Symptomatic Vs Asymptomatic Plaque Classification Biology Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Computer-aided diagnosis of carotid atherosclerosis into symptomatic or asymptomatic is useful in the analysis of cardiac health. . This paper describes a technique for symptomatic versus asymptomatic plaque classification of carotid ultrasound images. The classification system involves two steps (i) feature extraction using Discrete Wavelet Transform (DWT) and averaging algorithms and (ii) Support Vector Machine (SVM) classifier for automated decision making. Our results show that the proposed system is able to achieve a classification accuracy of 81.1%. Furthermore, we propose a novel Carotid Plaque Classification Index (CPCI) which is a single index that can be used to identify symptomatic and asymptomatic carotid plaque classes from carotid ultrasound images.

Key Words: Plaque; Carotid ultrasound; Discrete Wavelet Transform; Support Vector Machine (SVM); classification; Carotid Plaque Classification Index.


Stoke causes disruption of blood flow resulting in the blockage of oxygen supply to the part of the brain cells, and these cells will eventually begin to die.

In most cases, this disturbance is due to ischemia, which is caused by a blockage or leakage of blood (Sims and Muyderman 2009). Similarly life threatening are the symptoms of heart disease. In general, both heart disease and stroke are linked to atherosclerosis and high blood pressure (Labarthe 1998, Nagai et al. 2001). In this paper, we focus on atherosclerosis. The main symptom of atherosclerosis is artery thickening which is caused by the deposition of multiple plaques (Maton et al. 1993). Ultrasound and autopsy studies have shown the presence and extent of atherosclerotic lesions correlate with that of atherosclerosis elsewhere in the circulatory system, including coronary arteries (Solberg et al. 1968; Pancioli et al. 1998; Craven et al. 1990; Kallikazaros et al. 1999). Several studies have indicated that the presence of carotid stenosis (thickening of the carotid artery) is a strong predictor of death in the general population (Lerfeldt et al. 2002; Joakimsen et al. 1976).

Diagnosis systems can make a real difference, because treatment of atherosclerotic cardiovascular diseases is possible. For example, it has been shown that surgical removal of plaques reduced the risk of epsilateral stroke (European Carotid Surgery Trialists' Collaborative Group 1998; North American Symptomatic Carotid Endarterectomy Trial Collaborators 1991). However, the treatment success also depends on the quality of the diagnosis system, since not all carotid plaques are necessarily harmful and also carotid surgery carries a considerable risk for the patient. Carotid atherosclerosis detection is also important for other ways of treatment, apart from surgery. One important way of treatment was established after the completion of the human genome project. The genom identification led to the development of apolopoprotein A1 Milano which protects, to a certain extent, against vascular events in carriers of this disease (Chiesa and Sirtori 2003; Franceschini et al. 1985; Gualandri et al. 1985). The therapy, based on this protein, resulted in the regression of coronary atherosclerosis in a matter of weeks.

Atherosclerosis is a common disease therefore a diagnosis support system must be cost effective in order to keep the monetary burden on society as low as possible. In terms of cost effectiveness and availability, ultrasound imaging is a good choice for medical data acquisition. There is evidence that ultrasonographic B-mode characterization of plaque morphology may be useful in the assessment of the atherosclerotic lesion vulnerability (Gronholdt et al. 2001; AbuRahma et al. 2002; Sabetai et al. 2000). Despite the strong role of diagnostic ultrasound in cerebrovascular disease, significant deficits still exist. For example, a confident and reproducible classification of "dangerous" plaques is still not available. Furthermore, the correlation between ultrasonographic features and the histological evaluation of carotid plaques is often poor (Hatsukami et al. 1997; Droste et al. 1997). The limiting factors responsible for these deficits include low spatial resolution and ultrasound artifacts. Thus, improving ultrasonographic image quality, using adequate image pre-processing as well as the extraction of good features will improve the correlation with the histological evaluation of carotid plaques.

We have used state of the art image processing and classification techniques to deliver a reproducible classification of symptomatic and asymptomatic plaque. To demonstrate our ideas, we propose a low-cost non-invasive Computer-Aided Diagnosis (CAD) system which automatically classifies symptomatic and asymptomatic plaques. The system is based on B-mode ultrasound images from which features are extracted using Discrete Wavelet Transform (DWT). These extracted features were fed to the Support Vector Machine (SVM) classifier with a polynomial kernel of order 2 for automated decision making. We have also proposed a Carotid Plaque Classification Index (CPCI) which is a number to identify the symptomatic and asymptomatic carotid plaque classes. This CPCI can be used as an adjunct tool by the physicians during screening to cross check their diagnosis.

