Khan Amir A, Koudelka Christian, Goldstein Carly, Zhao Limin, Yokemick John, Dux Moira, Sikdar Siddhartha, Lal Brajesh K
Department of Bioengineering, George Mason University, Fairfax, Va; Division of Vascular Surgery, Center for Vascular Diagnostics, University of Maryland School of Medicine, Baltimore, Md.
Department of Bioengineering, George Mason University, Fairfax, Va.
J Vasc Surg. 2017 May;65(5):1407-1417. doi: 10.1016/j.jvs.2016.11.033. Epub 2017 Mar 6.
Vessel wall volume (VWV) assessed by three-dimensional duplex ultrasound (3DUS) imaging provides a more comprehensive measure of plaque burden than conventional two-dimensional measures of diameter stenosis. We previously demonstrated that manual outlining of the arterial lumen-intima boundary and outer wall boundary can be performed reliably on images obtained with a commercially available 3D-DUS transducer. Manual segmentation, however, is time consuming (∼45 minutes), limiting its clinical translation. We have developed a semiautomatic algorithm (manual selection of the carotid bifurcation image with subsequent automatic plaque outlining) to outline carotid plaques on 3DUS data sets. In this study, we investigated the accuracy, reproducibility, reliability, and time taken by this algorithm.
3DUS data sets from 30 patients with asymptomatic ≥50% carotid stenosis underwent manual outlining of lumen-intima boundary and outer wall boundary to measure VWV. Two observers implemented a semiautomatic segmentation algorithm. The algorithm's accuracy was compared with manual outlining using the Pearson correlation coefficient. The Dice similarity coefficient (DSC) and modified-Hausdorff distance (MHD) were used to quantify the geometric similarity of the outlines. We also compared results after an intermediate stage of the algorithm vs the complete algorithm. Reproducibility and the least amount of detectable change in plaque volume were computed for each method. Intraobserver and interobserver metrics for each method were computed using the intraclass correlation coefficient (ICC), coefficient of variability (CV), minimum detectable change (MDC), and standard error of measurement (SEM) of the VWV.
Plaque volume estimates obtained from the semiautomatic algorithm were accurate compared with manual outlining. The Pearson correlation coefficient was 0.76 (P < .001), and measurements were geometrically similar (DSC, 0.85; MHD, 0.48 mm). The algorithm was more reproducible and reliable and could detect smaller changes in plaque volume on repeat imaging (low interobserver variability: ICC, 0.9; CV, 8.22%; MDC, 5.57%; SEM, 1.45%; DSC, 0.88; MHD, 0.43 mm). Intraobserver variability was even lower (ICC, 0.9; CV, 8%; MDC, 3.62%; SEM, 1.31%; DSC, 0.89; MHD, 0.37 mm). Plaque volume estimates at the intermediate stage of the algorithm matched results from the full algorithm (Pearson correlation coefficient, 0.76; DSC, 0.84; MHD, 0.52 mm). The intermediate approach, however, was less reliable than the full algorithm (interobserver: ICC, 0.81; CV, 11.7%; MDC, 9.58%; SEM, 3.46%; DSC, 0.88; MHD, 0.42 mm; intraobserver: ICC, 0.87; CV, 8.6%; MDC, 4.55%; SEM, 1.64%; DSC, 0.89; MHD, 0.38 mm). The full algorithm required ∼14 minutes to implement. However, a quick (7 minutes) and accurate assessment of VWV can be obtained by running only the intermediate stage of the algorithm, although with a loss in repeatability and reliability.
We present a unique algorithm to perform semiautomatic quantification of carotid plaque volume using 3DUS imaging. It is quick (mean time, 14 minutes), accurate, repeatable, and implementable in a clinical environment and in longitudinal studies tracking plaque progression. It reliably detects plaque volume changes as low as 4% to 6% with 95% confidence.
与传统的二维直径狭窄测量方法相比，三维双功超声（3DUS）成像评估的血管壁体积（VWV）能更全面地衡量斑块负荷。我们之前证明，使用市售的3D-DUS换能器获得的图像上，可以可靠地手动勾勒动脉管腔 - 内膜边界和外壁边界。然而，手动分割耗时（约45分钟），限制了其临床应用。我们开发了一种半自动算法（手动选择颈动脉分叉图像，随后自动勾勒斑块），用于在3DUS数据集上勾勒颈动脉斑块。在本研究中，我们调查了该算法的准确性、可重复性、可靠性和所需时间。
对30例无症状且颈动脉狭窄≥50%的患者的3DUS数据集进行管腔 - 内膜边界和外壁边界的手动勾勒，以测量VWV。两名观察者实施半自动分割算法。使用Pearson相关系数将算法的准确性与手动勾勒进行比较。Dice相似系数（DSC）和改良豪斯多夫距离（MHD）用于量化轮廓的几何相似性。我们还比较了算法中间阶段与完整算法后的结果。计算每种方法的可重复性和斑块体积中可检测到的最小变化量。使用类内相关系数（ICC）、变异系数（CV）、最小可检测变化（MDC）和VWV测量的标准误差（SEM）计算每种方法的观察者内和观察者间指标。
与手动勾勒相比，半自动算法获得的斑块体积估计准确。Pearson相关系数为0.76（P <.001），测量结果在几何上相似（DSC，0.85；MHD，0.48毫米）。该算法更具可重复性和可靠性，并且在重复成像时能够检测到斑块体积的较小变化（观察者间变异性低：ICC，0.9；CV，8.22%；MDC，5.57%；SEM，1.45%；DSC，0.88；MHD，0.43毫米）。观察者内变异性更低（ICC，0.9；CV，8%；MDC，3.62%；SEM，1.31%；DSC，0.89；MHD，0.37毫米）。算法中间阶段的斑块体积估计与完整算法的结果匹配（Pearson相关系数，0.76；DSC，0.84；MHD，0.52毫米）。然而，中间方法不如完整算法可靠（观察者间：ICC，0.81；CV，11.7%；MDC，9.58%；SEM，3.46%；DSC，0.88；MHD，0.42毫米；观察者内：ICC，0.87；CV，8.6%；MDC，4.55%；SEM，1.64%；DSC，0.89；MHD，0.38毫米）。完整算法实施需要约14分钟。然而，仅运行算法的中间阶段可以快速（7分钟）且准确地评估VWV，尽管会损失重复性和可靠性。
我们提出了一种独特的算法，用于使用3DUS成像对颈动脉斑块体积进行半自动量化。它快速（平均时间14分钟）、准确、可重复，并且可在临床环境和跟踪斑块进展的纵向研究中实施。它能够以95%的置信度可靠地检测低至4%至6%的斑块体积变化。
