Automatic Vessel Wall Contour Detection and Quantification of Wall Thickness in In-Vivo MR Images of the Human Aorta

Transcription

1 JOURNAL OF MAGNETIC RESONANCE IMAGING 24: (2006) Original Research Automatic Vessel Wall Contour Detection and Quantification of Wall Thickness in In-Vivo MR Images of the Human Aorta Isabel M. Adame, MSc, 1 * Rob J. van der Geest, MSc, 1 David A. Bluemke, MD, PhD, 2 João A.C. Lima, MD, 2 Johan H.C. Reiber, PhD, 1 and Boudewijn P.F. Lelieveldt, PhD 1 Purpose: To develop an automated technique to trace the contours of the lumen and outer boundary of the aortic wall, and measure aortic wall thickness in axial MR images. Materials and Methods: The algorithm uses prior knowledge of vessel wall morphology. A geometrical model (ellipse) is deformed, translated and rotated to obtain a rough approximation of the contours. Model-matching is based on image gradient measurements. To enhance edges, the images were preprocessed using gray-level stretching. Refinement is performed by means of dynamic programming. Wall thickness is computed by measuring the distance between inner and outer contour of the aortic wall. Results: The algorithm has been tested on high-resolution axial MR images from 28 human subjects of the descending thoracic aorta. The results demonstrate: High correspondence between automatic and manual area measurements: lumen (r 0.99), outer (r 0.96), and wall thickness (r 0.85). Conclusion: Though further optimization is required, our algorithm is a powerful tool to automatically draw the boundaries of the aortic wall and measure aortic wall thickness in aortic wall devoid of major lesions. Key Words: aortic wall thickness; atherosclerosis; magnetic resonance imaging; model-based segmentation; contour detection J. Magn. Reson. Imaging 2006;24: Wiley-Liss, Inc. ATHEROSCLEROSIS IS A COMMON DISORDER of the arteries, and is the primary cause of heart disease and 1 Department of Radiology, Division of Image Processing, Leiden University Medical Center, Leiden, the Netherlands. 2 Department of Radiology, Division of Cardiology, Johns Hopkins Hospital, Baltimore, Maryland, USA. Contract grant sponsor: National Heart, Lung, and Blood Institute; Grant numbers: N01-HC-95159, N01-HC-95162, N01-HC *Address reprint requests to: I.M.A., Division of Image Processing (LKEB), Department of Radiology, Leiden University Medical Center (LUMC), Albinusdreef 2, P.O.Box 9600, 2300 RC Leiden, The Netherlands. I.M.Adame@lumc.nl Received July 28, 2005; Accepted May 5, DOI /jmri Published online 28 July 2006 in Wiley InterScience ( wiley.com). stroke. It affects and handicaps many individuals, and is responsible for 40% of all deaths in Western societies (1). Therefore, extensive research is being conducted to better understand the natural history of the disease, and find the best treatment to reduce its effects. Direct visualization of the arterial wall in atherosclerotic vessels by imaging techniques has the potential to provide insights into the pathophysiology of plaque progression or regression (2 4). Currently, the degree of lumen stenosis is regarded as the gold standard in clinical practice, especially to determine when surgical treatment is needed. However, it has its limitations when it comes to patients in an early stage of the disease. In fact, all methods that rely on reduction of lumen diameter will fail to detect the early stages of atherosclerosis, characterized by an inflammation of the vessel wall (5,6) (wall thickening) and normal lumen dimensions. Most of the standard imaging techniques identify the luminal diameter or stenosis (angiography) or wall thickness, and plaque volume (intravascular ultrasonography), but are invasive (7). A noninvasive method would be desirable for repeated measurement in the same individual. Percutaneous ultrasound techniques have also been applied successfully to the quantitative evaluation of atherosclerosis by measuring the intima-media thickness (IMT) (8). However, due to interface artifacts between the near wall and close surroundings (abdominal fat, bowel gas), measurements of IMT in the aortic wall present a higher variability than those of superficial arteries. Therefore, it is important to find quantitative methods that are noninvasive, reproducible, and reliable for aortic wall imaging. High-resolution MR has emerged as the leading noninvasive imaging modality for imaging the aortic wall in vivo (9). Aortic wall measurements have proven useful as predictors for cardiovascular disease. Different studies, for example the recently initiated Multi-Ethnic Study of Atherosclerosis (MESA) (10), have been carried out to evaluate objective measurements of subclinical cardiovascular disease, detecting risk associations and aiming in the early pathogenesis of clinical disease. Most of these studies have been conducted to determine aortic compliance and flow rate quantification Wiley-Liss, Inc. 595

