Computer aided diagnosis of Coronary Artery …...Review paper Computer aided diagnosis of Coronary...

Physica Medica 33 (2017) 1–15

Contents lists available at ScienceDirect

Physica Medica

journal homepage: ht tp : / /www.physicamedica.com

Review paper

Computer aided diagnosis of Coronary Artery Disease, MyocardialInfarction and carotid atherosclerosis using ultrasound images: A review

http://dx.doi.org/10.1016/j.ejmp.2016.12.0051120-1797/� 2016 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.

⇑ Corresponding author.E-mail address: [email protected] (K.H. Ng).

Oliver Faust a, U. Rajendra Acharya b,c,d, Vidya K. Sudarshan b, Ru San Tan b, Chai Hong Yeong f,Filippo Molinari e, Kwan Hoong Ng d,⇑aDepartment of Engineering and Mathematics, Sheffield Hallam University, United KingdombDepartment of Electronic & Computer Engineering, Ngee Ann Polytechnic, SingaporecDepartment of Biomedical Engineering, School of Science and Technology, SIM University, SingaporedDepartment of Biomedical Engineering, Faculty of Engineering, University of Malaya, Kuala Lumpur, MalaysiaeDepartment of Electronics and Telecommunications, Politecnico di Torino, Turin, ItalyfDepartment of Biomedical Imaging, Faculty of Medicine, University of Malaya, Kuala Lumpur, Malaysia

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

Article history:Received 25 August 2016Received in Revised form 21 November2016Accepted 4 December 2016Available online 20 December 2016

Keywords:Computer aided diagnosisCoronary Artery DiseaseMyocardial InfarctionCarotid atherosclerosisThoracic UltrasoundIntravascular Ultrasound

The diagnosis of Coronary Artery Disease (CAD), Myocardial Infarction (MI) and carotid atherosclerosis isof paramount importance, as these cardiovascular diseases may cause medical complications and largenumber of death. Ultrasound (US) is a widely used imaging modality, as it captures moving imagesand image features correlate well with results obtained from other imaging methods. Furthermore, USdoes not use ionizing radiation and it is economical when compared to other imaging modalities.However, reading US images takes time and the relationship between image and tissue composition iscomplex. Therefore, the diagnostic accuracy depends on both time taken to read the images and experi-ence of the screening practitioner. Computer support tools can reduce the inter-operator variability withlower subject specific expertise, when appropriate processing methods are used. In the current review,we analysed automatic detection methods for the diagnosis of CAD, MI and carotid atherosclerosis basedon thoracic and Intravascular Ultrasound (IVUS). We found that IVUS is more often used than thoracic USfor CAD. But for MI and carotid atherosclerosis IVUS is still in the experimental stage. Furthermore, tho-racic US is more often used than IVUS for computer aided diagnosis systems.

� 2016 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32. Background. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1. Coronary Artery Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.2. Myocardial Infarction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52.3. Carotid atherosclerosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3. Diagnosis support system design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.1. Ultrasound image acquisition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.1. Intravascular Ultrasound imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2. Feature extraction and assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73.3. Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4. Diagnosis support system realisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85. Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

5.1. Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

http://crossmark.crossref.org/dialog/?doi=10.1016/j.ejmp.2016.12.005&domain=pdf

http://dx.doi.org/10.1016/j.ejmp.2016.12.005

mailto:[email protected]

http://dx.doi.org/10.1016/j.ejmp.2016.12.005

http://www.sciencedirect.com/science/journal/11201797

http://http://www.physicamedica.com

Acronyms

3G international mobile telecommunications-20004g successor to 3GA accuracyA-V atrio-ventricularADA American diabetes associationAES advanced encryption standardAF atrial fibrillationAFL atrial flutterApEn approximate entropyANS autonomic nervous systemANN artificial neural networkANSI American national standards instituteAHI apnea/hypopnea indexAI apnea indexAI artificial intelligenceASM active shape modelANOVA analysis of varianceAPI application programming interfaceAR autoregressiveARMA autoregressive moving averageAT atrial tachycardiaATM automated teller machineAUC area under curveBA Bayesian averagingBCI brain computer interfaceBMI body mass indexBPA back-propagation algorithmBPSO binary particle swarm optimizationBPM beats per minuteBSN biomedical sensor networkCAD computer-aided diagnosisCAD Coronary Artery DiseaseCaD capacity dimensionCAN cardiovascular autonomic neuropathyCART classification and regression treeCD correlation dimensionCEUS contrast enhanced ultrasoundCI confidence intervalCHF congestive heart failureCHS community health centersCLDA clustering linear discriminant analysis algorithmCM clustered microcalcificationsCSP communicating sequential processesCS compressed samplingCT curvelet transformCPA communicating process architecturesCPC cardiopulmonary couplingCSA central sleep apneaCSME clinically significant macular edemaCSR cambridge silicon radioCVD cardiovascular diseaseCWT continuous wavelet transformD2H2 distributed diagnosis and home healthcareDAGSVM directed acyclic graph support vector machineDCT discrete cosine transformDET determinismDII diabetic integrated indexDM diabetes mellitusDN diabetic neuropathyDR diabetic retinopathyDT decision treeDWT discrete wavelet transformE externalE-M expectation-maximizationECG electrocardiography

EDR ECG-derived respirationECHONET echocardiographic healthcare online networking

expertise in TasmaniaEEG electroencephalographEM electromagneticEMG electromyographyEOG electrooculographyFAQ frequently asked questionsFB fusion beatFD fractal dimensionFDDI fiber distributed data interfaceFES functional electrical stimulationFCM fuzzy C-meansFIR finite impulse responseFN false negativeFNN false nearest neighborFP false positiveFSC fuzzy sugeno classifierGA genetic algorithmGLCM gray-level co-occurrence matrixGMM Gaussian mixture modelGPRS general packet radio serviceGSM global system for mobile communicationsGUI graphical user interfaceH hurst exponentHDNT hospital digital networking technologiesHIHM home integrated health monitorHIPAA health insurance portability and accountability actHMA haemorrhages and microaneurysmsHOS higher order spectraHR heart rateHRUS High Resolution UltrasoundHRV Heart Rate VariabilityHSDPA High-Speed Downlink Packet AccessI InternalICA independent component analysisIEEE institute of electrical and electronic engineersIIR infinite impulse responseIR infraredIT information technologyIVUS Intravascular UltrasoundJMMB journal of mechanics in medicine and biologyKS kolmogorov sinaiK-NN K-nearest neighbourLBP local binary patternLC linear classifierLCP local configuration patternLDA linear discriminant analysisLLE largest lyapunov exponetLTE long term evolutionLRNC lipid-rich necrotic coreLV left ventricularMA moving averageMC clustered microcalcificationsNCSME non-clinically significant macular edemaNDDF normal density discriminant functionMI Myocardial InfarctionML maximum likelihoodMLPNN multilayer perceptron neural networksMOH ministry of healthMP mobile phoneMR magnetic resonanceMRI magnetic resonance imagingMSE mean squared errorNASA national aeronautics and space administration

2 O. Faust et al. / Physica Medica 33 (2017) 1–15

NBC naive Bayes classifierNEB non-ectopic beatNPDR mon-proliferative diabetic retinopathyNSR normal sinus rhythmOFDM orthogonal frequency-division multiplexingOSA obstructive sleep apneaPCA principal Component AnalysisPDR proliferative diabetic retinopathyPET positron emission tomographyPET-MRI positron emission tomography - magnetic resonance

imagingPHC primary health centersPNN probabilistic neural networkPSD power spectrum densityPSG polysomnographyPPV positive predictive valuePVC premature ventricular contractionQoS quality of serviceRBF radial basis functionRBFNN radial basis function neural networkREC recurrence rateRF random forestRHN regional health networkRNN recurrent neural networkROC receiver operating characteristicROI region of interestRP recurrence plotsRSA respiratory sinus arrhythmiaRVM relevance vector machineSWT stationary wavelet transformSA sino-atrialSampEn sample entropySCD sudden cardiac deathSDNN standard deviation of normal to normal intervalsSe sensitivitySE standard error

ShanEn shannon entropySMDS switched multimegabit data serviceSNN spiking neural networkSOM self organizing mapSp specificitySQL structured query languageSVEB supraventricular ectopic beatSVM support vector machineT thoracicTIA transient ischemic attackTEE transesophageal echocardiogramTN true negativeTP true positiveTTE transthoracic echocardiographyUB unclassified beatUIT urinary tract infectionULP ultra-low powerUMTS universal mobile telecommunications systemUS ultrasoundUWB ultra widebandVEB ventricular ectopic beatVF ventricular FibrillationVFL ventricular flutterVT ventricular tachycardiaWBAN wireless body area networkWEP wired equivalent privacyWHO world health organizationWiMAX worldwide interoperability for microwave accessWLAN wireless local area networkWMAN wireless metropolitan area networkWPA Wi-Fi protected accessWPAN wireless personal area networkWWAN wireless wide area network

O. Faust et al. / Physica Medica 33 (2017) 1–15 3

1. Introduction

Cardiovascular disease has been a global public health problemfor the last 35 years [1]. Public health statistics show that the num-ber of patients, with some form of cardiovascular disease increasesteadily in countries with a low and middle gross national income,while a number of countries with a high gross national incomehave managed to diminish the incidence of cardiovascular disease[2,3]. On a global scale, cardiovascular disease is responsible foraround 30% of human mortality as well as 10% of the disease bur-den [4,5]. In 2005, 17 million out of 58 million deaths worldwidewere caused by cardiovascular disease [6–9]. According to thestatistics, published in 2013, the various risk factors causing deathare 40.6% due to high blood pressure, 13.7% came from smoking,13.2% resulted from a poorly balanced diet, 11.9% are attributedto insufficient physical exercise and 20.6% could not be attributed[10,11]. Another noteworthy result was the fact that 88% of all car-diovascular fatalities had abnormal glucose levels [12,13]. Coron-ary Artery Disease (CAD) is a specific cardiovascular disease,which affects the coronary arteries of human heart. Having CADcarries the risk that the patient may develop Myocardial Infarction(MI) [14]. The pathology of MI is characterised by an occlusion in acoronary artery which leads to the death of myocardium [15]. Eachyear in the United Kingdom, there are more than 250,000 docu-mented acute MI, with at least the same number of patients beingadmitted to hospital to rule out acute MI. The diagnosis of acute MIis difficult, because of time pressure and complexities of medicaldata interpretation. Traditionally, the diagnosis is based on a com-

bination of clinical history, Electrocardiography (ECG) findings andbiochemical tests [16]. To prevent or at least reduce the number ofMI events it is necessary to monitor the cardiac health of a patient.Studies show that carotid plaque is a good predictor of cardiovas-cular diseases [17,18]. Thoracic Ultrasound (US) is widely used todiagnose carotid atherosclerosis [19–21]. In recent years, computersupport systems have moved on from mere plaque detection toplaque characterization. Despite considering multiple data sourcesand improved image analysis methods, the diagnosis of CAD, MIand carotid atherosclerosis still lacks Sensitivity (Se) and Speci-ficity (Sp).

