Summary of a portion of a paper

Knowledge-Based Systems 52 (2013) 236–245

Contents lists available at ScienceDirect

Knowledge-Based Systems

journal homepage: www.elsevier .com/locate /knosys

A low-cost screening method for the detection of the carotid artery diseases

0950-7051/$ - see front matter � 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.knosys.2013.08.007

⇑ Corresponding author. Tel.: +20 1001426485. E-mail addresses: [email protected] (A.F. Seddik), doaashaw-

[email protected] (D.M. Shawky).

Ahmed F. Seddik a, Doaa M. Shawky b,⇑ a Biomedical Engineering Department, Faculty of Engineering, Helwan University, Cairo, Egypt b Engineering Mathematics Department, Faculty of Engineering, Cairo University, 12613 Giza, Egypt

a r t i c l e i n f o

Article history: Received 24 February 2013 Received in revised form 26 July 2013 Accepted 2 August 2013 Available online 13 August 2013

Keywords: Automatic diagnosis Carotid artery diseases Doppler signal classification Artificial neural networks K-nearest neighbor

a b s t r a c t

Carotid artery diseases are defined as the narrowing or the blockage of the carotid arteries. These two conditions are called carotid artery stenosis or occlusion respectively. Stenosis and occlusion are usually caused by cholesterol deposits and fatty substances which are called plaque. In addition, they represent significant causes of strokes. Thus, they should be a part of regular physical examinations. An important and preliminary diagnosis is to listen to the arteries in the neck using a stethoscope or a Doppler ultra- sound (US) device. However, it is sometimes very difficult for a non-professional physician to differenti- ate between a normal and an abnormal sound due to blood flow blockage.

This paper presents a low-cost efficient method that can be used in the automatic screening of carotid artery diseases, especially in areas with high population. Doppler US signals are preprocessed for noise elimination. Then, some features for normal, stenosis and occlusion signals are extracted from the fre- quency domain of these signals using their spectrograms. A multi-layer feed forward neural-network (MLFFNN) and a k-nearest neighbor (KNN) classifiers were used to automatically diagnose the input sig- nals. The approach is applied to 72 samples divided into three equal sets which represent the three main classes to be identified, i.e., normal, stenosis and occlusion patterns. We used in the training phase 75% of each set and the rest was used in the test phase. Experimental results show the simplicity and efficiency of the presented approach for automatic diagnosis of carotid artery diseases. The maximum obtained classification accuracies are 91.67%, 100%, and 95.89% for the normal, stenosis and occlusion patterns respectively when the MLFFNN classifier is used. In comparison with similar approaches, the proposed approach is less complex, hence runs faster which suggests its suitability as an efficient screening method for the detection of carotid artery diseases.

� 2013 Elsevier B.V. All rights reserved.

1. Introduction

Vascular diseases are guided primarily by the history and phys- ical examination. However, there is a need for non-invasive inves- tigations to compensate the lack of expert physicians and the high cost diagnosis methods. This is especially true in rural areas and developing countries, also for countries of high populations.

Carotid artery diseases are among the most common vascular diseases. A Carotid artery disease occurs when the major arteries in the neck become narrowed (stenosis) or blocked (occlusion). This problem happens frequently as people age and it is a common cause to strokes [1]. There are many techniques for the diagnosis of carotid artery diseases. Doppler ultrasonography is one of the sim- plest and low-cost methods for detecting carotid artery diseases.

This paper presents an approach to diagnosing carotid artery stenosis and occlusion using Doppler ultrasound signals. A data set that contains sound waves for normal carotid artery, carotid ar- tery stenosis and occlusion is used. Some features are extracted from the spectrogram of Doppler US signals. Then, using these fea- tures we trained a feed-forward neural network to classify the sound waves into normal, stenosis or occlusion. The classifier has a total accuracy of 95.48%. In addition, a KNN classifier is used, however, the classification accuracy is 94%. The Doppler sound waves rather than images are usually collected at a much lower cost. In addition, only five features are used and yet, the obtained results are very satisfactory. Thus, the proposed approach is less complex and less expensive than similar works which utilize Doppler ultrasonography and image processing techniques. Since a good screening method should be fast, affordable and available to a large number of populations [2], the simplicity of the proposed approach suggests its suitability as an efficient screening method especially in developing countries and rural areas where it is infea-

A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245 237

sible to perform other diagnostic tests that need to be done by medical experts.