The paper is organized as follows: The Materials and Methods section briefly describes the data used, presents brief descriptions of the features that were extracted and the SVM classifier, and also lists the other statistical techniques used in the work. The Results section presents the values of the extracted features and the classification results. It also describes the CPCI index. Discussion is provided in the Discussion section, and the paper concludes in the Conclusion section.

Materials and Methods

Figure.1 shows the block diagram of the proposed system. The ultrasonic images of carotid plaque are pre-processed and subjected to feature extraction using DWT technique. Subsequently, the significant features are extracted using Student's 't' -test . Then the selected features are then fed to the SVM classifier for classification. These techniques are briefly described in this section.

Insert Figure 1 here

Ultrasonic images of carotid plaques

150 asymptomatic and 196 symptomatic (total=346) carotid plaque ultrasound images were used in this work. These images were collected from patients referred to the Vascular Screening Diagnostic Center, Nicosia, Cyprus, for diagnostic carotid ultrasound to detect both presence and severity of internal carotid stenosis. Informed consent was obtained from the patients, and the approval from the Institutional Review Board was obtained prior to conducting the study. Patients with cardioembolic symptoms or distant symptoms (> six months) were not considered for this work. The plaque images were collected consecutively with elimination of plaques that produced less than 50% (European Carotid Surgery Trial) stenosis and emergency cases scanned after normal working hours as they were often performed by personnel not trained in plaque image capture as outlined in the methods. As a result, a database was created from images of symptomatic and asymptomatic plaques. The ultrasonographers that, performed the examinations and obtained the plaque images, knew the reason for referral, since they were performing routine diagnostic testing for the presence and grading of stenosis. However, for the purpose of studies such as the present, the database images were anonymized and the persons who subsequently did the image normalization and image analysis did not know whether the plaques were symptomatic or asymptomatic (Griffin et al. 2010). Image normalization was applied during duplex image acquisition as listed below.

(a) Dynamic range adoption;

(b) Frame averaging (persistence);

(c) The time gain compensation curve (TGC) was sloping through the tissues but was positioned vertically through the lumen of the vessel because the ultrasound beam was not attenuated as it passed through blood. This ensured that the adventitia of the anterior and posterior walls had similar brightness;

(d) Gain adjustment.

(e) Post processing with a linear transfer curve;

(f) The ultrasound beam was at 90° to the arterial wall;

(g) The minimum depth was used so that the plaque occupied a large part of the image;

(h) The probe was adjusted so that Adventitia adjacent to the plaque was clearly visible as a hyperechoic band that it could be used for normalization.

Those plaques associated with retinal or hemispheric symptoms (unstable plaques), such as stroke, Transient Ischemic Attack (TIA), and Amaurosis Fugax (AF) were categorized as symptomatic plaques (88 stroke, 70 TIA, and 38 AF). Asymptomatic plaques were truly asymptomatic if they had never been associated with symptoms in the past. As a pre-processing step, the region of interest (extracted by medical practitioner) was extracted from each of the studied images prior to feature extraction. Fig. 2(a) and Fig. 2(b) show the typical symptomatic and asymptomatic carotid images, respectively. Fig. 3(a) and Fig. 3(b) show the region of interest of symptomatic and asymptomatic carotid images.

Insert Figure 2 here

Insert Figure 3 here

For this work, the entire dataset was split into three equal parts (roughly). For classifier training, 242 images were used. The training dataset consisted of 137 images from symptomatic (26 AF, 49 TIA, and 62 stroke) and 105 images from asymptomatic. For testing, 104 images were used. The test dataset consisted of 59 images from symptomatic (12 AF, 21 TIA, and 26 stroke) and 45 images from asymptomatic.

Feature extraction

In this work, we used two dimensional (2D) DWT and averaging algorithms for feature extraction. We start the discussion of these methods by introducing the DWT, which analyzes one-dimensional (1D) signals. Later, the concept is extended to 2D signals. This is done by considering all degrees of freedom such 2D signals offer, we arrive at the definition of 2D DWT. Finally we define the averaging methods for the 2D DWT results which yield the feature vector elements.

The DWT transform of a signal is evaluated by sending it through a sequence of down-sampling high and low pass filters. The low pass filter is defined by the transfer function and the high pass filter is defined by the transfer function. The output of the high pass filtering is known as the detail coefficients. The following equation shows how these coefficients are obtained.


The output of the low pass filtering is known as the approximation coefficients. These coefficients are found by using the following equation.


The frequency resolution is further increased by cascading the two basic filter operations. To be specific, the output of the first level low pass filter is fed into the same low and high pass filter combination. The detailed coefficients are output at each level and they form the level coefficients. In general, each level halves the sample number and doubles the frequency resolution. Consequently, in the final level both detailed and approximation coefficients are obtained as level coefficients.

