A Review of Noninvasive Ultrasound Image Processing Methods in the Analysis of Carotid Plaque Morphology for the Assessment of Stroke Risk

Noninvasive ultrasound imaging of carotid plaques allows for the development of plaque-image analysis methods associated with the risk of stroke. This paper presents several plaque-image analysis methods that have been developed over the past years. The paper begins with a review of clinical methods for visual classification that have led to standardized methods for image acquisition, describes methods for image segmentation and denoizing, and provides an overview of the several texture-feature extraction and classification methods that have been applied. We provide a summary of emerging trends in 3-D imaging methods and plaque-motion analysis. Finally, we provide a discussion of the emerging trends and future directions in our concluding remarks.

cause. Stroke accounted for about one for every 16 deaths in the United States. A recent study by the World Health Organization revealed that by 2015 almost 20 milion people will die from CVDs, mainly from heart disease and stroke [2].
High-Resolution ultrasound has made the noninvasive visualization of the carotid bifurcation possible, and has thus been extensively used in the study of arterial wall changes. Studies include measurement of the thickness of the intima-media complex [intima-media thickness (IMT)], estimation of the severity of stenosis due to atherosclerotic plaques, and plaque characterization in order to assess the risk of stroke [3]- [5] (see Figs. 1 and 2).
Clinical applications of carotid-bifurcation ultrasound include the following: 1) Identification and grading of stenosis of extracranial carotid artery disease often responsible for ischemic strokes, transient ischemic attacks (TIAs), or amaurosis fugax (AF); 2) follow-up after carotid endarterectomy; 3) evaluation of pulsatile neck mass; 4) investigation of asymptomatic neck bruits, where severe internal carotid artery stenosis is used as a predictive factor for future stroke; 5) cardiovascular risk assessment where the presence of carotid bifurcation atherosclerotic plaques is associated with increased cardiovascular mortality [6]- [9]; 6) clinical studies on the effect of lipid lowering and other medications on carotid IMT, which includes plaque thickness [10]- [14]. During the past 20 years, the introduction of computer-aided methods and image standardization has improved the objective assessment of carotid plaque echogenicity [15], [16] and Fig. 2. Examples of segmented (a) asymptomatic and (b) symptomatic plaques. Under each plaque, the type of plaque according to [31] and selected texture features are given. (SGLDM: spatial gray-level-dependence matrices.) heterogeneity [15], [17] and has largely replaced subjective (visual) assessment [3], [18] that had been criticized for its relatively poor reproducibility [19].
In general, computer-aided diagnostic systems require the use of a wide variety of methods that are reviewed in this paper. First, a clinical-image-acquisition protocol is necessary for reducing variability during the acquisition process. Second, an imagenormalization procedure is needed to further standardize the images.
Plaque-image segmentation methods allow us to isolate the region of diagnostic interest. Noise from the extracted plaques can be removed with image despeckling methods. Texture features are subsequently computed over the segmented images. Texture features are then used as inputs to classifiers to provide an overall assessment of the input plaque images.
As an early indicator of CVD, we are also interested in the segmentation and characterization of the intima-media layer. We will provide an extensive summary of intima-media segmentation methods.
Emerging approaches in plaque-ultrasound-image analysis include the recent introduction of 3-D imaging methods, plaque motion analysis, stress and strain imaging, and the use of contrast agents.
A review of visual classification methods is given in Section II. Image segmentation and despeckling methods are reviewed in Section III. A summary of texture-feature extraction and classification methods is given in Section IV. Emerging methods are given in Section V. Concluding remarks on the emerging trends and future directions are given in Section VI.

