Validation of semiautomated and locally resolved aortic wall thickness measurements from computed tomography

Abstract

Objective

Aortic wall thickness (AWT) is important for anatomic description and biomechanical modeling of aneurysmal disease. However, no validated, noninvasive method for measuring AWT exists. We hypothesized that semiautomated image segmentation algorithms applied to computed tomography angiography (CTA) can accurately measure AWT.

Methods

Aortic samples from 10 patients undergoing open thoracoabdominal aneurysm repair were taken from sites of the proximal or distal anastomosis, or both, yielding 13 samples. Aortic specimens were fixed in formalin, embedded in paraffin, and sectioned. After staining with hematoxylin and eosin and Masson’s trichrome, sections were digitally scanned and measured. Patients’ preoperative CTA Digital Imaging and Communications in Medicine (DICOM; National Electrical Manufacturers Association, Rosslyn, Va) images were segmented into luminal, inner arterial, and outer arterial surfaces with custom algorithms using active contours, isoline contour detection, and texture analysis. AWT values derived from image data were compared with measurements of corresponding pathologic specimens.

Results

AWT determined by CTA averaged 2.33 ± 0.66 mm (range, 1.52–3.55 mm), and the AWT of pathologic specimens averaged 2.36 ± 0.75 mm (range, 1.51–4.16 mm). The percentage difference between pathologic specimens and CTA-determined AWT was 9.5% ± 4.1% (range, 1.8%–16.7%). The correlation between image-based measurements and pathologic measurements was high (R = 0.935). The 95% limits of agreement computed by Bland-Altman analysis fell within the range of −0.42 and 0.42 mm.

Conclusions

Semiautomated analysis of CTA images can be used to accurately measure regional and patient-specific AWT, as validated using pathologic ex vivo human aortic specimens. Descriptions and reconstructions of aortic aneurysms that incorporate locally resolved wall thickness are feasible and may improve future attempts at biomechanical analyses.

Aortic aneurysms, including abdominal aortic aneurysms (AAAs) and thoracic aortic aneurysms, are responsible for ~10,000 to 15,000 deaths yearly in the United States.^1,2 In large screening studies, AAAs were found in ~5% of all men and in 1% of all women aged >70 years,³ and the incidence of thoracic aortic aneurysms has been estimated at 10.4/100,000 patient-years.⁴ Aneurysm rupture is a major source of morbidity and mortality, with mortality rates >50% even in modern series.^5,6

Surgery—either the traditional open aneurysm repair or newer endovascular stent graft repair—is the treatment of choice to prevent or treat rupture. Preoperative planning, including the selection of the open or endovascular approach, is critically dependent on aneurysm geometry and, therefore, on radiologic imaging. The gold standard for preoperative imaging is computed tomography angiography (CTA), which has the capability of detecting the size and extent of the aneurysm, locations of critical branch vessels, and the presence and distribution of thrombus.

Patient-specific aneurysm geometries have been incorporated in improved models of aneurysm rupture risk stratification based on structural analysis. Biomechanical modeling of vascular structures is highly dependent on accurate geometric reconstructions of the vessel in question. Aortic wall stress analyses using finite element methods have shown to be significantly correlated to rupture status⁷ and expansion rate.^8,9 Peak wall stress in AAAs has been found to be a better predictor of rupture than diameter alone.¹⁰

Surprisingly, even studies demonstrating the relevance of biomechanical computational modeling to clinical scenarios do not incorporate patient-specific and regionally variable aortic wall thickness (AWT) measurements, because reliable AWT measurements have previously been difficult to obtain. Studies using excised specimens of ruptured and unruptured AAAs have shown significant variations in AWT, which are correlated with areas of degeneration and weakness.^11,12 The lack of locally resolved wall thickness in computational models has been emphasized as a significant barrier in the accurate computation of wall stress using current finite element methods.^13,14

Previous studies have reported attempts to quantify AWT. Arko et al¹⁵ used intravascular ultrasound to study the dynamic geometry and wall thickness in the neck region of infrarenal AAAs; however, that technique is invasive. Current surface ultrasound techniques are unable to provide the necessary resolution to discern AWT. Adame et al¹⁶ developed an algorithm for the determination of AWT in high-resolution magnetic resonance imaging (MRI) of the healthy descending aorta. Compared with MRI, CTA is more readily available, requires shorter scanning times, is better tolerated by patients, and is less subject to respiratory artifacts yielding images of poor image quality.

