Phương pháp chẩn đoán hình ảnh (Phần 4)

Trang 1

3 Texture and Morphological Analysis

of Ultrasound Images of the Carotid Plaque for the Assessment of Stroke

Christodoulos I Christodoulou, Constantinos S Pattichis, Efthyvoulos Kyriacou, Marios S Pattichis, Marios Pantziaris, and Andrew Nicolaides

CONTENTS

(SGLDM)

(NGTDM)

3.3.3.10 Morphological Features

2089_book.fm Page 87 Tuesday, May 10, 2005 3:38 PM

Trang 2

88 Medical Image Analysis

AcknowledgmentReferences

The objective of the work described in this chapter was to develop a aided system based on a neural network and statistical pattern recognition techniquesthat will facilitate the automated characterization of atherosclerotic carotid plaques,recorded from high-resolution ultrasound images (duplex scanning and color ﬂowimaging), using texture and morphological features extracted from the plaqueimages The developed system should be able to automatically classify a plaque into(a) symptomatic (because it is associated with ipsilateral hemispheric symptoms)2089_book.fm Page 88 Tuesday, May 10, 2005 3:38 PM

Trang 3

computer-Texture and Morphological Analysis of Ultrasound Images 89

and (b) asymptomatic (because it is not associated with ipsilateral hemisphericevents)

As shown in this chapter, it is possible to identify a group of patients at risk ofstroke based on texture features extracted from high-resolution ultrasound images ofcarotid plaques The computer-aided classification of carotid plaques will contributetoward a more standardized and accurate methodology for the assessment of carotidplaques This will greatly enhance the significance of noninvasive cerebrovasculartests in the identification of patients at risk of stroke It is anticipated that the systemwill also contribute toward the advancement of the quality of life and efficiency ofhealth care

An introduction to ultrasound vascular imaging is presented in Subsection 3.1.1,followed by a brief survey of previous work on the characterization of carotid plaque

In Section 3.2, the materials used to train and evaluate the system are described InSection 3.3, the modules of the multifeature, multiclassifiercarotid-plaque classifi-cation system are presented Image acquisition and standardization are covered inSubsection 3.3.1, and the plaque identification and segmentation module is described

in Subsection 3.3.2 Subsections 3.3.3 and 3.3.4 outline, respectively, the featureextraction and feature selection The plaque-classification module with its associatedcalculations of confidence measures is presented in Subsection 3.3.5, and the clas-sifier combiner is described in Subsection 3.3.6 In the following Sections 3.4 and3.5 the results are presented and discussed, and the conclusions are given in Section3.6 Finally, in the appendix at the end of the chapter, the implementation detailsare given for the algorithms used to extract texture features

3.1.1 U LTRASOUND V ASCULAR I MAGING

The use of ultrasound in vascular imaging became very popular because of its ability

to visualize body tissue and vessels in a noninvasive and harmless way and tovisualize in real time the arterial lumen and wall, something that is not possible withany other imaging technique B-mode ultrasound imaging can be used to visualizearteries repeatedly from the same subject to monitor the development of atheroscle-rosis Monitoring of the arterial characteristics like the vessel lumen diameter, theintima media thickness (IMT) of the near and far wall, and the morphology ofatherosclerotic plaque are very important in assessing the severity of atherosclerosisand evaluating its progression [7]

The arterial wall changes that can be easily detected with ultrasound are the endresult of all risk factors (exogenous, endogenous, and genetic), known and unknown,and are better predictors of risk than any combination of conventional risk factors.Extracranial atherosclerotic disease, known also as atherosclerotic disease of thecarotid bifurcation, has two main clinical manifestations: (a) asymptomatic bruitsand (b) cerebrovascular syndromes such as amaurosis fugax, transient ischemicattacks (TIA), or stroke, which are often the result of plaque erosion or rupture, withsubsequent thrombosis producing occlusion or embolization [8, 9]

Carotid plaque is deﬁned as a localized thickening involving the intima andmedia in the bulb, internal carotid, external carotid, or common femoral arteries

Trang 4

FIGURE 3.1 (Color figure follows p 274.) (a) An ultrasound B-scan image of the carotid artery bifurcation with the atherosclerotic plaque outlined; (b) the corresponding color image

of blood ﬂow through the carotid artery, which physicians use to identify the exact plaque region.

RT PROX ICA

Trang 5

Texture and Morphological Analysis of Ultrasound Images 91

thrombolytic therapy, plaque pathology, coagulation studies, and more recently,molecular biology have implicated atherosclerotic plaque rupture as a key mecha-nism responsible for the development of cerebrovascular events [10–12] Athero-sclerotic plaque rapture is strongly related to the morphology of the plaque [13].The development and continuing technical improvement of noninvasive, high-reso-lution vascular ultrasound enables the study of the presence of plaques, their rate ofprogression or regression, and most importantly, their consistency The ultrasoniccharacteristics of unstable (vulnerable) plaques have been determined [14, 15], andpopulations or individuals at increased risk for cardiovascular events can now beidentiﬁed [16] In addition, high-resolution ultrasound facilitates the identiﬁcation

of the different ultrasonic characteristics of unstable carotid plaques associated withamaurosis fugax, TIAs, stroke, and different patterns of computed tomography (CT)brain infarction [14, 15] This information has provided new insight into the patho-physiology of the different clinical manifestations of extracranial atheroscleroticcerebrovascular disease using noninvasive methods

Different classifications have been proposed in the literature for the ization of atherosclerotic plaque morphology, resulting in considerable confusion.For example, plaques containing medium- to high-level uniform echoes were clas-sified as homogeneous by Reilly [17] and correspond closely to Johnson’s [18] denseand calcified plaques, to Gray-Weale’s [19] type 3 and 4, and to Widder’s [20] type

character-I and character-Icharacter-I plaques (i.e., echogenic or hyperechoic) A recent consensus on carotidplaque characterization has suggested that echodensity should reﬂect the overallbrightness of the plaque, with the term “hypoechoic” referring to echolucent plaques[21] The reference structure to which plaque echodensity should be compared with

is blood for hypoechoic plaques, the sternomastoid muscle for the isoechoic, andthe bone of the adjacent cervical vertebrae for the hyperechoic ones

