Edge detected IMT measurements of the common carotid artery far wall can be consistently obtained in a large cross-sectional sample of the population. Edge detected IMT measurements have strong associations with cardiovascular risk factors, similar but slightly weaker than manual traced IMT measurements. Contrary to prior publications, we have not found that edge detected IMT measurements are more reproducible than manual measurements8, 11, 17–19
although we show that they decrease inter-reader differences.
Carotid IMT measurements have been proposed as a measure of cardiovascular risk and a means of possibly identifying individuals in need of pharmacotherapy or life style interventions20
. Edge detection offers the advantage of obtaining IMT measurements in a standardized fashion so that they can be compared against normative or calibrated values.
We show that IMT measurements made with an edge detector preserve key associations with cardiovascular risk factors () while decreasing reader bias (). As such, they could be substituted for manual traced measurements generating normative data for IMT risk assessment.
We used an algorithm based on dynamic programming that resembles one developed by Wendelhag et al11
. This algorithm processes data based on pixel intensity and gradients and is different than edge detector algorithms based on polynomial fitting of intensity curves perpendicular to interfaces or to algorithms using template matching21
. The results presented in this paper are specific to our implementation of a specific edge detector and do not apply to other edge detectors.
We were able to obtain automated IMT measurements in 99% (5574) of the 5633 individuals with manual determined IMT values. Wendelhag et al.11
reported that readers identified by hand a new start point for their edge detection process in 17% of cases. Our readers had the option to slightly alter the weighting factors used for edge detection. Pre-selected values were used in 93% of cases for the media-adventitia interface and 83% of cases for the lumen-intima interface (Please see http://stroke.ahajournals.org
). Despite these adjustments, shows outliers in individuals with low IMT values and where the algorithm failed (63 cases or 1.1% of individuals). We believe that this algorithm failure depends on the thickness of the media layer since the algorithm requires a minimum number of pixels between the lumen-intima and media-adventitia interfaces. Increasing image size (scaling in pixels/mm) could circumvent this limitation. We have included these 63 cases since they did not substantially alter our findings. By protocol, the same image was measured in order to reduce the variability inherent in image selection. This might bias our results in favor of manual tracings since selecting a different image might have improved edge detection, reduced variability and increased the predictive value of risk factors.
While edge detectors have been used in clinical trials6
, associations between risk factors and edge detected IMT measurements have not been studied. In our review of the literature, we have found studies with small groups of subjects where the reproducibility of edge detector data was evaluated or where edge detected IMT values were compared to manual tracings8, 17, 22–26
. These studies did not evaluate the associations of risk factors with common carotid IMT measurements.
The larger IMT values measured with the edge detector may be due to the mathematical process used to derive edges and the relative thickness of the intima and adventitia (). The mathematical location of these edges tends to be different than the line perceived by the human eye. For example, a human reader would tend to trace a line on the lumen-intima interface while the edge detector would place the edge above this line and cause an overestimation of 0.056 to 0.11 mm (one to two pixels at an image scale of 180 pixels/cm). The weighing function in our algorithm would also tend to “pull” the estimated interface towards the denser pixels in the adventitia, thereby further increasing the estimated IMT. The IMT differences between readers () are 0.09 to 0.14 mm, lower than the difference between manual traced and edge detected IMT values (0.19 mm). Edge detected mean IMT values show lack of inter-reader differences (). Contrary to previous studies, the variability of edge detected IMT was slightly higher than for manual tracing8, 11
. This may represent an image quality issue since we accepted images from six centers rather than from one laboratory.
Carotid artery far-wall interface differences. The lumen-intima interface (top arrow) is thinner than the media-adventitia interface(bottom arrow). This is in part due to the greater thickness of the adventitia.
Our observations apply to this ultrasound device with its presets for image texture and scale. This may limit the general applicability of our observations.
We conclude that edge detected IMT measurements can be substituted for manual traced IMT measurements in cross-sectional IMT studies. As implemented, this specific edge detector gives IMT measurements comparable to manual traced measurements in a large multi-ethnic patient population. Edge detected IMT values might be better suited for cardiovascular risk assessment since they do not seem to have the significant inter-reader differences seen with manually traced IMT measurements.