II. VISUAL CLASSIFICATION IN THE ASSESSMENT OF ATHEROSCLEROTIC PLAQUES IN ULTRASOUND IMAGING
High-resolution ultrasound provides information not only on the degree of carotid artery stenosis, but also on the characteristics of the arterial wall, including the size and consistency of atherosclerotic plaques. Several studies have indicated that "complicated" carotid plaques are often associated with ipsilateral neurological symptoms and share common ultrasonic  TABLE I  ULTRASOUND CAROTID PLAQUE HETEROGENEITY AND CLINICAL IMPLICATIONS characteristics, being more echolucent (weak reflection of ultrasound, and therefore, containing hypoechoic structures) and heterogeneous (having both hypoechoic and hyperechoic areas) [25]- [28]. In contrast, "uncomplicated" plaques that are often asymptomatic tend to be of uniform consistency (uniformly hypoechoic or uniformly hyperechoic) without evidence of ulceration [3], [20], [31].
Different classifications of plaque ultrasonic appearance have been proposed in the literature (see Fig. 2). Reilly et al. classified [3] carotid plaques as homogenous and heterogeneous, defining as homogeneous plaques those with "uniformly bright echoes" that are now known as uniformly hyperechoic (type 4) (see next). Johnson et al. classified plaques as dense and soft [32], Widder et al. classified plaques as echolucent and echogenic based on the their overall level of echo patterns [24], while Gray-Weale et al. described four types: type 1, predominantly echolucent lesions; type 2, echogenic lesions with substantial (>75%) components of echolucency; type 3, predominately echogenic with small area(s) of echolucency occupying less than a quarter of the plaque; and type 4, uniformly dense echogenic lesions [18]. Geroulakos et al. subsequently modified the Gray-Weale classification by using a 50% area cutoff point instead of 75% and by adding a fifth type, which, as a result of heavy calcification on its surface, cannot be correctly classified [31] (see also Fig. 2).
In an effort to improve the reproducibility of visual (subjective) classification, a consensus conference has suggested that echodensity should reflect the overall brightness of the plaque with the term hyperechoic referring to echogenic (white) and the term hypoechoic referring to echolucent (black) plaques [33]. The reference structure, to which plaque echodensity should be compared with, should be blood for hypoechoic, the sternomastoid muscle for isoechoic, and bone for hyperechoic plaques. More recently, a similar method has been used by Polak et al. [34].
In the past, a number of workers had confused echogenicity with homogeneity [3]. It is now realized that measurements of texture are different from measurements of echogenicity. The observation that two different atherosclerotic plaques may have the same overall echogenicity, but frequently have variations of texture within different regions of the plaque has been made, as early as 1983 [35]. The term homogeneous should therefore refer to plaques of uniform consistency, irrespective of whether they are predominantly hypoechoic or hyperechoic. The term heterogeneous should be used for plaques of nonuniform consistency, i.e., having both hypoechoic and hyperechoic components (Gray-Weale et al.  [20], [21], there has been a considerable degree of diversity in terminology used by others, as shown in Table I. Because of this confusion, frequently, plaques having intermediate echogenicity or being complex are inadequately described. For example, echolucent plaques have been considered as heterogeneous [23]. As a reflection of this confusion, a report from the committee on standards for noninvasive vascular testing of the Joint Council of the Society for Vascular Surgery and the North American Chapter of the International Society for Cardiovascular Surgery proposed that carotid plaques should be classified as homogeneous or heterogeneous [36]. Regarding the clinical significance of carotid plaque heterogeneity, it seems that the heterogeneous plaques described in the three studies published in the 1980s (see Table I) include hypoechoic plaques. Also, heterogeneous plaques in all studies listed in Table I contain hypoechoic areas (large or small) and appear to be the plaques that are associated with symptoms or, if found in asymptomatic individuals, they are the plaques that subsequently tend to become symptomatic [30].
An important feature of visual classification that has emerged as an important characteristic of unstable--symptomatic plaques in the last years is the juxtaluminal (the site of the plaque near the lumen) location of an echolucent region in heterogenous plaques, which was shown to be an additional marker of increased risk [37].