The rest of the paper is organized as follows. Section 2 intro- duces a background on some related concepts. In Section 3, the proposed approach is presented in detail. Section 4 presents the performed experimental study and discusses the obtained results. In Section 5, a survey on the related research is presented. Finally, Section 6 presents the conclusions and highlights some directions for the future work.

2. Carotid artery diseases: a background

As shown in Fig. 1, the carotid arteries are two large blood ves- sels that supply blood to the front part of the brain where major functions such as thinking, speech and motor functions reside [3]. Carotid artery disease is a disease in which a waxy substance called plaque builds up inside the carotid arteries. If plaque builds up in the arteries (the condition is called atherosclerosis) it hard- ens and narrows the arteries. This may limit the flow of oxygen- rich blood to the brain and affect major functionalities of the body. In some cases it may cause stroke which is a common cause of death all over the world, especially for people with age over 65 years [4].

Many diagnostic techniques are used to detect carotid artery stenosis at its early stages. These methods include CT scan and CT Angiography (CTA) which can show X-ray pictures of the arter- ies in the neck. Another method for diagnosis is to utilize magnetic resonance angiography (MRA) which uses radio waves and mag- netic fields to create detailed images. Some forms of this test can show moving blood flow and may help evaluate carotid artery dis- ease. To improve the test’s accuracy, physicians sometimes inject a material, called gadolinium, to make the arteries more visible [6]. Another dangerous test for diagnosing carotid artery stenosis is to inject a contrast dye through a catheter that is threaded into the arteries and then takes X-ray pictures. This test is called angi- ography. Angiography shows how blood flows through the arteries and whether they are narrowed. Moreover, carotid duplex ultra- sound can be used to determine if there is narrowing. Ultrasonog- raphy is a common method for detecting carotid artery diseases. In ultrasonography, a probe with two piezoelectric crystals is used. The transmitting crystal produces ultrasound at a fixed frequency and the receiving crystal vibrates in response to reflected waves and produces an output voltage. A two dimensional picture is built using ultrasound waves reflected from the tissues [7]. The simplest method for ultrasonography can be done by using a Doppler ultra- sound device. Ultrasound signals reflected off stationary surfaces retain the same frequency with which they were transmitted. However, the frequency of signals reflected from moving objects

Fig. 1. The carotid artery [5].

such as red blood cells shifts in proportion to the velocity of the target. The output from a continuous wave Doppler ultrasonogra- phy is usually presented as an audible signal. This signal is heard by an expert physician to be diagnosed. Although the previous diagnostic methods are efficient, however, there is a need for a sys- tem to aid the physician in detecting abnormality in the collected signals, especially for areas with large populations. Also, an auto- matic diagnosis is needed to compensate the lack of experts in some cases.

3. Proposed approach

The basic steps of the proposed approach are depicted in Fig. 2. The approach consists of the following main steps:

1. Reading the Doppler US files. 2. Pre-processing the analyzed signals. 3. Applying feature extraction techniques. 4. Classifying the analyzed signal in to normal, stenosis or

occlusion.

In the rest of this section, each step of the proposed approach will be explained in detail.

In the first step, as shown in Fig. 2, the files that contain Doppler US waves are read into Matlab software package [8]. In the second step, some preprocessing and noise removal is applied. This is done using high pass filters as Doppler signals are usually low-frequency signals. In the third step, the signal spectrogram is generated using Short-Time Fourier Transform (STFT). STFT is a sequence of fast Fourier transform (FFT) of windowed data segments, where the windows are usually allowed to overlap in time, typically by 25– 50%. STFT is a powerful general-purpose tool for audio signal pro- cessing [9]. The calculations of STFT can be done as follows [10]:

XmðxÞ ¼ X1

n¼�1 xðnÞxðn�mRÞe�jxn ð1Þ

where x(n) is the input signal at time n; m = 1, 2, . . ., N, where N is the total number of used segments. x(n) is a window function of length M; Xm(x) is the Discrete-Time Fourier Transform (DTFT) of windowed data centered about mR; And R is the hop size, in sam- ples between successive DTFTs.

The input signal x is divided into eight segments. If x cannot be divided exactly into eight segments, the rest which is at most se- ven samples is truncated. A Hamming window of length 256 is used. In addition, a hop size (R) of length 128 is used. After gener- ating the spectrogram of each signal, a set of features is extracted

Fig. 2. The basic steps of the proposed approach.

