Quantification of human atherosclerotic plaques using spatially enhanced cluster analysis of multicontrast-weighted magnetic resonance images

Authors

  • Vitalii V. Itskovich,

    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
    • Vitalii V. Itskovich and Daniel D. Samber contributed equally to this work.

  • Daniel D. Samber,

    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
    • Vitalii V. Itskovich and Daniel D. Samber contributed equally to this work.

  • Venkatesh Mani,

    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    2. Zena and Michael A. Wiener Cardiovascular Institute, Mount Sinai School of Medicine, New York, New York
    3. Marie-Josée and Henry R. Kravis Cardiovascular Health Center, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
  • Juan Gilberto S. Aguinaldo,

    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    2. Zena and Michael A. Wiener Cardiovascular Institute, Mount Sinai School of Medicine, New York, New York
    3. Marie-Josée and Henry R. Kravis Cardiovascular Health Center, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
  • John T. Fallon,

    1. Zena and Michael A. Wiener Cardiovascular Institute, Mount Sinai School of Medicine, New York, New York
    2. Marie-Josée and Henry R. Kravis Cardiovascular Health Center, Mount Sinai School of Medicine, New York, New York
    3. Department of Pathology, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
  • Cheuk Y. Tang,

    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
  • Valentin Fuster,

    1. Zena and Michael A. Wiener Cardiovascular Institute, Mount Sinai School of Medicine, New York, New York
    2. Marie-Josée and Henry R. Kravis Cardiovascular Health Center, Mount Sinai School of Medicine, New York, New York
    Search for more papers by this author
  • Zahi A. Fayad

    Corresponding author
    1. Imaging Science Laboratories, Department of Radiology, Mount Sinai School of Medicine, New York, New York
    2. Zena and Michael A. Wiener Cardiovascular Institute, Mount Sinai School of Medicine, New York, New York
    3. Marie-Josée and Henry R. Kravis Cardiovascular Health Center, Mount Sinai School of Medicine, New York, New York
    • FAHA, Imaging Science Laboratories, Mount Sinai School of Medicine, One Gustave L. Levy Place, Box 1234, New York, NY 10029-6574
    Search for more papers by this author

Abstract

One of the current limitations of magnetic resonance imaging (MRI) is the lack of an objective method to classify plaque components. Here we present a cluster analysis technique that can objectively quantify and classify MR images of atherosclerotic plaques. We obtained three-dimensional (3D) images from 12 human coronary artery specimens on a 9.4T imaging system using multicontrast-weighted fast spin-echo (T1-, proton density-, and T2-weighted) imaging with an isotropic voxel size of 39 μ. Spatially enhanced cluster analysis (SECA) was performed on multicontrast MR images, and the resulting segmentation was evaluated against histological tracings. To visualize the overall structure of plaques, the MR images were rendered in 3D. The specimens exhibited lesions of American Heart Association (AHA) plaque classification types I-VI. Both MR images and histological sections were independently reviewed, categorized, and compared. Overall, the classification obtained from the cluster-analyzed MR and histopathology images showed very good agreement for all AHA types (92%, Cohen's κ = 0.89, P < 0.0001). All plaque types were identified and quantified by SECA with a high degree of correlation between cluster-analyzed MR and manually traced histopathology data. MRI combined with SECA provides an objective method for atherosclerotic plaque component characterization and quantification. Magn Reson Med 52:515–523, 2004. © 2004 Wiley-Liss, Inc.

The characterization of atherosclerotic plaque structure in coronary arteries is an important factor in assessing a person's susceptibility to acute coronary syndromes (1). Recently, noninvasive high-resolution MRI has been used to monitor changes in the burden and composition of human carotid arteries and aortic plaques (2, 3). Multicontrast MRI has been used to classify atherosclerotic plaque components (2, 4–7). Shinnar et al. (4) used multicontrast MRI to identify fibrocellular tissue, calcium, and lipid core with the help of semiautomated algorithms. Pattern recognition has been employed to obtain qualitative assessments of plaque composition in human aortas with the use of T1- (T1W), T2- (T2W), and proton density- (PDW) weighted MR images. Normal walls (media), lipid-rich plaques, and fibrous plaques have been differentiated with the use of a clustering algorithm (8).

We hypothesized that color composite multicontrast MR images could enhance the differentiation of various tissue components based on their signal intensities in different MR weightings. We further hypothesized that spatially enhanced K-means cluster analysis (SECA) could improve the characterization of plaques in ex vivo three-dimensional (3D) MR images.