III. ULTRASOUND-IMAGE PREPROCESSING AND SEGMENTATION
Visual classification of atherosclerotic plaques on ultrasound is subjective in the sense that if scanning is performed in a relatively dark room, the sonographer reduces the image gain and vice versa. This may explain the relatively poor reproducibility results [19], [34], [38], [39]. In order to overcome this problem, some authors applied linear scaling for image normalization using blood and adventitia as reference points [40].

A. Image Acquisition
The use of a standardized acquisition protocol has been shown to result in reproducible measurements of overall plaque echogenicity with a high inter and intraobserver reproducibility [39], [41].
Essential guidelines for standardized image acquisition include: 1) maximum dynamic range; 2) low persistence level with high frame rates for improved temporal resolution; 3) timegain compensation curve (TGC) sloping through the tissues and vertically through the lumen of the vessel; 4) noise reduction using low gain increased until small amounts of noise appear in the lumen; 5) linear histogram stretching; 6) ultrasound-beam propagation at 90 • to the arterial wall; 7) minimum depth imaging for low attenuation; 8) normalization using clearly visible, hyperechoic adventitia adjacent to the plaque; and 9) acquisition that facilitates subsequent successful image normalization (described next).

B. Normalization
It has been shown that image normalization reduces variability caused by different gain settings, different operators, and different equipment thus allowing more reproducible gray-scale measurements [15], [16], [41], [42]. In a recent study by Griffin et al. [40], the ultrasound images of the common carotid artery (CCA) were normalized using linear histogram stretching. For 8-bit images, the gray-level value of blood was mapped to a value of 0, and the gray level of the middle 2/4th of the adventitia (artery wall) to a value of 190. Thus, ultrasound-image intensity throughout the image is readjusted according to the gray-scale values of two reference regions (blood and adventitia). Thus, to maintain high reproducibility using image normalization, we need to have a representative sample of the adventitia region. This is accomplished by imaging with the ultrasound-beam propagating at right angles to the adventitia so that it is visible adjacent to the plaque. It has been demonstrated that using this method reproducible measurements of grayscale can be obtained when the same subject is being scanned in different rooms by different ultrasonographers and scanners [39], [41].

C. Despeckling
Diagnostic ultrasound-image resolution is significantly limited by speckle noise. Speckle is not truly a noise in the typical engineering sense because its texture often carries useful information about the image being viewed. It is the primary factor that limits the contrast resolution in diagnostic ultrasound imaging, thereby limiting the detectability of small, low-contrast lesions, and making the ultrasound images generally difficult for the nonspecialist to interpret [43]- [45], [47], [56]. Due to the speckle presence, ultrasound experts with insufficient experience may not often draw useful conclusions from the images [6]. Speckle noise also limits the effective application of image processing and analysis algorithms (i.e., edge detection, segmentation) and display in 2-D and volume rendering in 3-D. Therefore, speckle is most often considered as a dominant source of noise in ultrasound imaging and should be filtered out [44]- [46] without affecting important features of the image. In [47], where a review on ultrasound-image segmentation methods has been presented, it was discussed whether speckle should be treated as noise or a feature. It was concluded that from a segmentation perspective, you may chose to remove it or utilize it for the information it contains.
Additionally, there are a number of recent research papers, where ultrasound-imaging despeckling was proposed (e.g., [48]).
As a result, speckle reduces detectability of small, lowcontrast lesions, thus making the ultrasound images generally difficult for the nonspecialist to interpret [43]- [47], [49]- [52], [56]. In a recent study [53], for images of the carotid artery, it was shown that despeckle filtering increases image quality. In addition, normalization combined with speckle reduction filtering also improved the performance of both automated as well as the manual segmentations of the IMT [54] and the plaque [52], [55], and enhanced computer-aided diagnosis [56].