3.1.2 P REVIOUS W ORK ON THE C HARACTERIZATION OF C AROTID P LAQUE

There are a number of studies trying to associate the morphological characteristics

of the carotid plaques as shown in the ultrasound images with cerebrovascularsymptoms A brief survey of these studies is given below

Salonen and Salonen [3], in an observational study of atherosclerotic sion, investigated the predictive value of ultrasound imaging They associated ultra-sound observations with clinical endpoints, risk factors for common carotid andfemoral atherosclerosis, and predictors of progression of common carotid athero-sclerosis On the basis of their ﬁndings, the assessment of common carotid athero-sclerosis using B-mode ultrasound imaging appears to be a feasible, reliable, valid,and cost-effective method

progres-Geroulakos et al [2] tested the hypothesis that the ultrasonic characteristics ofcarotid artery plaques are closely related to symptoms and that the plaque structuremay be an important factor in producing stroke, perhaps more than the degree ofstenosis In their work, they manually characterized carotid plaques into four ultra-sonic types: echolucent, predominantly echolucent, predominantly echogenic, andechogenic An association was found of echolucent plaques with symptoms andcerebral infarctions, which provided further evidence that echolucent plaques areunstable and tend to form embolisms

Trang 6

El-Barghouty et al [4], in a study with 94 plaques, reported an associationbetween carotid plaque echolucency and the incidence of cerebral computed tomog-raphy (CT) brain infarctions The gray-scale median (GSM) of the ultrasound plaqueimage was used for the characterization of plaques as echolucent (GSM ≤ 32) andechogenic (GSM > 32)

Iannuzzi et al [22] analyzed 242 stroke and 336 transient ischemic attack (TIA)patients and identiﬁed signiﬁcant relationships between carotid artery ultrasound plaquecharacteristics and ischemic cerebrovascular events The results suggested that thefeatures more strongly associated with stroke were either the occlusion of the ipsilateralcarotid artery or wider lesions and smaller minimum residual lumen diameter Thefeatures that were more consistently associated with TIAs included low echogenicity

of carotid plaques, thicker plaques, and the presence of longitudinal motion

Wilhjelm et al [23], in a study with 52 patients scheduled for endarterectomy,presented a quantitative comparison between subjective classiﬁcation of the ultra-sound images, ﬁrst- and second-order statistical features, and a histological analysis

of the surgically removed plaque Some correlation was found between the threetypes of information, where the best-performing feature was found to be the contrast.Polak et al [5] studied 4886 individuals who were followed up for an average

of 3.3 years They found that hypoechoic carotid plaques, as seen on ultrasoundimages of the carotid arteries, were associated with increased risk of stroke Theplaques were manually categorized as hypoechoic, isoechoic, or hyperechoic byindependent readers Polak et al also suggested that the subjective grading of theplaque characteristics might be improved by the use of quantitative methods.Elatrozy et al [24] examined 96 plaques (25 symptomatic and 71 asymptomatic)with more than 50% internal carotid artery stenosis They reported that plaques withGSM < 40, or with a percentage of echolucent pixels greater than 50%, were goodpredictors of ipsilateral hemispheric symptoms related to carotid plaques Echolucentpixels were deﬁned as pixels with gray-level values below 40

Furthermore, Tegos et al [25], in a study with 80 plaques, reported a relationshipbetween microemboli detection and carotid plaques having dark morphological char-acteristics on ultrasound images (echolucent plaques) Plaques were characterized usingﬁrst-order statistics and the gray-scale median of the ultrasound plaque image.AbuRahma et al [6], in a study with 2460 carotid arteries, correlated ultrasoniccarotid plaque morphology with the degreeof carotid stenosis As reported, thehigher the degree of carotid stenosis,the more likely it is to be associated withultrasonic heterogeneousplaque and cerebrovascular symptoms Heterogeneity ofthe plaquewas more positively correlated with symptoms than with any degreeofstenosis These ﬁndings suggest that plaque heterogeneityshould be considered inselecting patients for carotid endarterectomy

Asvestas et al [26], in a pilot study with 19 carotid plaques, indicated a icant difference of the fractal dimension between the symptomatic and asymptomaticgroups Moreover, the phase of the cardiac cycle (systole/diastole) during which thefractal dimension was estimated had no systematic effect on the calculations Thisstudy suggests that the fractal dimension, estimated by the proposed method, could

signif-be used as a single determinant for the discrimination of symptomatic and tomatic subjects

asymp-2089_book.fm Page 92 Tuesday, May 10, 2005 3:38 PM

Trang 7

In most of these studies, the characteristics of the plaques were usually tively defined or defined using simple statistical measures, and the association withsymptoms was established through simple statistical analysis In the work we areabout to describe in this chapter, a large number of texture and morphologicalfeatures were extracted from the plaque ultrasound image and were analyzed usingmultifeature, multiclassifier methodology

subjec-3.2 MATERIALS

A database of digital ultrasound images of carotid arteries was created such that foreach gray-tone image, there was also a color image indicating the blood flow The colorimages were necessary for the correct identification of the plaques as well as theiroutlines The carotid plaques were labeled as symptomatic after one of the followingthree symptoms was identified: stroke, transient ischemic attack, or amaurosis fugax.Two independent studies were conducted In the first study with Data Set 1, atotal of 230 cases (115 symptomatic and 115 asymptomatic) were selected Two sets

of data were formed at random: one for training the system and another for evaluatingits performance For training the system, 80 symptomatic and 80 asymptomaticplaques were used, whereas for evaluation of the system, the remaining 35 symp-tomatic and 35 asymptomatic plaques were used A bootstrapping procedure wasused to verify the correctness of the classiﬁcation results The system was trainedand evaluated using ﬁve different bootstrap sets, with each training set consisting

of 160 randomly selected plaques and the remaining 70 plaques used for evaluation