238 A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245

from the spectrogram of the analyzed Doppler signals. This set in- cludes the following five features:

� The maximum of the coefficients generated in the spectrogram of the signal (Max). � The minimum of the coefficients generated in the spectrogram

of the signal (Min). � The mean of the coefficients generated in the spectrogram of

the signal (Mean). � bRI which is defined as bRI = (Max �Min)/Min. � bPI which is defined as bPI = (Max �Min)/Mean.

The features bRI and bPI are related to the resistivity index (RI) and the pulsatility index (PI) respectively. These two indices are derived from Doppler sonograms [11] and they are reflections of the resistance to flow. PI is computed according to the method of Gosling and King [12] as follows:

PI ¼ ðVPS � VMDÞ=Vmean ð2Þ

where VPS is the blood flow velocity at peak systole, VMD is the blood flow velocity at maximum diastolic deflection, and Vmean is the velocity time averaged over the cardiac cycle. RI is calculated by the formula of Planiot et al. [13] as follows:

RI ¼ ðVPS � VEDÞ=VPS ð3Þ

where VED is the blood flow velocity at end-diastole. Although, as stated in [14], PI and RI values are not good predictors for Doppler waveforms, bPI and bRI which are analogous to PI and RI respectively have proven to be good features for Doppler spectrograms in the proposed approach.

After extracting the features, a classifier is trained to classify the signals which represent normal, stenosis and occlusion cases. Two different classifiers are used; a neural-network based classifier, and a KNN classifier. During the training phase, only 75% of the data is used. The other 25% of the data is used in the test phase.

In the last step of the proposed method, to assess the efficiency of the proposed approach, evaluation of the obtained results is done. Classification accuracy, sensitivity and specifity of the trained classifier are calculated. Accuracy is the ratio of correctly classified patterns over the total number of analyzed patterns. Meanwhile, sensitivity is the ratio of true positives over the total number of patterns representing abnormal patterns. The last used performance measure is the specifity which is the ratio of true neg- atives over the total number of patterns representing normal signals.

It should be mentioned that the suggested approach can be used as a screening method. Screening methods are different from diagnostic tests. For a screening method to be effective, it should be sensitive, specific, available to a large number of population, and acceptable in terms of cost, risk and patient tolerability [2,15–17]. Also, a screening method should be fast to be performed on a large number of patients. Thus, it should be simple and yet efficient.

4. Experimental study

4.1. Used data

The used data set includes 72 samples of complete cardiac cy- cles divided equally among the three classes; normal, stenosis and occlusion. The twenty four normal signals are equally divided between males and females. Meanwhile, the twenty four signals representing stenosis belong to 14 males and 10 females. Finally, the occlusion signals belong to 18 males and 6 females. The ages of the 72 subjects range from 56 to 73 years with an average of

66 years. This data set was collected in Helwan’s University medi- cal center using a pen probe Doppler US device at 4 MHz.

4.2. Results and discussion

We first perform some exploratory analysis to help us deter- mine whether the suggested features can be used as good predic- tors for the Normal, Stenosis and Occlusion classes or not. First, the wave forms of the three classes are plotted as shown in Figs. 3– 7. It is clear that the wave forms have different characteristics. Also, it should be mentioned that in the cases of stenosis or occlu- sion, the physician was able to distinguish whether the heard sig- nal is near or far from the location of the stenosis or occlusion. Thus, for Stenosis and Occlusion patterns, two wave forms are shown for each case since they produce different wave forms according to the location of the Doppler device. Although these dis- eases have different wave forms as shown in Figs. 4–7, however, the proposed classifiers are trained to detect the existence of the stenosis or occlusion irrespective of their exact location. Thus, in the evaluation step, we merged these patterns to obtain signals representing only the three main classes to be identified. The spec- trograms for the above wave forms are shown in Figs. 8–12.

Fig. 13 shows the mappings of the three the main classes to be identified using the five extracted features; bRI , bPI; Max, Min, and Mean respectively. It should be noticed that using the feature Min, for instance, a large number of patterns for the Normal and Occlusion classes are overlapped. Hence, the Min feature has a low separation power between these classes. On the other hand, the bRI feature has a better separation power as less number of pat- terns belonging to the different classes are overlapped.

To further assess the separation power of the used features, we performed t-test with significance level of 0.05 [18] for every pair of the three main classes using the five extracted features. The t- test examines the null hypothesis that the represented patterns using only feature i and feature j have the same mean against the hypothesis that the two means are different and hence the sep- aration power is higher. The lower the p-value, the better the sep- aration power is. Table 1 shows the obtained p-values of each different pair of the Normal, Stenosis, and Occlusion patterns. It should be noticed that the p-values are very small except for the Min feature for Normal–Occlusion pair; it is above the significance level. The same is true for the Mean feature for Normal–Stenosis pair. Otherwise, the zero or almost zero p-values for the other cases suggest the suitability of the extracted features as good predictors of the three classes to be identified.

