Secondary Logo

Journal Logo

Original Articles

Magnetic Resonance Imaging of Carotid Atherosclerosis

Plaque Analysis

Kerwin, William PhD*; Xu, Dongxiang PhD*; Liu, Fei PhD*; Saam, Tobias MD; Underhill, Hunter MD*; Takaya, Norihide MD; Chu, Baocheng MD*; Hatsukami, Thomas MD§∥; Yuan, Chun PhD*

Author Information
Topics in Magnetic Resonance Imaging: October 2007 - Volume 18 - Issue 5 - p 371-378
doi: 10.1097/rmr.0b013e3181598d9d
  • Free


The risk of a heart attack or stroke resulting from atherosclerotic plaque accumulation in an artery wall is traditionally assessed by angiographic measurements of stenosis. Although risk does increase with stenosis, the bulk of clinical events likely arise from low-to-moderate-grade stenoses as a result of their greater prevalence.1 Efforts to better stratify risk have led to the concept of a "vulnerable plaque" being defined by morphological features of the plaque itself, not simply its stenotic impact on the lumen.2,3 This, in turn, has led to efforts to image the features of vulnerable plaque.

In large arteries, particularly the carotid arteries, magnetic resonance imaging (MRI) has demonstrated an ability to characterize morphological (volume, area, thickness), structural (fibrous cap rupture), and compositional (necrotic core [NC], hemorrhage, calcification) features of atherosclerotic plaque in vivo.4-13 A retrospective MRI study established that fibrous cap rupture is highly associated with prior symptoms in advanced carotid plaques.14 Comparisons of ipsilateral and contralateral arteries in symptomatic subjects have shown that ipsilateral arteries have a higher prevalence of intraplaque hemorrhage and fibrous cap rupture.15-17 More rapid lesion progression has been shown to be associated with the presence of intraplaque hemorrhage.18

In the first comprehensive study to examine the risk of future events based on MRI, Takaya et al19 examined baseline plaque features in a group of 154 asymptomatic subjects with 50% to 79% carotid stenosis by duplex ultrasound. After a mean follow-up time of 38.2 months, they demonstrated that the presence of intraplaque hemorrhage and a thinned or ruptured fibrous cap were both significantly associated with a higher hazard ratio (HR) for subsequent clinical events. In addition to these dichotomous characteristics, any increases in several continuous measurements were each associated with increasing HRs. Specifically, each 1-mm increase in the maximal thickness of the artery wall was associated with an increased HR of 1.6, each 10-mm2 increase in average area of intraplaque hemorrhage was associated with an increased HR of 2.6, and every 10-mm2 increase in NC area was associated with an increased HR of 1.6.

These studies have shown that MRI has the potential to provide lesion-specific risk assessments. Furthermore, they indicate the importance of both morphological characteristics and plaque composition for assessing risk. Finally, they illustrate the importance of quantifying plaque composition and morphology with continuous measurements.

The importance of quantification in evaluating plaque features by MRI suggests a role for image postprocessing techniques. Highly automated measurement techniques can reduce analysis time, reduce reader-dependent bias, and improve measurement reproducibility. The purpose of this article is to present an overall framework for measuring atherosclerotic plaque characteristics in the carotid artery. The framework will be demonstrated with our software platform called the Computer-Aided System for CArdiovascular Disease Evaluation (CASCADE). This system includes boundary detection, registration of multiple contrast weightings, segmentation of internal plaque components, and 3-dimensional (3D) display of results. Among other uses, CASCADE is suitable for use in clinical trials to assess the effects of treatment on plaque composition. In this investigation, the performance of CASCADE is compared with manually outlined results.