In the second study, where the morphology features were investigated, a newData Set 2 of 330 carotid plaque ultrasound images (194 asymptomatic and 136symptomatic) were analyzed For training the system, 90 asymptomatic and 90symptomatic plaques were used; for evaluation of the system, the remaining 104asymptomatic and 46 symptomatic plaques were used

3.3 THE CAROTID PLAQUE MULTIFEATURE,

MULTICLASSIFIER SYSTEM

The carotid plaque classification system was developed following a multifeature,multiclassifier pattern-recognition approach The modules of the system aredescribed in the following subsections and are illustrated in Figure 3.2. In the firstmodule, the carotid plaque ultrasound image was acquired using duplex scanning,and the gray level of the image was manually standardized using blood and adventitia

as reference In the second module, the plaque region was identiﬁed and manuallyoutlined by the expert physician In the feature-extraction module, ten different textureand shape feature sets (a total of 61 features) were extracted from the segmented plaqueimages of Data Set 1 using the following algorithms: statistical features (SF), spatialgray-level-dependence matrices (SGLDM), gray-level difference statistics (GLDS),neighborhood gray-tone-difference matrix (NGTDM), statistical-feature matrix(SFM), Laws’s texture energy measures (TEM), fractal dimension texture analysis(FDTA), Fourier power spectrum (FPS), and shape parameters

Trang 8

FIGURE 3.2 Flowchart of the carotid plaque multifeature, multiclassiﬁer classiﬁcation system (From Christodoulou, C.I et al., IEEE Trans.

Medical Imaging, 22, 902–912, 2003 With permission.)

Feature set 1

Feature set n

Feature set 2

Diagnosis:

• Symptomatic

• Asymptomatic

Weighting Factors

THE CAROTID PLAQUE MULTIFEATURE MULTICLASSIFIER SYSTEM

Feature Extraction and Selection

Classifier 1

Classifier 2

Classifier n

Classifier Combiner

Image Acquisition and

Standardization

Plaque Identification &

Segmentation

Carotid plaque segmented image

• Identification and outline

of the plaque region manually by the human expert

Texture features using:

• Statistical Features

• Spatial Gray Level Dependence Matrices

• Gray Level Diff Statistics

• Neighborhood Gray Tone Difference Matrix

• Statistical Feature Matrix

• Laws Text Energy Meas.

• Fractal Dimension Texture

• Fourier Power Spectrum

• Shape Parameters

• Modular classifier system using the neural SOM classifier

multi-• Modular classifier system using the statistical KNN classifier

multi-• Combining using majority voting

• Combining by averaging the confidence measures

Trang 9

Following the feature extraction, several feature-selection techniques were used

to select the features with the greatest discriminatory power For the classiﬁcation,

a modular neural network using the unsupervised self-organizing feature map (SOM)classiﬁer was implemented The plaques were classiﬁed into two types: symptomatic

or asymptomatic For each feature set, an SOM classifier was trained, and tendifferent classification results were obtained Finally, in the system combiner, theten classification results were combined using: (a) majority voting and (b) weightedaveraging of the ten classification results based on a confidence measure derivedfrom the SOM For the sake of comparison, the above-described modular systemwas also implemented using the KNN statistical classifier instead of the SOM

3.3.1 I MAGE A CQUISITION AND S TANDARDIZATION

The protocols suggested by the ACSRS (asymptomatic carotid stenosis at risk ofstroke) project [1] were followed for the acquisition and quantiﬁcation of the imagingdata The ultrasound images were collected at the Irvine Laboratory for Cardiovas-cular Investigation and Research, Saint Mary’s Hospital, U.K., by two ultrasonog-raphers using an ATL (model HDI 3000, Advanced Technology Laboratories, Leich-worth, U.K.) duplex scanner with a 4- to 7-MHz multifrequency probe Longitudinalscans were performed using duplex scanning and color ﬂow imaging [27] B-modescan settings were adjusted so that the maximum dynamic range was used with alinear postprocessing curve The position of the probe was adjusted so that theultrasonic beam was vertical to the artery wall The time gain compensation (TGC)curve was adjusted (gently sloping) to produce uniform intensity of echoes on thescreen, but it was vertical in the lumen of the artery, where attenuation in blood wasminimal, so that echogenicity of the far wall was the same as that of the near wall.The overall gain was set so that the appearance of the plaque was assessed to beoptimal and noise appeared within the lumen It was then decreased so that at leastsome areas in the lumen appeared to be free of noise (black) The resolution of theimages was on the order of 700 × 500 pixels, and the average size and standarddeviation of the segmented images was on the order of 350 ± 100 × 100 ± 30 pixels.The scale of the gray level of the images was in the range from 0 to 255 Theimages were standardized manually by adjusting the image so that the median gray-level value of the blood was between 15 and 20 and the median gray-level value ofthe adventitia (artery wall) was between 180 and 200 [27] The image was linearlyadjusted between the two reference points, blood and adventitia This standardizationusing blood and adventitia as reference points was necessary to extract comparableresults when processing images obtained by different operators and equipment andvascular imaging laboratories

3.3.2 P LAQUE I DENTIFICATION AND S EGMENTATION

The plaque identification and segmentation tasks are quite difficult and were carriedout manually by the expert physician The main difficulties are due to the fact thatthe plaque cannot be distinguished from the adventitia based on brightness leveldifference, or using only texture features, or other measures Also, calcification and2089_book.fm Page 95 Tuesday, May 10, 2005 3:38 PM

