Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
JOURNAL OF MAGNETIC RESONANCE IMAGING 24:379 –387 (2006) Original Research Automated Measurement of Mean Wall Thickness in the Common Carotid Artery by MRI: A Comparison to Intima-Media Thickness by B-Mode Ultrasound Hunter R. Underhill, MD,1 William S. Kerwin, PhD,1 Thomas S. Hatsukami, MD,2,3 and Chun Yuan, PhD1 Purpose: To determine whether the mean wall thickness (MWT) of the common carotid artery (CCA) measured by MRI is comparable to B-mode ultrasound (US) measurement of the intima-media thickness (IMT), an established marker of cardiovascular risk. Materials and Methods: As part of the two-year ORION trial, 43 patients with 16 –79% stenosis by duplex US underwent high-resolution MRI and B-mode US examinations of their carotid arteries. Twenty-eight carotid arteries were identified as having both sufficient proximal coverage and adequate image quality of the CCA on MRI and a corresponding US. A novel algorithm utilizing statistical shape modeling was developed to automatically detect and measure MWT to within subpixel accuracy. The interrater and interscan reproducibility of the MWT measurement was computed as the rootmean-square (RMS) difference. The MWT and IMT measurements were compared via the Pearson correlation coefficient. Results: The MWT and IMT had a high Pearson correlation coefficient (r ⫽ 0.93; P ⬍ 0.001). The RMS difference between readers and between scans was 0.01 mm and 0.04 mm, respectively. Our automated algorithm correctly identified the lumen in 28 cases (100%) and the outer-wall boundary in 26 cases (93%). Conclusion: Automated measurements of the MWT by MRI are reproducible and have a high correlation with the IMT by B-mode US. Key Words: carotid artery MRI; atherosclerosis; intima-media thickness; statistical shape modeling; B-mode ultrasound J. Magn. Reson. Imaging 2006;24:379 –387. Published 2006 Wiley-Liss, Inc.† DUE TO THE BURGEONING COST of health care, greater emphasis is being placed on preventative med- 1 Department of Radiology, University of Washington, Seattle, Washington, USA. 2 Department of Surgery, University of Washington, Seattle, Washington, USA. 3 Veteran’s Administration (VA) Puget Sound Health Care System, Seattle, Washington, USA. Contract grant sponsor: AstraZeneca Pharmaceuticals; Contract grant number: ZD4544; Contract grant sponsor: Cardiovascular Research Training Program; Contract grant numbers: T-32; HL07828. *Address reprint requests to: H.R.U., Department of Radiology, Vascular Imaging Laboratory, University of Washington, 815 Mercer St., Box 358050, Seattle, WA 98109. E-mail: [email protected] Received November 28, 2005; Accepted April 27, 2006. DOI 10.1002/jmri.20636 Published online 19 June 2006 in Wiley InterScience (www.interscience. wiley.com). Published 2006 Wiley-Liss, Inc. †This article is a US Government work and, as such, is in the public domain in the United States of America. icine and early intervention in an attempt to reduce the morbidity and mortality associated with cardiovascular disease, the leading cause of death in the world (1). As a consequence, strategies for identifying individuals at risk for future events, and measuring response to therapy are in demand. B-mode ultrasound (US) is a noninvasive technique that has been shown to identify the intimal and medial layers of the carotid artery (2). In epidemiological studies the carotid intima-media thickness (IMT) has been associated with prevalence of cardiovascular disease and involvement of other arterial beds with atherosclerosis (3,4). Furthermore, in a landmark paper by O’Leary et al (5), increases in carotid IMT were directly associated with an increased risk of myocardial infarction and stroke in older adults with no history of cardiovascular disease. Consequently, carotid IMT has emerged as a marker for cardiovascular disease (6) and has been used as an endpoint in clinical trials assessing the effect of pharmacological treatment of systemic atherosclerosis (7–10). In this study we sought to determine whether MRI could make an equivalent measurement of carotid wall thickness. One reason for investigating MRI as an alternative to US is that the variability of IMT measurements by B-mode US limits its use in large, populationbased studies. Although measurement variability (interreader) has been reduced via the introduction of automated edge detection algorithms (11), interscan reproducibility errors persist due to the effects of variations in the incident angle and body habitus that occur during image acquisition (12). As such, individual or even small-population studies of risk assessment and response to therapy are currently not viable. The excellent anatomical detail afforded by the high spatial resolution of carotid MRI has established MRI as a highly reproducible option for studies of plaque size and composition in smaller populations (13). Comparisons of carotid MRI with histological ground truth in previous studies (14 –16) have also established its exceptional sensitivity and accuracy for measuring plaque size and identifying plaque components. However, these studies did not address measurements of the thin-walled region of the common carotid artery (CCA) proximal to the plaque. Thus, the ability of MRI to measure wall thickness in the proximal CCA, where IMT is measured, has not been established. 