MRI Protocol

Various studies have concluded that accurate tissue characterization requires information combined from multiple MRI contrast weightings.4-7,20-22 Additionally, submillimeter pixel sizes and thin image slices (≤2 mm) are critical to differentiate distinct components that may occupy volumes smaller than 1 mm3. A standard in vivo carotid MRI protocol now includes T1-weighted (T1W), T2-weighted (T2W), proton-density-weighted (PDW), time-of-flight (TOF), and, possibly, gadolinium contrast-enhanced T1W MRI. Diffusion-weighted images are also under investigation.23 To perform volumetric analysis of plaque composition, stacks of axial images of the carotid artery are obtained. A dedicated carotid surface coil is used to provide a longitudinal coverage of up to 5 cm and maximal signal-to-noise ratio.24

In this investigation, we used images from 26 subjects with 16% to 79% carotid stenosis by duplex ultrasound who underwent 2 carotid MRI examinations within a 2-week period at the University of Washington Medical Center. The images were collected as part of the Outcome of Rosuvastatin treatment on carotid artery atheroma: a magnetic resonance Imaging ObservatioN (ORION) trial.25 Both MRI examinations per subject were collected before treatment on a 1.5-T MRI scanner (GE Signa 5.8; Milwaukee, Wis). Each MRI scan included T1W double-inversion recovery, dual-echo PDW/T2W, and 3D TOF image sequences. All images were collected with a 2-mm slice thickness and a 13- to 16-cm field of view depending on the subject's neck size. Other imaging parameters are summarized in Table 1. A representative set of images showing the different contrast weightings is shown in Figure 1.

Multicontrast MRI protocol for the carotid artery at the carotid bifurcation: T1W (A), TOF (B), PDW (C), T2W (D). The internal (I) and external (E) carotid artery branches are indicated.
Imaging Parameters for ORION Trial

Vessel Boundary Detection

The image processing challenge for measuring morphological indices of atherosclerotic plaque is one of boundary detection. Boundary detection in medical imaging is typically performed using active contour methods, also known as "snakes."26 The basic active contour methodology seeks to identify a curve corresponding to edges in the image. Optimization of the curve is accomplished by minimizing an energy function that may depend on the image itself and the smoothness of the contour.

We have found that effective boundary detection in vessel wall imaging is afforded by B-spline snakes.27 In B-spline snakes, the boundary contour C(u) = (cx[u], cy[u]) is parameterized as a B-spline

where u ∈ (0, K),ξk and ψk are the spline coefficients, and cubic B-spline kernels are used:

The values of the spline coefficients are modified until an image-based energy functional is minimized. Our energy formulation is

where l(C) is the length of the contour, n(u) is the inward facing unit normal to the contour, and the image gradient is computed at the corresponding point on the contour. Identifying a contour where the interior is brighter than the exterior is accomplished by setting b = +1, and a darker interior is found by setting b = −1. Scaling by the magnitude of the derivative of C(u) prevents spline nodes from clustering in a region of high gradient and dividing by l(C) weights all curves equally, regardless of length, and avoids an infinitely expanding contour.

Such an approach has several advantages. First, the inherent smoothness of the spline means that no explicit smoothness term is required in the energy. Second, the node points of the spline, defined as C(k), where k is an integer, can be manually moved if necessary for rapid manual correction of boundaries. Third, optimization can be formulated as simple gradient descent. The major challenge for the B-spline snake is to initialize it near the true boundary to ensure convergence to the proper minimum energy point.

Lumen Detection

To identify the lumen boundary, the B-spline snake is iteratively propagated through each slice of a black-blood sequence, nominally the T1W sequence. Initialization of the B-spline snake is accomplished by first mapping the center of the lumen contour from the previous slice to each successive slice based on nonrigid registration of the 2 images.28 This center point is then used in a mean-shift segmentation algorithm29 to identify all pixels with similar intensity. The boundary of the connected region containing the seed point is used to initialize the B-spline snake. A user-placed seed point in the distal internal carotid artery is used to initialize the process. Starting in the distal internal carotid artery and propagating in the proximal direction avoids uncertainty at the lumen bifurcation. At each location, the contours can be manually adjusted before moving to the next location (Fig. 2A).

Results of B-spline snake detection of lumen (A) and wall boundaries (B). The points within the contours indicate nodes that can be manually adjusted to alter the contour shape.