In the present study, SECA on multicontrast 3D MR images of ex vivo human coronary arteries enabled the automated, objective classification and quantification of plaque components. This technique facilitates the task of analyzing large MR datasets, eliminates interobserver variability in atherosclerotic plaque characterization, and enables plaque visualization in 3D using cluster-classified data.

MATERIALS AND METHODS

This study was approved by our institutional review board.

Ex Vivo MRI

Imaging was performed on a Bruker 9.4T MR system (Bruker Instruments, Billerica, MA). Twelve human coronary artery specimens (1.28 cm long; six from the Pathobiological Determinants of Atherosclerosis in Young (PDAY) data library (9), and six from the Mount Sinai School of Medicine Autopsy Service) were imaged with the use of a 10-mm-diameter birdcage coil.

A thermocouple-based temperature-controlling element was used to maintain the samples at a physiologic temperature of 37°C. Paraformaldehyde-fixed, longitudinally cut and flattened PDAY specimens (2–5 months postmortem) were obtained in an airtight container. The PDAY specimens were refrigerated and rewarmed before imaging, while the coronary artery specimens from the autopsy service were imaged fresh. Vascular morphology and MR contrast are minimally affected by this preparation (5, 10, 11). Each specimen was placed in an 8-mm-diameter polyethylene tube that was filled with Fomblin solution (perfluoropolyether; Ausimont USA Inc., Morristown, NJ) and sealed to prevent air bubbles. Since Fomblin does not have a residual MR signal, it does not interfere with signals from various biological tissues (12). We imaged the specimens using a 3D fast spin-echo (FSE) sequence with an isotropic voxel of 39 × 39 × 39 μm3 (field of view (FOV) = 1.0 cm), echo-train length = 4, and acquisition matrix = 2563. Some larger samples required an FOV of 1.28 cm (50 × 50 × 50 μm3 voxel size). We used multicontrast-weighted MRI to characterize each specimen, using PDW, T1W, and T2W images with repetition time/echo time (TR/TE) = 2000/9 ms, 500/9 ms, and 2000/25 ms, respectively. The acquisition time was 8 hr for the PDW and T2W imaging, and 2 hr for the T1W imaging. The position and size of the 3D volume were not changed from one MR acquisition to another.

Color Composite MR Images

We created multicontrast, normalized grayscale images by linearly mapping the T1W, PDW, and T2W images to red, green, and blue channels, respectively. In each color channel, black was mapped to zero, and white was mapped to 255. The color composite images were additionally rendered in 3D with the use of Matlab (MathWorks, Natick, MA). Multicontrast images were normalized in the sense that pixel intensity values were adjusted to make use of all available bandwidth afforded by the pixel's data type. For example, if the image's pixel values ranged from 47 to 203 (the full range is 256 levels of gray), a linear operation was applied to map the original pixel values to span the entire range (0–255).

Histopathology

Each specimen underwent decalcification, paraffin embedding, and serial sectioning. Sections (5 μ thick) were selected and stained with a combined Masson's elastic (CME) stain. Two digitally imported photomicrographs of histopathology slices from each sample were randomly chosen for analysis, and tissue component areas were manually traced. All of the tissue components (loose fibrous, media, dense fibrous, fibrocellular (cap), thrombus, and lipid/necrotic core) present in the histopathology specimens were identified (13, 14). The analyzed MR images were compared with matched “gold standard” pathology tracings.

Cluster Analysis

Cluster analysis (15, 16) is an objective exploratory technique that partitions data into groups, or clusters, with strong associations between members of the same cluster, by iteratively minimizing a specified aspect of the dataset. In our study, two aspects of the image pixel data were examined: the image's color variance (cluster compactness in color space), and spatial information (local pixel discontiguity).

We minimized chromatic variance by finding the position of the cluster centroids, where the sum of the distances between a given centroid and all points within the cluster (in RGB color-space) was minimal. The image's color variance of a cluster K is quantified as follows:

equation image(1)

Where nk is the number of elements in the kth cluster, xi is the ith pixel value, and ck is the center of the kth cluster.

We minimized pixel discontiguity by iteratively assigning each pixel to the cluster in which the majority of its neighboring pixels belonged. The degree to which pixel intensity changes in the spatial domain in each cluster is symbolized by Dk, and is quantified as follows:

equation image(2)

Where Di is the percentage of pixels that are not in the same cluster as their neighboring pixels, and nk is the number of elements in the kth cluster. SECA minimized a linear combination of these two criteria weighted by 1-β and β, respectively (Fig. 1b) as follows:

equation image(3)

The value of β determines the degree to which the algorithm weights minimizing either the local pixel contiguity or the cluster's variance in color-space. Although the value for β was determined empirically, it reflected the fact that while some images could be adequately segmented by conventional cluster analysis, most images benefited from the incorporation of spatial constraints, and therefore β was typically set to 0.9.

Figure 1.

a: Spatially enhanced cluster analysis method. Block arrows denote the order of operations. Multicontrast color composite MR image of the coronary artery shows large necrotic core (dark area) in between the thin fibrous cap (blue) and the narrow media (bright stripe). Resulting cluster images are given in pseudo-color. R, G, and B represent red, green, and blue channels, respectively. An expanded view of the box is shown in part b. b: Parameters undergoing minimization. A weighted linear combination (E, β (weighting coefficient)) of cluster compactness (Vk) and pixel discontiguity (D) is minimized iteratively until convergence is reached. The variance is an average Euclidian distance from a point in a given cluster to its centroid, where Vk is the variance of cluster k, Nk is the number of samples in the kth cluster, xi is the data point in the Ith sample, and ck is the center of the kth cluster (centroid). The local spatial discontiguity (Di, i (data index)) of a pixel is the ratio of its neighboring pixels that belong to a different cluster to the total number of neighboring pixels. The global spatial discontiguity (D) is the averaged Di over the total number of samples in the data set (N).

In the MRI data sets, SECA (Fig. 1a) produced data clusters that delineated different tissue components. Fig. 2 shows the flow chart of the SECA algorithm. We determined the number of clusters, and their seed points and associated tissue type for each MR image individually using ranges of color values established for each segmented tissue type by a histopathology-verified representative specimen (Table 1). The presence of specific tissue components was determined automatically when pixel values within the image of interest were matched to the representative color ranges. If a specimen contained a color range that was not described by the representative sample, SECA would automatically incorporate it into a new cluster. It is important to note that the representative color ranges were only approximations that initialized the number of clusters and seed points for the SECA algorithm without reference to other image data or histopathology slices. The normality of the intensity distributions of pixels within each cluster was tested by means of the Kolmogorov-Smirnov test. All of our data satisfied this requirement.

Figure 2.

Flow chart of the SECA procedure. The formulae for computing variance, discontiguity, and energy are given in Fig. 1b.

Table 1. Tissue Response Characteristics to T1, T2 and PD Weighted MRI
Tissue typeLoose fibrousMediaFibrocellularLipid/necr. coreThrombusDense fibrous
WeightingsT1WT2WPDWT1WT2WPDWT1WT2WPDWT1WT2WPDWT1WT2WPDWT1WT2WPDW
  1. Signal intensities:–dark (4 times below noise level), + bright (greater than 16 times noise level), ± medium.

MR signal intensity±+++±++±±±±+

We explored the applicability of SECA to segmenting in vivo images by computationally reducing ex vivo images until their appearance approximated what is typically achieved clinically (0.4 × 0.4 × 2 mm3), and subjecting them to SECA.

The color composite MR images of all 256 slices from the 12 specimens and two randomly chosen histology sections per specimen were subjected to SECA algorithms developed with Matlab. The typical analysis time was 10 s per 2562 image on an AthlonXP (1400MHz) computer.

Data and Statistical Analysis

T1W, PDW, T2W MR, color composite MR, and cluster-analyzed MR images were independently classified according to AHA's plaque classification scheme by trained observers (J.G.S.A., Z.A.F., V.V.I., V.M., and D.D.S.) on the basis of the signal intensity in the images for each plaque component (4, 5, 7). Table 1 and Fig. 3 show the signal intensities of different tissue components in multicontrast MRI. For example, dense fibrous tissue has a medium signal intensity (±) in T1W and T2W images, and a high signal intensity (+) in PDW images, while fibrocellular tissue has a medium intensity in TIW images, and a high intensity in T2W and PDW images.

Figure 3.

Color composite method for MR images. Left: Multicontrast images. Right: Composite picture, pseudo-color cluster analysis image, and histopathology image. T1W, PDW, and T2W axial plane images were assigned red (R), green (G), and blue (B) colors, respectively, and combined to yield a composite image of the coronary artery with color distributions corresponding to different tissue types. Right panels show a type Vb-Vc lesion. This lesion exhibits eccentric fibrous deposits and a dark area of calcification. The identified areas include dense fibrous tissue (df), calcium deposits (cal), media (med), and fibrous cap (fc). The CME histopathology shows excellent agreement with the MR images.

An expert pathologist (J.T.F.) identified the plaque type on histology slices, and rated them according to AHA classification (17). Cohen's κ (18) and P-value (StatExact-5; SYTEL, Cambridge, MA) were used to determine the statistical significance of the observers' agreement for plaque-type classification of the independently analyzed MR and histopathology images (Table 2).

Table 2. Qualitative Analysis of Coronary Plaques (AHA Classification)
Comparison typeCohen's KappaP-valueContingency tables of AHA classification for each lesion type
CLMRHistopathology
I–IIIIV–VaVb–VcVITotal
  • Total number of specimens was 12. The lesions were classified as: Type I–III, isolated macrophage foam cells, intracellular lipid accumulation; Type IV–Va, lipid/necrotic core and fibrotic layer; Type Vb–Vc, calcific or fibrotic; Type VI, complex with surface defect; hematoma-hemorrhage, thrombus.

  • a

    Denotes statistically significant agreement

Histopathology tracing and cluster-analyzed MRI (CLMR)0.89<0.0001aI–III30003
   IV–Va04004
   Vb–Vc00202
   VI01023
   Total352212
   CCMRHistopathology
    I–IIIIV–VaVb–VcVITotal
Histopathology tracing and color composite MRI (CCMR)0.780.0001aIV–Va31004
   IV–Va03003
   Vb–Vc00202
   VI01023
   Total352212
   T2WHistopathology
    I–IIIIV–VaVb–VcVITotal
Histopathology tracing and T2W grayscale MRI0.420.0268aI–III22004
   IV–Va12014
   Vb–Vc00202
   VI01012
   Total352212
   PDWHistopathology
    I–IIIIV–VaVb–VcVITotal
Histopathology tracing and PDW grayscale MRI0.310.1034I–III22004
   IV–Va12014
   Vb–Vc00213
   VI01001
   Total352212
   T1WHistopathology
    I–IIIIV–VaVb–VcVITotal
Histopathology tracing and T1W grayscale MRI0.290.1873I–III22004
   IV–Va12115
   Vb–Vc00101
   VI01012
   Total352212

Quantitatively, areas of different tissue types obtained from operator-independent SECA of color composite high-resolution and simulated in vivo MR images were compared with manually traced plaque components from histopathology. The trained observers assessed the agreement between the relative locations of tissues. To determine the efficacy of the SECA algorithm compared to other approaches, the high-resolution MR images were also analyzed by a less computationally demanding simple color-matching method, whereby composite color values representing tissue types were subdivided into a predetermined (from histopathology images) number of bins.

Pearson's correlation coefficient and Bland-Altman analyses were performed on tissue regions to compare 1) histopathology tracing with cluster-analyzed MR images, 2) histopathology tracing with reduced-resolution cluster-analyzed MR images, and 3) histopathology tracing with color-space matched MR images (Table 3). Additionally, a one-way analysis of variance (ANOVA) was performed on manually traced histopathology tissue areas, cluster-analyzed histopathology tissue areas, and cluster-analyzed color MR images.

Table 3. Quantitative Analysis and Signal Intensity of Plaque Components
Tissue typeLoose fibrous, N = 22Media, N = 22Fibrocellular, N = 20Lipid core/necrotic core, N = 7Thrombus, N = 3Dense fibrous, N = 11
Comparison typeRbiaslimrbiaslimrbiaslimrbiaslimrbiaslimrbiasLim
  1. Individual components' areas were computed as a percentage of total wall area. Bland Altman and correlation analyses were performed. N, number of specimens; r, Pearson's Correlation coefficient; bias, mean value of differences between paired measurements; lim, limits of agreement between paired measurements (95% confidence interval).

Histopathology tracing and cluster-analyzed MRI.822%±7%.840%±8%.890%±6%.792%±6%.983%±3%.831%±5%
Histopathology tracing and cluster-analyzed MRI (reduced resolution).713%±8%.612%±11%.555%±13%.750%±7%.854%±4%.743%±9%
Histopathology tracing and color-space matching.584%±9%.514%±14%.402%±9%.376%±11%.604%±7%.232%±11%

RESULTS

The specimens exhibited lesions that spanned the entire range of AHA plaque types I–VI. Table 2 shows the classification results and agreement between histopathology and all employed techniques for all lesion types.

The qualitative comparison between cluster-analyzed MR data and histopathology yielded a P-value of <0.0001. This indicates the reliability of the results in comparison with the gold standard (i.e., histopathology). The Cohen κ values comparing cluster analysis and color composite MR to histopathology were 0.89 and 0.78, respectively. According to the Cohen κ table, they were classified as very good and good correspondence, respectively (18). Therefore, since the numbers fell in two different categories, the cluster analysis was superior and the difference was statistically significant. The results indicate that the T2W images showed better agreement with histopathology compared to the T1W and PDW images. An example of an AHA type Vb-Vc plaque is shown (Fig. 3). AHA types I–III (early-stage plaque), IV-Va, and VI plaque are shown in Fig. 4.

Figure 4.

Examples of color composite MR (left), pseudo-color MR cluster (center), and CME pathology (right) images of AHA type VI (top panel), AHA type I–III (middle panel, artery cut enface), and AHA type IV-Va (bottom panel, artery cut enface) lesions. The type VI lesion shows the presence of thrombus (block arrow), the type IV-Va lesion contains necrotic core, and the type I–III lesion exhibits intimal thickening (block arrow). The following wall layers were distinguished: media (med), lipid core/necrotic core (lc/n), fibrocellular (fc; fibrous cap), adventitia (adv; loose fibrous tissue), dense fibers (df), calcium (cal; noise level signal on MR image), and thrombus.

We quantified the wall components on the MR and histopathology images by calculating the surface area of their associated clusters as a percentage of total vessel area. ANOVA showed no statistically significant difference between the tissue areas manually traced on histopathology, cluster-analyzed MR images, or cluster-analyzed reduced-resolution MR images.

Quantitative comparison data for each tissue component are presented in Table 3. There was a high degree of correlation (average Pearson's correlation coefficient = 0.86) between the histopathology tracings and the cluster-analyzed high-resolution MR images for all identified tissue components (loose fibrous, media, dense fibrous, fibro-cellular (cap), thrombus, and lipid/necrotic core; six in total for all specimens). With the current multicontrast MRI method, the signal from calcium is indistinguishable from noise, and thus was not included in the quantitative analysis. The results of SECA on reduced-resolution MR images showed a decreased but still acceptable correlation (average Pearson's correlation coefficient = 0.70) between the histopathology tracings and the cluster-analyzed areas on MR images.

To enhance visualization, we used 2D color composite MR images to reconstruct a virtually dissectable 3D object, which provided a more accurate spatial representation of the plaque structure and surface morphology (Fig. 5).

Figure 5.

Color composite 3D MRI renderings of human coronary arteries (a and b) with corresponding cluster analysis (c and d). Dense eccentric fibrous tissue deposits surround the vessel lumen in a and c, while necrotic cores (dark areas) and a fibrous cap are present in b and d. Subjecting the images in a and b to SECA yielded the quantitative volumes displayed in c and d.

DISCUSSION AND CONCLUSIONS

Multicontrast-weighted MR images of coronary artery specimens were subjected to SECA and 3D rendering. All lesions were classified according to the AHA classification. We identified and characterized all of the plaque components, and confirmed the results by correlating the MR and histopathological findings. We performed SECA on the low-resolution aortic images to investigate the applicability of the technique to in vivo data.

The decision was made early in the study to use stained histopathology samples as the gold standard against which all segmentations would be compared. Although a measure of region overlap between the segmented and histopathology images would be a better indicator of correlation, nonlinear warping of histology images (due to tissue shrinkage during preparation) makes this approach problematic. Therefore, we established a correlation between the SECA images and their associated histological images by comparing the areas of segmented labeled tissues relative to the vessels' overall area for each image. While this technique does allow for quantitative comparisons, it is not sensitive to differences in segmented tissue morphology. In addition, this correlation technique is inefficient in that it requires tissues to be labeled prior to correlation. In a larger study, measures that do not require labeling could be employed to further automate the validation procedures.

In the ex vivo images, high spatial resolution provided a “3D MRI histology” and allowed early-stage plaque to be discriminated. Qualitatively, plaque component identification was improved on the cluster-analyzed color MR images compared to the grayscale and composite color images (Table 2). This improvement was due to more accurate discrimination of thrombus and lesion size on cluster-analyzed images.

The signal intensities of the plaque components in this study were in agreement with those determined in previous studies (4, 7). In contrast to the Shinnar et al. (4) method, our technique does not use a diffusion-weighted modality; instead, a 3D acquisition is employed, which results in improved spatial resolution in the slice direction by one order of magnitude. Qualitative analysis was performed in both studies; however, in the present work, color composite images were used. Fomblin eliminated the need for user interaction to remove background signal on the MR images. Qualitatively, there was very good agreement with histopathology in both studies; however, the present study also had a quantitative aspect, since the areas of different plaque components were determined by SECA and correlated with “gold standard” histopathology tracings.

While the benefit of classification by cluster-analysis MRI as compared to classification with color composite images is subtle, there is a significant advantage to the cluster analysis that is not reflected in this score. The well-defined cluster analysis technique produces a quantitative assessment that is amenable to further analysis. Evaluations of color composite images, on the other hand, are based on an effectively unknown algorithm, are highly subjective, and produce a largely qualitative result.

In the quantitative analysis, the SECA technique demonstrated a substantially higher correlation with histopathology compared to the composite color matching method (Table 3). In the computationally reduced MR images, the decreased spatial resolution resulted in tissue blurring and loss of discrimination of the fibrous cap in two specimens.

SECA is computationally less demanding than hierarchical clustering, simpler than density-based clustering, and more capable of incorporating spatial considerations. SECA does not require user-defined cluster description parameters, and, unlike model-based clustering, it can handle large datasets.

Despite the advantages offered by SECA, it is important to acknowledge that it is not a panacea for all of the difficulties involved in objective tissue segmentation. For example, K-means cluster analysis may not always yield the unbiased, objective results that we would expect for intensities with a normal distribution. Fortunately, a Kolmogorov-Smirnov analysis of our data indicates that the intensity values of the tissue regions do indeed appear to be normally distributed.

In addition, like many search algorithms, K-means cluster analysis suffers from a tendency to converge on local optima that may not necessarily be globally optimal. Although this deficiency is inherent to the algorithm, it is to some degree related to the initial seed points (or cluster centroids) provided at the onset of the analysis (Table 1). In fact, the traditional solution to this problem is to perform several cluster analyses using a variety of initial seedings, and choose the segmentation with the minimum cluster variance. Although this approach may reduce nonoptimal clustering, it does so at the expense of increased computation time. In our application, since the initial seed points are based on color values established for each segmented tissue type typically present in the image, non-optimal conversion should be reduced. In addition, SECA's incorporation of both spatial and chromatic information should help minimize any spurious convergences that might arise from the analysis based on only one criterion. However, it should be noted that the difficulty of eliminating this problem completely is not trivial. Whether plaque segmentation can be improved by the use of variants of cluster analysis alone will be the subject of further research.

An important limitation of this study is the poor calcium discrimination by MRI. Gradient-echo sequences are more sensitive to calcium imaging (19) because of the induced artifact of field inhomogeneity around the calcium deposits. Tissue shrinkage due to the histological process might contribute to lower correlation values between histopathology and MRI measurements (20, 21). Combining time-of-flight imaging and contrast-enhanced MRI (22) with multicontrast-weighted MRI can improve fibrous cap characterization in vivo (7, 23). Although SECA has significant potential for characterizing plaque in ex vivo coronary images, we cannot overlook the difficulties that arise when this technique is applied to in vivo data. Although we obtained encouraging results when we applied SECA to simulated (reduced-resolution) in vivo data, the limitations of the technique are significant. Future studies are needed to evaluate the effects of diminished in-plane resolution, reduced SNR, image artifacts, and other imaging parameters on the performance of SECA in terms of plaque characterization.

In addition, complications such as image artifacts and registration difficulties between different modalities must be addressed. Registration can be achieved through the use of the automated image registration (AIR) algorithm (24), while improvements in MRI (such as high-field clinical systems) may also lead to improvements in the signal-to-noise ratio (SNR). Such advances may make it feasible to use SECA for plaque characterization.

In conclusion, this study demonstrates that cluster analysis can objectively quantify ex vivo atherosclerotic plaque components, and suggests future directions of research to determine the clinical utility of cluster analysis methodology for evaluating atherosclerosis by in vivo MRI.

Acknowledgements

We thank Drs. Samuel S. Gidding (Alfred duPont Hospital for Children), Edward A. Fisher (Mount Sinai School of Medicine), C. Alex McMahan (University of Texas Health Science Center at San Antonio) for guidance, helpful discussions, and providing the specimens.

Ancillary