379 380 We hypothesized that cross-sectional MR images obtained from the nondiseased segment of the proximal CCA could be used to measure the mean wall thickness (MWT), which would then provide a direct correlate to carotid IMT measured by B-mode US. Additionally, the MRI measurements would have the potential to provide higher reproducibility. Furthermore, MRI could then be used to assess local plaque characterization in a single imaging session, using established techniques, and provide a global assessment of cardiovascular risk by measuring the MWT. To test this hypothesis we developed an automated algorithm for measuring MWT by MRI and compared the results with IMT measured by B-mode US within a clinical population. MATERIALS AND METHODS Subject Population As part of the two-year outcome of rosuvastatin treatment on carotid artery atheroma: a magnetic resonance imaging observation (ORION) (17) trial with rosuvastatin (4552IL/0044), a total of 43 subjects (30 men and 13 women, mean age ⫽ 65 years) with 16 –79% carotid stenosis by duplex US were recruited from the University of Washington Medical Center, VA Puget Sound Health Care System, and the University of Utah Medical Center for serial high-resolution MRI and B-mode US examinations of their carotid arteries. The study protocol and informed-consent forms were approved by the institutional review boards of each institution. From the initial population, 40 patients were identified as having both an MRI and B-mode US separated by no more than four weeks of at least one carotid artery. Images from these subjects were reviewed to identify those that were appropriate for a retrospective study comparing automated MWT by MRI to IMT by B-mode US. Exams were selected based on adequate image quality (IQ) and the presence of sufficient proximal coverage of the CCA that was comparable to the region covered by B-mode US (5–15 mm proximal to the bifurcation). Since the imaging protocol for the initial population under the ORION trial was not specifically designed for direct comparison of MWT and IMT in corresponding locations at the outset, not all cases were guaranteed to contain such coverage. Beginning with the first timepoint, each exam was successively reviewed until at least two adjacent slices, both of American Heart Association (AHA) lesion type I-II, were identified in the spec- Underhill et al. ified range on at least one carotid artery. Only one exam was selected from each patient. From the 40 available patients, 26 subjects were found to have sufficient coverage. Additionally, six of these patients had sufficient coverage bilaterally, for a total of 32 arteries. Due to poor IQ, four (12%) of these arteries were excluded, yielding a final total of 28 arteries for comparison with B-mode US. Finally, a subset of arteries (N ⫽ 8) had two MRI scans separated by no more than 14 days, which enabled a preliminary assessment of interscan reproducibility. B-Mode Ultrasound For each subject, anterior, posterior, and medial longitudinal views of the left and right common carotid walls were obtained by B-mode US. In each view a 1-cm segment of the CCA was identified, centered 1 cm proximal to the bifurcation. The intimal and medial echoes of the far wall were identified using Q-Lab (Philips Medical Systems). If plaque was identified at the distal end of the measurement near the bifurcation, the region of interest (ROI) for boundary detection in Q-Lab was reduced to include only the nondiseased segment. The average thickness along the identified segment was computed (Fig. 1). Finally, the average of the measurements from all three views, to account for eccentricity, was recorded as carotid IMT for the respective artery. MRI A standardized carotid imaging protocol (18) was used to acquire MRI data on a 1.5 T Scanner (Signa Horizon EchoSpeed; GE Medical Systems) using bilateral phased-array surface coils. The imaging protocol included a T1-weighted (T1W) black-blood (double inversion recovery, TI ⫽ 650 msec) fast spin-echo sequence that was deemed best for morphological assessment of the vessel (19). The imaging parameters were TR ⫽ 800 msec, TE ⫽ 9.3 msec, echo train length equal to 8, matrix size ⫽ 256 ⫻ 256, slice thickness ⫽ 2 mm, two averaged excitations, and fat suppression. The protocol did not include breath-holding, and the total time for acquisition of the T1W sequence was four minutes. The field of view (FOV), as determined by the technician at the time of the scan, was based on neck thickness and was equal to 1) 13 cm (N ⫽ 17), resulting in a 0.51-mm resolution; 2) 15 cm (N ⫽ 1), resulting in a 0.59-mm resolution; or 3) 16 cm (N ⫽ 10), resulting in a 0.62-mm resolution. The respective pixel sizes were 0.25, 0.29, Figure 1. a: Original B-mode US image of CCA centered 1 cm proximal to the bifurcation (*). b: The blue lines demonstrate automatic IMT detection by Q-Lab (Philips Medical Systems) over a 1-cm segment. Carotid Wall Thickness: MRI vs. US and 0.31 mm after zero-filled interpolation to an image size of 512 ⫻ 512. Axial images were acquired at 2-mm intervals for a total longitudinal coverage of 2 cm, centered at the carotid bifurcation of the side with greater stenosis. Immediately after acquisition, all images were rated for IQ as adequate (discernable boundaries/ structures) or poor (vessel boundaries obscured beyond recognition). Examples of each category are displayed in Fig. 2. Because the absence of discernable boundaries makes visual verification of boundaries impossible, subjects who exhibited poor IQ in the CCA were excluded from this study. MWT Measurement The technical challenge of measuring MWT is to automatically detect the lumen and outer-wall boundaries of the CCA within an axial T1W image. This can be difficult when incomplete blood-flow suppression leaves residual flow artifacts, which can obscure the border between the lumen and wall. Also, the signal intensity of some of the tissues surrounding the outerwall boundary is similar to that of the wall, which can make that boundary indistinct. Additionally, there are frequently structures, such as the jugular vein, in the area surrounding the carotid artery that produce additional edges, which can lead to erroneous solutions if traditional edge-detection techniques are implemented. Finally, noise secondary to random patient movement, swallowing, or arterial pulsation can lead to further border obscuration. We limited these obstacles by restricting the search space to shapes that resembled the predictable anatomic structure of the CCA in this region. This was accomplished by using active shape models (ASMs) in the detection of both boundaries. ASMs utilize a statistical model of shape to restrict the search space to anatomically reasonable solutions. These anatomically reasonable shapes, which are based on a set of training shapes, are defined in the ASM by a mean shape plus one or more modes of variation or “eigenshapes.” ASM Training The procedure used to train an ASM is identical for both the lumen and outer wall. To prevent redundancy, only the details of the steps used to create the lumen ASM will be presented, but the final model for each will be Figure 2. Right CCA, axial T1W image representations of adequate (a) and poor (b) IQ. An asterisk (*) is used to identify the lumen. 381 demonstrated. Details regarding the steps outlined below are available in Ref. 20. From a separate database of high-resolution carotid MRI that used similar imaging parameters, 20 carotid arteries with adequate coverage were selected. From each artery, a single slice 5–15 mm proximal to the bifurcation and with no evidence of disease was selected. These 20 images of different CCAs formed the training set. Given the inherently circular nature of the CCA, there were no definable “landmark points.” Therefore, the training set was labeled in the following manner: 1. Via graphical user interface, an expert in carotid MRI selected 15–20 points that marked the luminal boundary. 2. These points were then used to initialize a B-spline snake that detected the contour of the luminal boundary. 3. The center of gravity of this contour was calculated. 4. From this center point, 16 radial lines were spaced every 22.5°. 5. Each luminal shape was represented by its center point and the 16 points of intersection between the individual radial lines and the contour. These steps are graphically illustrated in Fig. 3. As a result of this procedure, each shape was represented by a 34-element vector xi consisting of the 16 radial points and the center of gravity as in x i ⫽ 共xi0,yi0,xi1,yi1L,xin⫺1,yin⫺1兲T (1) The goal in an ASM is to represent this object, which has 34° of freedom, more compactly with 冘 k x i ⫽ x ⫹ bikpk (2) k⫽1 where x is the mean shape, pk are K fixed vectors, and bik are variable weights that define the different shapes. In this approximate representation, each shape is represented with only K (⬍⬍34) degrees of freedom. The first step in the ASM method to determine x and pk is to align all the shapes in the training set. Any two similar shapes xi and xj are aligned by finding the rota- 382 Underhill et al. 冉冘 冊 ⫺1 n⫺1 wk ⫽ V Rkl . (5) l⫽0 Based on this method, the following algorithm is used to align a set of N shapes: 1. Rotate, scale, and translate each shape to align with the first shape in the set. 2. Repeat step 1. 3. Calculate the mean shape from the aligned shapes. 4. Normalize the orientation, scale, and origin of the current mean to the initial first shape. 