Outer Wall Detection

Outer wall boundary detection proceeds in a manner similar to lumen detection, iteratively proceeding from the proximal end to the distal end of the carotid artery segment. Initialization of the B-spline in each successive image is performed using a conditional shape model (CSM).30 This method is similar to active shape models31 in which a contour is parameterized as a series of points arrayed into a vector X given by

A set of training shapes is used to determine the mean shape μx and the eigenshapes pk, which are the K most significant eigenvectors of the covariance matrix Σx of the training shapes. Varying the weights bk steps through the set of all allowable shapes. In the CSM, an additional observed parameter, z, is used to further constrain the shapes by replacing the mean shape with the conditional mean shape μx|z and the set of eigenshapes with the eigenvectors of the conditional covariance matrix Σx|z. For outer wall detection, the conditional parameter is the ratio of the minor and major axes of the wall contour in the previous slice.

To identify the wall boundary in the first location, the model is run with z = 1 on the most proximal slice, which assumes a nearly circular wall boundary. For each successive location, the value of z is determined from the previous location, and the set of allowable shapes is obtained from the CSM. The shape that minimizes the contour energy function1 is used to initialize the B-spline snake. At each location, the contours can be manually adjusted before moving to the next location (Fig. 2B).

Multicontrast Registration

Vessel boundaries from a single contrast weighting are sufficient to compute plaque burden measurements, but to measure plaque composition, information must be integrated from multiple contrast weightings. This introduces a further processing challenge because patient motion between successive acquisitions misaligns the images. Thus, a first step for compositional analysis is to register the multiple contrast weightings to eliminate patient shifts. Prior determination of the vessel boundaries facilitates this step.

For the small motions that occur within a single imaging session, the atherosclerotic carotid artery behaves like a rigid body embedded in a deforming medium. Locally, an in-plane shift is sufficient to correct for motion. To determine the shift, we use an active edge map formulation.32 Specifically, we find a 2D displacement d that minimizes the energy

where xi ranges over the set of points in the lumen and outer wall contours, C is the set of points in the Canny edge map33 of the image to be registered, and g is a Gaussian function with a width (σ) of 2 pixels. The parameter a (xi) = 3 if xi is in the lumen contour and 1 if it is in the wall contour.

Use of a Gaussian kernel creates a continuous, differentiable image energy function for gradient-based optimization and extends the capture range of the registration algorithm. Use of the binary Canny edge map limits the influence of the surface coils, which produce stronger edges closer to the coils. The term a (xi) causes the lumen and wall contours to receive approximately equal weight despite the greater number of points in the wall contour. The range of possible shifts is restricted to ±2 mm in each direction. To overcome larger patient shifts, we introduced a feature we call "smart drag," in which the user can move any image relative to the contours, and upon releasing the image, the registration algorithm minimizes the energy within a 1-mm region surrounding that location. The registration procedure is illustrated in Figure 3.

Registration: lumen and wall boundaries mapped to TOF (A), PDW (B), and T2W (C) images show misalignment that is corrected by the registration algorithm (D-F).

Plaque Segmentation

In the next analysis step, regions within the artery wall are automatically detected, outlined, and assigned a tissue type by the morphology-enhanced probabilistic plaque segmentation (MEPPS) algorithm.34 The core task of the MEPPS algorithm is to assign a set of probabilities to each pixel, one for each possible tissue type. We base this probability on the pixel intensity in each contrast weighting-represented by the vector x-and on 2 morphological factors: the local wall thickness (WT; t) and the distance of the pixel from the lumen (d). Thus, we determine the probability Pr(Ti|t, d, x), where Ti corresponds to one of the possible tissue types. The 2 distances (t and d) capture information about the local plaque morphology typically used in manual review. For example, thin-walled regions are generally fibrous.

Under the assumption that the intensity of a given tissue does not vary with position in the plaque, Bayes rule yields:

The 2 conditionally independent probability density functions, p(x|Ti) and p(t, d, Ti), and the relative frequency of each of the 4 tissue types, Pr(Ti), were estimated from a training set of in vivo MRI with postsurgery histological confirmation.34 Applying these results to a new data set yields a probability map for each pixel in the artery wall, as illustrated in Figure 4A. Each pixel has been color coded to indicate the tissue with the highest probability. The intensity represents the difference between the highest and second highest probabilities, essentially providing a confidence metric in the classification. We refer to this image as "MRI histology."