Trang 10

acoustic shadows make the problem more complex The identiﬁcation and outlining

of the plaque were facilitated using a color image indicating the blood ﬂow (seeFigure 3.1) All plaque images used in this study were outlined using their corre-sponding color blood ﬂow images This guaranteed that the plaque was correctlyoutlined, which was essential for extracting texture features characterizing the plaquecorrectly

The procedure for carrying out the segmentation process was established by ateam of experts and was documented in the ACSRS project protocol [1] Thecorrectness of the work carried out by the single expert was monitored and veriﬁed

by at least one other expert However, the extracted texture features depend on thewhole of the plaque area and are not signiﬁcantly affected if a small portion of theplaque area is not included in the region of interest

and the corresponding color blood ﬂow image Figure 3.3 illustrates a number ofexamples of symptomatic and asymptomatic plaques that were segmented by anexpert physician

FIGURE 3.3 Examples of segmented symptomatic and asymptomatic plaques Selected ture values are given for the following features: median (2), entropy (14), and coarseness (36) (The numbers in parentheses denote the serial feature number as listed in Table 3.1.)

tex-Symptomatic Plaques

Median = 19.60, Entropy = 5.51, Coars = 8.56 Median = 1.43, Entropy = 3.65, Coarseness = 4.96

Asymptomatic Plaques

Median = 58.92, Entropy = 7.93, Coars = 20.76 Median = 50.79, Entropy = 7.65, Coarseness = 44.30 2089_book.fm Page 96 Tuesday, May 10, 2005 3:38 PM

Trang 11

of an image [28] Visually, these spatial interrelationships and arrangements of theimage pixels are seen as variations in the intensity patterns or gray tones Therefore,texture features have to be derived from the gray tones of the image Although it iseasy for humans to recognize texture, it is quite a difﬁcult task to deﬁne texture sothat it can be interpreted by digital computers

In this work, ten different texture-features sets were extracted from the plaquesegments using the algorithms described in Appendix 3.1 Some of the extractedfeatures capture complementary textural properties However, features that werehighly dependent on or similar to features in other feature sets were identiﬁed throughstatistical analysis and eliminated The implementation details for the texture-fea-ture-extraction algorithms can be found in Appendix 3.1 at the end of the chapter.

3.3.3.1 Statistical Features (SF)

The following statistical features were computed [29]: (1) mean value, (2) medianvalue, (3) standard deviation, (4) skewness, and (5) kurtosis

3.3.3.2 Spatial Gray-Level-Dependence Matrices (SGLDM)

The spatial gray-level-dependence matrices as proposed by Haralick et al [30] are based

on the estimation of the second-order joint conditional probability density functionsthat two pixels (k,l) and (m,n) with distance d in direction speciﬁed by the angle θ haveintensities of gray-level i and gray-level j Based on the probability density functions,the following texture measures [30] were computed: (1) angular second moment, (2)contrast, (3) correlation, (4) sum of squares: variance, (5) inverse difference moment,(6) sum average, (7) sum variance, (8) sum entropy, (9) entropy, (10) difference variance,(11) difference entropy, and (12, 13) information measures of correlation

For a chosen distance d (in this work d = 1 was used, i.e., 3 × 3 matrices) andfor angles θ = 0°, 45°, 90°, and 135°, we computed four values for each of the 13texture measures In this work, the mean and the range of these four values werecomputed for each feature, and they were used as two different feature sets

3.3.3.3 Gray-Level Difference Statistics (GLDS)

The GLDS algorithm [31] uses ﬁrst-order statistics of local property values based

on absolute differences between pairs of gray levels or of average gray levels toextract the following texture measures: (1) contrast, (2) angular second moment, (3)entropy, and (4) mean These features were calculated for displacements δ = (0, 1),(1, 1), (1, 0), (1, −1), where δ≡ (∆x, ∆y), and their mean values were taken.2089_book.fm Page 97 Tuesday, May 10, 2005 3:38 PM

Trang 12

3.3.3.4 Neighborhood Gray-Tone-Difference Matrix (NGTDM)

Amadasun and King [28] proposed the neighborhood gray-tone-difference matrix

to extract textural features that correspond to visual properties of texture The

following features were extracted, for a neighborhood size of 3 × 3: (1) coarseness,

(2) contrast, (3) busyness, (4) complexity, and (5) strength

3.3.3.5 Statistical-Feature Matrix (SFM)

The statistical-feature matrix [32] measures the statistical properties of pixel pairs

at several distances within an image, which are used for statistical analysis Based

on the SFM, the following texture features were computed: (1) coarseness, (2)

contrast, (3) periodicity, and (4) roughness The constants Lr, Lc, which determine

the maximum intersample spacing distance, were set in this work to Lr = Lc = 4

3.3.3.6 Laws’s Texture Energy Measures (TEM)

For Laws’s TEM extraction [33, 34], vectors of length l = 7, L = (1, 6, 15, 20, 15,

6, 1), E = (−1, −4, −5, 0, 5, 4, 1), and S = (−1, −2, 1, 4, 1, −2, −1) were used, where

L performs local averaging, E acts as edge detector, and S acts as spot detector If

we multiply the column vectors of length l by row vectors of the same length, we

obtain Laws’s l×l masks In order to extract texture features from an image, these

masks are convoluted with the image, and the statistics (e.g., energy) of the resulting

image are used to describe texture The following texture features were extracted:

(1) LL, texture energy from LL kernel, (2) EE, texture energy from EE kernel, (3)