D. Segmentation
Several ultrasound-segmentation algorithms have been reviewed in a recent survey by Noble and Boukerroui [47]. Here, we are primarily interested in reporting on recent algorithms that were used for the segmentation of the plaque and the intima layer.
The IMT of the CCA can serve as an early indicator of the development of CVD, like myocardial infarction and stroke. Previous studies indicated that increase in the IMT of the CCA is directly associated with an increased risk of myocardial infarction and stroke, especially in elderly adults without any history of CVD [57]. Therefore, the development and evaluation of new IMT segmentation techniques is of importance.   II  CAROTID IMT 2-D AND 3-D ULTRASOUND SEGMENTATION TECHNIQUES A gradient-based segmentation method proposed in [66] produced large variability in the measurements, whereas the methods in [67] and [68] were allowing manual corrections. Recent commercial systems, which support IMT segmentation, include snakes-based segmentation in [58] and using a contouring approach in [59].
In a recent study [54], we proposed a snake's segmentation method (Williams and Shah) with manual initialization [see Fig. 1(a)]. In Fig. 1(a), the despeckle filter LSMV (local statistics mean filter based on the mean and the variance of the pixel values in each sliding moving window) was iteratively applied for five times with a moving pixel window of 7 × 7 pixels [54], [56]. We did not find significant differences between the manual and automated IMT measurements. Furthermore, it was shown that when normalization and despeckle filtering are applied prior to IMT or plaque segmentation, the automated segmentation measurements are in better agreement with manual measurements [54]. Furthermore, the estimation and positioning of the initial snake contour may sometimes result to segmentation errors. This should be placed as close as possible to the area of interest, otherwise it may be trapped into local minima or false edges and converge to a wrong location. In the present study in less than 5% of the cases, the positioning of the initial snake contour was not calculated correctly. The applicability of the proposed snakes border detection in cases, where the IMT is larger than 1.4 mm, is not possible, i.e., because for larger IMT, a different initialization procedure, based on plaque segmentation, should be followed as proposed in [55].
The risk of stroke increases with the severity of carotid stenosis and is reduced after carotid endarterectomy [15], [16]. The degree of internal carotid stenosis is the only well-established measurement that is used to assess the risk of stroke [16]. Indeed, it is the only criterion at present used to decide whether carotid endarterectomy is indicated or not [22], [69]. The need for the accurate segmentation of the atherosclerotic carotid plague in ultrasound imaging in order to assess the degree of stenosis is, therefore, a very important task.
In Table III, we summarize several recent methods for segmenting carotid plaques. In summary, we have segmentation methods based on edge detection and snakes [55].
A user-independent plaque characterization and IMTsegmentation method was proposed in [72], based on interest-identification stage, gradient-segmentation stage, and a contour-refinement stage, using deformable parametric model. The overall accuracy of the system determined as normalized error was overall to 8%.
In a recent study [55], we proposed a snake-segmentation method based on the Lai and Chin snake for segmenting the atherosclerotic-carotid plaque. The initial contour estimation was carried out without user interaction using the blood-flow image. It was also shown that normalization and speckle-reduction filtering improves the outcome of the plaque segmentation. The user was able to interact and correct the segmentation results manually. A limitation of the proposed method includes the presence of acoustic shadowing together with strong speckle noise, which hinders the visual and automatic analysis in ultrasound images. Furthermore, only vessels without atherosclerotic plaques were segmented in this study. Fig. 1(b) shows segmentation results of the plaque at the far wall of the CCA, where normalization and speckle- reduction  TABLE III  RECENT SEGMENTATION TECHNIQUES FOR CAROTID PLAQUE 2-D AND 3-D ULTRASOUND filtering was applied prior to segmentation. There was no significant difference between the manual and the snakes segmentation regions.
In [73], a method based on a Star-Kalman algorithm was used to determine vessel contours and ellipse parameters using an extended Kalman filter with an elliptical model. The segmentation and tracking were implemented in real time and validated using simulated ultrasound data with known features and real data, for which expert segmentation was performed. Results indicate that mean errors between segmented contours and expert tracings are on the order of 1%-2% of the maximum feature dimension, and that the transverse cross-sectional vessel area, as computed from estimated ellipse parameters determined by the algorithm, is within 10% of that determined by experts.

