The accuracy of the stenosis measurement is important, seen its role in clinical decision making about carotid endarterectomy. Traditionally, the stenosis in the carotid artery was assessed using intra-arterial digital subtraction angiography (DSA), which is still considered the gold standard. The degree of stenosis was defined as the residual lumen at the stenosis as a percentage of the normal lumen in the distal internal carotid artery (according to the NASCET criteria) or as the residual lumen as a percentage of the estimated original diameter of the artery at the level of the stenosis (according to the ECST criteria). In the large symptomatic carotid surgery trials, conventional DSA was performed in two or three projections (lateral, posterio-anterior, and/or oblique) which were investigated for the most severe stenosis. Whereas rotational DSA, using multiple planes, showed to provide a benefit in detecting the smallest diameter in a stenosed artery compared to conventional DSA [25], the association between the severity of stenosis and stroke risk and therefore the indication for surgical intervention remained based on conventional DSA.

The volumetric CTA datasets allow for MPRs and MIPs in any plane and therewith provide much more information on the lumen and its morphology than conventional DSA. The residual lumen is almost never circular and DSA performed in a limited number of projections does not always reveal the narrowest lumen. Analysis of 3D information therefore may provide a more realistic way to assess the true maximum stenosis.

In case CTA replaces DSA in clinical decision making, stenosis measurements on CTA should be performed in a comparable way, i.e., measuring the diameter of the remaining lumen at the level of the maximal stenosis and of the normal lumen distal to the stenosis. This can be done in several ways using different post-processing techniques. Although with MIP reconstructions images comparable to those in DSA can be obtained, this technique is limited in calcified plaques and it is recommended not to use MIP images for stenosis measurements in arteries with calcifications (Fig. 11.4). Generally, one uses 3D software to create MPRs and/or CPRs in oblique planes parallel to the carotid lumen to seek the point of maximum stenosis and measures the smallest diameter in the cross-sectional plane perpendicular to the central lumen line at that level. Figure 11.5 shows this method of stenosis measurement using 3D software. The reference diameter is measured in the same way at a level above the carotid bulb where the lumen walls run parallel to each other (i.e., the healthy distal carotid artery).

When using the ECST criteria to assess the degree of stenosis, CTA directly enables visualization of the outer vessel wall, whereas on DSA the vessel diameter has to be estimated by delineating the projected lumen contour. Therewith, CTA takes into account the changes in vessel diameter caused by vascular remodeling, whereas this phenomenon is ignored when measured on DSA. This might cause differences in ECST stenosis measurements between CTA and DSA.

Several diagnostic studies have been performed which compared single slice CTA with DSA in the assessment of carotid stenosis. From a meta-analysis of studies published between 1990 and 2003, single slice CTA has been shown to have a pooled sensitivity of 85% and a pooled specificity of 93% for detection of a 70–99% stenosis. Sensitivity and specificity for detection of an occlusion were 97% and 99%, respectively [26]. Another systemic review reported a pooled sensitivity of 95% and a specificity of 98% for the detection of a 70–99% stenosis [27]. The latter study also found that CTA was sensitive (95%), but slightly less specific (92%) in depicting stenosis >30%. In 2006, Wardlaw and colleagues performed a meta-analysis comparing noninvasive imaging techniques with intra-arterial angiography. They found only 11 studies on CTA, published between 1980 and 2004, that explicitly met the Standards for Reporting of Diagnostic Accuracy (STARD) criteria [28] and they reported a sensitivity of 77% and a specificity of 95% for diagnosing 70–99% stenosis using CTA [29]. The authors warned for the methodological shortcomings of many studies evaluating diagnostic imaging. They concluded that the existing data might support the cautious use of noninvasive imaging to diagnose 70–99% stenosis, but that more data are needed from carefully designed trials to determine true sensitivity and specificity of noninvasive imaging techniques in routine clinical practice, especially for 50–69% stenosis, or when used in combination [29]. In 2009, Chappell and colleagues performed an individual patient data meta-analysis to find clinically significant estimates of the accuracy of non-invasive imaging in diagnosing severe and moderate symptomatic artery stenosis [30]. They also concluded that existing primary studies provide limited data and that the literature overestimates the accuracy of noninvasive imaging techniques. The small CTA dataset included in this analysis revealed a sensitivity and specificity of 65% and 56% for detection of 70–99% stenosis, respectively [30].

A difficulty in the evaluation of the accuracy of stenosis measurement using noninvasive imaging techniques is that both acquisition and post-processing procedures evolve rapidly. Although multidetector CTA is now widespread and is expected to improve diagnostic accuracy, this has barely been tested. Only one study compared MDCTA with DSA and found MDCTA to have a high specificity and a high negative predictive value for significant carotid disease [31]. Since DSA is not routinely used anymore in clinical practice, the assessment of new noninvasive imaging techniques against DSA can not be justified ethically anymore. Therefore there is an increasing need for practical, reliable methods for evaluating new technologies, for example, by standardized comparison with other non-invasive tests or test phantoms.