SS, texture energy from SS kernel, (4) LE, average texture energy from LE and EL

kernels, (5) ES, average texture energy from ES and SE kernels, and (6) LS, average

texture energy from LS and SL kernels

3.3.3.7 Fractal Dimension Texture Analysis (FDTA)

Mandelbrot [35] developed the fractional Brownian motion model to describe the

roughness of natural surfaces The Hurst coefﬁcient H(k) [34] was computed for

image resolutions k = 1, 2, 3, 4 A smooth surface is described by a large value of

the parameter H, whereas the reverse applies for a rough surface

3.3.3.8 Fourier Power Spectrum (FPS)

The radial sum and the angular sum of the discrete Fourier transform [31] were

computed to describe texture

3.3.3.9 Shape Parameters

The following shape parameters were calculated from the segmented plaque image:

(1) X-coordinate maximum length, (2) Y-coordinate maximum length, (3) area, (4)

perimeter, and (5) perimeter2/area

Trang 13

3.3.3.10 Morphological Features

Morphological image processing allows the detection of the presence of speciﬁc

patterns, called structural elements, at different scales The simplest structural

ele-ment for near-isotropic detection is the cross ‘+’ consisting of ﬁve image pixels

Using the cross ‘+’ as a structural element, pattern spectra were computed for each

plaque image as deﬁned in the literature [36–38] After computation, each pattern

spectrum was normalized

All features of the ten feature sets were normalized before use by subtracting their

mean values and dividing by their standard deviations

3.3.4 FEATURE SELECTION

The selection of features with the highest discriminatory power can reduce the

dimensionality of the input data and improve the classiﬁcation performance A

simple way to identify potentially good features is to compute the distance between

the two classes for each feature as

(3.1)

where m1 and m2 are the mean values, and σ1 and σ2 are the standard deviations of

the two classes [39] The best features are considered to be the ones with the greatest

distance The mean and standard deviation for all the plaques, as well as for the

symptomatic and asymptomatic groups, were computed, and the distance between

the two classes for each feature was calculated as described in Equation 3.1 The

features were ordered according to their interclass distance, and the features with

the greatest distance were selected to be used for the classiﬁcation

Another way to select features and reduce dimensionality is through principal

component analysis (PCA) [40] In PCA, the data set is represented by a reduced

number of uncorrelated features while retaining most of its information content

This is carried out by eliminating correlated components that contribute only a small

amount to the total variance in the data set In this study, the 61-feature vector was

reduced to nine transformed parameters by retaining only those components that

contributed more than 2% to the variance in the data set A new feature set comprising

the nine PCA parameters was used as input to the SOM and the KNN classiﬁers

3.3.5 PLAQUE CLASSIFICATION

Following the computer-aided feature extraction and selection, feature classiﬁcation

was implemented based on multifeature, multiclassiﬁer analysis The SOM classiﬁer

and the KNN classiﬁer were used to classify the carotid plaques into one of the

following two types:

dis= −

+

1 2 2 2

Trang 14

100 Medical Image Analysis

1 Symptomatic because of ipsilateral hemispheric symptoms

2 Asymptomatic because they were not connected with ipsilateral spheric events

hemi-The different features sets described in Subsection 3.3.3 were used as input tothe classiﬁer

3.3.5.1 Classification with the SOM Classifier

The SOM was chosen because it is an unsupervised learning algorithm where theinput patterns are freely distributed over the output-node matrix [41] The weightsare adapted without supervision in such a way that the density distribution of theinput data is preserved and represented on the output nodes This mapping of similarinput patterns to output nodes that are close to each other represents a discretization

of the input space, allowing a visualization of the distribution of the input data Theoutput nodes are usually ordered in a two-dimensional grid, and at the end of thetraining phase, the output nodes are labeled with the class of the majority of theinput patterns of the training set assigned to each node In the evaluation phase, aninput pattern is assigned to the output node with the weight vector closest to theinput vector, and it is said to belong to the class label of the winning output nodewhere it has been assigned

Beyond the classification result, a confidence measure was derived from theSOM classifier characterizing how reliable the classification result was The confi-dence measure was calculated based on the classes of the nearest neighbors on theself-organizing map For this purpose, the output nodes in a neighborhood windowcentered at the winning node were considered The confidence measure was com-puted for five different window sizes: 1 × 1, 3 × 3, 5 × 5, 7 × 7, and 9 × 9 For eachone of the ten feature sets, a different SOM classifier was trained The implemen-tation steps for calculating the confidence measure were as follows:

Step 1: Train the classifier An SOM classiﬁer is trained with the training set,

using as input one of the ten feature sets

Step 2: Label the nodes on the SOM Feed the training set to the SOM

classiﬁer again and label each output node on the SOM with the number

of the symptomatic or asymptomatic training input patterns assigned to it

Step 3: Apply the evaluation set In the evaluation phase, a new input pattern is

assigned to a winning output node The number of symptomatic and tomatic training input patterns assigned to each node in the given neighborhoodwindow (e.g., 1 × 1, …, 9 × 9) around the winning node are counted

asymp-Step 4: Compute the confidence measure and classify plaque Calculate the

confidence measure as the percentage of the majority of the training inputpatterns to the total number of the training input patterns in the given neigh-borhood window To set its range from 0 to 1 (0 = low confidence, 1 = highconfidence), the confidence measure is calculated more specifically as

conf = 2 (max{SN1, SN2}/(SN1 + SN2)) − 1 (3.2)

Trang 15

Texture and Morphological Analysis of Ultrasound Images 101

where SN m is the number of the input patterns in the neighborhood window

for the two classes m = {1, 2}:

(3.3)

where L is the number of the output nodes in the R × R neighborhood dow with L = R2 (e.g., L = 9 using a 3 × 3 window), and N mi is the number

win-of the training patterns win-of the class m assigned to the output node i W i =

1/(2 d i ), is a weighting factor based on the distance d i of the output node i

to the winning output node W i gives the output nodes close to the winningoutput node a greater weight than the ones farther away (e.g., in a 3 × 3

window, for the winning node W i = 1, for the four nodes perpendicular to

the winning node W i = 0.5 and for the four nodes diagonally located around

W i = 0.3536, etc) The evaluation input pattern was classiﬁed to the class

m of the SN m with the greatest value as symptomatic or asymptomatic

3.3.5.2 Classification with the KNN Classifier

For comparison reasons, the KNN classiﬁer was also used for the carotid plaque

classiﬁcation To classify a new pattern in the KNN algorithm, its k nearest neighbors

from the training set are identified The new pattern is classified to the most frequentclass among its neighbors based on a similarity measure that is usually the Euclideandistance In this work, the KNN carotid plaque classification system was imple-

mented for values of k = 1, 3, 5, 7, and 9, and it was tested using for input the ten

different feature sets In the case of the KNN, the conﬁdence measure was simply

computed as given in Equation 3.2 and Equation 3.3, with SN m being the number

of the nearest neighbors per class m.