Segmentation results showing MRI histology in which the tissue with maximal probability is color coded (yellow, NC; gray, fibrous; blue, calcification), and the intensity reflects the difference in probability from the next highest value (A); contours delineating the boundaries of the regions are shown on the corresponding T1W image (B).

Once the probabilities for each pixel are determined, the final step is to identify contours delineating each tissue region for which we use the active region method.35 Each contour seeks 1 preassigned tissue. Figure 4B shows the results of segmentation of an example based on probabilities.

Plaque Measurement and Display

Once the boundaries of the vessel wall and internal plaque components are available, the information must be recorded or presented in a meaningful way. The total volume of the wall or components can be determined by adding the total areas in each cross section and multiplying by the slice thickness. Care must be taken, however, when comparing measurements to account for differences in body size and longitudinal coverage of the artery. A larger volume may be attributable to a longer extent of coverage or a larger vessel size as opposed to a larger disease burden.

The normalized wall index (NWI), defined as

where A denotes area, has been proposed as a measurement of plaque burden that is independent of vessel size. For plaque components, the average area per slice and the maximal area over all slices have been used as measurements that are less sensitive to differences in coverage than are volume measurements.

Maximal WT and mean WT36 are also measurements of plaque burden that are relatively insensitive to artery size or coverage. Measuring thickness in the atherosclerotic carotid artery can be challenging given the unusual shapes of the lumen and outer wall boundaries in heavily diseased arteries and near the bifurcation (Fig. 5). We have proposed a thickness measurement algorithm37 based on DeLaunay triangulation38 that achieves intuitive thickness measurements for any morphology.

Example of wall thickness measurement showing the minimal and maximal values detected.

Finally, display of the entire plaque extent can be helpful for observing subtle differences over time or for surgical planning. To convert cross-sectional contours into a 3D model of the plaque, we have proposed a method39 whereby a membership function (equal to 1 within a contour and 0 outside a contour) is interpolated onto a dense grid for each contour type. An isosurface algorithm is then used to extract the surface corresponding to the boundary of the tissue type. This approach robustly handles changes in morphology (eg, the bifurcation), isolated contours, and physically separated regions of the same type. An example from the ORION study showing plaque changes from baseline to 2 years after start of rosuvastatin treatment is shown in Figure 6.

Three-dimensional rendering of plaque segmentation results showing a substantial decrease in NC (yellow) from baseline (A) to 2 years after start of rosuvastatin treatment (B). Development of a small calcification (blue) is also apparent. The vessel lumen is shown in red, and the outer wall is semitransparent orange.

Validation Study

To validate CASCADE, images from the 26 subjects included in this investigation were loaded into the CASCADE software package and vertically aligned by matching the location of the carotid bifurcation across all weightings and both time points. Cases were randomly assigned to 4 reviewers (B.C., T.S., N.T., and H.U.) with expertise in carotid MRI, and the manual outlining tools of CASCADE were used to identify the lumen boundary, outer wall boundary, calcifications, and NC (including intraplaque hemorrhage) for all locations common to both scans. After a period of 6 months, the images were randomly reassigned to the same 4 reviewers and analyzed using the automated tools of CASCADE. Manual readjustment of lumen and outer wall boundaries and registration results was permitted as needed. Segmentation results were not adjusted.

For both the manual and automated results, 4 representative metrics were computed: maximal NWI, maximal WT, average NC, and average calcified area per slice (CA). Association between manual and automated results was assessed by computing Pearson correlation coefficient (R) for the first time point only. Cohen κ was used to assess the agreement between manual and automated reviews for the presence of NC and calcification. Bias was assessed by Bland-Altman analysis and the paired t test.40 Reproducibility of both manual and automated results was assessed by computing the intraclass correlation coefficient (ICC) for each measurement.