US images can provide vital information for the diagnosis ofCAD, MI and carotid atherosclerosis. However, they are often dis-torted and incomplete which makes them open to multiple inter-pretations [22]. Therefore, the diagnostic relevance of a randomlyselected US image falls within a wide range. For a significant num-ber of images the diagnostic relevance is below a threshold whererelevant features are ambiguous, and hence the practitioner has todepend on experience to establish reasonable inference [23,24]. AUS scan is a gradual process; hence the image interpretation isrevised as soon as new information is revealed [25]. Consequently,there are distinct hierarchical levels of interpretation [26]. A read-ing radiologist must have expert knowledge in the language usedto describe the images as well as an excellent understanding ofthe relationship between pixel grey levels and the anatomicalobjects, to ensure that all interpretations are uniform and are ofhigh quality [27]. Apart from the inherent difficulty to extract diag-nostic information from US images, there is also the problem of


data overloading [28]. Progress in US imaging means to producemore data during scanning [29,30]. A human screener might spenda significant amount of time to get the required information. Forcomputer aided diagnosis, data overloading is not a problem, asdigital processing and data storage will take care of this [31]. Con-sequently, computer support systems handle the increasing datavolumes much better than human practitioners. Hence, the ques-tion shifts from whether there is computer support, for US basedcardiovascular disease diagnosis, to how reliable is the computersupport. The most advanced computer support systems incorpo-rate machine learning and artificial decision making algorithms[32,33]. Potential application areas for computer aided diagnosissystems are treatment monitoring [34], drug efficacy tests [35]and most importantly disease detection [36].

A well-designed computer support system can reduce the costand improve the quality of US based diagnosis. Replacing humanwork with computer processing yields these benefits. Computersupport systems can be mass produced; hence manufacturerscan improve their sales and thrive in a competitive market. Usingcomputing technology keeps the systems flexible and it is easy toincorporate the latest progress in the relevant fields of medicine,computing and engineering through hardware and softwareupgrades. In this review, of US based diagnosis of CAD, MI and car-otid atherosclerosis, we highlight that these diseases are linkedand hence the processing techniques, used for diagnostic support,should be similar as well. We found, that the processing techniquesdepend on whether thoracic US or Intravascular Ultrasound (IVUS)is used for image acquisition. During our review, we learned thatIVUS is predominantly used for CAD diagnosis, whereas thoracicUS is mainly used for MI. For the diagnosis of carotid atherosclero-sis, the invasive nature of IVUS is a problem. Therefore, theintravascular method is still experimental, hence it is mostly usedfor post mortem studies. We are surprised to find only one com-puter aided diagnosis system based on IVUS. We suspect that thelack of IVUS based computer aided diagnosis systems comes fromthe fact that the imaging modality is relatively new, hence theresearch focuses on image analysis rather than diagnosis support.We adopt the position that there is a need for IVUS based computeraided CAD and MI diagnosis. Creating such IVUS based diagnosissupport tools is a logical progression from focused imaging modal-ities. These systems are very relevant, because they can improvethe efficiency of practitioners by reducing the amount of timespent in scanning the US images. In the long run, computer aideddiagnosis systems reduce the cost and increase the diagnosticaccuracy.

The article is organized as follows. The next section provides thenecessary background on CAD, MI and carotid atherosclerosis. Sec-tion 3 introduces the materials and methods used to design com-puter support systems for US based cardiovascular diseasediagnosis. The results section lists diagnostic support systems forCAD, MI and carotid atherosclerosis. We focus on discriminatingArtificial Intelligence (AI) based computer aided diagnosis systemsfrom non AI based computer support systems. Practical settingsand limitations are covered in the discussion section. Section 6concludes the review.

2. Background

2.1. Coronary Artery Disease

Coronary arteries provide oxygen and nutrients to heart mus-cles. They are composed of three basic layers: the intima, media,and adventitia which are arranged into three concentric layers.Any disease that affects the coronary arteries causes systemic dis-ability and in some cases death. The major cause of CAD is

atherosclerosis, which is characterized by the deposition of choles-terol and lipids, predominantly within the intimal wall of theartery [37]. CAD is a progressive condition that takes several yearsto develop [38]. The phases of atherosclerosis progression are fattystreak, fibrous plaque, and complicated lesion. The initial fattystreaks are marked by lipid-filled smooth muscle cells [37]. Later,a yellow shade appears when fatty streaks progress within thesmooth muscle cells. The subsequent fibrous plaque phase exhibitsthe onset of continuous ongoing alterations in the endothelium ofthe arterial wall. The fatty streak is eventually covered by collagenforming a fibrous plaque that appears greyish or whitish [39].These plaques can develop on one section of the artery or in a cir-cular manner, comprising the entire lumen. The last phase, in theprogression of the atherosclerotic lesion, is the most threatening.The continuous growth of plaque, which causes inflammation,results in plaque instability and rupture [39]. Later, platelets gatherin large numbers, resulting in the formation of a thrombus. Thethrombus sticks to the arterial wall, causing to additional narrow-ing or complete occlusion of the artery. During the early stages, apatient experiences little or no symptoms, but if the causes ofCAD remain, the disease progresses. Hence, treatment will startwith eliminating the causes and in more serious cases medicationwill be administered. Without intervention, a CAD patient candevelop ischemia (intermittent blood supply) and then infarctions(loss of blood supply). These serious conditions can lead to SuddenCardiac Death (SCD) [40].

Forbes et al. found a statistically relevant correlation betweenplaque in the coronary arteries and a greater long term risk of con-tracting CAD [41]. With current IVUS technology it is possible todifferentiate four different plague types: necrotic core; dense cal-cium; fibro-fatty; and brous [42,43]. Coronary artery plaques witha thin fibrous cap over a large necrotic core are more prone to rup-ture [44,45]. In a focused review, Banach et al. analysed the effi-ciency of statin therapy on fibrous plaque[46,47]. 830 subjectsparticipated in the reviewed clinical studies, 737 were given statinand 93 were in the placebo control group [48–50]. Their analysisshows that statin therapy reduces fibrous plaque, but the necroticcore volume remains unchanged [51,52]. Nozue et al. used virtualhistology IVUS in a clinical study about the effect of fluvastatin oncoronary plaque [53]. It is found that, apart from evaluating theplaque morphology, the mechanical properties of the coronaryarteries are also important in evaluating the CAD [54,55]. Estab-lishing the mechanical strain pattern of the artery wall can helpto determine whether a lesion is unstable and prone to rupture.Changes in the coronary artery blood flow can be interpreted assigns of CAD [56]. Therefore, features based on fluid dynamic algo-rithms carry diagnostically important information which can beused in future computer aided diagnosis systems.

Contrast angiography is the gold standard for diagnosing theextent of CAD [57,58]. But, its limitation is that visually analysedangiography results in underestimating atherosclerosis in ‘‘nor-mal” coronary artery segments [59,60]. Govindaraju et al. reviewedthe functional severity of Coronary Artery Disease with fluiddynamics [61]. They found that fractional flow reserve is a stan-dard index to identify the severity of CAD. It can be used to avoidsurgical complications by estimating pressure drop and flowreduction caused by invasive interventions. Flow models were alsoused to estimate artery wall stress, which appears to contributesignificantly to CAD [62]. US is frequently used to assess the coro-nary arteries. Fig. 3 shows thoracic images of the human heart,which can be used to assess the coronary arteries. Fig. 4 showsIVUS images from within a coronary artery. The three-layers ofthe coronary artery appear as the bright inner layer (intima), mid-dle echo-lucent zone (media), and outer bright layer (adventitia) ina cross-sectional view using IVUS. Detecting these layers of vesselprovides orientation in the IVUS image. Assessing the coronary


artery is important, because CAD can progress towards more dan-gerous cardiovascular diseases, such as MI. The next sectionexplores the relationship between CAD and MI.

2.2. Myocardial Infarction

MI manifests in a wide range of symptoms. The spectrumreaches from an absence of symptoms over a minor event to amajor event, which can result in SCD or severe haemodynamicdeterioration [63]. MI results from a progressive collection ofatherosclerotic plaque on the walls of coronary arteries [64]. Theaccumulated plaque gets ruptured initiating a clot which resultsin complete blockage of the artery. Angiography is performed todetect narrowing or complete blockage of the infarct-related coro-nary artery [65,66]. The artery occlusion decreases the blood flowto the myocardium, which damages the heart muscle [67]. Ifreduced blood supply to the myocardium persists long enough, aprocess called ischemic cascade is initiated [68] whereby the heartcells begin to die, causing a condition called MI. Accordingly, thepatient’s heart will be permanently (irreversibly) damaged [69].

For many cases, CAD progression occurs in previously insignifi-cant lesions [70]. It is very often observed that lesions of acute MIhave severe stenosis. Even though the stenosis is severe, it is sec-ond to the superimposed thrombus. However, before the event,severe stenosis might not occur [71]. Acute MI is caused by oneof two events. The first event is a sudden rupture or ulcer in a coro-nary artery [72]. The second cause is the formation of vulnerableplaque which leads to a thrombus blocking the coronary artery.The speed with which the catastrophic event unfolds leaves littleor no time for counter measures. Vulnerable plaques can be char-acterized by a fibrous cap with macrophage infiltration and a largelipid pool [73]. Cardiologists examine different cross-sectionalplanes with various US transducer positions to assess left ventric-ular wall segments. They use the same technique to view anddetect MI [74]. Fig. 3b shows a thoracic US image of an MI affectedheart. Plaque and plaque formation analysis is also of paramountimportance for diagnosing carotid atherosclerosis.

2.3. Carotid atherosclerosis

The carotid arteries are the two large blood vessels (internal andexternal carotid arteries) that supply oxygenated blood to thelarge, front part of the brain [75]. Like coronary arteries, carotidarteries are also susceptible to atherosclerosis, an inflammatoryaccumulation of plaques. These plaques contain Lipid-Rich Necro-tic Core (LRNC), which is enclosed by depleted smooth muscle cellsand a thin fibrous cap [76]. Thinning of these fibrous cap is a dis-tinct risk indicator for underlying or forthcoming ischemic neuro-logical abnormalities [77,78]. Emboli or thrombus may break off,due to artery wall stress, from plaque having a thin fibrous capand join the blood circulation towards the brain [79]. As the vesselbecomes narrower, the thrombus gets attached to the vessel walland cause carotid artery stenosis [80]. The formation of carotidartery stenosis either diminishes or restricts blood flow to brainregions which are supplied by the vessel, thereby causes TransientIschemic Attacks (TIAs). TIAs are warning signs, frequently fol-lowed by severe permanent (irreversible) thromboembolic stroke[81–83]. Further extension of this condition may lead to loss ofbrain function or even death.

Carotid stenosis, due to atherosclerosis, is grouped into asymp-tomatic [84] and symptomatic [85]. Asymptomatic carotid stenosisdenotes around 60% narrowing of proximal carotid artery in theabsence of earlier history of stroke or TIA [86]. Symptomatic caro-tid stenosis is frequently linked with type 5 (having more extracel-lular lipids, hematoma) or 6 (having surface defect, hematoma andthrombosis) plaques [76]. The presence of larger LRNC, in both

symptomatic and asymptomatic carotid stenosis a thin or rupturedfibrous cap [87]. Thus, the accurate identification or classificationof symptomatic and asymptomatic carotid stenosis is importantfor selecting appropriate treatment [88]. The ability of a particulartreatment to prevent stroke, in both symptomatic and asymp-tomatic patients having chronic carotid stenosis, is the topic ofongoing research [89–91]. Fig. 1a shows a US scan of the dangeroussystematic plaque. Fig. 1b shows a section of the carotid arterywith asymptomatic plaque. In contrast, Fig. 1c shows a US imageof a normal carotid region.

The minimal invasive method of contrast enhanced CurveletTransform (CT) has been used to quantify carotid plaque [92].These systems are used for carotid plaque segmentation. US is apreliminary non-invasive imaging technique that can be used toassess carotid artery stenosis [93]. In clinical settings, high-resolution, B-mode US together with Doppler flow is often usedfor assessing carotid arteries [94]. Furthermore, Doppler US is alsoused for the characterization of high risk plaques and thus help infor assessing the severity of stenosis [95]. A main restriction of thisimaging modality is that it is highly user dependent. Therefore,advancements in non-invasive imaging technology, usingcomputer-aided methods, have enhanced the acquisition of dataand significantly improved the diagnostic accuracy. In the next sec-tion, we explore the design of US based diagnosis systems for car-diovascular diseases.