2 596 Adame et al. Figure 1. Enhancement of the edge between lumen and vessel wall. Each pixel in the original image domain (abscissa) is mapped to the modified image domain (lumen-vessel wall edge-enhanced) (ordinate) according to this graph. The intensity values are normalized against the average intensity value of fat. However, aortic wall thickness has not been as extensively investigated. MR can generate high-resolution cross-sectional images of an artery, and, as it does not involve ionizing radiation, can be repeated sequentially over time, which is particularly important to monitor progression and regression of atherosclerotic plaque. A recent study by Lima et al (11) has shown the relevance of aortic wall thickness measurements of MR images for monitoring the effects of drug treatments (statin therapy). Another study, by Jaffer et al (12) reported a cardiovascular MR population-based study of subclinical aortic atherosclerosis in a free-living population. However, all reported measurements were performed manually by expert radiologists using an analytical software system, which is subjected to inter- and intraobserver variability. There have been few reports that discuss automatic approaches to measure the wall thickness of the human aorta in in vivo MR images. The work of Hänni et al (13) proves the capability of MRI to measure wall thickness in the rabbit aorta using a semiautomatic method. However, extensive manual interaction is needed, making the process quite subjective. Therefore, the goal in this study was to develop a fully-automated algorithm to trace the contours of the vessel wall and measure wall thickness of the human descending aorta in in-vivo MR images. Image Data MR images of the human descending aorta were acquired with a 1.5 T whole-body MRI system Signa CV/I (General Electric Medical Systems, Waukesha, WI, USA). A four-element phased array torso coil was placed anteriorly and posteriorly. Images were obtained using a double inversion recovery black-blood fast spinecho sequence (9) with electrocardiogram (ECG) gating. Axial images of the descending thoracic aorta were obtained at the level of the right pulmonary artery. Imaging parameters were TR 2 R-R intervals (R-R: interval from the onset of one QRS complex to the onset of the next QRS complex, in an ECG tracing); TE 42 msec; field of view 40 cm; scan matrix size ; pixel size 0.78 mm; slice thickness 4 mm; echotrain length 32; and receiver bandwidth 62.5 khz. Description of the Algorithm We report an automated approach to measure wall thickness in in-vivo MR images of the human descending aorta. To achieve a fully automated process, the first step is to locate the descending aorta within the rest of vessel-like structures in the image. The aorta at the level of the right pulmonary artery is circular, so a method to find circular structures within the image, namely the Hough transform (14,15), was used. Next, prior knowledge on vessel wall morphology was incorporated into the algorithm to detect lumen and outer wall contours, by means of an ellipse-fitting procedure. Localization of the Descending Aorta: Hough Transform The first step is to locate the center point of the vessel of interest. We use the Hough transform (14,15) to find circular structures in the image, which yields several points, each one indicating the center of a different circular structure in the image. In our case, the radius was constrained to a range between 6 and 15 mm to MATERIALS AND METHODS Subjects For this study, 28 subjects, aged between 45 and 84 years old, for which manual contours drawn by experts were available, were selected from a population-based screening study, carried out throughout the United States. Informed consent to be scanned was obtained from all study subjects. These subjects had no prior clinical evidence of cardiovascular disease. Figure 2. Image enhancement to suppress periaortic fat. Each pixel in the original image domain (abscissa) is mapped to the modified image domain (vessel wall-periaortic fat edgeenhanced) (ordinate) according to this graph. The intensity values are normalized against the average intensity value of fat.