Correlation between manual and automated reviews was strong for all 4 parameters evaluated (Table 2). Correlation coefficients ranged from 0.84 to 0.96. Bland-Altman analysis (Fig. 7) and paired t tests showed significant bias between the measurement techniques. Both maximal WT and NWI were significantly smaller in the automated review as a result of both the lumen region being drawn smaller and the outer wall region being drawn larger in manual review. The difference was attributable to the fact that the automated boundary detection technique places the boundary at the point between 2 regions where the intensity is halfway between the intensities of the structures forming the boundary. Visually, however, the boundary tends to appear further into the darker of the 2 regions and depends on the window and level settings chosen for viewing. Thus, the boundary of the lumen in dark-blood images tends to be placed further inside the lumen, and the outer wall boundary tends to be placed further outside the brighter wall in manual review.

Bland-Altman plots comparing manual and automated analysis results for maximum NWI (A), maximum WT (B), NC (C), and CA (D). Difference values represent the results of manual review subtracted from the results of automated review. Lines denoting the mean (solid) difference and ±1.96 SDs (dashed) are also shown.
Comparison of Manual and Automated Measurements of Plaque Morphology and Composition

For the compositional measurements of NC and CA, we also computed Cohen κ for detecting lesions containing each component. For NC, κ was 0.62. Although this indicates substantial agreement, the somewhat low value was attributed to the automated algorithm detecting several small (<1 mm2 per slice) regions of NC that were not identified in the manual review. If regions less than this threshold were rejected, κ was 0.77. For CA, κ was 0.62. In several cases, the automated algorithm failed to detect small (<1 mm2 per slice) regions of CA that were identified in the manual review. If values less than this threshold were rejected, κ increased to 0.74.

Both NC and CA also showed significant bias between manual and automated reviews, with both appearing smaller in the automated review. The difference can, in part, be explained by the smaller wall area in the automated review, which left less area to be classified as CA or NC. Additionally, residual misregistration and slice offsets meant that corresponding regions in different contrast weightings did not align exactly. The segmentation algorithm tended to include only the areas common to multiple contrast weightings. In contrast, manual review tended to outline a single region in a single contrast weighting that provided the best depiction. In general, the common region identified by the segmentation algorithm tended to be smaller.

Reproducibility results are provided in Table 3. The ICC values for WT and NWI were higher for automated review, suggesting that the automated boundary detection methods improved reproducibility for plaque burden measurements. In contrast, ICC values for plaque composition-NC and CA-were marginally worse for the automated segmentation algorithm but still high.

Scan-Scan Reproducibility (ICC) of Manual and Automated Measurements of Plaque Morphology and Composition


Together, the image processing techniques presented in this article describe the process flow in Figure 8, which have been programmed into the software platform CASCADE. With CASCADE, the user is guided through the interactive procedures to perform lumen detection via the registration-driven snake, then wall detection using a CSM,30 registration with active edge maps,32 segmentation with MEPPS,34 and finally, volume rendering. Each of the principal steps in the analysis can be accomplished with other techniques, and improvements are an active area of research. Other researchers have proposed a number of alternative techniques for lumen and wall boundary detection,41-44 for plaque segmentation,21,44-46 and for plaque visualization.47

Process flow in automated plaque analysis with CASCADE.

In this overview, we focused on 4 measurements of plaque composition: NWI, WT, NC, and CA. Normalized wall index has emerged as the preferred index of plaque burden that is largely independent of vessel size and is closely related to the "percent atheroma volume" proposed for intravascular ultrasound.48 Even a normal carotid artery will have an NWI near 0.4. Maximal WT is another indicator of plaque size irrespective of vessel size and represents the maximal extent of plaque separating the lumen from the outer wall boundary. In the prospective study of Takaya et al,19 maximal WT was the one measurement of burden that was significantly associated with the development of symptoms. In that study, NC also showed a significant association with the development of symptoms. Finally, CA has been explored as a risk factor in CT studies.49

In addition to these characteristics, significant interest exists in characterizing the fibrous cap and intraplaque hemorrhage, which have both been tied to symptoms in subjects with carotid atherosclerosis. Recently, the MEPPS segmentation algorithm was enhanced to separate the NC into hemorrhage and lipid components.50 Evaluation of this feature in this study was not possible because the ORION study was conducted using a population with relatively small lesions that have a very low prevalence of intraplaque hemorrhage. Assessment of the fibrous cap has been performed by automatically detecting the region between the lumen boundary and NC. However, only manual delineations and classifications of the cap have been histologically validated.7,8 Automated characterization of the fibrous cap is an area for future work.