5. Realign every shape with the current mean until the process converges. After the set of N shapes is aligned, the mean shape x is determined by Figure 3. The different steps used in labeling a shape for the training set. a: Original image. b: LC via graphic user interface (central asterisk marks the center). c: The 16 radial lines extending every 22.5° from the center. The 17 asterisks in d represent the final labeled luminal shape—the center point along with the 16 points of intersection between the luminal contour and radial lines. 1 x ⫽ N 冘 N xi and the vectors pk are determined as follows. Let D be a 2n ⫻ N matrix with the following columns: D ⫽ (x 1,x 2,L x N). tion j scaling factor sj, and translation (txj,tyj) that minimize the weighted sum E j ⫽ 共xj ⫺ M共sj,j兲关xj兴 ⫺ tj兲tW(xj ⫺ M共sj,j兲关xj兴 ⫺ tj), (3) where M共s,兲 ⫺ 共ssin兲y 冏 yx 冏 ⫽ 冉 共scos兲x 共ssin兲x ⫹ 共scos兲y 冊 jk jk jk jk t j ⫽ 共t xj ,t yj ,L,txj,tyj兲T, jk jk (4) and W is a diagonal matrix of weights for each point defined as follows: let Rkl be the distance between points k and l in a shape, and let VRkl be the variance in this distance over the set of shapes. The weight, wk, is chosen for the kth point using (6) i⫽1 (7) Let the covariance matrix, S, be defined as S⫽ 1 DDT N (8) Then, the N unit orthogonal eigenvectors of S are pi(1,L,N), with corresponding eigenvalues i, ordered from largest to smallest. The vectors pk are commonly referred to as “eigenshapes.” For our lumen model we assume that any allowable shape can be represented by Eq. [2] with K ⫽ 2, and bik within ⫾2√. The influence of each eigenvector pk on the mean luminal shape x is represented in Fig. 4 because they are ranged over ⫾2√. After an identical process, the mean outer-wall shape x is obtained. The influence of its two dominant eigenvectors is represented in Fig. 5 as they are ranged over ⫾2√. Figure 4. Effects of the two most prominent eigenvectors (p1 and p2 on the luminal mean shape (solid). The left graph depicts p1 as it is ranged from ⫺2√1 (dashed) to ⫹2√1 (dotted). Notice that p1 purely regulates size. The right graph depicts the effect of p2 on x as it is ranged in a similar fashion over ⫾2√2. Note that p2 has a distinctly different, albeit minor, effect on x . Carotid Wall Thickness: MRI vs. US 383 Figure 5. Effects of the two most prominent eigenvectors (p1 and p2) on the outer-wall mean shape (solid). The left graph depicts p1 as it is ranged from ⫺2√1 (dashed) to ⫹2√1 (dotted). Note that p1 purely regulates size. The right graph depicts the effect of p2 on x as it is ranged in a similar fashion over ⫾2√2. Note that p2 has a distinctly different, albeit minor, effect on x . Carotid Identification A user identifies a point, P, via graphical user interface within the lumen of the CCA. From a 9 ⫻ 9 pixel region about P, the maximum grayscale value, which represents luminal noise, is found. From P, search-vectors are passed bidirectionally along the x-axis. Each vector roughly identifies the lumen boundary by identifying the point at which the intensity exceeds the maximum grayscale value of the 9 ⫻ 9 region. The midpoint is determined and then two more search-vectors are passed bidirectionally along the orthogonal axis (yaxis). In a similar fashion the lumen boundary is roughly identified. The midpoint is calculated and, given the inherent circular nature of the CCA, serves as an estimation of the center of the lumen. From this center point, all additional steps begin. Lumen Boundary Detection Our luminal ASM is then released and allowed to pass through all possible shapes under the constraints of only the first two eigenshapes, since they represent 99% of all the training shapes, and a shift vector that accounts for errors in the detected lumen center. The use of an exhaustive search, instead of having the shape iteratively deform based on local edge information, prevents the contour from being trapped by a local minimum. As the ASM steps through each successive shape, fitness is determined by combining local edge information and area. At each of the outer 16 points, the local edge is calculated by averaging the gradient of the two nearest pixels along a line radiating outward from the center point. For each point, if the result exceeds a predefined optimized threshold based on the training set, a point is awarded. Two points are awarded if the result is greater than twice the threshold. The shape with the highest number of points and largest area is declared the optimal solution of the luminal ASM. To finalize the detected boundary, we utilized a B-spline snake (21) that parameterized the boundary contour as a closed, cubic B-spline with six knots. The B-spline was initialized by finding the least-squares approximation of the 16 points from the ASM. The B-spline was then modified using gradient ascent to maximize the average of the squared image gradient along the length of the contour. Outer-Wall Boundary Detection In a similar