3. Diagnosis support system design

On a conceptual level, we discuss two different design strategiesfor CAD, MI and carotid atherosclerosis diagnosis support systems.The first system design approach yields implementations, whichprovide just analytical support. This support can range from simpleUS image enhancements to 3D reconstruction of coronary arteries[96]. In general, these systems extract relevant information whichhelps cardiologists with their diagnosis [97,98]. Computer aideddiagnosis systems aim to provide efficient diagnosis. Hence, thesesystems can be used to automate the diagnosis process, which hasimportant benefits including but not limited to a large cost savingpotential. Therefore, computer aided diagnosis became the focus ofmajor research work in medical imaging. Furthermore, these sys-tems are of interest in diagnostic radiology [99]. All diagnostic sup-port systems assist practitioners by extracting features fromunderlying data [100]. The underlying data comes from physiolog-ical measurements or medical images [101]. A computer aideddiagnosis system uses these features as input to decision makingprocesses. The system communicates the decision results in theform of a diagnosis to the practitioner. Computer aided diagnosissystems with low complexity or systems which must deal withparticularly difficult data, offer only disease or non-disease diagno-sis. In contrast, more sophisticated systems, which might be basedon the combination of different imaging modalities and physiolog-ical measurements, may have to diagnose number of classes(stages of diseases). Regardless of the system complexity, the clas-sification performance determines the computer aided diagnosisquality. Hence, the classification performance is used to selectand compare the systems [102]. To achieve a high classificationperformance, it is necessary to carry out statistical performancetests of features during the training phase.

In the sections below, we adopt the design perspective toexplain diagnostic support systems used for CAD, MI and carotidatherosclerosis based on US images. On the highest level ofabstraction, the design is partitioned into an online and an offlinesystem. Such a conceptual split is very important, because the off-line system allows the designer to focus on creating, benchmarkingand selecting the most appropriate algorithm structure. In con-

Fig. 1. Thoracic US images of carotid arteries.


trast, an online system uses the selected algorithm structure toprovide diagnosis support. The left part of Fig. 2 shows both algo-rithm structure and statistical tests carried out in the offline sys-tem. The online system deals with unknown US images. Theoffline system uses known or labelled US images as input. Theinput images are subjected to pre-processing and feature extrac-tion. In the offline system, a range of methods are tested and onlythe most efficient algorithms are used in the online system. Simi-larly, the designer tests several classification algorithms in the off-line system and the online system uses only the best decisionmaking method. Simple computer support systems do not have aclassification step, only computer aided diagnosis systems includesuch algorithms. In general, testing more algorithms creates morecompetition which improves the system performance. The featureextraction algorithms take images from thoracic US or IVUS. Theseimage acquisition methods are discussed below.

Fig. 2. A block diagram depicting the highest level of abstraction for the design of a comThe diagram highlights the fact that, unlike normal support systems, computer aided di

3.1. Ultrasound image acquisition

Thoracic echocardiographic technology has the advantages ofportability, mature technology and low image acquisition cost[103]. Furthermore, the cardiovascular system assessment results,for both doppler and 2D thoracic US imaging modalities, have beenvalidated with other diagnostic imaging techniques, such as CT[104]. Echocardiography is most commonly used to assess the car-diac chamber and establish the extent of its functionality. Espe-cially for analysing the cardiac chamber functionality, therealtime nature of echocardiography is beneficial, because movingimages report chronological as well as spatial information. Stan-dardization of the methodology, used to assess cardiac chambers,is established by collecting, creating and disseminating scientificresearch results, which, when followed by practitioners, provideuniformity and facilitate unambiguous communication [105].

puter support systems for CAD, MI and carotid atherosclerosis based on US images.agnosis systems provide decision support through automated classification.


Despite the efforts of standardisation, US imaging poses a num-ber of challenges for image analysis and feature extraction [26].First and foremost, the relation between pixel intensity and tissueproperties is complex, because of the acoustic phenomena used forimage acquisition. US systems send high frequency sound waves(typically 1 MHz to 18 MHz) through human tissue and recordreflected, as well as scattered, signal components. The US systemdisplays these received signals on a screen. Tissue transitions arerepresented by reflections. Wave scattering results in interferencepattern, known as speckle pattern. Hence, US images are an overlayof speckle pattern and sharper reflection structures. In many cases,different tissues can only be distinguished based on minutechanges in the speckle pattern. These changes might even be tran-sient, which makes it necessary to observe a sequence of US imagesover a period of time. Furthermore, US images depend on operatorspecific properties, like angle and depth of the US beam [106]. Inaddition, the images might include artefacts and image noise[107]. Another big problem is missing information, such as drop-outs, shadowing, scan sector limitations and restricted echo win-dows [108]. Fast moving structures can also cause aliasing effectswhich result in spatial distortion [109]. In general, the relationshipbetween tissue formation and image texture is better defined formedical imaging modalities that use ionizing radiation, such ascomputed tomography (Hounsfield units) and X-ray (Lambert–Beer law) [110]. It takes an experienced practitioner to overcomethe problems of US image interpretation [111]. Computer supportsystems must incorporate functionality which mimics that creativeprocess.

US image acquisition is a gradual process; therefore, an imagerepresents a snapshot of the information available. Fig. 3a showsa typical US image of a normal human heart while Fig. 3b showsan MI affected heart. For the untrained eye, the features, which dis-tinguish normal and MI images are not clear. It requires experttraining to build up the knowledge to distinguish MI from normal.To establish such an expert knowledge is the challenge for thedesign centric offline system. The online system must providethe features that give the diagnosis in a practical setting.

3.1.1. Intravascular Ultrasound imagingThe first IVUS images of normal and atherosclerotic arterial wall

thickness is published by Mallery et al. in 1988 [112]. Hence, it is arelatively new medical imaging modality. IVUS delivers precisetomographic images which enable the reading radiographer toassess coronary arteries in vivo [113]. Clinical studies establishedthat IVUS is sensitive in detecting atherosclerosis and quantifyingboth plaque geometry as well as structure. However, the effective-

Fig. 3. Thoracic US imag

ness for providing diagnostic support is limited due to the two-dimensional signal representation, which is still used in most sys-tems [114]. Analytical computer support systems aim to improvethe information presentation by making it more accessible, forexample through 3D representation.

The physical IVUS set up is established by mounting a US trans-ducer on the tip of a catheter. The catheter is inserted into an arterywhere the US sensor captures intravascular images. These imagesdepict the morphology of both plaque and arterial wall [115]. IVUScan help to detect the presence and to determine the atherosclero-sis composition in angiographically normal reference sites [116].Computer support systems extract information bearing featuresfrom these images. In contrast to thoracic US, IVUS is invasive[117]. The act of inserting the US transducer can, amongst othercomplications, cause plaque to come loose from the artery wall[118]. Plaque debris in the carotid artery can cause stroke and pla-que derbies in the coronary artery can cause MI.

Like thoracic US images, IVUS images are difficult to interpret.Fig. 4a and b show typical IVUS images of normal and diseasedcoronary arteries, respectively. Detecting vessel intima, adventitiaand media layer provides orientation in the IVUS image. However,the challenge is to detect and characterize plaque. The first step totackle that challenge is to employ effective feature extractionmethods.

3.2. Feature extraction and assessment

Feature extraction is a process which determines one or multi-ple information bearing properties from an US image to form a fea-ture vector [119]. US image interpretation requires the conversionof image textures into features. The underlying assumption is thatimage textures contain diagnostically relevant information [120].The features extract as well as condense the information and pre-sent it as a parameter value [121]. Statistical methods, such asAnalysis Of Variance (ANOVA) with p- and F-values, can be usedto select and rank the features. However, statistical tests assumethat the known test data has specific statistical properties. Forexample, ANOVA assumes a Gaussian distribution. As a conse-quence, these statistical test methods provide just an indicationof what features should be used for classification. The final featureselection is based on classification results. The classification perfor-mance depends, to a large extent, on the selected features. Hence,feature extraction and selection are crucial processes in the designof computer aided diagnosis systems. The following list details fea-ture extraction methods used in the reviewed computer supportsystems:

es of human hearts.

Fig. 4. IVUS of coronary artery segments – Left Anterior Descending view.


� Texture features, such as the Gray-Level Co-occurrence Matrix(GLCM) [122], are widely used for MI diagnosis. The strengthof texture features lies in extracting information which relatesto the spatial entanglement of intensity values within theRegion of Interest (ROI) of an US image. A significant weaknessof texture features comes from the fact that most textureextraction algorithms depend on both image and grey scaleresolution.

� Statistical features, such as Principal Component Analysis (PCA)and Higher Order Spectra (HOS) [123] are found to be effective.The advantage of statistical features is that they are robust inthe presence of noise. That robustness depends on the lengthof data that can be averaged. Being and essentially linear meth-ods, statistical methods fail to capture nonlinear informationcontained in US images.

� Transform domain features, such as DiscreteWavelet Transform(DWT) [124], Stationary Wavelet Transform (SWT) [125] and CT[126]. As such, CT generalizes wavelet transform results to rep-resent US image structures in terms of scale, orientation andlocation [127,128]. Transform domain features detect subtlechanges in the US image. The methods perform well in the pres-ence of noise and low computational complexity algorithmsexist for standard feature extraction. A disadvantage of thesemethods is that establishing the transform domain is insuffi-cient, another processing step is needed to extract a specificfeature.

� Features based on configuration information, such as Local Con-figuration Pattern (LCP) combines local structural [129] andmicroscopic [130] configuration information which can be usedfor image classification. That hybrid feature extraction methodhas potential. However, practical computer support systemsshould use more than two feature extraction methods, to har-vest the benefits and at the same time minimize the disadvan-tages of the individual methods.

3.3. Classification

Only computer aided diagnosis systems incorporate a classifica-tion step. The classification is established through artificial learn-ing and decision making algorithms. All, except the study byThangavel et al. [131], reviewed computer aided diagnosis systemsused supervised learning algorithms. Unsupervised differ fromsupervised learning algorithms in the way the clusters are estab-lished [132]. Unsupervised learning algorithms establish clustersin the data [133]. Unknown feature vectors are labelled as belong-ing to one of these clusters. Supervised learning is based on classi-

cal ideas of a student and teacher relationship [134]. In terms ofmachine learning, that relationship is expressed as extracting algo-rithm parameter from labelled training data. Once the machinelearning algorithm is trained, it is tested with labelled data thatis not used during the training phase. The test results are expressedin terms of Accuracy (A), Se and Sp [135].

Only the design centric offline system incorporates an assess-ment step. The most suitable classification algorithm, for the prac-tical online system, is selected based on the assessment results[135]. The following classification algorithms were used in thereviewed US based CAD, MI and carotid atherosclerosis computeraided diagnosis systems:

� Threshold classification is a simple algorithm which classifiesbased on whether or not a feature value has crossed a specificthreshold [136,137]. The threshold value is established duringthe training phase.

� Unsupervised learning, such as Self Organizing Map (SOM)[138]. The clusters are established through weight updates ina local area of the map space.

� Classical supervised machine learning, such as Relevance VectorMachine (RVM) [139], K-Nearest Neighbour (K-NN), GaussianMixture Model (GMM), Decision Tree (DT), Naive Bayes Classi-fier (NBC), Probabilistic Neural Network (PNN) [140], and Sup-port Vector Machine (SVM) [141,142] are used in decisionsupport systems. For many classification tasks, SVM outper-forms other methods due to nonlinear kernel functions.

� Classification problem formulated as an optimization questionand solved with Binary Particle Swarm Optimization (BPSO)[143,144].

� Classification problem formulated as an optimization questionand solved with Genetic Algorithm (GA) [145].

� Minimum distance classifier uses an algorithm that compares afeature vector with two or more vectors which represent theindividual signal classes. The method is often used in imageprocessing [146].