Unlike other vessels, such as the carotid artery, where duplex measurements of arterial wall thickness have been rigorously validated, no universally accepted noninvasive method of measuring AWT exists. Shum et al¹⁷ presented a segmentation package capable of semiautomatic vessel wall detection for AAAs based on CTA data. Although their AWT measurements show good correlation of 10% to 15% with manual segmentation, neither the automatic nor the manual segmentations were validated against measurements from pathologic specimens.

This study presents a series of algorithms to segment axial slices of the thoracic and abdominal aortas. These algorithms use multiple image segmentation techniques, including level set methods, intensity-based and gradient-based segmentation, and texture analysis. AWT measurements derived from these image analysis techniques are validated against pathologic specimens from the abdominal and thoracic aortas.

METHODS

This study was approved by the Institutional Review Board, and patients provided informed consent.

Subjects and image data

Ten patients undergoing open thoracoabdominal aneurysm repair for aneurysms of atherosclerotic etiology were considered. Each patient’s most recent preoperative CTA examination was used for image segmentation. CTA examinations were performed using multidetector CT scanners (≥64 slice) with arterial-phase intravenous injection of 80 to 120 mL of nonionic iodinated contrast. Digital Imaging and Communications in Medicine (DICOM; National Electrical Manufacturers Association, Rosslyn, Va) images had a scan matrix size of 512 × 512, pixel size of 0.65 to 0.82 mm, and slice thickness of 1.2 to 1.5 mm.

Pathologic specimens

Full-thickness aortic wall specimens of ~15 mm (axial) × 15 mm (longitudinal) were taken from grossly normal-appearing aorta at the proximal or distal, or both, anastomotic sites. The excised aortic tissue was immediately placed in Krebs-Ringer bicarbonate buffer (Sigma-Aldrich, St. Louis, Mo). Orientation and specific location, including the level of the aortic specimen and its clock position, was noted. Aortic specimens were fixed overnight in a solution of 10% formalin and dehydrated in 70% ethanol solution. Subsequently, specimens were embedded in paraffin, sectioned, and stained with hematoxylin and eosin and Masson’s trichrome. The sections were digitally scanned and measured using ImageScope (Aperio Inc, Vista, Calif). Wall thickness measurements were taken in a 10-mm × 10-mm central area of the specimen in portions that contained intact adventitial, medial, and intimal layers (Fig 1, A). A mean wall thickness was then calculated for this area of the aortic wall and compared with wall thickness derived from aortic segmentation. Mean wall thicknesses were measured by two observers to assess interobserver variability.

Fig 1.

A, An ex vivo aortic specimen, stained with hematoxylin and eosin, shows the wall thickness measurement (arrow). B, Result of axial image segmentation shows outer adventitial (red), inner arterial (blue), and luminal (green) boundaries.

Segmentation overview

The image segmentation required to determine wall thickness involves the delineation of three separate surfaces: the lumen, the adventitial (outer) wall, and the inner arterial wall. By generating these surfaces, the arterial wall and any intraluminal thrombus (ILT) can be identified. The segmentation process involves a series of routines written in MATLAB software (MathWorks, Natick, Mass) and is capable of semiautomated as well as manual segmentation of axial CTA slices.

Lumen segmentation

Accurate luminal segmentation requires the administration of intravenous contrast to maximize the image intensity gradient between the lumen and arterial or ILT structures. Automatic windowing is initially applied to the stack of DICOM images to enhance the contrast between the aorta and other surrounding structures. The windowing procedure can be done automatically, or user-specified thresholds can be selected. Subsequently, an anisotropic diffusion filter is used to further sharpen edges.¹⁸

The user is asked to identify a single point inside the lumen for the first slice, and a small circle is generated around this point. A two-dimensional level-set method described by Chan and Vese¹⁹ uses this curve to automatically detect the luminal border in each slice. The series of luminal curves is finally presented to the user. Infrequently, the segmentation will encroach on surrounding high-intensity structures, such as the spine, and will require manual adjustment.