In comparing automated and manual measurements, results from the 2 techniques can be viewed as largely equivalent despite significant bias between measurements. Because this study was done on subjects with stenosis that did not meet the criteria for surgery, no histological verification of the accuracy of either method was possible. In independent analyses, both manual and automated reviews showed excellent agreement with histological criterion standards.4,34 Nevertheless, the high correlation between methods in this study suggests that either approach would be suitable for measuring plaque burden and composition provided the bias was accounted for. It is important to note that in either case, oversight by trained experts in carotid MRI is of paramount importance to obtain these results. In particular, manual adjustment of lumen and wall boundaries is frequently required to recognize and correct for flow artifacts or poorly defined boundaries.

In serial studies, such as in clinical trials of pharmaceuticals, reproducibility is also highly important. This study suggests that lumen and wall boundaries generated with automated techniques improve reproducibility compared to manual drawings. In contrast, automated segmentation techniques were found to be moderately less reproducible. Thus, for applications where reproducibility is of paramount importance, use of manual outlining or adjustment may be preferable to automated segmentation.

Overall, CASCADE was found to provide quantitative information regarding plaque burden and composition that was highly similar to manual outlining for sensitivity and specificity of detecting NC or CA, correlation between measurements, and measurement reproducibility. This means that similar analysis results can be obtained with considerably less analysis time because CASCADE substantially reduces the analysis burden compared with manual review. Although not investigated in this study, automated tools also have the potential to reduce interrater variability.