3 Automated Detection of Aortic Wall Contours 597 Table 1 Descriptive Statistics of Aortic Wall Thickness and Area as Measured by Two Observers and the Automatic Measurements Resulting from Our Algorithm Wall thickness (mm) Lumen area (cm2) Outer wall area (cm2) Vessel wall area (cm2) Observer Observer Automatic limit the candidate points. Besides, the only candidate points considered are the ones corresponding to the three maximum values (superior vena cava, ascending aorta, and descending aorta; as can be seen in Fig. 6) in the Hough-transformed domain. To obtain the point that corresponds to the center of the descending aorta, we selected the most posterior one, the closest to the spine of the patient. Lumen Contour Detection Once the center point of the descending aorta is found, the lumen and outer wall contours have to be traced. Prior to lumen detection, the original image is preprocessed, performing image enhancement in the spatial domain. Image enhancement techniques (16) are used to emphasize and sharpen image features. In our case, the edge between lumen and vessel wall is not very well depicted. To facilitate the ellipse-fitting procedure, that edge needs to be enhanced. To accomplish this, a graylevel stretching is performed according to Fig. 1. Afterward, taking into account that the shape of the vessel is approximately elliptic, a geometrical model (ellipse) is created around the center point obtained in the previous step. This ellipse is resized, translated, deformed, and rotated, following an iterative procedure, to match the lumen in the enhanced image, according to gradient measurements. The greater the average gradient along the ellipse, the better it approximates the lumen contour. The direction of the gradient is also taken into account because, at this point in the process, we are only interested in the internal contour of the aortic wall. The preprocessing of the images is the first step for the ellipse-fitting. Without the preprocessing, the ellipse would be driven toward the boundary between the fat layer and the artery, as that edge is stronger than the edge between lumen and vessel wall. This first step provides a rough estimation of the contours, which are further refined (deformed and translated) based on the original image. The ellipse-fitting algorithm is fully described in the work of Adame et al (17). The ellipse with the largest average gradient is taken as a rough approximation of the lumen contour. To get a more accurate contour, a minimum cost approach (based on dynamic programming (18)) is performed on the original image using the elliptical contour as initial model: at each point on the model a scan line perpendicular to the contour is constructed, with the image intensity values on this line derived from the original image. The length of the scan lines is chosen such that only a small neighborhood around the model contour is included. Those values are annotated in a row of a scan matrix ; each row corresponds to a point in the contour. In addition, a cost matrix is also computed, each element of which represents the cost that involves selecting the corresponding element in the scan matrix as the new point Figure 3. Regression plots showing the comparison between automated measurements and observer 1 (a), automated and observer 2 (b), automated and the average contours obtained from both observers (c), and measurements according to observer 1 and observer 2 (d).

4 598 Adame et al. Table 2 Interobserver Reproducibility SD COV (%) Luminal area 0.17 cm Outer wall area 0.77 cm Vessel wall area 0.15 cm for the contour at that position. The values of the cost matrix are derived from the intensity changes in the scan matrix and are influenced by a few parameters In our case, those parameters are: 1) side step size : maximum allowed displacement (perpendicular to the contour) for a point in the contour, with respect to its original position; 2) side step cost : cost associated to the displacement of a point in the contour; and 3) model position factor : cost associated with the model. The higher this value, the less variation permitted from the original rough contour. These parameters describe geometric properties of the aorta shape, which makes the parameter selection largely independent of the scanning protocol, therefore generalizing well toward other MR acquisition protocols not tested here (8,11,13). Outer Wall Contour Detection The technique to detect the outer wall of the aorta is similar to that followed to detect the lumen, but using a different enhancement to suppress periaortic fat. In this step a gray-level stretching is performed on the original image: all pixels within the intensity range corresponding to the vessel wall (averaged for all images in the study) are enhanced, while the pixels corresponding to lumen and periaortic fat (low- and high-intensity values, respectively) are not modified (see Fig. 2). A new ellipse is created to match the outer wall in the new enhanced image. As in lumen detection, an iterative procedure is carried out to find the ellipse that represents the best match to the outer contour. However, there are some restrictions, the outer wall surrounds the lumen contour, which means that the short-axis of the outer wall ellipse must be larger than the long-axis of the lumen ellipse, provided that the center of gravity of both contours coincides. In vessels devoid of major lesions, lumen and outer wall are concentric, which makes this statement valid. Again, the sign of the gradient plays an important role to detect the external contour of the aortic wall. Finally, refinement of the contour is performed on the original image by means of dynamic programming (18) in the same manner as for the lumen contour. Figure 4. The Bland-Altman plot of the differences between automatic (auto wall thickness) and manual measurements from observer 1 (ob1) (a); between automatic and measurements from observer 2 (ob2) (b); between automatic and average manual measurements (c); and between observers 1 and 2 (d).