A. Feature Extraction
Following segmentation, we extract features from the region of interest (see Figs. 1(b) and 2). In what follows, we provide a summary of the extracted feature sets.
Earlier studies have been primarily focused on basic statistical features, such as the gray-scale medial (GSM), the mean, the median, the standard deviation, skewness, and kurtosis [16], [37], [75]- [77]. In these earlier studies, the GSM was found to be very successful in differentiating between symptomatic and asymptomatic cases [16], [76]. Depending on the image preprocessing method, threshold values for the GSM were provided for differentiating between symptomatic and asymptomatic cases. Here, brighter plaques tended to be asymptomatic (see Table IV and Section IV-B, see also Fig. 2).
More extensively, histogram features were later used to provide plaque-signature vectors [57], [78]- [80]. Similarly, histograms of grayscale occurrences at different angles and distances (correlograms, not the same as used in spatial statistics) were reported in [57].
Standard texture features have been extensively used for the classification of carotid plaques [57], [78], [81]- [83]. An early discussion of standard texture features can be found in [84]. The most commonly used texture features include: 1) spatial graylevel-dependence matrices (SGLDM); 2) gray-level-difference statistics; 3) neighborhood gray-tone difference matrix; 4) statistical feature matrix; 5) laws texture energy measures; and 6) fractal-dimension texture analysis. A summary of the basic differences between texture characteristics from symptomatic and asymptomatic cases is discussed in [78] and also illustrated in Fig. 2.
More recently, we have used morphological features for plaque-image characterization [57], [82], [83]. The most successful morphological features were based on a multilevel decomposition, associated with different plaque image components. In the multilevel approach, each normalized plaque is thresholded at three different intensity ranges (low, medium, and high). The darkest (low) components are associated with unstable plaque components, such as lipid and hemorrhages. On the other hand, more stable plaque components are captured at higher brightness levels. For each one of the three cases, we compute pattern spectra to provide size distributions of the plaque components [85], [86]. We then use the pattern spectra as texture features for classification.

