Greedy Technique and Its Validation for Fusion of Two Segmentation Paradigms Leads to an Accurate Intima–Media Thickness Measure in Plaque Carotid Arterial Ultrasound

Abstract

The intima–media thickness (IMT) of carotid arteries is a reliable indicator of cardiovascular risk. IMT is usually manually measured on longitudinal B-mode ultrasound images. Many computer-based techniques for IMT measurement have been proposed to overcome the limits of manual segmentation. Most of these, however, require a certain degree of user interaction. We recently developed two automated techniques for segmentation of carotid arteries in ultrasound images and its intima–media thickness (IMT) measurement. The first technique was determined on the basis of local statistics and signal analysis (we named it CULEXsa, ie, Completely User-independent Layer EXtraction based on signal analysis). The second technique was determined on the basis of an integrated approach consisting of feature extraction, line fitting, and classification (or CALEXia, Completely Automated Layers EXtraction based on integrated approach). Both the techniques automatically traced the lumen—intima (LI) and the media—adventitia (MA) boundaries of the carotid wall. IMT was measured as the distance between LI and MA. We observed that CULEXsa offers better performance in the LI segmentation and CALEXia offers better performance in MA segmentation. The goal of this research was to estimate the IMT measurement by using a greedy approach by fusing CULEXsa and CALEXia segmentations, given the ground truth (manually traced by experts). Starting from the technique with the lower system error (CULEXsa for LI and CALEXia for MA), we iteratively swapped the vertices of the profiles until we minimized its overall distance with respect to manual boundary. The fusion boundary, consisting of points of CULEXsa and points of CALEXia, was the Greedy boundary. We used the polyline distance as a metric for both performance evaluation and error minimization. We ran the segmentation protocol over the database of 200 carotid longitudinal B-mode ultrasound images and compared the performance of all the three techniques. The mean error of the greedy technique yielded 0.42 ± 0.89 pixel (26.3 ± 55.6 μm) for the LI boundary (a 12.5 ± 5.6% improvement over CULEXsa) and 0.26 ± 0.56 pixel (16.2 ± 31.3 μm) for MA boundary (a 16.1 ± 6.7% improvement over CALEXia). IMT measurement error is 1.33 ± 0.99 pixel (83.1 ± 61.8 μm), a 3.6 ± 1.4% improvement over CULEXsa.