4.2.1. Neural network based classifier The implemented network employs multi-layer feed-forward

back propagation (MLFBP) with one hidden layer of five neurons. The MLFFBP network is trained using the training set which con- tains 75% of the total dataset. In comparison to similar approaches e.g. [14], the training phase is very efficient (has low complexity) as it reaches the minimum mean square error (MSE) between the net- work output and the desired output in only 14 epochs. As shown in Fig. 14, after 14 epochs, the test error starts to increase due to over fitting, thus the training phase should be stopped at this point. Of course part of this is due to lower training set size. However, the low dimensionality of the features set, in addition to good features selections are two other important factors for the efficiency of any classification problem.

Table 2 presents the confusion matrix for the classification problem. As shown in the table, only three patterns are misclassi- fied. Also, the classification accuracies for Normal, Stenosis and Occlusion classes are 91.67%, 100%, and 95.83% respectively. In addition, the total classification accuracy is 95.48% which is a very

0 0.5 1 1.5 2 2.5 3 3.5 4 x 105

-0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

Samples

Fr eq

ue nc

y Sh

if t (

H z)

Fig. 3. The normal Doppler US wave form.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x 105

-0.3

-0.25

-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

0.2

Samples

F re

qu en

cy S

hi ft

( H

z)

Fig. 4. The over-stenosis Doppler US wave form.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 x 105

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

Samples

F re

qu en

cy S

hi ft

( H

z)

Fig. 5. The stenosis-distant Doppler US wave form.

A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245 239

promising accuracy for such a low-cost and a fast screening method.

Furthermore, as shown in the confusion matrix, the sensitivity is 97.9%. Meanwhile, the specifity is 91.7%.

4.2.2. KNN classifier KNN is a statistical-based classifier that is used to predict the

response of an observation using a non-parametric estimate of

the response distribution of its k nearest neighbors [19]. It is based on the assumption that the characteristics of members of the same class should be similar and thus, observations located close to- gether in covariate (statistical) space are members of the same class or at least have the same posterior distributions on their respective classes. The choice of k affects the performance of the KNN algorithm. To determine the nearest neighbors to a sample, the relative distance between instances is determined by using a

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x 105

-0.6

-0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

Samples

F re

qu en

cy S

hi ft

( H

z)

Fig. 6. The over-occlusion Doppler US wave form.

0 1 2 3 4 5 6 7 x 105

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Samples

F rq

ue nc

y Sh

if t

(h z)

Fig. 7. The occlusion-distant Doppler US wave form.

0 200 400 600 800 1000 1200 1400 1600 1800 2000 0

50

100

150

200

250

300

350

Frequency (Hz)

Fig. 8. The normal Doppler US spectrogram.

240 A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245

distance metric. We used KNN classifier with k equals to 3 and Euclidean distance metric to classify the analyzed patterns. A clas- sification accuracy of 94% is obtained with 94% and 83% for the sen- sitivity and the specifity respectively. Thus, for the studied problem, the neural network based classifier had better perfor- mance than the KNN classifier.

To study the effect of the sample size over the generalization power, we performed t-test using each feature with the null hypothesis that the samples come from normal distribution with population mean that is equal to the sample mean. The test fails to reject the null hypothesis for the five used features at a signifi- cance level of 0.05. Moreover, there is a 95% chance of having the

0 500 1000 1500 2000 2500 0

50

100

150

200

250

300

Frequency (Hz)

Fig. 9. The over-stenosis Doppler US spectrogram.

0 500 1000 1500 2000 2500 0

50

100

150

200

250

Frequency (Hz)

Fig. 10. The stenosis-distant Doppler US spectrogram.

0 500 1000 1500 2000 2500 0

100

200

300

400

500

600

Frequency (Hz)

Fig. 11. The over-occlusion Doppler US spectrogram.

A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245 241

population mean belonging to the confidence interval [364.37, 398.04], [364.36, 398.03], [365.42, 388.41], [0.09, 0.13] and [13.14, 15.69] for the features bRI , bPI Max, Min and Mean respec- tively. To evaluate to what degree the used samples represent the real population, we propose using the following measure:

Degree of Confidence ¼ 1� Length of CI Mean of CI

� � � 100% ð4Þ

where CI is the confidence interval generated by the t-test for each feature. Thus, we are confident by 91.17%, 91.18%, 91.27%, 60.85% and 82.33% that the features bRI , bPI Max, Min and Mean represent

the real population respectively. To obtain a narrower confidence interval, more data samples should be used. Hence, more confi- dence degrees can be obtained.