3.3.6 C LASSIFIER C OMBINER

In the case of difﬁcult pattern-recognition problems, the combination of the outputs

of multiple classifiers, using for input multiple feature sets extracted from the rawdata, can improve the overall classification performance [42] In the case of noisydata or of a limited amount of data, different classifiers often provide differentgeneralizations by realizing different decision boundaries Also, different featuresets provide different representations of the input patterns containing different clas-sification information Selecting the best classifier or the best feature set is notnecessarily the ideal choice, because potentially valuable information contained inthe less successful feature sets or classifiers may not be taken into account Thecombination of the results of the different features and the different classifiersincreases the probability that the errors of the individual features or classifiers will

be compensated by the correct results of the rest Furthermore, according to Perrone[43], the performance of the combiner is never worse than the average of the

Trang 16

individual classifiers, but it is not necessarily better than the best classifier Also,the error variance of the final result is reduced, making the whole system morerobust and reliable The use of a confidence measure to establish the reliability ofthe classification result can further improve the overall performance by weightingthe individual classification results before combining

In this work, the usefulness of combining neural-network classifiers was tigated in the development of a decision-support system for the classification ofcarotid plaques Two multifeature modular networks, one using the SOM classifierand one using the KNN classifier, were implemented The first ten feature sets,described in Subsection 3.3.3, were extracted from the plaque ultrasound images ofData Set 1 and were inputted into ten SOM or KNN classifiers The ten classificationresults were combined using: (a) majority voting and (b) weighted averaging based

inves-on a cinves-onﬁdence measure

3.3.6.1 Majority Voting

In majority voting, the input plaque under evaluation was classiﬁed as symptomatic

or asymptomatic by the ten classiﬁers using as input the ten different feature sets.The plaque was assigned to the majority of the symptomatic or asymptomatic votes

of the ten classiﬁcation results obtained at the end of step 4 of the algorithm described

in Subsection 3.3.5 The diagnostic yield was computed for the ﬁve window sizes:

1 × 1, 3 × 3, 5 × 5, 7 × 7, and 9 × 9

3.3.6.2 Weighted Averaging Based on a Confidence Measure

In combining with the use of a confidence measure, the confidence measure wascomputed from the ten SOM classifiers, as given in Equation 3.2 When combining,the confidence measure decided the contribution of each feature set to the final result.The idea is that some feature sets may be more successful for specific regions ofthe input population The implementation steps for combining using weighted aver-aging were as follows:

Step 1: Assign negative confidence measure values to the symptomatic plaques If an input plaque pattern was classiﬁed as symptomatic, as given

in step 4 of the algorithm described in Subsection 3.3.5, then its conﬁdencemeasure is multiplied by −1, whereas the asymptomatic plaques retain theirpositive values

Step 2: Calculate the average confidence Calculate the average of the n

conﬁdence measures that is the ﬁnal output of the system combiner as

1

j n

confconf

Trang 17

The final output of the system combiner is the average confidence, , and itsvalues are ranging from −1 to 1 Values of close to zero mean low confidence

of the correctness of the final classification result, whereas values close to −1 or 1indicate a high confidence

In the case of the KNN classiﬁer the n classiﬁcation results were combined in

a similar way to that of the SOM classiﬁer, i.e., (a) with majority voting and (b) by

averaging of the n conﬁdence measures The algorithmic steps described in the

previous subsections for the SOM classiﬁer apply for the KNN classiﬁer as well

When averaging, the ﬁnal diagnostic yield was the average of the n conﬁdence measures obtained when using the n different feature sets.

3.4 RESULTS

3.4.1 FEATURE EXTRACTION AND SELECTION

In Data Set 1, a total of 230 (115 symptomatic and 115 asymptomatic) ultrasoundimages of carotid atherosclerotic plaques were examined Ten different texture-feature sets and shape parameters (a total of 61 features) were extracted from themanually segmented carotid plaque images as described in Subsection 3.3.3 [39, 44].The results obtained through the feature-selection techniques described in Sub-section 3.3.4 and the selected features with the highest discriminatory power aregiven in Table 3.1 [39] The mean and standard deviation for all the plaques, andfor the symptomatic and asymptomatic groups, were computed for each individualfeature Furthermore, the distance between the two classes was computed asdescribed in Subsection 3.3.4 in Equation 3.1, and the features were ordered accord-ing to their interclass distance The best features were the ones with the greatestdistance As shown in Table 3.1, for all features the distance was negative, whichmeans that the feature values of the two groups overlapped The high degree ofoverlap in all features makes the classiﬁcation task of the two groups difﬁcult.The best texture features, as tabulated in Table 3.1, were found to be: thecoarseness of NGTDM, with average and standard deviation values for the symp-tomatic plaques 9.3 ± 8.2 and for the asymptomatic plaques 21.4 ± 14.9; the range

of values of angular second moment of SGLDM with 0.0095 ± 0.0055 and 0.0050

± 0.0050 for the symptomatic and the asymptomatic plaques, respectively; and therange of values of entropy also of SGLDM with 0.28 ± 0.11 and 0.36 ± 0.11 forthe symptomatic and the asymptomatic plaques, respectively Features, from otherfeature sets that also performed well were: the median gray-level value (SF), withaverage values for the symptomatic plaques 15.7 ± 16.6 and for the asymptomaticplaques 29.4 ± 22.9; the fractal value H1, with 0.37 ± 0.08 and 0.42 ± 0.07 for thesymptomatic and the asymptomatic plaques, respectively; the roughness of SFM,with 2.39 ± 0.13 and 2.30 ± 0.10 for the symptomatic and the asymptomatic plaques,respectively; and the periodicity also of SFM, with 0.58 ± 0.08 and 0.62 ± 0.06 forthe symptomatic and the asymptomatic plaques, respectively