Outer wall segmentation

The adventitial wall is segmented by analyzing isointensity contours. The stack of windowed and filtered images is cropped to an area immediately surrounding the lumen. A contour function is then applied that generates a series of isolines of image intensity. Isolines that do not fully enclose the luminal boundary are eliminated.

For the initial slice, the appropriate isoline representing the adventitial border can be manually selected by the user or an automatic algorithm can be used. In the manual option, the remaining isolines are presented to the user as a topographic map, and the user selects an isoline corresponding to his or her interpretation of the outer adventitial border. The automatic method initially selects the isoline with minimal area that encloses the luminal boundary. Subsequent isolines of decreasing intensity are iteratively compared by area and curvature. Large changes in area and curvature were found to likely reflect the transition from aortic structures to surrounding tissues.

For subsequent slices, manual segmentation proceeds in a manner similar to the initial slice. The automatic method uses the same technique described above, but the candidate contour is compared with the previous slice by curvature and area as well. In the event of large differences in either of these features, a new contour is selected. Through iterative comparison, the contour that minimizes these differences while still fulfilling the criteria of enclosing the luminal boundary is selected. In the event that such a contour cannot be generated or the contour generated was deemed unacceptable by visual inspection, manual segmentation is used. Manual segmentation was most often necessary in areas where soft tissues of very similar pixel intensity to the arterial wall (eg, abdominal viscera) were immediately adjacent to the aorta.

Inner arterial wall segmentation

Our methods for the detection of the inner arterial surface are based on texture feature differences between the ILT and the arterial wall. Areas that were outside the luminal surface and inside the adventitial surface were considered as candidates for ILT. The user then selects regions of both thrombus and background at varying locations within the image set to serve as a reference for feature based discrimination.

Numerous image texture features were subsequently calculated for the candidate ILT area. A candidate ILT border was generated using common histogram features such as mean and standard deviations of image intensity, image entropy, and image intensity ranges. Preliminary evaluation showed that histogram-based texture analysis alone was insufficient for generating reasonable ILT boundary curves because they tended to conservatively estimate the ILT area. Therefore, a second approach was selected based on the calculation of the gray level co-occurrence matrix (GLCM). Before the application of our GLCM method, the windowed and cropped CTA images were coarsened to 256 gray levels. The GLCM is an estimation of a second-order joint conditional probability density function that seeks to characterize the spatial inter-relationships of the gray values in an image.²⁰ Textural features of energy, variance, and entropy were used by our algorithm.

Using this series of texture-based features, we formulated a nearest neighbor decision rule using the user-selected regions of thrombus as the training set to determine if a region was classified as thrombus or background. An averaged curve combining the outer border of this region and the border generated using this histogram-based method was generated as the inner arterial border. A sample segmentation is shown in Fig 1, B.

Postprocessing and AWT calculation

The luminal, inner arterial, and outer arterial surfaces underwent Gaussian curvature based-smoothing to produce smooth boundaries. AWT was calculated at 72 points along the circumference of the aorta in each axial slice by determining the minimal distance from the inner arterial surface to the outer adventitial surface. Because AWT was calculated (and not measured) from the difference between two mathematical constructs, it was reported as two significant figures, matching that of the pathologic specimens. A sample wall thickness map is shown in Fig 2. An average wall thickness was calculated for the area of the aorta corresponding to the one from which the pathologic specimen was taken.

Fig 2.

Wall thickness map generated from the segmentation of a proximal descending thoracic aneurysm.

