Although measurement of calcification using CT is commonly used as a measure of atherosclerosis, non-calcified plaque quantification has been mainly limited to ultrasound and MRI because of the higher contrast resolution allowing better visualization of soft tissue [20
]. However, the carotid and coronary arteries are smaller vascular beds, whereas quantification of systemic atherosclerosis requires a much larger scan range, especially if the lower extremities are included, to assess risk for peripheral artery disease. Ultrasound is necessarily a local imaging tool, and a whole body MRI may not give sufficient resolution for effective quantification within a reasonable scan time. Calcification is also better visualized and quantified on CT. Although detection of non-calcified plaque has been explored using computed tomography [6
], no automated quantitative method has yet been developed for this modality. One method proposed by Shum et al. [48
] quantifies wall thickness in the presence of thrombus in abdominal aortic aneurysms. This method relies on a median filter edge detector and extensive manual interaction to set thresholds for the detection of the outer wall. The paper is mainly concerned with the detection of the inner wall, which is obscured by intraluminal thrombus. The algorithm applies only to abdominal aortic aneurysms with intraluminal thrombus, whereas our algorithm is designed to calculate wall thickness in all clinically significant arteries in patients without severe vascular disease. When combined with previous research [11
], our method has the potential to quantify both non-calcified and calcified plaques in all clinically significant systemic arteries, from the thoracic aorta to the arteries of the calf, over a wide range of diameters.
Automated and semi-automated in vivo methods of quantifying aortic wall thickness have been studied before using both MRI and transdermal and intravascular ultrasound [20
]. More detailed characterization of individual plaques by MRI has also been validated against pathological specimens [24
]. However, characterization by both MRI and ultrasound has been largely limited to individual plaques or smaller vascular beds such as the carotid or the coronary arteries, and there is no existing literature describing a method for global characterization of systemic plaque burden that includes the lower extremity arteries. One prior method reported by Adame et al. [20
] relied heavily on detecting periaortic fat using axial MRI slices and an ellipse fitting procedure to calculate the outer wall contour. Although the method has excellent correlation with manual measurement, this method does not take into account artery orientation, which is important in smaller arteries and older patients with more tortuous large arteries. The method also requires an MRI sequence that is not currently in clinical practice, assumes that the wall is almost circular, and assumes that the outer wall is largely parallel to the inner wall. Our method is independent of artery orientation and uses a standard CT angiography protocol that is routinely used for other indications.
Adjacent soft tissue or vascular structures that abut the outer wall of a blood vessel pose the greatest challenge for vascular segmentation, even for experienced human reviewers. Common adjacent structures include the diaphragmatic crura, bowel, bone, musculature, and other arteries and veins. Our algorithm, by its nature, follows the unobscured wall closely and uses a tuned weighting factor to follow the obscured outer wall with accuracy similar to human observers in a variety of vascular beds and adjacent soft tissue structures, as exemplified in Fig. .
Fig. 3 Oblique reformats perpendicular to the course of the artery showing examples of algorithm performance when the outer wall border is obscured by adjacent soft tissues. The white outline represents the outer wall contour calculated by the algorithm. Crosshairs (more ...)
The precision of our algorithm could be affected by differences in user selection of the arterial points. While we did not assess this directly, these user inputs only affect the location of the centerline used to create orthogonal cross-sections. In a previous study [49
], it was shown that the centerline algorithm that we use was highly insensitive to these inputs, and so we expect our results to be similarly precise.
The results of the sensitivity analysis (Table ) indicate that errors in thickness measurement are insensitive to choice of the distance weight, α, over a large range of α. This indicates that a single optimal value of α can be calculated from a representative set of scans. Also, since our algorithm is deterministic, the outer wall contour that is calculated is guaranteed to be the one that conforms to the global minimum cost.
Our algorithm was designed to reduce user interaction time to a few seconds and achieves an average time of 0.04 s per image, enabling quantification of a scan of the whole body in about 1 min, which makes automated quantification clinically viable. One important factor reducing processing time is that edges of the graph are constrained to a course around the circumference of the vessel. Thereby, the number of possible paths is reduced.
While there is no imaging gold standard with which to assess accuracy, our method gives results that are within the range of human precision. Validation with manual determination on 20 CT angiograms showed that the measured error of the automated algorithm was comparable to the inter-observer variability. Although the error does increase (though not significantly) with smaller arteries, it is still within the range of human variation. The measured maximum wall thickness in images with atherosclerotic plaque is significantly greater than the maximum wall thickness in images without atherosclerotic plaques, indicating that the algorithm has potential to be used for detecting and quantifying atherosclerotic regions of the arterial wall.