Table 3 shows our approach in comparison with similar ap- proaches. In order to be able to compare the complexity of the pre- sented approaches, we used the number of features which are used as the inputs to the classifier as an indicator. However, if it was not available, we used the number of the training epochs instead. In [21], the author has used the US Doppler signals that were acquired from mitral valve of the subjects in order to detect vascular heart diseases. The STFT of the sonograms of the obtained signals were

0 1000 2000 3000 4000 5000 6000 7000 8000 9000 0

1000

2000

3000

4000

5000

6000

Frequency (Hz)

Fig. 12. The occlusion-distant Doppler US spectrogram.

(b)(a)

(c) (d)

(e)

0 5 10 15 20 25 200

250

300

350

400

450

500

550

600 First Feature (Ri)

0 5 10 15 20 25 200

250

300

350

400

450

500

550

600 Second Feature (Pi)

0 5 10 15 20 25 0

5

10

15

20

25

30 Fifth Feature (Mean)

0 5 10 15 20 25200

250

300

350

400

450

500

550

600 Third Feature (Max)

0 5 10 15 20 250

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45 Fourth Feature (Min)

Fig. 13. Mappings of the data using the five features (a) bRI (b) bPI (c) Mean (d) Max (e) Min.

242 A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245

calculated and 100 sample points were used as inputs to neural network based classifier with two hidden layers consisting of 25 and 12 neurons. Thus, the objective of the work proposed in [21] was to detect heart diseases not the carotid artery diseases. Also, although the approach was based on a neural network classifier,

however the topology consists of 2 hidden layers (in contrast with the proposed network which consists of only 1 hidden layer) which adds extra complexity. Also, in [22], the authors employed US Doppler images rather than audio signals which is costly due to the needed expensive device. Moreover, the process is computa-

Table 1 The obtained p-values of the five features for different pairs of the three classes.

Class pair/feature Max Min Mean bPI bRI

Normal–Stenosis 0.0000 0.0000 0.0000 0.0002 0.0632 Normal–Occlusion 0.0000 0.0000 0.0000 0.3167 0.0000 Stenosis–Occlusion 0.0000 0.0000 0.0000 0.0008 0.0000

Fig. 14. The training phase curve.

Table 2 The confusion matrix of the used network.

Output/desired Normal Stenosis Occlusion

Normal 22 0 1 Stenosis 0 24 0 Occlusion 2 0 23

Table 3 A comparison between proposed approach and similar works.

Proposed approach

Uguz et al. [20]

Kara [21] Hassan et al. [22]

Detected classes

Normal, Stenosis, Occlusion

Normal, Stenosis

Normal, Mild Stenosis, Severe Stenosis

Normal, abnormal

Accuracy (%) 95.48 97.38 96.7 98.4 Sensitivity (%) 97.9 98.53 100 99.15 Specifity (%) 91.7 94.55 93.3 93.94 Input Signals Doppler US

audio Doppler US audio

Doppler US audio

Doppler US images

Complexity indicator

5 inputs extracted from US audio signal

Training in 500 epochs

100 inputs extracted from US audio signal

Min # of features used is 48 extracted from US images

A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245 243

tionally complex. Firstly, a preprocessing stage is performed. Sec- ondly, spatial, wavelets and gray level co-occurrence matrix fea- tures are extracted from carotid artery ultrasound images. Thirdly, redundant and less important features are removed from the features set using genetic search process. Finally, segmentation process is performed on the reduced features. Thus, although the proposed approach in [22] gives the best accuracy, it is very com- plicated which makes it inefficient as a screening method.

It should be noticed that the approach which utilizes Doppler US images rather than audio signals provided the highest perfor-

mance. However, it is usually very costly in terms of complexity and money as it requires Doppler ultrasonography imaging device followed by image processing techniques. This suggests the suit- ability of these approaches as diagnostic rather than screening tests. On the other hand, for the techniques which utilize the Dopp- ler US audio signals, the performance is slightly degraded. Further- more, as shown in Table 3, there is a slight sacrifice of performance of the proposed approach. However, this is compensated by the less complexity and hence the faster to be run test. This makes the proposed approach more suitable to be used for screening a large number of population followed by diagnostic tests for only positive subjects.