RESULTS

Sixteen aortic samples were taken from 10 patients (six men and 4 women) undergoing open thoracoabdominal aneurysm repair. The patients were an average age of 71.2 ± 5.4 years. After tissue fixation, staining, and sectioning, 13 specimens were used for comparison with results from image-based AWT calculations. Three specimens were excluded for excessive architectural distortion arising as an artifact from the sectioning of highly calcified aortic tissue that rendered wall thickness measurements infeasible. Six (46.1%) of the remaining 13 samples were from locations in the thoracic aorta, and seven (53.9%) originated in the abdominal aorta. Further details regarding the location of specimens are given in the Table. Small amounts of thrombus were found in four of the 13 specimens; the rest were free of thrombus.

Table.

Specific locations of pathologic specimens with corresponding wall thickness measurements

Patient	Specimen location	Image segmentation wall thickness, mm	Pathologic specimen wall thickness, mm	Difference, %a
1	2 cm distal to LSCA	2.16	2.12	+1.8
2	3 cm distal to LSCA	1.64	1.51	+9.3
3	1 cm proximal to celiac	3.07	2.78	+10.4
4	2 cm proximal to celiac	2.15	2.30	−6.5
5	5 cm distal to LSCA	1.52	1.74	−12.6
5	1 cm proximal to celiac	2.60	2.48	+4.8
6	4 cm proximal to left renal	2.10	1.80	+16.7
6	Aortic bifurcation	1.54	1.75	−12.0
7	2 cm distal to LSCA	2.39	2.25	+6.2
7	2 cm above celiac	1.77	1.98	−10.6
8	4 cm distal to LSCA	3.55	3.76	−5.6
9	1 cm above to celiac	3.24	3.51	−7.7
10	Level of pulmonary trunk	2.57	2.34	9.8

LSCA, Left subclavian artery.

^aCalculated as the percentage difference from the pathologically determined wall thickness.

AWT determined by CTA was an average of 2.33 ± 0.66 mm (range, 1.52–3.55 mm), and the average wall thickness of pathologic specimens was 2.36 ± 0.75 mm (range, 1.51–4.16 mm). Because the pathologic specimens were taken from proximal and distal anastomotic sites, there was relatively little ILT. In addition, the aorta in these locations was relatively free from adjacent organs. In these areas, the image segmentation routines were run free of user intervention so the calculation of interobserver reproducibility was unnecessary. The average difference between observers for pathologic measurements was small (mean, 2.5% ± 1.33%; range, 0.3%–5.2%), and the correlation coefficient was excellent (R = 0.998; Fig 3, A).

Fig 3.

A, Plot shows interobserver reliability of wall thickness measurements derived from excised aortic specimens. B, Plot shows correlation between pathologic specimen wall thickness measurements and image segmentation measurements.

The percentage difference between pathologic specimens was 9.5% ± 4.1% (range, 1.8%–16.7%), and the correlation between image-based measurements and pathologic measurements was favorable (R = 0.935; Fig 3, B). The highest percentage differences between the two wall thickness measurement methods (patients 3 and 6) were found in scans of the coarsest pixel density (0.80 and 0.82 mm, respectively).

Bland-Altman analysis revealed some variability between the two methods, but no systematic bias was detected (Fig 4), although most of the wall thickness measurements were between 1.5 and 2.5 mm. Although the sample size was small, no significant difference was noted when comparing absolute errors of samples with thrombus and those without (7.3% ± 5.4% vs 10.4% ± 3.3%; P = .22). The 95% limits of agreement, which indicate how far apart measurements by the two methods are likely to fall for each comparison, fell in the range of −0.42 to 0.42 mm.

Fig 4.

Bland-Altman plot shows the difference between wall thickness measurements derived from image analysis and wall thickness measured in pathologic specimens.

DISCUSSION

In this study, we describe a method by which AWT can be measured from CTA DICOM data in the abdominal and descending thoracic aorta. This was accomplished through a series of semiautomated image-processing routines that are capable of detecting luminal, inner arterial, and outer arterial boundaries. Wall thickness measurements derived from image segmentation were then compared with thickness measurements derived from pathologic specimens excised during open aortic repair, and good correlation (R = 0.935) was obtained. Although previous studies have reported techniques to measure AWT in CT, MRI, and ultrasound data,^13–15 none have validated these measurements against ex vivo pathologic specimens.