4. Diagnosis support system realisations

The review of computer support systems for CAD, MI and caro-tid atherosclerosis based on US images was structured according tothe design principles outlined in the previous section. During thereview, we established whether Internal (I) or Thoracic (T) USwas used for image acquisition. Having that information is impor-tant, because the feature extraction methods depend on the imagetype. For example, a 3D coronary artery wall reconstruction feature


[96] makes sense only for IVUS. Next, we determined which classi-fier was used for diagnostic support. Simple analysis systems lackautomated classification, hence these systems offer only analyticalresults and they depend entirely upon human expertise to reachthe diagnosis. The design of all reviewed systems was based uponlabelled US images, usually taken from patients and normal sub-jects. As part of the review, we determined the number of datasetsused in each work, which helps to establish the performanceresults, such as A, Se and Sp.

We present the review results in three tables. Table 1 focuses on17 computer support systems for CAD diagnosis. The table details

Table 1Summary of the review results for CAD diagnosis support systems based on T ultrasound

Authors USType

Features C

Nissen et al. [147] I Lumen Size and Wall Morphology –Goar et al. [148] I Artery diameter –Mintz et al. [149] I Lesion calcification –Potkin et al. [150] I Stenosis diameter –Hermiller et al. [151] I Arterial narrowing –Alfonso et al. [152] I Plaque area –Mintz et al. [153] I Artery cross section –Mintz et al. [154] I Plaque mass –Klingensmith et al. [155] I Image segmentation –

Cothren et al. [96] I 3D reconstruction of the coronary arterywall

–

Schoenhagen et al. [156] I Remodelling ratio –Schartl et al. [157] I Scale value analysis –Çomak et al. [158] T DWT; Spectrograms S

Babaoglu et al. [159] T PCA SShalbaf et al. [136] T Nonrigid image registration TMukhopadhyay et al. [160] T Epicardial and Endocardial Border

Detection–

Araki et al. [32] I Grayscale features S

Table 2Summary of the review results for MI diagnosis support systems.

Authors USType

Features Classifiers N

Skorton et al. [161] T Grey level distribution,kurtosis, skewness

Manualclassification

7

Kamath et al. [162] T Pixel intensity Statisticalclassification

1

Tak et al. [163] T Pixel intensity Statisticalclassification

1

Mojsilovic et al. [164] T DWT energy Unsupervisedclassification

1

Neskovic et al. [165] T DWT energy Distanceclassifier

1

Kotani et al. [166] I Lesion as well asproximal and distalreference

– 7

Ehara et al. [167] I Artery dimension – 1Hong et al. [168] I Computerized planimetry – 2Skorton et al. [169] T Grey level distribution,

kurtosis, skewnessRVM 1

Moldovanu et al. [170] T Mean grey level,skewness, kurtosis andentropy

– 1

Agani et al. [171] T GLCM and DWT Distanceclassifier

1

Acharya et al. [172] T BPSO and GA SVM 1Sudarshan et al. [173] T DWT SVM 1Sudarshan et al. [33] T Texton features PNN 1Sudarshan et al. [174] T CT, LCP SVM 1Sudarshan et al. [175] T SWT and relative wavelet

energy and entropySVM 1

the review results in accordance with the assessment rubrics dis-cussed above. The materials section provides background on thespecific features and classifiers used in the investigated systems.Table 2 details the review results of computer based MI diagnosissupport. The table introduces the review results for 16 diagnosissupport systems. The columns of this table convey the same infor-mation as Table 1. Table 3 presents the review results for 11 caro-tid atherosclerosis computer support systems in a similar way. Inthe discussion section, we compared the review results for the con-sidered cardiovascular diseases.

and I ultrasound – in most cases IVUS.

lassifiers No. of subjects Performance measure

51 (8 N and 43 CAD) Descriptive comparison20 IVUS vs. angiography110 First order statistics51 Description38 Visual inspection27 Vessel compliance884 Angiography vs. IVUS comparison209 Angiography vs. IVUS comparison.Three motorizedpullbacks

Description

Description

85 Angiography vs. IVUS comparison.131 Drug efficiency statistics

VM 215 (95 N, 120 CAD) Se = 94.5%, Sp = 90.0%, five fold crossvalidation

VM 480 A = 76.67%hreshold 14 72% to 92%

44 –

VM 15 A = 94.95%

o. of subjects Performance measure

adult dogs Se = 90%, Sp = 70%

5 Not mentioned

7 (5 N, 12 MI) Date based MPI statistics

5 A = 96%

8 On day 2: Se = 73%, Sp = 86%, A = 78%; Week 1: Se = 91%,Sp = 86%, A = 89%; Week 3: Se = 100%, Sp = 100%, A = 100%

8 Culprit plaques statistics

78 Frequency of calcium deposits35 Plaque rupture frequency73 A = 84.17%

2 (6 N 6 MI) Entropy: 6.23 ± 0.52 vs. 9.95 ± 0.17

7 A = 91.32%

5 A = 81.46%, three fold cross validation60 (80 N, 80 MI) A = 99.5%, Se = 99.75%, Sp = 99.25%, ten fold cross validation60 (80 N, 80 MI) A = 94.37%, Se = 91.25%, Sp = 97.50%20 (40 N, 80 MI) A = 98.99%, Se = 98.48%, Sp = 100%60 (80 N, 80 MI) A = 96.20%, Se = 97.5%, Sp = 95.0%

Table 3Summary of the review results for carotid atherosclerosis diagnosis support systems.

Authors USType

Features Classifiers No. of subjects Performance measure

Tsiaparas et al. [176] T DWT, Discrete curvelettransform, Finite ridgelettransform

SVM 20 atheromatousplaques

Discrete curvelet transform: A = 84.9%

Thangavel et al. [131] T SOM ADAboost Contourlet transform 257 carotid USimages

SVM: A = 85.6%

Christodoulou et al. [177] T Texture, Chaos, Spectrum,Shape

SOM, K-NN 230 US images SOM: A = 73.1%

Acharya et al. [178] T Texture SVM 99 SVM with RBF: A = 91.7%, Se = 97% and Sp = 80%.Acharya et al. [179] T Texture SVM, DT, GMM, K-NN,

NBC, PNN, SugenoFuzzy

346 carotid plaqueultrasound image

A 89.5%.

Acharya et al. [180] T Texture, DWT, HOS SVM 146 carotidbifurcation plaquesin 99 patients

A = 91.7%, Se = 97%, Sp = 80%

Acharya et al. [181] T Texture, Trace transform,HOS, Spectrum

SVM 160 plaques A = 90.66%, Se = 83.33%, Sp = 95.39%

Acharya et al. [182] T Texture, DWT SVM, GMM, PNN, DT,K-NN, NBC

SVM: A = 88%, Se = 90.2%, Sp = 86.5%

Nair et al. [183] I Spectrum DT 51 post-mortem A = 79.7% for fibrous, A = 81.2% for fibrolipidic,A = 92.8% for calcified, and A = 85.5% for calcified-necrotic regions

Prati et al. [137] I Plaque measurements Threshold 2 carotid arteriespost-mortem

Se = 65%, Sp = 95%

Irshad et al. [184] I Visual inspection – 5 Cases –


5. Discussion

It is evident from this review that several computer-aidedmethods are developed using texture features in order to investi-gate the health of myocardium [174,33,175] and carotid artery[180,178,182,179]. Acharya et al., [180] applied DWT, HOS and tex-ture methods to analyse the US images of carotid artery to charac-terize symptomatic and asymptomatic carotid atherosclerosis.Studies by Sudarshan et al., used second-order spatial statisticalanalysis together with wavelet, HOS, DWT [173], SWT [175], tex-tons [33], Local Binary Pattern (LBP) [174], and LCP [174] to extractechocardiogram image textural properties within the myocardium.These studies have unveiled further useful details of carotid arteryand myocardium compared to the previous manual and semi-automated techniques. Wavelet based methods (DWT[158,164,165,171,173,176,182], SWT [175]) and HOS [180,181]have shown to extract most of the information including subtlechanges, occurring in the carotid artery and myocardium with sig-nificantly high accuracies. Furthermore, LBP and LCP methods,applied on the echocardiogram images, can identify the minutevariations in the myocardium and thereby classify them into mod-erately and severely infarcted myocardium [174].

However, it is also evident from the literature survey that onlyfew studies have used computer aided methods to support thediagnosis for CAD and carotid atherosclerosis using IVUS. Otherthan Araki et al. and Nair et al., not much research work is reportedon CAD and carotid atherosclerosis using computer-aided analysisof IVUS images. Araki et al., [32] used grayscale features to analysethe texture properties of IVUS images to identify the CAD risk strat-ification. Nair et al., [183] performed spectral analysis of IVUSimages for real time in vivoplaque classification [183]. Althoughthe first-order [32] and time–frequency domain [183] texture anal-ysis techniques achieved good results in the identification of CAD[32] and carotid atherosclerosis [183], second-order statistics,HOS, DWT, and other nonlinear methods can substantially improvethe performance of computer support systems based on IVUSimages. These methods may extract additional higher-order andphase information from IVUS image texture, compared to first-

order and linear methods with better accuracy and may even aidin identifying the severity of the conditions.

The most important aspect of expert systems, such as computeraided diagnosis, is to replace human labour with machine work[185]. Throughmass production, it is possible to harvest the econo-mies of scale, which brings down the unit price and revenue is gen-erated via sales volume. However, there is an inherent danger inthis approach, namely the problem of monoculture. To be specific,mass produced systems tend to be well engineered, but their num-ber may become so large that even small probabilities for an indi-vidual safety failure can become a major concern [186]. Safetyconcerns are important aspects for computer support in medicine[187]. As described in materials section, the offline system is usedto establish the functionality aspects of a proposed system. Theperformance measures, reported in Tables 1–3, the system perfor-mance under well controlled conditions. However, mass produc-tion and subsequent mass deployment demand considerations onhow to ensure reliability and safety. These considerations are chal-lenging, because safety is a multifaceted concept. One of thesefacets is active safety [188]. For example, a diagnosis, suggestedby the computer system must be confirmed by the clinician.Another aspect passive safety [189], which demands that safetycritical system failures must be kept at a minimum. The onlyway to meet the required passive safety levels is to improve thedesign process. Splitting the design into online and offline systemsis one step in the right direction. However, the concept of havingan online and an offline system is not sufficient to achieve therequired passive safety levels, because the design steps, describedin Section 3, are only concerned with requirements capturing. Withthe offline system we find out what system we are going to build.To design safe and reliable systems, we need another step, namelyspecification refinement. In other words, we need to find out howwe build the system. The specification refinement step must beexecuted as formal as possible, because logical errors in this stepmay impact negatively on passive safety. Formal logic can be usedto ensure that no setting in the user interface has an impact on thesystem functionality which might compromise patient wellbeing[190,191].


The considerations outlined above, lead to the postulation ofthe formal and model driven design methodology for biomedicalsystems [192]. The design steps follow the systems engineeringmeta model: need definition, requirements capturing – with theoffline system, implementation of the online system, deploymentand decommissioning [193]. In the specification refinement phase,formal methods, such as Communicating Sequential Processes(CSP) [194], B-method [195] and petri nets, can be used to establishbeyond reasonable doubt specific safety and reliability aspects ofthe diagnosis support system [196].

5.1. Future work

The following list introduces areas for future work:

� Deep learning concept can be used extract features and improvethe decision making performance of computer aided diagnosissystems.

� The accuracy of cardiovascular disease diagnosis, based on USimages, may improve by segmenting the ROI and extracting fea-tures from these ROI. More work is needed to segment the ROIfrom IVUS and US images without losing much information.

� Although the MI detection algorithms (DWT, HOS, Textons,LCP), proposed by Sudarshan et al., achieved high accuracy, sim-ilar investigation need to be carried out for diagnosis of CADusing IVUS [174].