5. Related research

The literature contains a lot of research for the detection of car- otid artery diseases employing various techniques. For instance, in [23], the authors proposed a model for carotid artery stenting (CAS) devices using finite element analysis. However, as the authors mentioned, its results depend on the proper selection of patients and devices. Moreover in [24], the authors presented an approach for improving the ability of the identification of plaque that may produce strokes. The study combines some extracted tex- ture features form duplex images of 1121 subjects with asymp- tomatic internal carotid artery stenosis with clinical factors such as age, BMI, etc. In addition, some morphological features (dark, bright and medium-brightness regions) associated with plaque composition are also considered. In [25], a reference data for the common carotid artery is generated using the envelope waveforms of blood velocity of 202 healthy subjects. Moreover, in [26], data mining techniques were employed to explore the relationships among some factors such as hypertension, cardiac morbidity, smoking, diabetes, and physical inactivity that can be used in asymptomatic carotid stenosis. More specifically, genetic algo- rithms, logistic regression, and Chi-square tests have been applied to 372 patient subjects. Logistic regression yielded better results than genetic algorithms. Chen et al. [27], proposed a scheme for the detection of common carotid artery using some waveforms fea- tures from ultrasound images. Extracted features were previously used by physicians to differentiate between normal blood flow and five types of abnormal blood flow. The automated approach has an accuracy of 97%. In addition, in [28], a computer model based on an flow velocity distribution is proposed to generate Doppler ultrasound signals from blood flow in the vessels with var- ious stenosis degrees. The factors included in the analysis include the velocity field from pulsatile blood flow in the stenosed vessels, sample volume shape and acoustic factors that affect the Doppler signals. Navier–Stokes equations are analytically solved to calcu- late the velocity distributions of pulsatile blood flow in the vessels with various stenosis degrees. Then, power spectral density of the Doppler signals is estimated. Estimated Doppler signals were close to the theoretical ones.

Gupta et al. [29] stated that carotid image registration has the potential to improve the monitoring, quantification and character- ization of the disease. The authors conclude that ultrasonography is an important cost-effective and non-invasive imaging modality which can be used to screen a large number of symptomatic and asymptomatic patients suspicious of having vulnerable plaques.

Other different modelling techniques also exist in the literature. For example, in [30], Kefayati et al. provided quantitative measures of instabilities and turbulence in the carotid artery bifurcation using proper orthogonal decomposition to visualize complex blood flow patterns with different stenosis-severity levels. The study used anatomically generated flow models for Doppler ultrasound and multi-modality flow studies using developed life-sized phantoms.

244 A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245

Being a low-cost non-invasive method to screen patients suspi- cious of having carotid artery stenosis or occlusion, many studies in the literature focused on the automatic classification of carotid artery Doppler signals. For example, in [31] complex-valued Artifi- cial Neural Network structure is used to classify carotid artery Doppler signals using Principal Component Analysis and Fuzzy c- means Clustering as feature extraction methods. Moreover, an ap- proach that is based on Fast Fourier Transform, Hilbert Transform, and Welch Method with different window types is proposed in [32]. The authors investigated effects of window types on classifi- cation of carotid artery Doppler signals. Many spectral analysis methods were used by Guler et al. [33]-[34] to extract features from Doppler signals. These features were then classified using neural networks. In addition, Latfaoui et al. [35] have used Fourier transform, wavelet transform and S-transform for the analysis of carotid and femoral arteries Doppler signal. In [36], the diagnostic accuracy of US in the detection of high-grade stenosis or occlusion of the celiac artery was investigated. The results of Doppler US were compared with those of lateral aortography. The authors con- cluded that Doppler US can be used as a screening method and that it can reduce the use of unnecessary, invasive angiography. In addi- tion, in [37], a review of studies that compared Doppler ultrasound to CT angiography is performed. Also, in [38], the authors devel- oped a Discrete Hidden Markov Model system to classify the inter- nal carotid artery Doppler signals. The proposed system reached 97.38% of classification accuracy. In comparison with the presented related works, the proposed method is less expensive and less complex which makes it more suitable as an efficient screening method for large populations.

6. Conclusion and future work

In this paper, a method for the screening of carotid artery dis- eases (stenosis and occlusion) is proposed.