For 2D signals, the 2D DWT can be used. Our discussion focuses on Wavelet packets (WP) for images. These images are represented as an gray scale matrix where each element of the matrix represents the intensity of one pixel. All non-border pixels in, where and, have eight immediate neighboring pixels. These eight neighbors can be used to traverse through the matrix. However, changing the direction with which the matrix is traversed just inverts the sequence of pixels and the 2D DWT coefficients are the same. For example, the WP result is the same when the matrix is traversed from left to right as from right to left. Therefore, we are left with four possible directions, which are known as decomposition corresponding to 0° (horizontal, Dh), 90° (vertical, Dv) and 45° or 135° (diagonal, Dd) orientation. The implementation of this algorithm follows the block diagram shown in Fig. 4. The diagram shows the dimensional input image and the results for level 1. The results from level 1 were sufficient to obtain significant features.

Insert Figure 4 here

In this work, we have selected three different wavelet functions. Each of these wavelet functions has both a unique low pass filter transfer function g[n] and a unique high pass filter transfer function h[n]. Fig. 5 shows the transfer functions for the Biorthogonal 3.1 (bior3.1) family.

Insert Figure 5 here

The 1st level 2D DWT yields 4 result matrices, namely Dh1, Dv1, Dd1 and A1, whose elements are intensity values. Fig. 6 shows a schematic representation of these result matrixes. Unfortunately, these matrixes cannot be used for classification directly, because the number of elements is too high. Therefore, we defined two averaging methods which represent a result matrixes with just one number. The first method is used to extract average measures from 2D DWT result vectors.



The final averaging method uses averages not the intensity values as such; it averages the energy of the intensity values.


These three elements form the feature vector.

Insert Figure 6 here

Statistical tests

We have used t-test and Receiver Operating Characteristic (ROC) to analyze the feature extraction and classification results respectively. A t-test is known as a statistical hypothesis test which follows a Student's t distribution if the null hypothesis is supported. In our case, we assume a normal distribution of the data, but the scaling term in the test statistic is known. Therefore, it is replaced by an estimate based on the data. If the assumption holds, the test statistic follows a Student's t distribution. ROC tests are widely used in biomedical informatics. The ROC curve is a plot of sensitivity versus (1 - specificity) (Lasko et al., 2005). Sensitivity indicates the probability that a test result is positive when the disease is present and specificity denotes the number of negatives correctly identified as negatives. A good classifier will have an area closer to 1 (Downey et al. 1999; DeLeo 1993 ).

Support Vector Machine (SVM)

In SVM, a separating hyperplane that maximizes the margin between the input data classes which are plotted in the feature space is calculated. In order to determine the margin, two parallel hyperplanes are constructed, one on each side of the separating hyperplane with the help of the training data. Once the separating hyperplane is thus formed, the test data are classified using it. In the case of linearly non-separable data, the features are first mapped to a higher dimensional space using kernel functions, and then the hyperplane is determined.



The features selected were: energy, average horizontal and vertical DWT coefficients. These features were fed to the SVM (with polynomial kernel of order 2) classifier for automatic detection of the unknown class. Table 1 presents the features obtained using DWT with the bior3.1 wavelet. A p-value of less than 0.0001 indicates that the features are significant.

Insert Table 1 here

Classification results

In this work, we have compared the performance of 54 different wavelet functions and 15 classifiers. The mother wavelet families that were analyzed are reverse biorthogonal wavelet family (rbio), Daubechies wavelet (db), Biorthogonal 3.1 wavelet (bior), Coiflets, Symlets, Discrete Meyer (FIR approximation) (dmey) and Haar family. The different classifiers studied were Perceptron_VIM, Nearest_Neighbor, Store_Grabbag, Perceptron_Batch, Batch Variable Increment Perceptron, Reduced Coulomb Energy (RCE), Fixed Margin Perceptron, Regularized Discriminant Analysis (RDA), Batch Relaxation with Margin, Classification And Regression Tree (CART), Single-Sample Relaxation with Margin, Parzen, Probabilistic Neural Network (PNN), AdaBoost, and Support Vector Machine (SVM).

The biorthogonal (bior3.1) performed better compared to the other wavelet functions. SVM with polynomial of order 2 performed better among 15 classifiers used. Three-fold stratified cross validation method was used to test the classifiers. The data was split into three parts, two parts were used for training and the other one part was used for testing (i.e. training consists of 242 images and testing consists of 104 images each time). To test different sections of the test data each time, we have repeated the procedure three times. Our results show that the proposed system achieved accuracy of 81.1%, sensitivity of 80%, and specificity of 81.9% for SVM with polynomial of order 2. These results are presented in Table 2. TN stands for the number of True Negatives, FN for False Negatives, TP for True Positives and FP for False Positives. Fig. 7 illustrates the ROC curves obtained for the various configurations of the SVM classifier.