B. Classification
Several classification techniques were used for the classification of the carotid plaques. We have neural classifiers, such as the self-organizing map (SOM) [57], [78], the back propagation (BP) [81], the radial basis function (RBF), the probabilistic neural network (PNN) [82], [83]. More recently, we see classification based on support vector machines (SVMs) [82], [83]. In addition, we also have research done with statistical classifiers, such as the K-nearest neighbor (KNN) [57], [78] or simple statistical analysis of the plaque characteristics [75]- [77], [87]. For measuring performance, the leave-one-out method has been commonly used together with receiver-operating characteristic (ROC) analysis [78], [81]. We provide a brief survey of a number of classification studies and comment on the association between the extracted plaque characteristics and cerebrovascular symptoms. These studies are further tabulated in Table IV.  TABLE IV  ULTRASOUND CAROTID PLAQUE CLASSIFICATION STUDIES Geroulakos et al. [75] tested the hypothesis that the ultrasonic characteristics of carotid artery plaques are closely related to symptoms and the plaque structure may be an important factor in producing stroke, perhaps more than the degree of stenosis. In this paper, the authors categorized the carotid plaques into four ultrasonic types: echolucent, predominately echolucent, predominately echogenic, and echogenic. An association was found of echolucent plaques with symptoms and cerebral infractions, which provided further evidence that echolucent plaques are unstable and tend to embolize.
El-Barghouty et al. [76] in a study with 94 plaques reported an association between carotid plaque echolucency and the incidence of cerebral computed tomography (CT) brain infractions. The GSM of the ultrasound-plaque image was used for the characterization of plaques as echolucent (GSM ≤ 32) and echogenic (GSM > 32).
Iannuzzi et al. [77] analyzed 242 stroke and 336 TIA patients and identified significant relationships between carotid artery ultrasound-plaque characteristics and ischemic cerebrovascular events, respectively. The results suggested that the features more strongly associated with stroke were either the occlusion of the ipsilateral carotid artery or wider lesions and smaller minimum residual lumen diameter. The features, which were more consistently associated with TIAs, were low echogenicity of carotid plaques, thicker plaques, and the presence of longitudinal motion.
Elatrozy et al. [16] examined 96 plaques (25 symptomatic and 71 asymptomatic) from 80 patients with more than 50% internal carotid artery stenosis. They reported that plaques with GSM < 40, or with a percentage of echolucent pixels greater than 50% were good predictors of ipsilateral hemispheric symptoms related to carotid plaques. As echolucent pixels were defined, pixels with gray level values below 40.
Wilhjelm et al. [87] in a study with 52 patients scheduled for endarterectomy, presented a quantitative comparison between subjective classifications of the ultrasound images, first-and second-order statistical features, and a histological analysis of the surgically removed plaque. Some correlation was found between the three types of information, where the best-performing feature was found to be the contrast.
Rakebrandt et al. [79] demonstrated that texture analysis of B-mode ultrasound images of carotid plaques using histogram features (in conjunction with co-occurrence matrices, fractal models, and first-order statistics) can predict plaque composition using histological methods.
Asvestas et al. [88] in a pilot study with 19 carotid plaques indicated a significant difference of the fractal dimension between the symptomatic and asymptomatic groups. Moreover, the phase of the cardiac cycle (systole/diastole) during which the fractal dimension was estimated had no systematic effect on the calculations. This study suggests that the fractal dimension could be used as a single determinant for the discrimination of symptomatic and asymptomatic subjects.
Kakkos et al. [80] used histogram measures and found out that the percentage of pixels below 10 and gray scale median between 1 to 25 (i.e., the darkest parts of the plaque) were associated with symptoms (AF, TIA, and stroke).
In most of the aforementioned studies, the characteristics of the plaques were usually defined subjectively or using simple statistical measures, and the association with symptoms was established through simple statistical analysis. In the following studies, intelligent diagnostic systems were developed for the automatic classification of plaques into symptomatic or asymptomatic.
Christodoulou et al. [78] extracted a total number of 61 texture and shape features from 230 ultrasound-plaque images, and these features were analyzed using a multifeature multiclassifier methodology. A diagnostic yield of 73.1% was reported, thus indicating that it is possible to identify a group of patients at risk of stroke based on texture features and neural networks. In content-based image-retrieval study, Christodoulou et al. [57] showed that correlograms gave slightly better performance than traditional texture features.
Mougiakakou et al. [81], in a study with 108 plaque images, extracted first-order statistical features and Laws' texture energy measures that were classified with a neural network BP algorithm. They claimed an overall accuracy of 99.1% in the classification of symptomatic and asymptomatic plaques.
Holdfeldt [89] developed an automated system for the classification of echogenic versus echolucent plaques using heuristics and an adaptive threshold. The plaques were labeled as echogenic or echolucent by a human expert. Holdfieldt reported that the system could correctly identify the plaques with a success rate of 91%.
Kyriacou et al. [83], [86] describe an integrated system for the assessment of the risk of stroke based on clinical risk factors (noninvasive) and carotid plaque texture analysis. The system was validated on 274 images. It included semiautomatic plaque segmentation, morphological image analysis, and classification using multiple classifiers. For image features, the system compared the use of a new multilevel morphological decomposition (see Section IV-A) versus standard gray-scale morphological analysis. For classification, comparisons were made between a PNN and a SVM with RBF kernels. The best classification result was at 73.4% using the SVM classifier with multilevel morphological features.

