Clinical Physics Laboratory, Department of Pediatrics, Radboud University Nijmegen Medical Center, Geert Grooteplein Zuid 10 6500 HB Nijmegen, The Netherlands.
Ultrason Imaging. 2010 Jul;32(3):143-53. doi: 10.1177/016173461003200303.
The aim of this study was to test the hypothesis that automatic segmentation of vessels in ultrasound (US) images can produce similar or better results in grading fatty livers than interactive segmentation. A study was performed in postpartum dairy cows (N=151), as an animal model of human fatty liver disease, to test this hypothesis. Five transcutaneous and five intraoperative US liver images were acquired in each animal and a liverbiopsy was taken. In liver tissue samples, triacylglycerol (TAG) was measured by biochemical analysis and hepatic diseases other than hepatic lipidosis were excluded by histopathologic examination. Ultrasonic tissue characterization (UTC) parameters--Mean echo level, standard deviation (SD) of echo level, signal-to-noise ratio (SNR), residual attenuation coefficient (ResAtt) and axial and lateral speckle size--were derived using a computer-aided US (CAUS) protocol and software package. First, the liver tissue was interactively segmented by two observers. With increasing fat content, fewer hepatic vessels were visible in the ultrasound images and, therefore, a smaller proportion of the liver needed to be excluded from these images. Automatic-segmentation algorithms were implemented and it was investigated whether better results could be achieved than with the subjective and time-consuming interactive-segmentation procedure. The automatic-segmentation algorithms were based on both fixed and adaptive thresholding techniques in combination with a 'speckle'-shaped moving-window exclusion technique. All data were analyzed with and without postprocessing as contained in CAUS and with different automated-segmentation techniques. This enabled us to study the effect of the applied postprocessing steps on single and multiple linear regressions ofthe various UTC parameters with TAG. Improved correlations for all US parameters were found by using automatic-segmentation techniques. Stepwise multiple linear-regression formulas where derived and used to predict TAG level in the liver. Receiver-operating-characteristics (ROC) analysis was applied to assess the performance and area under the curve (AUC) of predicting TAG and to compare the sensitivity and specificity of the methods. Best speckle-size estimates and overall performance (R2 = 0.71, AUC = 0.94) were achieved by using an SNR-based adaptive automatic-segmentation method (used TAG threshold: 50 mg/g liver wet weight). Automatic segmentation is thus feasible and profitable.
本研究旨在验证如下假设，即自动分割超声（US）图像中的血管可以在脂肪肝分级中产生与交互分割相似或更好的结果。为此，我们以产后奶牛（N=151）为动物模型进行了一项研究，以验证这一假设。在每只动物身上采集了 5 个经皮和 5 个术中 US 肝脏图像，并进行了肝活检。在肝组织样本中，通过生化分析测量三酰甘油（TAG），并通过组织病理学检查排除除脂肪变性以外的其他肝脏疾病。使用计算机辅助 US（CAUS）协议和软件包提取超声组织特征（UTC）参数-平均回波水平、回波水平标准差（SD）、信噪比（SNR）、残余衰减系数（ResAtt）以及轴向和侧向散斑大小。首先，两名观察者对肝组织进行了交互分割。随着脂肪含量的增加，超声图像中可见的肝血管越来越少，因此需要从这些图像中排除的肝脏比例更小。实施了自动分割算法，并研究了与主观且耗时的交互分割过程相比，是否可以获得更好的结果。自动分割算法基于固定和自适应阈值技术，结合“散斑”形状的移动窗口排除技术。所有数据均在包含 CAUS 的情况下，以及使用不同的自动分割技术进行了分析。这使我们能够研究应用的后处理步骤对各种 UTC 参数与 TAG 的单和多元线性回归的影响。使用自动分割技术可提高所有 US 参数的相关性。推导了逐步多元线性回归公式，并用于预测肝脏中的 TAG 水平。应用接收器操作特征（ROC）分析来评估预测 TAG 的性能和曲线下面积（AUC），并比较这些方法的敏感性和特异性。使用基于 SNR 的自适应自动分割方法（使用的 TAG 阈值：50mg/g 肝湿重）可获得最佳的散斑尺寸估计值和整体性能（R2=0.71，AUC=0.94）。因此，自动分割是可行且有利的。