� IVUS and US images are associated with high speckle noise[197,198]. Therefore, an efficient algorithm for speckle noisereduction as a pre-processing technique before the featureextraction may improve the performance of the computer aideddiagnosis.

� Using computer-aided methods, there is a possibility of explor-ing the detection of mild stage of MI using US images; similarly,the extent of CAD can be detected using IVUS.

� Carotid atherosclerosis diagnosis may benefit from automatingthe ROI segmentation. Such automation can help to improve theclassification accuracy of asymptomatic plaque.

We predict that the role of computer aided diagnosis will con-tinue to expand at an accelerating speed. Once the systems aredesigned, they can be mass produced with minimum cost. Further-more, hardware and software updates can be used to incorporatethe latest research results from relevant fields of medicine andtechnology in an installed system. At the moment, we are at anearly stage of such a design revolution. By using a meta model,which structures the design process into an online and an offlinesystem, we have the flexibility to explore novel research on algo-rithms even if the online system is already deployed. The offlinedesign process will look more and more like a big data task wherealgorithms inspect huge data and extract useful information fromthe training images. These tasks will be executed with an ever-increasing level automation and the processing will migratetowards virtualized bulk processing environments, such as cloudcomputing [199]. As a consequence, local processing and decisionmaking will be superseded by distributed processing systemsincorporating the latest information in the decision-makingprocess.

6. Conclusion

This paper reviews computational and system design aspects ofcomputer support for cardiovascular disease diagnosis based on USimages. Initially we provided background information on CAD, MIand carotid atherosclerosis. The three investigated diseases arelinked. CAD usually precedes MI and carotid atherosclerosis is a

good indicator of plaque in the coronary arteries. A major problemfor the design of such systems comes from the fact that it is impos-sible to know a priori which algorithm structure works best for agiven problem. Therefore, the design process follows a well-defined meta model which structures the design efforts into creat-ing an offline and an online system. The offline system is used toresearch, assemble and test a wide range of algorithm structures.Only the best algorithms are used in the online system to providethe best possible diagnostic support for practitioners. Hence, thekey concept during the design of diagnostic support systems iscompetition facilitated through quality measurements. Conse-quently, we reviewed several algorithms used in the proposeddiagnosis support systems and presented their performances.

During the review, we placed emphasis on discriminatingbetween basic analysis systems and computer aided diagnosis sys-tems. We found that, there is many thoracic US based systemswhich provide diagnostic support through artificial decision mak-ing. In contrast, only one IVUS based system provided that typeof support. The rest of the reviewed IVUS based systems providedonly clinical analysis. Such analytical support is important; dataanalysis has progressed more towards diagnostic support. There-fore, the next logical step for IVUS based computer support sys-tems is to incorporate a classification step. The resultingcomputer aided diagnosis systems will be more useful as theycan improve the diagnosis quality and traceability. Another impor-tant point of this review is that IVUS is predominantly used forautomated CAD diagnosis. In contrast, thoracic US is most oftenused for MI diagnosis support systems. IVUS is used for post-mortem analysis of carotid plaque and specialized carotid arteryassessment. However, that imaging methodology is consideredunsuitable for screening applications.

References

[1] Anand SS, Yusuf S. Stemming the global tsunami of cardiovascular disease.Lancet 2011;377(9765):529–32.

[2] Reddy KS. Cardiovascular disease in non-western countries. New England JMed 2004:2438–510.

[3] Ford ES, Ajani UA, Croft JB, Critchley JA, Labarthe DR, Kottke TE, et al.Explaining the decrease in us deaths from coronary disease, 1980–2000. NewEngland J Med 2007;356(23):2388–98.

[4] World Health Organization. The world health report 2002: reducing risks,promoting healthy life. World Health Organization; 2002.

[5] WHO, World health statistics 2008. World Health Organization; 2008.[6] World Health Organization and International Society of Hypertension Writing

Group, Prevention of cardiovascular disease: guidelines for assessment andmanagement of total cardiovascular risk. WHO. Geneva.

[7] World Health Organization. Preventing chronic diseases: a vital investment:who global report. WHO report.

[8] Mendis S, Thygesen K, Kuulasmaa K, Giampaoli S, Mähönen M, Blackett KN,Lisheng L. World health organization definition of myocardial infarction:2008–09 revision. Int J Epidemiol 2011;40(1):139–46.

[9] Reichlin T, Hochholzer W, Bassetti S, Steuer S, Stelzig C, Hartwiger S, et al.Early diagnosis of myocardial infarction with sensitive cardiac troponinassays. New England J Med 2009;361(9):858–67.

[10] Go AS, Mozaffarian D, Roger VL, Benjamin EJ, Berry JD, Borden WB, et al. Heartdisease and stroke statistics–2013 update: a report from the american heartassociation. Circulation 2013;127(1):e6.

[11] Martis RJ, Acharya UR, Adeli H. Current methods in electrocardiogramcharacterization. Comput Biol Med 2014;48:133–49.

[12] Bastien M, Poirier P, Lemieux I, Després J-P. Overview of epidemiology andcontribution of obesity to cardiovascular disease. Prog Cardiovasc Dis2014;56(4):369–81.

[13] Aje TO, Miller M. Cardiovascular disease: a global problem extending into thedeveloping world. World J Cardiol 2009;1(1):3–10.

[14] Acharya UR, Faust O, Sree V, Swapna G, Martis RJ, Kadri NA, Suri JS. Linear andnonlinear analysis of normal and cad-affected heart rate signals. ComputMethods Programs Biomed 2014;113(1):55–68.

[15] Otto CM. Textbook of clinical echocardiography: expert consult-online. Elsevier Health Sciences; 2013.

[16] Thygesen K, Alpert JS, Jaffe AS, White HD, Simoons ML, Chaitman BR, et al.Third universal definition of myocardial infarction. J Am Coll Cardiol 2012;60(16):1581–98.

[17] Stein JH, Korcarz CE, Hurst RT, Lonn E, Kendall CB, Mohler ER, et al. Use ofcarotid ultrasound to identify subclinical vascular disease and evaluatecardiovascular disease risk: a consensus statement from the american society

http://refhub.elsevier.com/S1120-1797(16)31101-2/h0005





































of echocardiography carotid intima-media thickness task force endorsed bythe society for vascular medicine. J Am Soc Echocardiogr 2008;21(2):93–111.

[18] Peters SA, Den Ruijter HM, Bots ML, Moons KG. Improvements in riskstratification for the occurrence of cardiovascular disease by imagingsubclinical atherosclerosis: a systematic review. Heart 2012;98(3):177–84.

[19] Crouse JR, Harpold GH, Kahl FR, Toole JF, McKinney W. Evaluation of a scoringsystem for extracranial carotid atherosclerosis extent with b-modeultrasound. Stroke 1986;17(2):270–5.

[20] Evensen K, Dahl A, Ronning OM, Russell D. The assessment of carotidatherosclerosis using a new multipurpose ultrasound probe. J Neuroimaging2015;25(2):232–7.

[21] Wyman RA, Mays ME, McBride PE, Stein JH. Ultrasound-detected carotidplaque as a predictor of cardiovascular events. Vasc Med 2006;11(2):123–30.

[22] Acharya RU, Faust O, Alvin APC, Sree SV, Molinari F, Saba L, Nicolaides A, SuriJS. Symptomatic vs. asymptomatic plaque classification in carotid ultrasound.J Med Syst 2012;36(3):1861–71.

[23] Acharya UR, Faust O, Sree SV, Molinari F, Suri JS. Thyroscreen system: highresolution ultrasound thyroid image characterization into benign andmalignant classes using novel combination of texture and discrete wavelettransform. Comput Methods Programs Biomed 2012;107(2):233–41.

[24] Acharya UR, Faust O, Sree SV, Molinari F, Saba L, Nicolaides A, Suri JS. Anaccurate and generalized approach to plaque characterization in 346 carotidultrasound scans. IEEE Trans Instrum Meas 2012;61(4):1045–53.

[25] Acharya U, Faust O, Sree SV, Molinari F, Garberoglio R, Suri J. Cost-effectiveand non-invasive automated benign & malignant thyroid lesion classificationin 3d contrast-enhanced ultrasound using combination of wavelets andtextures: a class of thyroscanTM algorithms. Technol Cancer Res Treat2011;10(4):371–80.

[26] Bosch HG, van Burken G, Nijland F, Reiber JHC. Overview of automatedquantitation techniques in 2d echocardiography. In: What’s new incardiovascular imaging? Springer; 1998. p. 363–76.

[27] Acharya UR, Faust O, Sree SV, Molinari F, Saba L, et al. Symptomatic versusasymptomatique plaque classification in carotid ultrasound. In: Multi-modality atherosclerosis imaging and diagnosis. Springer; 2014. p. 399–408.

[28] Ng K, Faust O, Sudarshan V, Chattopadhyay S. Data overloading in medicalimaging: Emerging issues, challenges and opportunities in efficient datamanagement. J Med Imaging Health Inf 2015;5(4):755–64.

[29] Fenster A, Downey DB. 3-d ultrasound imaging: a review, engineering inmedicine and biology magazine. IEEE 1996;15(6):41–51.

[30] Frinking PJ, Bouakaz A, Kirkhorn J, Ten Cate FJ, De Jong N. Ultrasound contrastimaging: current and new potential methods. Ultrasound Med Biol 2000;26(6):965–75.

[31] Faust O, Yu W, Acharya UR. The role of real-time in biomedical science: ameta-analysis on computational complexity, delay and speedup. Comput BiolMed 2015;58:73–84.

[32] Araki T, Ikeda N, Shukla D, Londhe ND, Shrivastava VK, Banchhor SK, et al. Anew method for ivus-based coronary artery disease risk stratification: a linkbetween coronary & carotid ultrasound plaque burdens. Comput MethodsPrograms in Biomed 2015:161–79.

[33] Sudarshan VK, Acharya UR, Ng E, San Tan R, Chou SM, Ghista DN. Anintegrated index for automated detection of infarcted myocardium fromcross-sectional echocardiograms using texton-based features (part 1).Comput Biol Med 2016:231–40.

[34] Barnett GO. Computers in patient care. New England J Med 1968;279(24):1321–7.

[35] Jick H, Vasilakis C, Weinrauch LA, Meier CR, Jick SS, Derby LE. A population-based study of appetite-suppressant drugs and the risk of cardiac-valveregurgitation. New England J Med 1998;339(11):719–24.

[36] Bruce RA, Yarnall SR. Computer-aided diagnosis of cardiovascular disorders. JChronic Dis 1966;19(4):473–84.

[37] Huether SE, McCance KL. Understanding pathophysiology. Elsevier HealthSciences; 2013.

[38] Steinberg D, Gotto Jr AM. Preventing coronary artery disease by loweringcholesterol levels: fifty years from bench to bedside. Jama 1999;282(21):2043–50.

[39] American Heart Association, Statistical fact sheet 2012 update: women andcardiovascular disease; 2012.

[40] Lown B, Wolf M. Approaches to sudden death from coronary heart disease.Circulation 1971;44(1):130–42.

[41] Forbes C, Quek RG, Deshpande S, Worthy G, Ross J, Kleijnen J, et al.Relationship between changes in coronary atherosclerotic plaque burdenmeasured by intravascular ultrasound and cardiovascular disease outcomes:a systematic literature review. Curr Med Res Opin 2016:1–8.

[42] Hwang DS, Shin ES, Kim SJ, Lee JH, Kim JM, Lee S-G. Early differential changesin coronary plaque composition according to plaque stability following statininitiation in acute coronary syndrome: classification and analysis byintravascular ultrasound-virtual histology. Yonsei Med J 2013;54(2):336–44.

[43] Chiocchi M, Chiaravalloti A, Morosetti D, Loreni G, Gandini R, Mancino S, et al.Virtual histology-intravascular ultrasound as a diagnostic alternative formorphological characterization of carotid plaque: comparison with histologyand high-resolution magnetic resonance findings. J Cardiovasc Med2014:1–17.

[44] Pedrigi RM, de Silva R, Bovens SM, Mehta VV, Petretto E, Krams R. Thin-capfibroatheroma rupture is associated with a fine interplay of shear and wallstress. Arteriosclerosis, Thrombosis Vasc Biol 2014;34(10):2224–31.

[45] Guo S, Wang R, Yang Z, Li K, Wang Q. Effects of atorvastatin on serum lipids,serum inflammation and plaque morphology in patients with stableatherosclerotic plaques. Exp Ther Med 2012;4(6):1069–74.

[46] Banach M, Serban C, Sahebkar A, Mikhailidis DP, Ursoniu S, Ray KK, et al.Impact of statin therapy on coronary plaque composition: a systematicreview and meta-analysis of virtual histology intravascular ultrasoundstudies. BMC Med 2015;13(1):1–20.

[47] Lee SWL, Hau WKT, Kong SL, Chan KK, Chan P-H, Lam SC, et al. Virtualhistology findings and effects of varying doses of atorvastatin on coronaryplaque volume and composition in statin-naive patients. Circulation J2012;76(11):2662–72.

[48] Nasu K, Tsuchikane E, Katoh O, Tanaka N, Kimura M, Ehara M, et al. Effect offluvastatin on progression of coronary atherosclerotic plaque evaluated byvirtual histology intravascular ultrasound. JACC: Cardiovasc Interventions2009;2(7):689–96.

[49] Puri R, Libby P, Nissen SE, Wolski K, Ballantyne CM, Barter PJ, et al. Long-termeffects of maximally intensive statin therapy on changes in coronaryatheroma composition: insights from saturn. Eur Heart J CardiovascImaging 2014;15(4):380–8.

[50] Taguchi I, Oda K, Yoneda S, Kageyama M, Kanaya T, Toyoda S, et al. Evaluationof serial changes in tissue characteristics during statin-induced plaqueregression using virtual histology-intravascular ultrasound studies. Am JCardiol 2013;111(9):1246–52.

[51] Eshtehardi P, McDaniel MC, Dhawan SS, Binongo J, Krishnan SK, Golub L, et al.Effect of intensive atorvastatin therapy on coronary atherosclerosisprogression, composition, arterial remodeling, and microvascular function. JInvasive Cardiol 2012;24(10):522–9.

[52] Hong M-K, Park D-W, Lee C-W, Lee S-W, Kim Y-H, Kang D-H, et al. Effects ofstatin treatments on coronary plaques assessed by volumetric virtualhistology intravascular ultrasound analysis. JACC: Cardiovasc Interventions2009;2(7):679–88.

[53] Nozue T, Yamamoto S, Tohyama S, Umezawa S, Kunishima T, Sato A, et al.Statin treatment for coronary artery plaque composition based onintravascular ultrasound radiofrequency data analysis. Am Heart J 2012;163(2):191–9.

[54] De Korte CL, Van Der Steen AF. Intravascular ultrasound elastography: anoverview. Ultrasonics 2002;40(1):859–65.

[55] Brusseau E, Perrey C, Delachartre P, Vogt M, Vray D, Ermert H. Axial strainimaging using a local estimation of the scaling factor from rf ultrasoundsignals. Ultrason Imaging 2000;22(2):95–107.

[56] Meijboom WB, Van Mieghem CA, van Pelt N, Weustink A, Pugliese F, MolletNR, et al. Comprehensive assessment of coronary artery stenoses: computedtomography coronary angiography versus conventional coronaryangiography and correlation with fractional flow reserve in patients withstable angina. J Am Coll Cardiol 2008;52(8):636–43.

[57] Ropers D, Baum U, Pohle K, Anders K, Ulzheimer S, Ohnesorge B, et al.Detection of coronary artery stenoses with thin-slice multi-detector rowspiral computed tomography and multiplanar reconstruction. Circulation2003;107(5):664–6.

[58] Schwitter J, Nanz D, Kneifel S, Bertschinger K, Büchi M, Knüsel PR, et al.Assessment of myocardial perfusion in coronary artery disease by magneticresonance a comparison with positron emission tomography and coronaryangiography. Circulation 2001;103(18):2230–5.

[59] Marcus ML, Armstrong ML, Heistad DD, Eastham CL, Mark AL. Comparison ofthree methods of evaluating coronary obstructive lesions: postmortemarteriography, pathologic examination and measurement of regionalmyocardial perfusion during maximal vasodilation. Am J Cardiol 1982;49(7):1699–706.

[60] Harrison DG, White CW, Hiratzka LF, Doty DB, Barnes DH, Eastham C, MarcusML. The value of lesion cross-sectional area determined by quantitativecoronary angiography in assessing the physiologic significance of proximalleft anterior descending coronary arterial stenoses. Circulation 1984;69(6):1111–9.

[61] Govindaraju K, Badruddin IA, Viswanathan GN, Ramesh S, Badarudin A.Evaluation of functional severity of coronary artery disease and fluiddynamics’ influence on hemodynamic parameters: a review. Phys Med2013;29(3):225–32.

[62] Efstathopoulos EP, Patatoukas G, Pantos I, Benekos O, Katritsis D, Kelekis NL.Measurement of systolic and diastolic arterial wall shear stress in theascending aorta. Phys Med 2008;24(4):196–203.

[63] Thygesen K, Alpert JS, White HD. Universal definition of myocardialinfarction. J Am Coll Cardiol 2007;50(22):2173–95.

[64] Davies MJ, Thomas AC. Plaque fissuring–the cause of acute myocardialinfarction, sudden ischaemic death, and crescendo angina. Br Heart J 1985;53(4):363–73.

[65] Rentrop P, Blanke H, Karsch K, Kaiser H, Köstering H, Leitz K. Selectiveintracoronary thrombolysis in acute myocardial infarction and unstableangina pectoris. Circulation 1981;63(2):307–17.

[66] Burke AP, Virmani R. Pathophysiology of acute myocardial infarction. MedClin North Am 2007;91(4):553–72.

[67] Sudarshan V, Acharya UR, Ng EY-K, Meng CS, San Tan R, Ghista DN, et al.Automated identification of infarcted myocardium tissue characterizationusing ultrasound images: a review. IEEE Rev Biomed Eng 2015;8:86–97.

[68] Nesto RW, Kowalchuk GJ. The ischemic cascade: temporal sequence ofhemodynamic, electrocardiographic and symptomatic expressions ofischemia. Am J Cardiol 1987;59(7):C23–30.












































































































































































[69] Aaronson PI, Ward JP, Connolly MJ. The cardiovascular system at aglance. John Wiley & Sons; 2012.

[70] Moise A, Théroux P, Taeymans Y, Descoings B, Lespérance J, Waters DD, et al.Unstable angina and progression of coronary atherosclerosis. New England JMed 1983;309(12):685–9.

[71] Falk E. Plaque rupture with severe pre-existing stenosis precipitatingcoronary thrombosis. characteristics of coronary atherosclerotic plaquesunderlying fatal occlusive thrombi. Br Heart J 1983;50(2):127–34.

[72] DeWood MA, Spores J, Notske R, Mouser LT, Burroughs R, Golden MS, LangHT. Prevalence of total coronary occlusion during the early hours oftransmural myocardial infarction. New England J Med 1980;303(16):897–902.

[73] Ambrose JA, Tannenbaum MA, Alexopoulos D, Hjemdahl-Monsen CE, Leavy J,Weiss M, et al. Angiographic progression of coronary artery disease and thedevelopment of myocardial infarction. J Am Coll Cardiol 1988;12(1):56–62.

[74] Lang RM, Badano LP, Mor-Avi V, Afilalo J, Armstrong A, Ernande L, et al.Recommendations for cardiac chamber quantification by echocardiographyin adults: an update from the american society of echocardiography and theeuropean association of cardiovascular imaging. J Am Soc Echocardiogr2015;28(1):1–39.

[75] Maton A. Human biology and health. Prentice Hall; 1997.[76] Muller JE, Abela GS, Nesto RW, Tofler GH. Triggers, acute risk factors and

vulnerable plaques: the lexicon of a new frontier. J Am Coll Cardiol 1994;23(3):809–13.

[77] Kolodgie FD, Nakazawa G, Sangiorgi G, Ladich E, Burke AP, Virmani R.Pathology of atherosclerosis and stenting. Neuroimaging Clin North Am2007;17(3):285–301.

[78] Falk E, Shah PK, Fuster V. Coronary plaque disruption. Circulation 1995;92(3):657–71.

[79] Chaniotis A, Kaiktsis L, Katritsis D, Efstathopoulos E, Pantos I, Marmarellis V.Computational study of pulsatile blood flow in prototype vessel geometries ofcoronary segments. Phys Med 2010;26(3):140–56.

[80] Paraskevas KI, Mikhailidis DP, Moore WS, Veith FJ. Optimal contemporarymanagement of symptomatic and asymptomatic carotid artery stenosis.Vascular 2011;19(3):117–20.

[81] Milei J, Parodi JC, Alonso GF, Barone A, Grana D, Matturri L. Carotid ruptureand intraplaque hemorrhage: immunophenotype and role of cells involved.Am Heart J 1998;136(6):1096–105.

[82] Takaya N, Yuan C, Chu B, Saam T, Underhill H, Cai J, et al. Association betweencarotid plaque characteristics and subsequent ischemic cerebrovascularevents a prospective assessment with mri–initial results. Stroke 2006;37(3):818–23.

[83] Milei J, Parodi J, Fernandez AG, Barone A, Beigelman R, Ferreira L, et al. Carotidatherosclerosis. immunocytochemical analysis of the vascular and cellularcomposition in endarterectomies. Cardiologia (Rome, Italy) 1996;41(6):535–42.

[84] Johnson JM, Kennelly MM, Decesare D, Morgan S, Sparrow A. Natural historyof asymptomatic carotid plaque. Arch Surgery 1985;120(9):1010–2.

[85] Golledge J, Greenhalgh RM, Davies AH. The symptomatic carotid plaque.Stroke 2000;31(3):774–81.

[86] Mughal MM, Khan MK, DeMarco JK, Majid A, Shamoun F, Abela GS.Symptomatic and asymptomatic carotid artery plaque. Expert RevCardiovasc Ther 2011;9(10):1315–30.

[87] Ota H, Yu W, Underhill HR, Oikawa M, Dong L, Zhao X, et al. Hemorrhage andlarge lipid-rich necrotic cores are independently associated with thin orruptured fibrous caps an in vivo 3t mri study. Arteriosclerosis ThrombosisVasc Biol 2009;29(10):1696–701.

[88] Petty GW, Brown RD, Whisnant JP, Sicks JD, O’Fallon WM, Wiebers DO.Ischemic stroke subtypes a population-based study of incidence and riskfactors. Stroke 1999;30(12):2513–6.

[89] Roubin GS, New G, Iyer SS, Vitek JJ, Al-Mubarak N, Liu MW, et al. Immediateand late clinical outcomes of carotid artery stenting in patients withsymptomatic and asymptomatic carotid artery stenosis a 5-yearprospective analysis. Circulation 2001;103(4):532–7.

[90] Leng GC, Lee AJ, Fowkers FGR, Whiteman M, Dunbar J, Housley E, Ruckley CV.Incidence, natural history and cardiovascular events in symptomatic andasymptomatic peripheral arterial disease in the general population. Int JEpidemiol 1996;25(6):1172–81.

[91] Geroulakos G, Ramaswami G, Nicolaides A, James K, Labropoulos N, Belcaro G,Holloway M. Characterization of symptomatic and asymptomatic carotidplaques using high-resolution real-time ultrasonography. Br J Surgery1993;80(10):1274–7.