Arterial wall thickness plays a critical role in computational vascular biomechanics, because structural analysis depends on high-fidelity geometric representations. Variable wall thicknesses have not routinely been used in the calculation of biomechanical stresses in arterial structures. In fact, prior studies of the influence of assumptions and modeling sophistication on the accuracy and results of AAA computational stress prediction have sometimes ignored AWT altogether.²¹ Noninvasive techniques for the determination of AWT may improve the prognostic capabilities of finite element models of aortic aneurysms.

Although our motivation for this research concerns efforts to model the biomechanics of aortic aneurysms with optimum fidelity, other clinical and research uses for the techniques we have described can be imagined. AWT may itself be correlated with aneurysm rupture, because the yield or failure strength of the aortic wall is likely intimately related to the thickness of the aorta. However, prior efforts at quantifying the wall strength of the aorta have ignored AWT.^22,23 Di Martino et al,¹² in fact, showed greater AWT in ex vivo pathologic specimens from ruptured AAAs than from electively repaired AAAs. Finally, AWT may be indicative of certain disease processes or reflective of their severity, in analogy to carotid intimamedial thickness being a marker for peripheral and coronary arterial diseases.²⁴

This study used formalin-fixed aortic specimens as the reference standard to which our image segmentation results were compared. Other studies examining the local properties of aortic tissue have used laser micrometry to measure fresh samples from the operating room.¹² Although formalin fixation significantly alters the mechanical properties of aortic tissue, the process does not significantly alter tissue thickness.²⁵ We therefore felt that the potential error introduced by the tissue fixation process was minimal.

Nevertheless, the current work is subject to some important limitations. The algorithm presented is sensitive to the quality of CTA image data. In particular, preoperative scans used for image segmentation were of relatively high in-plane resolution (range, 0.65–0.82 mm). As pixel size increases, wall thickness measurements became more uniform in distribution and less accurate.

Although there were small amounts of thrombus in four of 13 specimens, most of the sections used for wall thickness validation were thrombus-free. This was because specimens were harvested from anastomotic margins to facilitate spatial correlation between CTA images and the anatomic location of the specimen and to avoid fracture and disintegration of the specimen in areas of severely degenerated aneurysmal and atherosclerotic aortic wall. The described segmentation procedures are capable of detecting the inner arterial wall in areas with large amounts of ILT, but the accuracy of the wall thickness measurements in these regions is still largely unproven.

From a biomechanics standpoint, this is unlikely to be of significance because previous aneurysm stress analyses have shown that there is little wall stress on thrombus-laden sections of the aorta.^14,26 Regional maxima in wall stress have been localized to abrupt changes in aneurysm morphology, such as in aneurysm necks, which often have minimal thrombus burden.^7–9 However, the evolution of wall thickness in regions of thrombus may play an important role in the understanding of the pathophysiology and natural history of aneurysmal disease. Further work in validating wall thickness maps over entire aortic aneurysms, including areas with significant thrombus, will be necessary to address this issue.

Although we found a favorable correlation between image-based and pathologic measurements of wall thickness, with small absolute percentage errors, there is clear room for improvement. This is reinforced by the relatively coarse 95% limits of agreement determined by Bland-Altman analysis of ± 0.4 mm. A possible source of error arises in the use of CTAs that were not electrocardiogram-gated, because the aortic wall in these studies is an average of the systolic and diastolic configurations. Furthermore, the pathologic specimens represent a “zero-stress” configuration, whereas the CTA images are under physiologic blood pressure, possibly injecting some bias into our measurements. Unfortunately, there are scarce data in the human aorta regarding the effects of removing residual stress on wall thickness. Improvements in scanning resolution combined with electrocardiogram-gated studies may provide higher-quality image data and improve accuracy.

CONCLUSIONS

CTA can be used to accurately measure regional and patient-specific AWT. Descriptions and reconstructions of aortic aneurysms that incorporate locally resolved wall thickness are feasible and may improve future attempts at biomechanical analyses.