In general, texture in symptomatic plaques tends to be darker, with highercontrast, greater roughness, and with less local uniformity in image density andbeing less periodical In asymptomatic plaques, texture tends to be brighter, with

confconf

Trang 18

TABLE 3.1

Statistical Analysis of 61 Texture and Shape Features Computed from 230 (115 Symptomatic and 115

Asymptomatic) Ultrasound Images of Carotid Atherosclerotic Plaques of Data Set 1

Symptomatic Asymptomatic Distance

Spatial Gray-Level-Dependence Matrices (SGLDM): Mean Values

6 Angular second moment 0.1658 0.1866 0.0646 0.1201 0.456 11

9 Sum of squares: variance 1315.2 1081.3 1621.8 957.5 0.212 48

10 Inverse difference moment 0.4856 0.1827 0.3545 0.1613 0.538 6

X X

Trang 19

Spatial Gray-Level-Dependence Matrices (SGLDM): Range of Values

19 Angular second moment 0.0095 0.0055 0.0050 0.0050 0.611 2

22 Sum of squares: variance 42.06 30.97 29.77 15.68 0.354 28

23 Inverse difference moment 0.090 0.029 0.098 0.025 0.196 49

Trang 20

TABLE 3.1

Statistical Analysis of 61 Texture and Shape Features Computed from 230 (115 Symptomatic and 115

Asymptomatic) Ultrasound Images of Carotid Atherosclerotic Plaques of Data Set 1 (continued)

Symptomatic Asymptomatic Distance

45 LL: texture energy from LL kernel 113,786 57,837 139,232 53,432 0.323 37

46 EE: texture energy from EE kernel 1,045.3 534.0 1,090.4 489.9 0.062 57

47 SS: texture energy from SS kernel 131.82 64.53 110.14 53.64 0.258 41

48 LE: average texture energy from LE and EL kernels 8,369.1 3754.8 9,514.1 3,639.9 0.219 46

49 ES: average texture energy from ES and SE kernels 335.64 174.69 312.85 149.85 0.099 55

50 LS: average texture energy from LS and SL kernels 1,963.5 1,008.5 2,054.6 907.2 0.067 56

Fractal Dimension Texture Analysis (FDTA)

57 X-coord max length 349.24 110.89 354.27 95.92 0.034 60

X X

Trang 21

less contrast, greater smoothness, and with large areas with small gray-tone tions and being more periodical These results are in agreement with the originalassumption that smooth surface, echogenicity, and a homogeneous texture are char-acteristics of stable plaques, whereas irregular surface, echolucency, and a hetero-geneous texture are characteristics of potentially unstable plaques Table 3.2 gives

varia-a verbvaria-al interpretvaria-ation of the varia-arithmeticvaria-al vvaria-alues of some of the fevaria-atures from Tvaria-able3.1 for the symptomatic vs the asymptomatic plaques [39] Figure 3.4 illustratesseveral box plots of some of the best features as selected with Equation 3.1.Principal component analysis (PCA) was also used as a method for featureselection and dimensionality reduction [40] The 61-feature vector was reduced tonine transformed parameters by retaining only those components that contributedmore than 2% to the variance in the data set The nine PCA parameters were used

as a new feature set for classiﬁcation

In Data Set 2, where the usefulness of the morphological features was gated, a total of 330 ultrasound images of carotid atherosclerotic plaques wereanalyzed [45] The morphological algorithm extracted 98 features from the plaqueimages Using the entire pattern spectra for classiﬁcation yielded poor results UsingEquation 3.1, the number of features used was reduced to only ﬁve, which proved

investi-to yield satisfacinvesti-tory classification results The selected features represent the mostsignificant normalized pattern spectra components We determined that small fea-tures due to: , and (see Equation 3.60 in Appendix 3.1)yield the best results Table 3.3 shows the good performance of , which may besusceptible to noise However, it is also the feature that is most sensitive to turbulentflow effects around the carotid plaques Table 3.3 tabulates the statistics for the fiveselected morphological features for the two classes and their interclass distance ascomputed with Equation 3.1 Additionally, for Data Set 2, the SF, the SGLDM, andthe GLDS texture-feature sets were computed and compared with the morphologicalfeatures [45]

TABLE 3.2

Verbal Interpretation of Arithmetic Values of Some Features

from Table 3.1 for Symptomatic vs Asymptomatic Plaques

Texture Feature

Symptomatic Plaques Asymptomatic Plaques Value Interpretation Value Interpretation

Median gray scale Low Darker High Brighter

Contrast High More local variations

present in the image

Low Fewer local variations present

in the image Entropy Low Less local uniformity in

image density

High Image intensity in neighboring

pixels is more equal Roughness High More rough Low More smooth

Periodicity Low Less periodical High More periodical

Coarseness Low Less local uniformity in

intensity

High Large areas with small

gray-tone variations Fractals H1, H2 Low Rough texture surface High Smooth texture surface

P1,' '+,P2,' '+,P3,' '+,P− +4,' ' P− +5,' '