Insert Table 2 here

Insert Figure 7 here

Carotid Plaque Classification Index

We have developed a Carotid Plaque Classification Index (CPCI) using the features listed earlier (Equation 3-5). The range of this index for the two classes is shown in Table 3. The box plot depicting these ranges is shown in Fig. 8. It is evident that the CPCI has an almost distinct range for the two classes. This index may be used by the clinicians for a more objective classification of plaques as it is only a single number and requires no subjective prediction.


Insert Table 3 here

Insert Figure 8 here


The classification of carotid plaque is a difficult and multifaceted problem. In this discussion section, we present published results from related studies, which were conducted recently,, and compare them to the results documented in this work. A scale/frequency approach, based on the wavelet transform, was used in an attempt to characterize carotid atherosclerotic plaque from B-mode ultrasound (Tsiaparas et al. 2009). Two wavelet decomposition schemes, namely the DWT and wavelet packets (WP), and three basis functions, namely Haar, symlet3 and biorthogonal3.1, were investigated in terms of their ability to discriminate between symptomatic and asymptomatic cases. A total of 12 detail sub-images were extracted using the DWT and 255 using the WP decomposition schemes. It was shown that WP analysis by the use of Haar filter and the 1-1 norm as texture descriptor could reveal differences not only in high but also in low frequencies, and therefore characterize efficiently the atheromatous tissue. They concluded that additional studies applying and further extending the above methodology were required to ensure the usefulness of wavelet-based texture analysis of carotid atherosclerosis.

Normalized pattern spectra computed for both a structural, multilevel binary morphological model were used as classification features with two different classifiers, the Probabilistic Neural Network (PNN) and the Support Vector Machine (SVM) for automated diagnosis of symptomatic and asymptomatic classes (Kyriacou et al. 2009). Using SVM classifier, they achieved classification efficiency of 73.7% for multilevel binary morphological image analysis and 66.8% for gray scale morphological analysis.

The texture and motion patterns of carotid atherosclerosis features were selected and fed fuzzy c-means classifier for classification in to symptomatic and asymptomatic plaques(Stoitis et al., 2004). Their algorithm was able to classify correctly 74% of plaques based on texture features only, and 79% of plaques based on motion features only. Classification performance reached 84% using combination of motion and texture features.

The bootstrap method was used to compare the mean values of the frequency-based texture features extracted from asymptomatic and symptomatic plaque images (Stoitsis et al., 2006). The application of bootstrapping in the small group of images were able to extract the discriminatory value of the texture features. They also showed that, Gabor filter- based texture analysis in combination with bootstrapping, may provide valuable information to discriminate the symptomatic and asymptomatic plaque tissue type (Stoitsis et al, 2009).

The texture features coupled with SVM classifiers were used for the automated identification of symptomatic and asymptomatic plaque images (Acharya et al., 2011). They obtained the highest classification accuracy of 82.4% for SVM with radial basis function kernel.

In this paper, a CAD system based on DWT was designed to analyze carotid plaque ultrasound images in order to automatically classify them into symptomatic or asymptomatic classes. Our proposed method resulted in accuracy, sensitivity, and specificity of 81.1%, 80%, and 81.9%, respectively. The accuracy of our work is comparable to the other previous studies. It can be further improved by taking other feature like texture parameters inaddition to the DWT fatures and better classiifers. Moreover, we have proposed a novel Carotid Plaque Classification Index (CPCI) which is a single number that effectively separates the two classes.


It is very difficult to diagnose the symptomatic and asymptomatic plaques from ultrasound images using image processing. . Currently, only experienced physicians or vascular ultrasonographers can detect these differences during ultrasound scans. In this paper, we proposed an automated classification system which overcomes some of these problems. The proposed system allows a) objective characterization of carotid atheromatous plaque based on echogenic measures and b) automatic classification into symptomatic or asymptomatic classes. The proposed system may be a valuable tool in modern clinical practice, since it can provide decision support with regard to carotid plaque treatment. The system we propose rests on DWT for feature extraction. It is able to diagnose the two classes automatically with an accuracy, sensitivity and specificity of more than 80%. Carotid Plaque Classification Index (CPCI) has been proposed to identify the symptomatic and asymptomatic carotid plaque ultrasound images to support the diagnosis using a single number immediately. Using DWT features, we have achieved a classification accuracy of 81.1%, sensitivity of 80% and specificity of 81.9%. These accuracies may not be sufficient enough for the system to be incorporated into routine clinical work flow. More research is needed to improve the classification results. Since we have tested a wide range of classifiers in this study, future work would include studying more feature extraction techniques in order to improve the accuracy. Specifically, we intend to study texture based methods and the combination of various feature extraction methods in our future studies.