fashion, the outer-wall ASM is released, but with the following necessary variations. First, it is constrained by the previously detected lumen contour (LC). Possible outer-wall boundary solutions with a maximal point ⬎5 mm or ⬍0.25 mm from the LC are rejected. Second, the fitness of each possible solution is based on both local and global edge information. For each point of the possible solution, a greater fitness value is awarded if the point represents not only the desired gradient, but also the first large negative gradient along a radial line extending outward from a corresponding point along the LC. This rule reflects the relatively homogenous, hyperintense appearance of the common carotid wall in this region. Once an optimal ASM solution is obtained, the individual 16 points representing the outer-wall boundary are used to initialize a B-spline snake on the image for final contour detection. Because of anatomic similarities, the final eigenvalues (1) that control size for the lumen and outer-wall active-shape models at the initial slice are used to significantly refine the respective searches in the adjacent more-distal slice. Instead of iterating over the entire range of 1, the search is localized to within a radius of ⫾ 0.25 mm of the respective proximal contours. Otherwise, the identical process for boundary detection is repeated. MWT Once the LC and outer-wall contour (OW) are established, the mean distance between the inner and outer contours is calculated at each slice level. To accurately calculate mean distance, we implemented an algorithm that utilizes the concentric circular structure of the CCA. The LC and OW can be represented by the points determined from the B-spline snake, which are at intervals of approximately 1 pixel giving the sets LC ⫽ 关 LC1 LC2 ... LCn 兴 384 Underhill et al. OW ⫽ 关 OW1 OW2 ... OWn 兴, 共m ⬎ n兲. (9) The distance from a point in LC to all points in OW is given by the Euclidean distance d i,j ⫽ 储LC i ⫺ OW j 储, (10) and the local thickness for a given point LCi is given by D i ⫽ min dij. 1ⱕjⱕm Once D1 is determined, D2 is determined in a similar manner with the next point from LC, except that the corresponding point from OW is removed to prevent the use of duplicate points. In this fashion, thickness is determined from the minimal distance for each LC point to the outer wall. MWT is then determined by the mean of all the n individual thickness measurements from each axial location. Data Analysis Two reviewers (one with experience in carotid MRI, the other an ABR board-certified radiologist without experience in carotid MRI) were trained in the use of our automated MWT algorithm and performed measurements on the data set. Successful detection of boundaries was evaluated qualitatively for each subject by assessing agreement with perceived boundaries. For correctly identified boundaries, interreader reproducibility was computed as the root-mean-square (RMS) difference between independent reviewer results. For the subset of arteries with two scans separated by no more than 14 days, a single reviewer applied the automated technique to both scans. Interscan reproducibility was computed as the RMS difference between results, and a Bland-Altman analysis was applied to assess for bias. Quantitative comparisons with US IMT measurements were conducted using the Pearson correlation coefficient and Bland-Altman analysis. The data were also partitioned based on the FOV, and quantitative comparisons were again made with US using the Pearson correlation coefficient. RESULTS Of the 28 cases available for comparison, the automated lumen and outer-wall boundaries detected by our algorithm visually coincided with the apparent lumen and outer-wall boundaries in 28 (100%) and 26 (93%) cases, respectively. The following figures demonstrate a sample of the results from successful (Fig. 6) and unsuccessful (Fig. 7) automated lumen and outer-wall boundary detection in the CCA on MRI. A repeated application of our algorithm on the same image set by a separate trained observer demonstrated the boundary detection to be highly reproducible. The RMS difference in MWT between reviewers was 0.01 Figure 6. Demonstration of successful lumen and outer-wall boundary detection of the CCA on axial T1W images with both “adequate” (top row) and “marginal” (bottom row) IQ. mm. The algorithm detected the incorrect border in the same two cases as the first trial. Comparing MWT results from the subset of arteries with two scans (N ⫽ 8) demonstrated an RMS difference of 0.04 mm. Bland-Altman analysis (Fig. 8) did not reveal a significant bias. When compared to IMT by B-mode US, MWT by MRI for the entire image set had a Pearson correlation coefficient equal to 0.93 (P ⬍ 0.001; Fig. 9). The results of the Bland-Altman analysis are demonstrated in Fig. 10. This shows a substantial and statistically significant upward bias in the MWT measurement (P ⬍ 0.001 in a paired t-test). An apparent