P1,' '+

Trang 22

3.4.2 C LASSIFICATION R ESULTS OF THE SOM C LASSIFIERS

For the classiﬁcation task, the unsupervised SOM classiﬁer was implemented with

a 10 × 10 output node architecture, and it was trained for 5000 learning epochs Fortraining the classifier, 80 symptomatic and 80 asymptomatic plaques were used,whereas for evaluation of the system, the remaining 35 symptomatic and 35 asymp-tomatic plaques were used To estimate more reliably the correctness of the classi-fication results, a bootstrapping procedure was followed The system was trainedand evaluated using five different bootstrap sets where, in each set, 160 differentplaques were selected at random for training, and the remaining 70 plaques wereused for evaluation The SOM classifier yielded a confidence measure (see Subsec-tion 3.3.5) on how reliable the classification result was, based on the number of thenearest neighbors on the self-organizing map Five different neighborhood windowswere tested: 1 × 1, 3 × 3, 5 × 5, 7 × 7, and 9 × 9 The confidence measure wascalculated using a weighting mask giving the output nodes nearest to the winningoutput node a greater weight than the ones farther away

FIGURE 3.4 Box plots of the features gray-scale median (2), entropy (14), and coarseness (36) for the symptomatic and asymptomatic plaques (The numbers in parentheses denote the

serial feature number as listed in Table 3.1.) The notched box shows the median, lower and upper quartiles, and conﬁdence interval around the median for each feature The dotted line connects the nearest observations within 1.5 of the interquartile range (IQR) of the lower and upper quartiles Crosses (+) indicate possible outliers with values beyond the ends of the 1.5

× IQR.

Trang 23

set of Data Set 1 [39] The best feature sets in average for all windows were: the SGLDM(range of values) with 65.3%, the TEM with 63.0%, followed by the NGTDM with62.2%, the SGLDM (mean values) with 61.7%, and the GLDS with 61.5% The worstfeature set was the shape parameters, with an average diagnostic yield of only 49.2%.The best SOM window sizes in average were the large ones 5 × 5, 7 × 7, and 9 × 9,with an average DY of about 65% The worst window size was the 1 × 1, with anaverage DY of only 43.3% As given in Table 3.4, the best individual DY was 70%,and it was obtained by the SGLDM (range of values) using a 5 × 5 neighborhoodwindow and by the NGTDM with a 9 × 9 window size Figure 3.5 illustrates thedistribution of 160 carotid plaques of the training set (80 symptomatic and 80 asymp-tomatic) on a 10 × 10 SOM using as input all the 61 features (* = symptomatic, o =asymptomatic) Similar plaques are assigned to neighboring SOM matrix nodes Thefigure demonstrates the overlap between the two classes and the difficulty of theproblem For comparison reasons, the diagnostic yield was computed using as a separatefeature set the first 15 best features that were selected through univariate selection, asdescribed in Subsection 3.3.4 using Equation 3.1 Using the first 15 best features yielded

an average DY for the ﬁve windows of 63.0%, with the highest DY of 68.5% obtainedwith the 7 × 7 window size This was better than the average success rate of theindividual feature sets but lower than the diagnostic yield of the best feature set, and itwas much worse than the overall success rate of the combiner

Furthermore, the 15 best features selected through multivariate selection werealso used for classiﬁcation The average diagnostic yield was poor (about 50%), and

it was much lower than the diagnostic yield obtained by the univariate selection.These results show the high degree of overlap between the two classes, demonstratingthe difﬁculty of using the search algorithms to identify feature combinations with

TABLE 3.3

Statistical Analysis of the Five Best Morphological Features Computed from

330 (194 Asymptomatic and 136 Symptomatic) Ultrasound Images of Carotid Plaques of Data Set 2

Symptomatic Plaques Asymptomatic Plaques Distance

Note: For each feature, the mean and standard deviation were computed for the asymptomatic group and

for the symptomatic group The distance between the symptomatic and the asymptomatic groups was computed as described in Equation 3.5.

+

1 2 2 2

X X

Trang 24

good class separability The nine parameters obtained through principal componentanalysis (PCA) were also used as input to the SOM classiﬁer The average diagnosticyield was about 64%, which was slightly better than the average DY of the best 15features obtained by the univariate feature selection but still much lower than thediagnostic yield obtained by combining the ten feature sets

In the second data set, where the usefulness of the morphological features wasinvestigated, 90 asymptomatic and 90 symptomatic plaques were used for trainingthe classiﬁer, whereas for evaluation of the system the remaining 104 asymptomaticand 46 symptomatic plaques were used [45] Table 3.5 tabulates the diagnostic yieldfor the SOM classiﬁer for the different feature sets and for different neighborhoodwindow sizes on the self-organizing map The highest diagnostic yield was 69.6%,and it was obtained with a 9 × 9 window size, using as input the GLDS feature set

On average, the results with the highest diagnostic yield were obtained by the GLDSfeature set, which was 64.6%, followed by the morphological feature set with adiagnostic yield of 62.9%, the SGLDM with 62.2%, and the SF with 59.9%

TABLE 3.4

Average Diagnostic Yield (DY) of the Self-Organizing Map (SOM) Classifier System for the Evaluation Set of Data Set 1 (35 Symptomatic and 35 Asymptomatic Plaques) of the Modular Neural Network Diagnostic System after Bootstrapping Available Data for Five Different Sets of Plaques

Diagnostic Yield (%) Window Size Feature Set 1 ×××× 1 3 ×××× 3 5 ×××× 5 7 ×××× 7 9 ×××× 9 Average

Note: DY is given for the ten feature sets, their average, and when combined using (a) majority voting and

(b) by averaging the ten confidence measures DY is also given for the first 15 best features as selected using Equation 3.5 DY was computed for five different SOM neighborhood windows: 1 × 1, 3 × 3, 5 × 5, 7 × 7, and 9 × 9.

Định dạng
Số trang	49
Dung lượng	5,03 MB