A. 3-D Studies
The use of 3-D techniques has primarily focused on measuring volume changes through time to monitor disease progression [90]- [92]. Landry et al. [90] demonstrated that plaque volume change can be reliably measured using 3-D ultrasound. They showed that a 20%-35% change can be measured with 95% confidence for plaques of volume <100 mm 3 . For larger plaques (volume >100 mm 3 ), with 95% confidence, they showed that we can measure finer changes of the order of 10%-20%, respectively. Chiu et al. [91] developed a 3-D segmentation method for measuring the combined thickness of the plaque, the intima, and media (vessel wall plus plaque thickness or VWT). The authors proposed the use of 3-D VWT and VWT-change maps for identifying disease progression in relations to disturbances of flow. The authors extended their work in [92], where they measured VWT volume changes for assessing and monitoring carotid artery disease.

B. Motion
The majority of motion studies are focused on the use of 2-D ultrasound. However, Meairs and Hennerici [93] reported on an important early study on the use of 4-D ultrasound for motion estimation. Using 45 patients, the authors showed that asymptomatic plaques maintained plaque-surface motion vectors that were approximately equal to motion vectors of the internal carotid artery. In contrast, plaques from symptomatic patients exhibited independent motion with larger surface motion.
For B-mode ultrasound, Murillo et al. [94] computed motion trajectories over the plaques and used the results to develop realistic, synthetic models of plaque motion. Murray et al. [95] developed a new motion estimation model and demonstrated that it can provide for dense estimates over the entire cardiac cycle. In [96], Golemati et al. used tissue mimicking phantoms and synthetic motion models to verify motion estimation using block matching and optical flow methods. A mathematical model was developed by Golemati et al. in [97], and the results were compared with the block-matching software developed in [98].

C. Stress-Strain
Scmitt et al. [99] proposed the use of elastography for measuring stiffening and mechanical interactions between plaque structures. Here, we note that research on improving motion estimation can also benefit future work in this area.

VI. CONCLUDING REMARKS AND FUTURE DIRECTIONS
The majority of plaque-image analysis studies are focused on the development of 2-D ultrasound systems. In general, these studies present effective methods for image segmentation, image despeckling, and texture feature extraction. Due to the large number of parameters involved, we can see many variations in the results. Thus, even for 2-D systems, there will always be an interest in developing more robust segmentation methods, new multiscale texture features and the application of innovative classification techniques. Some of the most interesting challenges are associated with emerging studies (see Section V). These methods can be used in order to focus on new 2-D problems associated with early plaque formation.
The extraction of 3-D shape and structure information can be used to further for our understanding of carotid plaque morphology. We believe that research will continue in establishing 3-D volumetric changes and their associations with atherosclerosis.
Plaque-motion analysis holds significant promise as well. The relative motion among different plaque components should be investigated. The future development of very accurate 3-D/4-D systems will also help with the development of accurate motion analysis systems.
It would be interesting to develop noninvasive, multimodality plaque-image analysis systems. The advancement of 3-D ultrasound will help these efforts. Basically, high-resolution 3-D ultrasound reconstructions will be much easier to fuse with 2-D slices from other modalities. To do this, we would need to register the geometric features of the 2-D slice to the 3-D volume or to use a mutual information registration method. It would also be interesting to examine how 2-D histological studies match 3-D ultrasound reconstructions. From 1996 to 2008, he was involved in a number of collaborations in both Greece and Cyprus. He is currently an Associate Professor in the Department of Computer Science and Engineering, Frederick University, Limassol, Cyprus. He has authored or coauthored 20 refereed journal, 64 conference papers, and 13 chapters in books in these areas. His research interests include e-health systems, medical imaging systems, and intelligent systems applications in medicine. He has been involved in numerous projects in these areas funded by EU, the National Research Foundation of Cyprus, the INTERREG, and other bodies. He was a Reviewer in many journals related to his research fields.