[92] Hemmati H, Kamli-Asl A, Talebpour A, Shirani S. Semi-automatic 3dsegmentation of carotid lumen in contrast-enhanced computedtomography angiography images. Phys Med 2015;31(8):1098–104.

[93] Gompels B. High definition imaging of carotid arteries using a standardcommercial ultrasound b scanner. a preliminary report. Br J Radiol 1979;52(620):608–19.

[94] Wolff T, Guirguis-Blake J, Miller T, Gillespie M, Harris R. Screening for carotidartery stenosis: an update of the evidence for the us preventive services taskforce. Ann Internal Med 2007;147(12):860–70.

[95] Beletsky VY, Kelley RE, Fowler M, Phifer T. Ultrasound densitometric analysisof carotid plaque composition pathoanatomic correlation. Stroke 1996;27(12):2173–7.

[96] Cothren RM, Shekhar R, Tuzcu EM, Nissen SE, Cornhill JF, Vince DG. Three-dimensional reconstruction of the coronary artery wall by image fusion of

intravascular ultrasound and bi-plane angiography. Int J Cardiac Imaging2000;16(2):69–85.

[97] Acharya UR, Faust O, Molinari F, Sree SV, Junnarkar SP, Sudarshan V.Ultrasound-based tissue characterization and classification of fatty liverdisease: a screening and diagnostic paradigm. Knowl Based Syst2015;75:66–77.

[98] Jenny NZN, Faust O, Yu W. Automated classification of normal and prematureventricular contractions in electrocardiogram signals. J Med Imaging HealthInf 2014;4(6):886–92.

[99] Doi K. Computer-aided diagnosis in medical imaging: historical review,current status and future potential. Comput Med Imaging Graphics 2007;31(4):198–211.

[100] Faust O, Acharya R, Ng EY-K, Ng K-H, Suri JS. Algorithms for the automateddetection of diabetic retinopathy using digital fundus images: a review. J MedSyst 2012;36(1):145–57.

[101] Nayak J, Bhat PS, Acharya UR, Faust O, Min LC. Computer-based identificationof cataract and cataract surgery efficacy using optical images. J Mech MedBiol 2009;9(04):589–607.

[102] Faust O, Acharya UR, Ng E, Hong TJ, Yu W. Application of infraredthermography in computer aided diagnosis. Infrared Phys Technol2014;66:160–75.

[103] Kremkau FW. Diagnostic ultrasound: principles and instruments. WBSaunders Company; 2001.

[104] Rosenfield K, Kelly SM, Fields CD, Pastore JO, Weinstein R, Palefski P, et al.Noninvasive assessment of peripheral vascular disease by color flow doppler/two-dimensional ultrasound. Am J Cardiol 1989;64(3):247–51.

[105] Lang RM, Bierig M, Devereux RB, Flachskampf FA, Foster E, Pellikka PA, et al.Recommendations for chamber quantification. Eur Heart J-CardiovascImaging 2006;7(2):79–108.

[106] Shutliov V, Alferieff M, Beyer RT. Fundamental physics of ultrasound. J AcoustSoc Am 1990;88(6). 2906–2906.

[107] Kremkau F, Taylor K. Artifacts in ultrasound imaging. J Ultrasound Med1986;5(4):227–37.

[108] Wang Y, Georgescu B, Comaniciu D, Houle H. Learning-based 3d myocardialmotion flowestimation using high frame rate volumetric ultrasound data. In:IEEE international symposium on biomedical imaging: from nano to macro,2010. IEEE; 2010. p. 1097–100.

[109] Elen A, Choi HF, Loeckx D, Gao H, Claus P, Suetens P, et al. Three-dimensionalcardiac strain estimation using spatio–temporal elastic registration ofultrasound images: A feasibility study. IEEE Trans Med Imaging 2008;27(11):1580–91.

[110] Buzug TM. Computed tomography: from photon statistics to modern cone-beam CT. Springer Science & Business Media; 2008.

[111] Paragios N, Jolly M-P, Taron M, Ramaraj R. Active shape models andsegmentation of the left ventricle in echocardiography. In: Scale space andPDE methods in computer vision. Springer; 2005. p. 131–42.

[112] Mallery J, Griffith J, Gessert J, Morcos N, Tobis J, Henry W. Intravascularultrasound imaging catheter assessment of normal and atheroscleroticarterial wall thickness. J Am Coll Cardiol 1988;11:22A.

[113] Gussenhoven EJ, Essed CE, Lancée CT, Mastik F, Frietman P, van Egmond FC,et al. Arterial wall characteristics determined by intravascular ultrasoundimaging: an in vitro study. J Am Coll Cardiol 1989;14(4):947–52.

[114] Sheikh KH, Davidson CJ, Kisslo KB, Harrison JK, Himmelstein SI, Kisslo J,Bashore TM. Comparison of intravascular ultrasound, external ultrasoundand digital angiography for evaluation of peripheral artery dimensions andmorphology. Am J Cardiol 1991;67(9):817–22.

[115] Mallery JA, Tobis JM, Griffith J, Gessert J, McRae M, et al. Assessment ofnormal and atherosclerotic arterial wall thickness with an intravascularultrasound imaging catheter. Am Heart J 1990;119(6):1392–400.

[116] Tobis JM, Mallery J, Mahon D, Lehmann K, Zalesky P, Griffith J, et al.Intravascular ultrasound imaging of human coronary arteries in vivo.Analysis of tissue characterizations with comparison to in vitro histologicalspecimens. Circulation 1991;83(3):913–26.

[117] de Vries BMW, van Dam GM, Tio RA, Hillebrands J-L, Slart RH, Zeebregts CJ.Current imaging modalities to visualize vulnerability within theatherosclerotic carotid plaque. J Vasc Surgery 2008;48(6):1620–9.

[118] Diethrich EB, Margolis MP, Reid DB, Burke A, Ramaiah V, Rodriguez-Lopez JA,et al. Virtual histology intravascular ultrasound assessment of carotid arterydisease: the carotid artery plaque virtual histology evaluation (capital) study.J Endovasc Ther 2007;14(5):676–86.

[119] Übeyl ED, Güler _I. Feature extraction from doppler ultrasound signals forautomated diagnostic systems. Comput Biol Med 2005;35(9):735–64.

[120] Park S-Y, Park JM, Kim J-I, Kim H, Kim IH, Ye S-J. Textural feature calculatedfrom segmental fluences as a modulation index for vmat. Physica Med2015;31(8):981–90.

[121] Faust O, Bairy MG. Nonlinear analysis of physiological signals: a review. JMech Med Biol 2012;12(04):1240015.

[122] Haralick RM, Shanmugam K, Dinstein IH. Textural features for imageclassification. IEEE Trans Syst Man Cybern 1973;6:610–21.

[123] Krishnan MMR, Faust O. Automated glaucoma detection using hybrid featureextraction in retinal fundus images. J Mech Med Biol 2013;13(01):1350011.

[124] Jensen A, la Cour-Harbo A. Ripples in mathematics: the discrete wavelettransform. Springer Science & Business Media; 2001.

[125] Nason GP, Silverman BW. The stationary wavelet transform and somestatistical applications. Lecture Notes In Statistics-new York-springer Verlag1995. 281–281.















































































































































































[126] Candes EJ, Donoho DL. Curvelets: a surprisingly effective nonadaptiverepresentation for objects with edges. Tech. rep, DTIC Document; 2000.

[127] Candes EJ, Donoho DL. Recovering edges in ill-posed inverse problems:optimality of curvelet frames. Ann Stat 2002:784–842.

[128] Candès EJ, Donoho DL. New tight frames of curvelets and optimalrepresentations of objects with piecewise c2 singularities. Commun PureAppl Math 2004;57(2):219–66.

[129] Guo Y, Zhao G, Pietikäinen M. Texture classification using a linearconfiguration model based descriptor. In: BMVC. Citeseer; 2011. p. 1–10.

[130] Roweis ST, Saul LK. Nonlinear dimensionality reduction by locally linearembedding. Science 2000;290(5500):2323–6.

[131] Thangavel M, Chandrasekaran M, Madheswaran M. Carotid plaqueclassification using contourlet features and support vector machines. JComput Sci 2014;10(9):1642.

[132] Weber M, Welling M, Perona P. Unsupervised learning of models forrecognition. Springer; 2000.

[133] Hastie T, Tibshirani R, Friedman J. Unsupervised learning. Springer; 2009.[134] Kotsiantis S. Supervised machine learning: a review of classification

techniques. In: Proceedings of the 2007 conference on emerging artificialintelligence applications in computer engineering: real word ai systems withapplications in ehealth, hci, information retrieval and pervasivetechnologies. IOS Press; 2007. p. 3–24.

[135] Baldi P, Brunak S, Chauvin Y, Andersen CA, Nielsen H. Assessing the accuracyof prediction algorithms for classification: an overview. Bioinformatics2000;16(5):412–24.

[136] Shalbaf A, Behnam H, Alizade-Sani Z, Shojaifard M. Automatic classification ofleft ventricular regional wall motion abnormalities in echocardiographyimages using nonrigid image registration. J Digital Imaging 2013;26(5):909–19.

[137] Prati F, Arbustini E, Labellarte A, Dal Bello B, Sommariva L, Mallus M, et al.Correlation between high frequency intravascular ultrasound andhistomorphology in human coronary arteries. Heart 2001;85(5):567–70.

[138] Kohonen T. Self-organized formation of topologically correct feature maps.Biol Cybern 1982;43(1):59–69.

[139] Tipping ME. Sparse bayesian learning and the relevance vector machine. JMach Learn Res 2001;1:211–44.

[140] Specht DF. Probabilistic neural networks. Neural Networks 1990;3(1):109–18.

[141] Vapnik V. The nature of statistical learning theory. Springer Science &Business Media; 2013.

[142] Song Z, Ji Z, Ma J-G, Sputh B, Acharya UR, Faust O. A systematic approach toembedded biomedical decision making. Comput Methods Programs Biomed2012;108(2):656–64.

[143] Kennedy J. Particle swarm optimization. In: Encyclopedia of machinelearning. Springer; 2011. p. 760–6.

[144] Kennedy J, Eberhart RC. A discrete binary version of the particle swarmalgorithm. IEEE international conference on systems, man, and cybernetics,1997. Computational cybernetics and simulation., 1997, vol. 5. IEEE; 1997. p.4104–8.

[145] Guvenir HA, Erel E. Multicriteria inventory classification using a geneticalgorithm. Eur J Oper Res 1998;105(1):29–37.

[146] Mahalanobis A, Kumar BV, Sims S. Distance-classifier correlation filters formulticlass target recognition. Appl Opt 1996;35(17):3127–33.

[147] Nissen SE, Gurley JC, Grines CL, Booth DC, McClure R, Berk M, et al.Intravascular ultrasound assessment of lumen size and wall morphology innormal subjects and patients with coronary artery disease. Circulation1991;84(3):1087–99.

[148] Goar FGS, Pinto FJ, Alderman EL, Fitzgerald PJ, Stadius ML, Popp RL.Intravascular ultrasound imaging of angiographically normal coronaryarteries: an in vivo comparison with quantitative angiography. J Am CollCardiol 1991;18(4):952–8.

[149] Mintz GS, Douek P, Pichard AD, Kent KM, Satler LF, Popma JJ, Leon MB. Targetlesion calcification in coronary artery disease: an intravascular ultrasoundstudy. J Am Coll Cardiol 1992;20(5):1149–55.

[150] Potkin BN, Keren G, Mintz GS, Douek PC, Pichard AD, Satler LF, et al. Arterialresponses to balloon coronary angioplasty: an intravascular ultrasoundstudy. J Am Coll Cardiol 1992;20(4):942–51.