1. Falk E, Shah PK, Fuster V. Coronary plaque disruption. Circulation. 1995;92:657-671.
2. Virmani R, Kolodgie FD, Burke AP, et al. Lessons from sudden coronary death: a comprehensive morphological classification scheme for atherosclerotic lesions. Arterioscler Thromb Vasc Biol. 2000;20:1262-1275.
3. Falk E. Why do plaques rupture? Circulation. 1992;86:III30-III42.
4. Saam T, Ferguson MS, Yarnykh VL, et al. Quantitative evaluation of carotid plaque composition by in vivo MRI. Arterioscler Throm Vasc Biol. 2005;25:234-239.
5. Chu B, Kampschulte A, Ferguson MS, et al. Hemorrhage in the atherosclerotic carotid plaque: a high-resolution MRI study. Stroke. 2004;35:1079-1084.
6. Yuan C, Mitsumori LM, Ferguson MS, et al. In vivo accuracy of multispectral magnetic resonance imaging for identifying lipid-rich necrotic cores and intraplaque hemorrhage in advanced human carotid plaques. Circulation. 2001;104:2051-2056.
7. Mitsumori LM, Hatsukami TS, Ferguson MS, et al. In vivo accuracy of multisequence MR imaging for identifying unstable fibrous caps in advanced human carotid plaques. J Magn Reson Imaging. 2003;17:410-420.
8. Cai J, Hatsukami TS, Ferguson MS, et al. In vivo quantitative measurement of intact fibrous cap and lipid-rich necrotic core size in atherosclerotic carotid plaque: comparison of high-resolution, contrast-enhanced magnetic resonance imaging and histology. Circulation. 2005;112:3437-3444.
9. Moody AR, Murphy RE, Morgan PS, et al. Characterization of complicated carotid plaque with magnetic resonance direct thrombus imaging in patients with cerebral ischemia. Circulation. 2003;107:3047-3052.
10. Luo Y, Polissar N, Han C, et al. Accuracy and uniqueness of three in vivo measurements of atherosclerotic carotid plaque morphology with black blood MRI. Magn Reson Med. 2003;50:75-82.
11. Cappendijk VC, Cleutjens KB, Heeneman S, et al. In vivo detection of hemorrhage in human atherosclerotic plaques with magnetic resonance imaging. J Magn Reson Imaging. 2004;20:105-110.
12. Cappendijk VC, Cleutjens KB, Kessels AG, et al. Assessment of human atherosclerotic carotid plaque components with multisequence MR imaging: initial experience. Radiology. 2005;234:487-492.
13. Trivedi RA, U-King-Im JM, Graves MJ, et al. MRI-derived measurements of fibrous-cap and lipid-core thickness: the potential for identifying vulnerable carotid plaques in vivo. Neuroradiology. 2004;46:738-743.
14. Yuan C, Zhang SX, Polissar NL, et al. Identification of fibrous cap rupture with magnetic resonance imaging is highly associated with recent transient ischemic attack or stroke. Circulation. 2002;105:181-185.
15. Saam T, Cai J, Ma L, et al. Comparison of symptomatic and asymptomatic atherosclerotic carotid plaque features with in vivo MR imaging. Radiology. 2006;240:464-472.
16. Murphy RE, Moody AR, Morgan PS, et al. Prevalence of complicated carotid atheroma as detected by magnetic resonance direct thrombus imaging in patients with suspected carotid artery stenosis and previous acute cerebral ischemia. Circulation. 2003;107:3053-3058.
17. Yamada N, Higashi M, Otsubo R, et al. Association between signal hyperintensity on T1-weighted MR imaging of carotid plaques and ipsilateral ischemic events. AJNR Am J Neuroradiol. 2007;28:287-292.
18. Takaya N, Yuan C, Chu B, et al. Presence of intraplaque hemorrhage stimulates progression of carotid atherosclerotic plaques: a high-resolution magnetic resonance imaging study. Circulation. 2005;111:2768-2775.
19. Takaya N, Yuan C, Chu B, et al. Association between carotid plaque characteristics and subsequent ischemic cerebrovascular events: a prospective assessment with MRI-initial results. Stroke. 2006;37:818-823.
20. Shinnar M, Fallon JT, Wehrli S, et al. The diagnostic accuracy of ex vivo MRI for human atherosclerotic plaque characterization. Arterioscler Thromb Vasc Biol. 1999;19:2756-2761.
21. Clarke SE, Hammond RR, Mitchell JR, et al. Quantitative assessment of carotid plaque composition using multicontrast MRI and registered histology. Magn Reson Med. 2003;50:1199-1208.
22. Yuan C, Kerwin WS, Ferguson MS, et al. Contrast-enhanced high resolution MRI for atherosclerotic carotid artery tissue characterization. J Magn Reson Imaging. 2002;15:62-67.
23. Qiao Y, Ronen I, Viereck J, et al. Identification of atherosclerotic lipid deposits by diffusion-weighted imaging. Arterioscler Thromb Vasc Biol. 2007;27:1440-1446.
24. Hayes CE, Mathis CM, Yuan C. Surface coil phased arrays for high-resolution imaging of the carotid arteries. J Magn Reson Imaging. 1986;6:109-112.