reduction in the bias at higher values is also apparent. Separating the images based on FOV, we found that images obtained with an FOV ⫽ 13 cm (N ⫽ 16) had a minimal recorded thickness of 0.80 mm and the strongest correlation (r ⫽ 0.95) with IMT. Although images that used an FOV of 16 cm (N ⫽ 9) had only a minimal recorded thickness of 0.91 mm, there was still a strong correlation (r ⫽ 0.82) with IMT. The difference between these two correlations was not statistically significant. DISCUSSION This study represents the first comparison between wall thickness measurements by MRI and B-mode US. We found a very high correlation between IMT by Bmode US and automated MWT by MRI, which strongly indicates that it is possible to use carotid MRI as a tool for assessing systemic atherosclerotic disease. The consistently larger MRI wall thickness measurements compared to the US results is probably a consequence of two factors. First, MWT may include the adventitia. IMT by B-mode US has been pathologically Carotid Wall Thickness: MRI vs. US 385 Figure 9. Comparison of IMT by B-mode US to MWT by MRI. Although MWT was generally thicker than IMT, the data had a very high correlation of r ⫽ 0.93. Figure 7. The borders of the proximal slice (upper panel) were correctly identified, but automated border detection in the adjacent more-distal slice failed. The indistinct border (arrow) resulted in an outer-wall boundary solution that did not visually coincide. validated to measure only the combined thickness of the intima and media (2). However, the medial-adventitial border is not readily apparent on MRI. Validation results performed on carotid endarterectomy subjects show larger volume measurements on in vivo MRI than the corresponding volume measurements of the specimens, which lack adventitia (19). Unfortunately, the average thickness of the adventitia has not been well defined, because it blends with supporting soft tissues and may be thickened in diseased arteries (29). However, our findings are consistent with a study by Crowe et al. (22) in which they concluded that wall area mea- Figure 8. Bland-Altman plot of interscan reproducibility data. No significant bias is present. surements by MRI were greater than US area measurements due to the inclusion of the adventitia by MRI. Resolution may be a second factor in the discrepancy. With increasing IMT and subsequent decreasing demand on resolution, a discrepancy persisted but was considerably smaller. The highest available effective resolution for this study was 0.51 mm (N ⫽ 16) with a minimum recorded thickness of 0.80 mm. The lowest effective resolution was 0.61 mm (N ⫽ 9) with a minimum thickness measurement of 0.91 mm. Together these facts suggest that for relatively thin arterial walls, resolution may contribute more significantly than adventitial inclusion to the measurement of MWT. Therefore, techniques that further improve resolution should have a profoundly positive influence on the results. Possibilities include the use of a smaller FOV to boost in-plane resolution, possibly with thicker image slices to increase the signal-to-noise ratio (SNR). Although thicker images are undesirable in the rapidly varying plaque region, they would be appropriate for the rela- Figure 10. Bland-Altman analysis demonstrates an overmeasurement of IMT by MWT that decreased as IMT increased. 386 tively uniform region where common carotid IMT is traditionally measured. Alternatively, SNR in the proximal common carotid may be increased with improved surface coils that provide greater coverage of the carotid artery. Additionally, modification of the image acquisition parameters may also enhance SNR by decreasing flow artifact (30), improving fat suppression, or enhancing tissue boundaries. Our described technique for automatic boundary identification correctly identified the nondiseased segment of the CCA wall. Restricting solutions via an ASM to only biologically reasonable results is an effective way to overcome many of the technical difficulties involved in automatically detecting shapes in MRI. The ASM can be limited by the training set; however, the relatively consistent shape of this segment of the CCA yielded a very robust model. Although our technique requires the user to identify a point within the lumen to initialize boundary detection, point selection had a minimal effect on algorithm performance since we found a high level of reproducibility between two independent reviewers. The implementation of a B-spline snake for final boundary detection produced very refined measurements. Although absolute resolution limits the minimum thickness measurement, the B-spline snake provides highly accurate measurements of any thickness greater than the minimum because of its subpixel accuracy (23). This feature is similar to strategies employed by automated measurements of IMT (24), and is crucial for detecting changes in MWT since the annual rate of change is less than the available resolution. Although interscan reproducibility could only be tested on a limited number of arteries, the results are similar to the highest level of reproducibility available with US (12,25). However, the reproducibility of IMT by B-mode US is well established, which has enabled IMT to be used as a surrogate for coronary artery disease (6) and a biomarker in the evaluation of new medications (7–10) in large population-based studies. Before further investigations are conducted, the interscan reproducibility for MWT should be carefully evaluated in a large population with a protocol directed toward maximizing SNR and resolution, as discussed above. With these potential improvements in study design, we would expect a higher level of interscan reproducibility compared to that achieved in this study. It would then be possible to reduce the population size necessary to detect change during clinical trials. Additionally, because of the possible inclusion of the microvascular-laden and dynamic adventitia in the MRI wall thickness measurement, the amount of acceptable reproducibility error required to detect change may be larger than that for IMT. Although the results for MWT by MRI are promising, several areas for potential improvement exist within the imaging protocol. The current protocol is focused at levels immediately adjacent to the bifurcation, where most local atherosclerotic disease occurs. If common carotid MWT is to be measured by MRI, the protocol should be adjusted to ensure sufficient proximal coverage of the CCA. Another potential improve- Underhill et al. ment is the use of contrast-enhanced MRI, which has been shown to provide better delineation of vessel boundaries (26). The adventitial enhancement with contrast administration may provide better characterization of adventitial involvement across various thicknesses to determine its role in MWT changes. Finally, the automated MWT by MRI algorithm should be evaluated on other contrast weightings typically obtained in carotid MRI, specifically T2- and protondensity-weighted MRI. In related studies, these weightings were found to be interchangeable with T1W MRI for morphological measurements, which suggests that MWT by MRI should be applied to the weighting with the best image quality (27). Several limitations of this study need to be considered. First, the IMT is based on the average thickness along a longitudinal segment of the artery, whereas the MWT is based on the average thickness within a cross section of the artery. Although these two views are different, they represent the viewing orientations that best depict the vessel within each modality. It is unclear, however, which orientation is best for assessing systemic atherosclerotic disease. Second, the MRI protocol used in the ORION study imaged a 2-cm segment of carotid artery centered at the bifurcation of the vessel with greater disease (the index artery), while US evaluated a segment centered 1 cm proximal to the bifurcation for both arteries. As such, the coverage of the index artery frequently did not include more proximal segments of the artery, as desired in the present study, which is why many cases were excluded. However, the bifurcation of the non-index artery was frequently located at a different level, which resulted in greater coverage of the proximal non-index CCA since images of both arteries were acquired simultaneously. Only arteries with overlapping MRI and US coverage were used for comparison, and future investigations should include a protocol that specifically covers the proximal CCA bilaterally to increase the number of arteries available for study. Third, IQ prevented the inclusion of 12% of arteries in this study. Although this was similar to other carotid MRI studies (14,28), the design of protocols specifically oriented toward MWT measurements, as previously discussed, should significantly reduce this percentage. In conclusion, we have shown in the CCA a high correlation between IMT by B-mode US and MWT measurements by MRI. The automated technique we developed affords high interreader reproducibility and measurements to within subpixel accuracy. Additionally, the preliminary results we obtained for interscan reproducibility appear promising. With additional development, it may be possible to replace large populationbased IMT studies with small- or individual-based MWT studies. ACKNOWLEDGMENTS The authors thank Michelle Bittle, M.D., for her assistance in determining the interreader variability. Carotid Wall Thickness: MRI vs. US REFERENCES 1. World Health Organization. World Health Report 2004. 2. Pignoli P, Tremoli E, Poli A, Oreste P, Paoletti R. Intimal plus medial thickness of the arterial wall: a direct measurement with ultrasound imaging. Circulation 1986;74:1399 –1406. 3. Allan PL, Mowbray PI, Lee AJ, Fowkes FG. Relationship between carotid intima-media thickness and symptomatic and asymptomatic peripheral arterial disease. The Edinburgh Artery Study. Stroke 1997;28:348 –353. 