[151] Hermiller JB, Buller CE, Tenaglia AN, Kisslo KB, Phillips HR, Bashore TM, et al.Unrecognized left main coronary artery disease in patients undergoinginterventional procedures. Am J Cardiol 1993;71(2):173–6.

[152] Alfonso F, Macaya C, Goicolea J, Hernandez R, Segovia J, Zamorano J, et al.Determinants of coronary compliance in patients with coronary arterydisease: an intravascular ultrasound study. J Am Coll Cardiol 1994;23(4):879–84.

[153] Mintz GS, Painter JA, Pichard AD, Kent KM, Satler LF, Popma JJ, et al.Atherosclerosis in angiographically normal coronary artery referencesegments: an intravascular ultrasound study with clinical correlations. JAm Coll Cardiol 1995;25(7):1479–85.

[154] Mintz GS, Popma JJ, Pichard AD, Kent KM, Satler LF, Wong SC, et al. Arterialremodeling after coronary angioplasty a serial intravascular ultrasoundstudy. Circulation 1996;94(1):35–43.

[155] Klingensmith JD, Shekhar R, Vince DG. Evaluation of three-dimensionalsegmentation algorithms for the identification of luminal and medial-adventitial borders in intravascular ultrasound images. IEEE Trans MedImaging 2000;19(10):996–1011.

[156] Schoenhagen P, Ziada KM, Kapadia SR, Crowe TD, Nissen SE, Tuzcu EM. Extentand direction of arterial remodeling in stable versus unstable coronarysyndromes an intravascular ultrasound study. Circulation 2000;101(6):598–603.

[157] Schartl M, Bocksch W, Koschyk DH, Voelker W, Karsch KR, Kreuzer J, et al. G.A.I.U.S. Investigators, Use of intravascular ultrasound to compare effects ofdifferent strategies of lipid-lowering therapy on plaque volume andcomposition in patients with coronary artery disease. Circulation 2001;104(4):387–92.

[158] Çomak E, Arslan A, Türkoglu _I. A decision support system based on supportvector machines for diagnosis of the heart valve diseases. Comput Biol Med2007;37(1):21–7.

[159] Babaoglu I, Fndk O, Bayrak M. Effects of principle component analysis onassessment of coronary artery diseases using support vector machine. ExpertSyst Appl 2010;37(3):2182–5.

[160] Mukhopadhyay A, Qian Z, Bhandarkar SM, Liu T, Voros S, Rinehart S.Morphological analysis of the left ventricular endocardial surface using abag-of-features descriptor. IEEE J Biomed Health Inf 2015;19(4):1483–93.

[161] Skorton DJ, Melton H, Pandian NG, Nichols J, Koyanagi S, Marcus ML, et al.Detection of acute myocardial infarction in closed-chest dogs by analysis ofregional two-dimensional echocardiographic gray-level distributions.Circulation Res 1983;52(1):36–44.

[162] Kamath MV, Way RC, Ghista DN, Srinivasan TM, Wu C, Smeenk S, et al.Detection of myocardial scars in neonatal infants from computerizedechocardiographic texture analysis. Eng Med 1986;15(3):137–41.

[163] Tak T, Visser C, Rahimtoola SH, Chandraratna PAN. Detection of acutemyocardial infarction with digital image processing of two-dimensionalechocardiograms. Am Heart J 1992;124(2):289–93.

[164] Mojsilovic A, Popovic MV, Neškovic AN, Popovic AD. Wavelet imageextension for analysis and classification of infarcted myocardial tissue. IEEETrans Biomed Eng 1997;44(9):856–66.

[165] Neskovic AN, Mojsilovic A, Jovanovic T, Vasiljevic J, Popovic M, Marinkovic J,et al. Myocardial tissue characterization after acute myocardial infarctionwith wavelet image decomposition a novel approach for the detection ofmyocardial viability in the early postinfarction period. Circulation 1998;98(7):634–41.

[166] Kotani J-I, Mintz GS, Castagna MT, Pinnow E, Berzingi CO, Bui AB, et al.Intravascular ultrasound analysis of infarct-related and non–infarct-relatedarteries in patients who presented with an acute myocardial infarction.Circulation 2003;107(23):2889–93.

[167] Ehara S, Kobayashi Y, Yoshiyama M, Shimada K, Shimada Y, Fukuda D, et al.Spotty calcification typifies the culprit plaque in patients with acutemyocardial infarction an intravascular ultrasound study. Circulation2004;110(22):3424–9.

[168] Hong M-K, Mintz GS, Lee CW, Kim Y-H, Lee S-W, Song J-M, et al. Comparisonof coronary plaque rupture between stable angina and acute myocardialinfarction a three-vessel intravascular ultrasound study in 235 patients.Circulation 2004;110(8):928–33.

[169] Chykeyuk K, Clifton DA, Noble JA. Feature extraction and wall motionclassification of 2d stress echocardiography with relevance vector machines.In: IEEE international symposium on biomedical imaging: from nano tomacro, 2011. IEEE; 2011. p. 677–80.

[170] Moldovanu S, Moraru L, Bibicu D. Characterization of myocardium musclebiostructure using first order features. Dig J Nanomater Bios2011;6:1357–65.

[171] Agani N, Abu-Bakar S, Salleh SS. Application of texture analysis inechocardiography images for myocardial infarction tissue. J Teknol 2012;46(1):61–76.

[172] Acharya UR, Sree SV, Krishnan MMR, Krishnananda N, Ranjan S, Umesh P, SuriJS. Automated classification of patients with coronary artery disease usinggrayscale features from left ventricle echocardiographic images. ComputMethods Programs Biomed 2013;112(3):624–32.

[173] Vidya KS, Ng EYK, Acharya UR, Chou SM, San Tan R, Ghista DN. Computer-aided diagnosis of myocardial infarction using ultrasound images with dwt,glcm and hos methods: a comparative study. Comput Biology Med2015;62:86–93.

[174] Sudarshan VK, Acharya UR, Ng E, San Tan R, Chou SM, Ghista DN. Data miningframework for identification of myocardial infarction stages in ultrasound: ahybrid feature extraction paradigm (part 2). Comput Biol Med2016;71:241–51.

[175] Sudarshan VK, Ng E, Acharya UR, Tan RS, Chou SM, Ghista DN. Infarcted leftventricle classification from cross-sectional echocardiograms using relativewavelet energy and entropy features. J Mech Med Biol 2016;1640009.

[176] Tsiaparas NN, Golemati S, Andreadis I, Stoitsis J, Valavanis I, Nikita KS.Assessment of carotid atherosclerosis from b-mode ultrasound images usingdirectional multiscale texture features. Meas Sci Technol 2012;23(11):114004.

[177] Christodoulou CI, Pattichis CS, Pantziaris M, Nicolaides A. Texture-basedclassification of atherosclerotic carotid plaques. IEEE Transac Med Imaging2003;22(7):902–12.

[178] Acharya UR, Faust O, Sree SV, Alvin APC, Krishnamurthi G, Seabra JC, et al.Atheromatictm: Symptomatic vs. asymptomatic classification of carotidultrasound plaque using a combination of hos, dwt & texture. In:Engineering in medicine and biology society, EMBC, 2011 annualinternational conference of the IEEE. IEEE; 2011. p. 4489–92.














































































































































































[179] Acharya UR, Sree SV, Krishnan MMR, Molinari F, Saba L, Ho SYS, et al.Atherosclerotic risk stratification strategy for carotid arteries using texture-based features. Ultrasound Med Biol 2012;38(6):899–915.

[180] Acharya UR, Faust O, Alvin A, Krishnamurthi G, Seabra JC, Sanches J, Suri JS.Understanding symptomatology of atherosclerotic plaque by image-basedtissue characterization. Comput Methods Programs Biomed 2013;110(1):66–75.

[181] Rajendra Acharya U, Muthu Rama Krishnan M, Vinitha Sree S, Sanches J,Shafique S, Nicolaides A, et al. Plaque tissue characterization andclassification in ultrasound carotid scans: a paradigm for vascular featureamalgamation. IEEE Trans Instrum Meas 2013;62(2):392–400.

[182] Acharya UR, Mookiah MRK, Sree SV, Afonso D, Sanches J, Shafique S, et al.Atherosclerotic plaque tissue characterization in 2d ultrasound longitudinalcarotid scans for automated classification: a paradigm for stroke riskassessment. Med Biol Eng Comput 2013;51(5):513–23.

[183] Nair A, Kuban BD, Tuzcu EM, Schoenhagen P, Nissen SE, Vince DG. Coronaryplaque classification with intravascular ultrasound radiofrequency dataanalysis. Circulation 2002;106(17):2200–6.

[184] Irshad K, Millar S, Velu R, Reid AW, Diethrich EB, Reid DB. Virtual histologyintravascular ultrasound in carotid interventions. J Endovasc Ther 2007;14(2):198–207.

[185] Rose J, Nestleroth J, Balasubramaniam K. Utility of feature mapping inultrasonic non-destructive evaluation. Ultrasonics 1988;26(3):124–31.

[186] Faust O, Acharya UR, Tamura T. Formal design methods for reliablecomputer-aided diagnosis: a review. IEEE Rev Biomed Eng 2012;5:15–28.

[187] Fei B, Ng WS, Chauhan S, Kwoh CK. The safety issues of medical robotics.Reliab Eng Syst Saf 2001;73(2):183–92.

[188] Lin C-C, Lin P-Y, Lu P-K, Hsieh G-Y, Lee W-L, Lee R-G. A healthcare integrationsystem for disease assessment and safety monitoring of dementia patients.IEEE Trans Inf Technol Biomed 2008;12(5):579–86.

[189] Elder A, Paterson C. Sharps injuries in uk health care: a review of injury rates,viral transmission and potential efficacy of safety devices. Occup Med2006;56(8):566–74.

[190] Margaria T, Steffen B. Leveraging applications of formal methods, verification,and validation. 4th international symposium on leveraging applications,ISoLA 2010, Heraklion, Crete, Greece, October 18–21, 2010, proceedings, vol.6415. Springer; 2010.

[191] Muradore R, Bresolin D, Geretti L, Fiorini P, Villa T. Robotic surgery. IEEE RobAutom Mag 2011;18(3):24–32.

[192] Acharya UR, Faust O, Ghista DN, Sree SV, Alvin APC, Chattopadhyay S, et al. Asystems approach to cardiac health diagnosis. J Med Imaging Health Inf2013;3(2):261–7.

[193] Faust O, Shetty R, Sree SV, Acharya S, Acharya R, Ng E, et al. Towards thesystematic development of medical networking technology. J Med Syst2011;35(6):1431–45.

[194] Sputh BH, Faust O, Allen AR. Portable csp based design for embedded multi-core systems. In: CPA; 2006, p. 123–34.

[195] Faust O, Yu W. Formal and model driven design of the bright light therapysystem luxamet. J Mech Med Biol 2015;1650065.

[196] Faust O, Acharya R, Sputh BH, Min LC. Systems engineering principles for thedesign of biomedical signal processing systems. Comput Methods ProgramsBiomed 2011;102(3):267–76.

[197] Sudha S, Suresh G, Sukanesh R. Speckle noise reduction in ultrasound imagesby wavelet thresholding based on weighted variance. Int J Comput TheoryEng 2009;1(1):7.

[198] Burckhardt CB. Speckle in ultrasound b-mode scans. Sonics Ultrason 1978;25(1):1–6.

[199] Rosenthal A, Mork P, Li MH, Stanford J, Koester D, Reynolds P. Cloudcomputing: a new business paradigm for biomedical information sharing. JBiomed Inf 2010;43(2):342–53.



























































Computer aided diagnosis of Coronary Artery …...Review paper Computer aided diagnosis of Coronary...

Documents

Transcript of Computer aided diagnosis of Coronary Artery …...Review paper Computer aided diagnosis of Coronary...