROI) of the IVUS frame. This approach is
implemented in the iMap Plaque Characterization software provided with the iLab
IVUS equipment (Boston Scientific).
Wavelet-Based Approach
The wavelet analysis of RF signals has been also studied in recent years [24, 31].
A wavelet is mainly a waveform of limited duration and zero average amplitude
[32]. The wavelet analysis, applied to IVUS-RF signal, consists in the computation
of wavelet coefficients, localized by the amplitude and the position of wavelets.
The hypothesis in this kind of approach is that the IVUS-RF signal belonging
to different tissues exhibits different behavior when analyzed by wavelet, and that
the wavelet coefficients, computed at different scales, could be the discriminative
measures to be used in the tissue characterization [26].
This approach has been proved to be appropriate in discriminating the fibrous
from the fatty areas within atherosclerotic plaques [24].
Joining Textural and Spectral Features
Recently [29], an approach that proposes to combine both textural and spectral
features provided interesting results in the field of plaque characterization. The used
feature vector is constructed by concatenating image textural and spectral features
with the aim of blending the appearance of the tissues with information obtained
from their spectral content. Themain idea is thatimage-based andRF signal analysis
provide complementary data for the accurate description of tissue composition and
leads to a rich set of features to be selected and weighted by the classifier. The
approach leads to an overall accuracy A = 89.93% in classifying fibrotic, lipidic,
and calcified plaques in vitro [29].
Coronary Atherosclerotic Plaque Characterization By Intravascular Ultrasound 185
2000
1800
1600
3
1400
1200
feature
1000
800
600
1
0.8
Fig. 5 Feature space. A set of point acquired from four different tissue types are represented in
the space of three discriminant features
tissue1
tissue2
tissue3
tissue4
0.6
feature
2
0.4
0.2
20
15
10
5
0
0
feature
1
3.2 Tissue Classification
A classifier is a probabilistic model (discriminative or generative) with a set of
parameters determined (learned) during the training process. The model is in fact
able to learn the features of the input data, and to tune its parameters according
to data. When the training process is over, the set of learned parameter is finally
saved and the classifier is ready to be used to characterize unknown examples.
The atherosclerotic plaque characterization problem implies the definition of a
multi-class classifier. Except for some approaches that discriminate for example
fibrotic from fatty (lipidic) plaques, the main scenario implies in fact the discrimination among fibrotic, lipidic, and calcified plaques. Furthermore, the necrotic
core, fibrolipidic, and lipidic with calcifications tissues are also considered in some
approaches. The characterization problem consists in separating, in the feature
space, clouds of points belonging to multiple classes (Fig.5).
Several approaches have been presented to solve the multi-class tissue characterization problem: we will refer to them with the word “architecture,” since they often
consist in combining a set of simple classifiers into a more complex framework.
3.2.1 Decision Tree
One simple architecture for the discrimination of different tissues is the decision
tree [16,17]. Starting from the main node (root), during the training process, at each
node the input data are split into sub-groups according to the most discriminative
feature in that node.