5 Automated Detection of Aortic Wall Contours 599 COV is the SD of the data divided by the mean of the data. Agreement with Radiologists To assess the accuracy of the algorithm, all contours, wall thickness, and area measurements were compared to manual contours drawn by two expert radiologists blinded to the results of the algorithm. To get a reference standard to assess the accuracy of the algorithm, average contours of lumen and outer wall are obtained from those drawn by observer 1 and 2. We will refer to these contours as average manual contours. The statistical methods described by Bland and Altman (19,20) for comparing paired data were used to study the differences between manual and automatic contours, to determine if there were significant differences, differences as a function of wall thickness, or any bias of the algorithm. RESULTS Figure 5. Scatterplot shows outer wall area (a) and luminal area (b) according to the automatic measurements (aut contour) against those according to the average manual contours (ref contours). Wall Thickness Measurement The contours obtained from the previous steps are used to measure the aortic wall thickness. To accomplish that, for each point in the lumen contour, the closest point in the outer wall contour is found, and the distance between them is computed. Afterward, the maximum and mean value from all those measurements are computed and taken as the maximum and average vessel wall thickness, respectively. In addition, the wall area is also computed as the difference between the area within the outer wall contour and the area within the lumen contour. The contour detection and vessel wall thickness measurement is fully automated and the computation time is around eight seconds on a standard PC workstation. Interobserver Reproducibility Two observers, blinded from the result of the study, traced the lumen and outer wall boundaries for all subjects. Reproducibility was expressed as standard deviations (SDs) and coefficients of variation (COV). The Table 1 provides descriptive statistics for the measurements by observer and automatic results. Average wall thickness measurements ( mm) are comparable to those previously reported by Li et al (21) ( mm) and Weiss et al (22) ( mm). From Table 1 and the regression plots in Fig. 3 we can observe that the agreement with radiologists of the automated method is on the same order of interobserver reproducibility (Table 2). Moreover, in comparison with the results from our automatic method, observer 1 had a tendency to overestimate the wall thickness, while observer 2 tended to underestimate it. This tendency can also be observed in the Bland and Altman plots (Fig. 4), in which the automated measurements are compared to those derived from observer 1 (Fig. 4a) and observer 2 (Fig. 4b). In Fig. 4c, observer 1 is compared to observer 2. We can see that the average difference between observer 1 and 2 ( mm) is larger than the differences between the automatic measurements and each observer ( mm for observer 1 and mm for observer 2, respectively). The Bland-Altman plots also show that there was a small random error as evidenced by the SDs. Figures 3 and 5 show area measurements according to the average contours plotted against those derived from the automated process (Fig. 3: vessel wall area; Fig. 5a: aortic outer wall area; Fig. 5b: aortic lumen area). Good agreement between automatic and manual area measurements is apparent for lumen (r 0.99), outer wall (r 0.96), and vessel wall (r 0.85). The average paired difference between the automatic/manual measurement pairs was mm 2 ( %; P NS) for the lumen area, and mm 2 ( %; P NS) for the outer wall area. Figures 6 and 7 present qualitative results. Figure 6 presents an example of the different steps of the algorithm. In Fig. 7a d, we can see the small but detectable differences between the automatic contours and those manually drawn (see arrows). This is mainly caused by the fact that the border between the aortic vessel wall and the surrounding tissue is very blurred. Note the