Both aforementioned systemic reviews [26, 27] did not provide enough evidence to draw robust conclusions about the diagnostic accuracy of the different post-processing techniques, although stenosis assessment using axial slices and MIPs seemed to be better than when using VR and SSD [27]. Most studies did not report on the exact—combinations of—reformatting techniques used, which hampers a solid meta-analysis. More recent studies comparing the post-processing techniques in MDCTA revealed that stenosis measurements on axial source images are highly reproducible and accurate and that the additional use of MPRs or other reconstructions is not necessary, but might aid in finding the location of the maximum stenosis [32–34].

Manual lumen segmentation and stenosis quantification are laborious and suffer from interobserver and intraobserver variability. Consequently much work has been performed on the development of (semi)automated lumen quantification. The majority of publications with respect to lumen quantification focus on the segmentation of the lumen while the assessment of the severity of luminal stenosis is addressed by few.

Lumen segmentation methods have been reviewed and grouped according to the mathematical framework used [35] or categorized with respect to (1) the way vessel geometry and appearance are modeled, (2) the image features which are used for vessel extraction, and (3) the methodology used in vessel extraction [36].

Some of the published methods have been tailored to or evaluated on carotid CTA images [37–40]. Reported values vary highly. However the comparison of these methods is hampered by the fact that they all use different imaging data and evaluation measures, like Dice similarity coefficient, mean surface distances, or visual inspection. In addition, most studies were performed on small and selected data sets. Figure 11.6 illustrates a 3D lumen segmentation of a carotid bifurcation.

To facilitate an objective comparison of carotid artery segmentation and stenosis quantification algorithms, the Carotid Bifurcation Algorithm Evaluation Framework was set up in 2009 (http:​/​/​cls2009.​bigr.​nl/​) [41]. This framework consists of a publically available image database, annotated data for training and evaluation, and standardized evaluation measures. Till date nine algorithms have been evaluated by the framework, of which only one is fully automatic, whereas the others require three initialization points. The three best performing methods evaluated by the framework have dice similarity coefficients of 0.92, 0.88, and 0.90, mean surface distances of 0.18, 0.54, and 0.17 mm and Hausdorff distances of 1.5, 4.4, and 1.7 mm, respectively [41]. Figure 11.7 shows three examples of lumen segmentations with three different dice values.

These three best performing methods are based on three different approaches, i.e., graph cut, level set, and active surface algorithms [41].

In the graph cut framework voxels are assigned to vessel lumen or background by considering all image voxels as nodes in a three-dimensional graph and creating an optimal surface which separates (cuts) the foreground (lumen) from the background. To compute this optimal cut the image gradient can be used.

In the level set framework, the vessel surface is represented implicitly by the zero level lines (zero level set) of an embedding function (similar as, e.g., sea level in a height map). This embedding function is then changed (evolved), implicitly resulting in deformation of the zero level set. This representation has the advantage that the zero level set can change topology (Fig. 11.8). The evolution of the embedding function should ensure that the zero level set halts at the vessel lumen boundary. This is achieved by defining a speed function derived from the image data. Both the initial segmentation and the design of the speed image are the key ingredients in the design of a level set-based segmentation method.

Active surfaces are a generalization of active contours (also called snakes). Using active surfaces the segmentation is also the result of the evolution of an initially segmented surface. However changing topology is much harder to model in this framework. The segmentation is modeled as a surface on which forces are acting which causes the evolution of the segmentation. This evolution can be constrained by properties of the used surface representation.

Although considerable research has been performed on vessel lumen segmentation, only few researchers have published on automatic vessel stenosis grading [42–44]. Also, approaches differ widely in the evaluation that has been performed, both with respect to evaluation measures and number of data sets used.

The evaluation framework discussed previously also allows objective comparison of performance in stenosis quantification. To date only three stenosis grading methods have been evaluated using this framework, also indicating that this field has received less attention [41]. Clinically, the minimal diameter is often used to calculate the stenosis degree. However, the minimal diameter of a non-elliptical shape is not uniquely defined and is therefore prone to measurement errors and is hard to measure automatically. The evaluation framework evaluates two stenosis measures: an area-based measure which compares the area of the lumen at the stenosis to the area of a distal vessel part and a measure that compares the minimal diameter at the two positions. In the framework, the diameter-based stenosis degree is defined by the smallest line that divides the cross-sectional area in two equal parts. Using automated lumen segmentation, the minimal diameter can easily be replaced by the lumen cross-sectional area. This is a much more accurate measure for the obstruction of the blood flow as carotid arteries, especially at the site of atherosclerotic plaque, generally do not have circular luminal cross sections and also do not run exactly perpendicular to the axial plane of the CT scan. Zhang et al. investigated the use of area measurements and found that assessment of area stenosis on CTA was highly reproducible and showed a satisfactory agreement with diameter stenosis on DSA, although it provides a less-severe estimate of the degree of carotid stenosis, especially in noncircular lumens [45]. The average error in assessing carotid artery stenosis of the best stenosis grading method according to the evaluation framework was 16.9% for area-based and 17.0% for diameter-based measurements [41].

Besides stenosis grading, lumen segmentation also enables the extraction of other quantitative measures. The extracted lumen model can, e.g., be used for Computation Fluid Dynamic calculations to assess the shear stress in the atherosclerotic carotid bifurcation [46] and the quantification of geometric parameters such as vessel tortuosity and bifurcation angles [47].