The suggested method extracts some features from Doppler US audio directly. The extracted features are then used to train a neu- ral network and a KNN classifiers that can classify the US audio as normal, stenosis or occlusion. The proposed approach reached a maximum total classification accuracy of 95.48% on the used data set when a neural network is employed. Although the used data set in the experimental study is not large enough to generalize the ob- tained results, however, obtained results are promising and indi- cate the suitability of the proposed approach as a low-cost efficient screening method for carotid artery stenosis and occlusion.

The main contribution of the proposed work is the utilization of the low-cost pen probe Doppler US audio device instead of the expensive imaging one. Also, we employed only five features which reduces the time of the test to the minimum. This makes the proposed approach more suitable as a screening method for the carotid artery diseases in comparison with similar approaches.

Future work includes the application of the suggested system on a larger data set to add more power to the generalization. In addition, other features may be added to the used features set in order to further enhance the classification results.

References

[1] A. De Fabritiis, E. Conti, S. Coccheri, Management of patients with carotid stenosis, Pathophysiol Haemost Thromb, Switzerland, 2002, pp. 381–385.

[2] C. Herman, What Makes a Screening Exam ‘‘Good’’? American Medical Association Journal of Ethics 8 (2006) 34–37. <http://virtualmentor.ama- assn.org/2006/01/cprl1-0601.html> (accessed 22.05.13).

[3] An Overview of Carotid Artery Disease. <http://www.webmd.com/heart- disease/carotid-artery-disease-causes-symptoms-tests-and-treatment> (accessed 03.01.13).

[4] Carotid Artery Disease – Symptoms, Risk Factors, Treatment Options. <http:// www.texheartsurgeons.com/CarotidDisease.htm> (accessed 03.01.13).

[5] Carotid artery surgery – series – Normal anatomy. <http://www.umm.edu/ patiented/articles/carotid_artery_surgery_normal_anatomy_000124.htm> (accessed 03.01.13).

[6] Carotid Artery Disease, Stroke, Transient Ischemic Attacks (TIAs). <http:// www.vascularweb.org/vascularhealth/Pages/carotid-artery-disease-,-stroke-,- transient-ischemic-attacks-(-tias-)-.aspx> (accessed 03.01.13).

[7] D. Richard, H. David, L. Nick, Non-invasive methods of arterial and venous assessment, ABC of Arterial and Venous Disease 320 (2000) 698–701.

[8] MATLAB – The Language of Technical Computing. <http:// www.mathworks.com/products/matlab/>.

[9] J. Allen, Applications of the short time Fourier transform to speech processing and spectral analysis, acoustics, speech, and signal processing, in: IEEE International Conference on ICASSP ‘82, 1982, pp. 1012–1015.

[10] J.B. Allen, L.R. Rabiner, A unified approach to short-time Fourier analysis and synthesis, Proceedings of the IEEE 65 (1977) 1558–1564.

[11] R.S. Thompson, B.J. Trudinger, C.M. Cook, A comparison of Doppler ultrasound waveform indices in the umbilical artery–II. Indices derived from the mean velocity and first moment waveforms, Ultrasound in Medicine and Biology 12 (1986) 845–854.

[12] R.G. Gosling, D.H. King, Arterial assessment by Doppler-shift ultrasound, in: Proceedings of the Royal Society of Medicine, 1974, pp. 447–449.

[13] T. Planiol, L. Pourcelot, J.M. Pottier, E. Degiovanni, Étude de la circulation carotidienne par les methods ultrasoniques et la thermographie, Review Neurology (Paris), 1974, pp. 127–141.

[14] E.D. Übeyli, _I. Güler, Neural network analysis of internal carotid arterial Doppler signals: predictions of stenosis and occlusion, Expert Systems with Applications 25 (2003) 1–13.

[15] R.L. Braham, M. Rabkin, Screening, Department of Medicine Columbia University. <http://www.medicineclinic.org/> (accessed 22.05.13).

[16] S. Kanchanaraksa, Evaluation of Diagnostic and Screening Tests: Validity and Reliability, Johns Hopkins University. <http://ocw.jhsph.edu/courses/fundepi/ PDFs/Lecture11.pdf> (accessed 22.05.13).

[17] N.A. Obuchowski, R.J. Graham, M.E. Baker, K.A. Powell, Ten criteria for effective screening their application to multislice CT screening for pulmonary and colorectal cancers, American Journal of Roentgenology 176 (2001) 1357–1362.

[18] K.S. Trevedi, Probability and Statistics with Reliability, Queuing and Computer Science Application, John Wiley and Sons, New York, 2001.

[19] T. Cover, P. Hart, Nearest neighbor pattern classification, IEEE Transactions on Information Theory 13 (1967) 21–27.

[20] H. Uğuz, G.E. Güraksın, U. Ergün, R. Saraçoğlu, Biomedical system based on the Discrete Hidden Markov Model using the Rocchio–Genetic approach for the classification of internal carotid artery Doppler signals, Computer Methods and Programs in Biomedicine 103 (2011) 51–60.

[21] S. Kara, Classification of mitral stenosis from Doppler signals using short time Fourier transform and artificial neural networks, Expert Systems with Applications 33 (2007) 468–475.

[22] M. Hassan, A. Chaudhry, A. Khan, J.Y. Kim, Carotid artery image segmentation using modified spatial fuzzy c-means and ensemble clustering, Computer Methods and Programs in Biomedicine 108 (2012) 1261–1276.

[23] F. Auricchio, M. Conti, M. De Beule, G. De Santis, B. Verhegghe, Carotid artery stenting simulation: from patient-specific images to finite element analysis, Medical Engineering & Physics 33 (2011) 281–289.

[24] E.C. Kyriacou, S. Petroudi, C.S. Pattichis, M.S. Pattichis, M. Griffin, S. Kakkos, A. Nicolaides, Prediction of high-risk asymptomatic carotid plaques based on ultrasonic image features, IEEE Transactions on Information Technology in Biomedicine 16 (2012) 966–973.

[25] A. Azhim, A. Ueno, M. Tanaka, M. Akutagawa, Y. Kinouchi, Evaluation of blood flow velocity envelope in common carotid artery for reference data, Biomedical Signal Processing and Control 6 (2011) 209–215.

[26] U. Bilge, S. Bozkurt, S. Durmaz, Application of data mining techniques for detecting asymptomatic carotid artery stenosis, Computers & Electrical Engineering (2011).

[27] C.-M. Chen, H.-C. Wu, C.-S. Tsai, Common carotid artery condition recognition technology using waveform features extracted from ultrasound spectrum images, Journal of Systems and Software 86 (2013) 38–46.

[28] L. Gao, Y. Zhang, K. Zhang, G. Cai, J. Zhang, X. Shi, A computer simulation model for Doppler ultrasound signals from pulsatile blood flow in stenosed vessels, Computers in Biology and Medicine 42 (2012) 906–914.

[29] A. Gupta, H.K. Verma, S. Gupta, Technology and research developments in carotid image registration, Biomedical Signal Processing and Control 7 (2012) 560–570.

[30] S. Kefayati, T.L. Poepping, Transitional flow analysis in the carotid artery bifurcation by proper orthogonal decomposition and particle image velocimetry, Medical Engineering & Physics (2012).

[31] Y. Özbay, S. Kara, F. Latifoğlu, R. Ceylan, M. Ceylan, Complex-valued wavelet artificial neural network for Doppler signals classifying, Artificial Intelligence in Medicine 40 (2007) 143–156.

[32] Y. Özbay, M. Ceylan, Effects of window types on classification of carotid artery Doppler signals in the early phase of atherosclerosis using complex-valued artificial neural network, Computers in Biology and Medicine 37 (2007) 287– 295.

[33] N.F. Güler, E.D. Übeylı, Wavelet-based neural network analysis of ophthalmic artery Doppler signals, Computers in Biology and Medicine 34 (2004) 601– 613.

A.F. Seddik, D.M. Shawky / Knowledge-Based Systems 52 (2013) 236–245 245

[34] _I. Güler, E.D. Übeyli, Implementing wavelet/probabilistic neural networks for Doppler ultrasound blood flow signals, Expert Systems with Applications 33 (2007) 162–170.

[35] M. Latfaoui, F.B. Reguig, Time-frequency analysis of femoral and carotid arterial Doppler signals, Procedia Engineering 29 (2012) 3434–3441.

[36] H.K. Lim, W.J. Lee, S.H. Kim, S.J. Lee, S.H. Choi, H.S. Park, Y.S. Do, S.W. Choo, I.W. Choo, Splanchnic Arterial Stenosis or Occlusion: Diagnosis at Doppler US1, 1999.

[37] Z. Chiara, R. Emma, S. Yves, Concordance rates of Doppler ultrasound and CT angiography in the grading of carotid artery stenosis: a systematic literature review, Journal of Neurology 259 (2012) 1015–1018.

[38] H. Uğuz, H. Kodaz, Classification of internal carotid artery Doppler signals using fuzzy discrete hidden Markov model, Expert Systems with Applications 38 (2011) 7407–7414.