6 600 Adame et al. Figure 6. Example showing the different steps of the algorithm for two different subjects. a d: Subject #1. e h: Subject #2. a,e: Original image showing the centers of the circular structures found by the Hough transform algorithm (the one depicted as a square corresponds to the descending aorta). b,f: Gray-level stretched for enhancement of the edge lumen-vessel wall. c,g: Gray-level stretched for enhancement of the outer boundary of the vessel wall. d,h: Final segmentation: vessel wall inner and outer boundaries. disagreement between observers (see arrows in Fig. 7b and c), and the similarity between the automatic and the average manual contours (compare Fig. 7a and d). In Fig. 7e h, this border appears better depicted, so a high degree of agreement is apparent in all four cases. DISCUSSION In this work, an automated method was developed to identify the boundaries of the aortic wall and quantify the wall thickness. The reference standard is based on measurements obtained from manual contours drawn by two expert observers. Li et al (21) and Weiss et al (22) have shown the adequacy of using manual measurements of aortic diameter for determining truth. However, there are considerable differences between the contours drawn by each of these observers, as can be inferred from Figs. 3, 4, and 7 (see arrows in Fig. 7). In fact, it can be observed that one of the observers has a tendency to overestimate the contours with respect to the other observer (see Table 1). The main reason for this disagreement is that the boundary between vessel wall and surrounding tissue is quite blurred, which makes the manual delineation of the contours a difficult task. Nevertheless, those contours were the only available gold standard to assess the accuracy of the algorithm. In order to get a more objective reference standard, an average of the manual contours was also computed. When compared to these average contours, our automated approach showed a high degree of agreement. However, when compared to the results from each observer separately, our algorithm yielded results that lay in between those obtained from observer 1 and those from observer 2, as can be seen in Table 1 and Fig. 3. Although all methods have variability, it can be inferred that our approach has roughly the same variability as the human readers, gaining in reproducibility. This will play a crucial role in

7 Automated Detection of Aortic Wall Contours 601 Figure 7. Contours of the vessel wall of the human descending aorta. a,e: Contours automatically-detected. b,f: Manuallytraced contours by observer 1. c,g: Manually-traced contours by observer 2. d,h: Average from contours in b and c. Arrows point to the part of the contour where the differences between contours drawn by observer 1 and 2 are more significant. follow up studies. Besides this, it is much faster than the manual tracing of the contours. Assessment of wall thickness parameters is derived using an automated algorithm without any manual interaction (for all images analyzed in this work there was no error in the detection of the aorta), but, if necessary, the contours may be adjusted manually after the automatic detection. The algorithm was integrated within an in-house developed software package (vessel wall MASS) for quantitative analysis of vessel wall MR studies. This software is equipped with drawing tools that allow setting of vertices, continuous drawing, or vertex shifting. Most previous works in aortic wall studies have focused on flow, strain, compliance, and stenosis measurements, which rely only on the lumen contour. Nevertheless, not much has been done to automatically draw both boundaries of the aortic wall and measure wall thickness. However, this is particularly important as early stages of atherosclerosis are characterized by inflammation of the vessel wall, without reduction of lumen dimensions. In addition, the area lumen measurements provided by our automated approach may also be used to study compliance. A recently published study by Li et al (21) classified the differences in wall thickness of the human aorta according to age, race, or sex. However, all the measurements are derived from manual delineation of contours, which makes it subjective, even when the reported agreement between observers was high. The study from which our images were obtained was designed as a pilot study to determine optimal methods, reproducibility, and potential areas of future investigation. Hence, the number of study subjects was small and only one slice of the midthoracic aorta was acquired, rather than a more representative sampling of both the thoracic and abdominal aorta. However, as the main core of the detection algorithm is based on a geometrical model (ellipse-fitting), and the shape of a cross-section of the vessel is similar in the abdominal and thoracic aorta, we expect the algorithm to perform adequately with minimum tuning (range of radii for Hough-transform, range of intensities for gray-levelstretching given that the images are non fat-suppressed). Another limitation is that the images were acquired without fat saturation which makes the preprocessing