4. Bots ML, Hoes AW, Koudstaal PJ, Hofman A, Grobbee DE. Common carotid intima-media thickness and risk of stroke and myocardial infarction: the Rotterdam Study. Circulation 1997;96:1432–1437. 5. O’Leary DH, Polak JF, Kronmal RA, Manolio TA, Burke GL, Wolfson Jr SK. Carotid-artery intima and media thickness as a risk factor for myocardial infarction and stroke in older adults. Cardiovascular Health Study Collaborative Research Group. N Engl J Med 1999; 340:14 –22. 6. Bots ML, Grobbee DE. Intima media thickness as a surrogate marker for generalised atherosclerosis. Cardiovasc Drugs Ther 2002;16:341–351. 7. Taylor AJ, Kent SM, Flaherty PJ, Coyle LC, Markwood TT, Vernalis MN. ARBITER: Arterial Biology for the Investigation of the Treatment Effects of Reducing Cholesterol: a randomized trial comparing the effects of atorvastatin and pravastatin on carotid intima medial thickness. Circulation 2002;106:2055–2060. 8. Smilde TJ, van Wissen S, Wollersheim H, Trip MD, Kastelein JJ, Stalenhoef AF. Effect of aggressive versus conventional lipid lowering on atherosclerosis progression in familial hypercholesterolaemia (ASAP): a prospective, randomised, double-blind trial. Lancet 2001;357:577–581. 9. Crouse 3rd JR, Grobbee DE, O’Leary DH, et al. Measuring effects on intima media thickness: an evaluation of rosuvastatin in subclinical atherosclerosis—the rationale and methodology of the METEOR study. Cardiovasc Drugs Ther 2004;18:231–238. 10. Terpstra WF, May JF, Smit AJ, Graeff PA, Meyboom-de Jong B, Crijns HJ. Effects of amlodipine and lisinopril on intima-media thickness in previously untreated, elderly hypertensive patients (the ELVERA trial). J Hypertens 2004;22:1309 –1316. 11. Dwyer JH, Sun P, Kwong-Fu H, Dwyer KM, Selzer RH. Automated intima-media thickness: the Los Angeles Atherosclerosis Study. Ultrasound Med Biol 1998;24:981–987. 12. Kanters SD, Algra A, van Leeuwen MS, Banga JD. Reproducibility of in vivo carotid intima-media thickness measurements: a review. Stroke 1997;28:665– 671. 13. Saam T, Ferguson MS, Yarnykh VL, et al. Quantitative evaluation of carotid plaque composition by in vivo MRI. Arterioscler Thromb Vasc Biol 2005;25:234 –239. 14. 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. 387 15. Chu B, Kampschulte A, Ferguson MS, et al. Hemorrhage in the atherosclerotic carotid plaque: a high-resolution MRI study. Stroke 2004;35:1079 –1084. 16. Hatsukami TS, Ross R, Polissar NL, Yuan C. Visualization of fibrous cap thickness and rupture in human atherosclerotic carotid plaque in vivo with high-resolution magnetic resonance imaging. Circulation 2000;102:959 –964. 17. 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. 18. Yuan C, Beach KW, Smith Jr LH, Hatsukami TS. Measurement of atherosclerotic carotid plaque size in vivo using high resolution magnetic resonance imaging. Circulation 1998;98:2666 –2671. 19. Yuan C, Miller ZE, Cai J, Hatsukami T. Carotid atherosclerotic wall imaging by MRI. Neuroimaging Clin N Am 2002;12:391– 401, vi. 20. Cootes TF, Taylor CJ, Cooper DH, Graham J. Active shape models—their training and application. Comput Vis Image Understand 1995;61:38 –59. 21. Bigger P. HJ, Unser M. B-spline snakes: a flexible tool for parametric contour detection. IEEE Trans Image Process 2000;9:1484 – 1496. 22. Crowe LA, Ariff B, Keegan J, et al. Comparison between threedimensional volume-selective turbo spin-echo imaging and twodimensional ultrasound for assessing carotid artery structure and function. J Magn Reson Imaging 2005;21:282–289. 23. Kisworo M. Venkatesh S, West GAW. Detection of curved edges at subpixel accuracy using deformable models. IEEE Proc Vis Image Signal Process 1995;142:304 –311. 24. Selzer RH, Hodis HN, Kwong-Fu H, et al. Evaluation of computerized edge tracking for quantifying intima-media thickness of the common carotid artery from B-mode ultrasound images. Atherosclerosis 1994;111:1–11. 25. Sramek A, Bosch JG, Reiber JH, Van Oostayen JA, Rosendaal FR. Ultrasound assessment of atherosclerotic vessel wall changes: reproducibility of intima-media thickness measurements in carotid and femoral arteries. Invest Radiol 2000;35:699 –706. 26. Zhang S, Cai J, Luo Y, et al. Measurement of carotid wall volume and maximum area with contrast-enhanced 3D MR imaging: initial observations. Radiology 2003;228:200 –205. 27. Zhang S, Hatsukami TS, Polissar NL, Han C, Yuan C. Comparison of carotid vessel wall area measurements using three different contrast-weighted black blood MR imaging techniques. Magn Reson Imaging 2001;19:795– 802. 28. Kang X, Polissar NL, Han C, Lin E, Yuan C. Analysis of the measurement precision of arterial lumen and wall areas using highresolution MRI. Magn Reson Med 2000;44:968 –972. 29. Silver MD. Cardiovascular pathology. New York: Churchill Livingstone; 1991. 30. Yarnykh VL, Yuan C. T1-insensitive flow suppression using quadruple inversion-recovery. Magn Reson Med 2002;48:899 –905.