25. Chu B, Hatsukami TS, Polissar NL, et al. Determination of carotid artery atherosclerotic lesion type and distribution in hypercholesterolemic patients with moderate carotid stenosis using noninvasive magnetic resonance imaging. Stroke. 2004;35:2444-2448.
26. Kass M, Witkin A, Terzopoulos D. Snakes: active contour models I. J Comput Vis. 1987;1:321-331.
27. Brigger P, Hoeg J, Unser M. B-spline snakes: a flexible tool for parametric contour detection. IEEE Trans Image Process. 2000;9:1484-1496.
28. Liu F, Xu D, Yuan C, et al. Elastic registration based interpolation of MRI images in 3D. Paper presented at: Proceedings of the 15th International Society of Magnetic Resonance in Medicine Meeting; May 15, 2007; Berlin, Germany; p. 2591.
29. Comaniciu D, Meers P. Mean shift: a robust approach toward feature space analysis. IEEE Trans PAMI. 2002;24:603-619.
30. Underhill HR, Kerwin WS. Markov shape models: object boundary identification in serial magnetic resonance images. Proceedings of the 15th International Society of Magnetic Resonance in Medicine Meeting; May 6-12, 2006; Seattle, WA; p. 829.
31. Cootes TF, Taylor CJ, Cooper DH, et al. Active shape models-their training and application. Comput Vis Image Underst. 1995;61:38-59.
32. Kerwin WS, Yuan C. Active edge maps for medical image registration. Proc SPIE. 2001;4322:516-526.
33. Canny J. A computational approach to edge detection. IEEE Trans Pattern Anal. 1986;8:679-698.
34. Liu F, Xu D, Ferguson MS, et al. Automated in vivo segmentation of carotid plaque MRI with morphology-enhanced probability maps. Magn Reson Med. 2006;55:659-668.
35. Paragios N, Deriche R. Coupled geodesic active regions for image segmentation: a level set approach. In: Computer Vision-ECCV 2000. Lecture Notes in Computer Science. Berlin, Germany: Springer; 2000:224-240.
36. Underhill HR, Kerwin WS, Hatsukami TS, et al. Automated measurement of mean wall thickness in the common carotid artery by MRI: a comparison to intima-media thickness by B-mode ultrasound. J Magn Reson Imaging. 2006;24:379-387.
37. Han C, Yuan C. Plaque morphological quantitation. In: Suri J, Laximinarayan S, eds. Angiography and Plaque Imaging. Boca Raton, FL: CRC Press; 2003:365-396.
38. Schumaker LL. Triangulation methods. In: Chui CK, Schumaker LL, Utreras FI, eds. Topics in Multivariate Approximation. Boston, MA: Academic Press; 1987:219-232.
39. Kerwin W, Yuan C. Three dimensional surface models of carotid atherosclerosis. In: Younan N, ed. Signal and Image Processing. Calgary, Alberta, Canada: ACTA Press; 2002:589-594.
40. Bland JM, Altman DG. Measuring agreement in method comparison studies. Stat Methods Med Res. 1999;8:135-160.
41. Han C, Hatsukami TS, Hwang JN, et al. A fast minimal path active contour model. IEEE Trans Image Process. 2001;10:865-873.
42. Han C, Kerwin WS, Hatsukami TS, et al. Detecting objects in image sequences using rule-based control in an active contour model. IEEE Trans Biomed Eng. 2003;50:705-710.
43. Mansard CD, Canet Soulas EP, Anwander A, et al. Quantification of multicontrast vascular MR images with NLSnake, an active contour model: in vitro validation and in vivo evaluation. Magn Reson Med. 2004;51:370-379.
44. Adame IM, van der Geest RJ, Wasserman BA, et al. Automatic segmentation and plaque characterization in atherosclerotic carotid artery MR images. Magma. 2004;16:227-234.
45. Itskovich VV, Samber DD, Mani V, et al. Quantification of human atherosclerotic plaques using spatially enhanced cluster analysis of multicontrast-weighted magnetic resonance images. Magn Reson Med. 2004;52:515-523.
46. Hofman JM, Branderhorst WJ, ten Eikelder HM, et al. Quantification of atherosclerotic plaque components using in vivo MRI and supervised classifiers. Magn Reson Med. 2006;55:790-799.
47. Adame IM, de Koning PJ, Lelieveldt BP, et al. An integrated automated analysis method for quantifying vessel stenosis and plaque burden from carotid MRI images: combined postprocessing of MRA and vessel wall MR. Stroke. 2006;37:2162-2164.
48. Nissen SE, Tsunoda T, Tuzcu EM, et al. Effect of recombinant ApoA-I Milano on coronary atherosclerosis in patients with acute coronary syndromes: a randomized controlled trial. JAMA. 2003;290:2292-2300.
49. Miralles M, Merino J, Busto M, et al. Quantification and characterization of carotid calcium with multi-detector CT-angiography. Eur J Vasc Endovasc Surg. 2006;32:561-567.
50. Liu F, Ferguson M, Chu B, et al. Intra-plaque hemorrhage quantification using morphology-enhanced probabilistic plaque segmentation (MEPPS) for in vivo carotid MRI. J Cardiovasc Magn Reson. 2007;9:389-390.

MRI; carotid artery; atherosclerosis; post-processing

© 2007 Lippincott Williams & Wilkins, Inc.