necessary. However, for images acquired with fat suppression, the preprocessing step would not be necessary. Fat saturation could be especially useful when characterizing components of plaque in the arterial wall. Nevertheless, our approach has proven to perform adequately in a cohort study of 28 human subjects, but some fine-tuning is still required and high-quality MR images will certainly improve the postprocessing. In conclusion, our method is a powerful tool to assist the radiologists in the detection of the contours of the aortic wall and measure the average and maximum wall thickness in aortic walls devoid of major lesions. These measurements were found to be accurate when compared to those derived from manually traced contours. The comparison to results from two different observers has also proved the good performance of our method in terms of reproducibility and objectivity. However, further developments in MR images will improve image quality and, subsequently, image postprocessing. In addition, combination with longitudinal studies may help draw a conclusion about whether aortic wall thickness can be used as a predictor of cardiovascular disease. ACKNOWLEDGMENTS We thank the other investigators, the staff, and the participants of the MESA study for their valuable contributions. A full list of participating MESA investiga-

8 602 Adame et al. tors and institutions may be found at mesa-nhlbi.org. REFERENCES 1. American Heart Association. Heart and stroke statistical update. Dallas, TX: American Heart Association; p Fayad ZA, Fuster V. Atherothrombotic plaques and the need for imaging. Neuroimaging Clin N Am 2002:12: Salonen JT, Salonen R. Ultrasound B-mode imaging in observational studies of atherosclerotic progression. Circulation 1993:87: Fayad ZA, Fuster V. Characterization of atherosclerotic plaques by magnetic resonance imaging. Ann NY Acad Sci 2000:902: Ross R. Atherosclerosis an inflammatory disease. N Engl J Med 1999:340: Glagov S, Weisenberg E, Zarins CK, Stankunavicius R, Kolettis GJ. Compensatory enlargement of human atherosclerotic coronary arteries. N Engl J Med 1987: Nair A, Kuban BD, Tuzcu EM, Schoenhagen P, Nissen SE, Vince DG. Coronary plaque classification with intravascular ultrasound radiofrequency data analysis. Circulation 2002:106: Astrand H, Sandgreen T, Ahlgren AR, Lanne T. Noninvasive ultrasound measurements of aortic intima-media thickness: implications for in-vivo study of aortic wall stress. J Vasc Surg 2003;37: Fayad ZA, Nahar T, Fallon JT, et al. In vivo magnetic resonance evaluation of atherosclerotic plaques in the human thoracic aorta. Circulation 2000;101: Bild DE, Bluemke DA, Burke GL, et al. Multi-Ethnic Study of Atherosclerosis: objectives and design. Am J Epidemiol 2002:156: Lima JAC, Desai MY, Steen H, Warren WP, Gautam S, Lai S. Statin-Induced cholesterol lowering and plaque regression after 6 months of magnetic resonance imaging-monitored therapy. Circulation 2004:110: Jaffer FA, O Donnell CJ, Larson MG, et al. Age and sex distribution of subclinical aortic atherosclerosis: a magnetic resonance imaging examination of the Framingham Heart Study. Arterioscler Thromb Vasc Biol 2002;22: Hänni M, Lekka-Banos I, Nilsson S, Haggroth L, Smedby Ö. Quantitation of atherosclerosis by magnetic resonance imaging and 3D morphology operators. Magn Reson Imaging 1999;17: Kimme C, Ballard DH, Sklansky J. Finding circles by an array of accumulators. Commun ACM 1975:18: Hough PVC. Method and means for recognizing complex patterns. US Patent 3,069,654, Umbaugh SE. Computer vision and image processing. Indianapolis: Prentice Hall PTR; p Adame IM, van der Geest RJ, Wasserman BA, Mohamed M, Reiber JHC, Lelieveldt BPF. Automatic segmentation and plaque characterization in atherosclerotic carotid artery MR images. MAGMA 2004:16: Sonka M, Hlavac V, Boyle R. Object recognition: fuzzy systems: In: Jeans S, McGee K, eds. Image processing, analysis, and machine vision, 2nd edition. Pacific Grove, CA: Brooks/Cole Publishing Company; p Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1: Bland JM, Altman DG. A note on the use of the intraclass correlation coefficient in the evaluation of agreement between two methods of measurement. Comput Biol Med 1990:20: Li AE, Kamel IR, Lima J, Rando F, Anderson M, Bluemke D. Using MRI to assess aortic wall thickness in the multiethnic study of atherosclerosis: distribution by race, sex and age. AJR Am J Roentgenol 2004:182: Weiss CR, Arai AE, Bui MN, et al. Arterial wall MRI characteristics are associated with elevated serum markers of inflammation in humans. J Magn Reson Imaging 2001;14: