School of Computer and Electronic Information, Guangxi University, Nanning 530004, Guangxi, PR China; IPILab, Biomedical Engineering Department, University of Southern California, Los Angeles, CA 90033, USA; Department of Medical Imaging, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, Jiangshu, PR China.
Department of Medical Imaging, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, Jiangshu, PR China.
Comput Med Imaging Graph. 2015 Dec;46 Pt 2:227-36. doi: 10.1016/j.compmedimag.2015.09.003. Epub 2015 Sep 18.
Texture patterns of hepatic fibrosis are one of the important biomarkers to diagnose and classify chronic liver disease from initial to end stage on computed tomography (CT) or magnetic resonance (MR) images. Computer-aided diagnosis (CAD) of liver cirrhosis using texture features has become popular in recent research advances. To date, however, properly selecting effective texture features and image parameters is still mostly undetermined and not well-defined. In this study, different types of datasets acquired from CT and MR images are investigated to select the optimal parameters and features for the proper classification of fibrosis.
A total of 149 patients were scanned by multi-detector computed tomography (MDCT) and 218 patients were scanned using 1.5T and 3T superconducting MR scanners for an abdominal examination. All cases were verified by needle biopsies as the gold standard of our experiment, ranging from 0 (no fibrosis) to 5 (cirrhosis). For each case, at least four sequenced phase images are acquired by CT or MR scanners: pre-contrast, arterial, portal venous and equilibrium phase. For both imaging modalities, 15 texture features calculated from gray level co-occurrence matrix (GLCM) are extracted within an ROI in liver as one set of input vectors. Each combination of these input subsets is checked by using support vector machine (SVM) with leave-one-case-out method to differentiate fibrosis into two groups: non-cirrhosis or cirrhosis. In addition, 10 ROIs in the liver are manually selected in a disperse manner by experienced radiologist from each sequenced image and each of the 15 features are averaged across the 10 ROIs for each case to reduce the validation time. The number of input items is selected from the various combinations of 15 features, from which the accuracy rate (AR) is calculated by counting the percentage of correct answers on each combination of features aggregated to determine a liver stage score and then compared to the gold standard.
According to the accuracy rate (AR) calculated from each combination, the optimal number of texture features to classify liver fibrosis degree ranges from 4 to 7, no matter which modality was utilized. The overall performance calculated by the average sum of maximum AR value of all 15 features is 66.83% in CT images, while 68.14%, and 71.98% in MR images, respectively; among the 15 texture features, mean gray value and entropy are the most commonly used features in all 3 imaging datasets. The correlation feature has the lowest AR value and was removed as an effective feature in all datasets. AR value tends to increase with the injection of contrast agency, and both CT and MR images reach the highest AR performance during the equilibrium phase.
Comparing the accuracy of classification with two imaging modalities, the MR images have an advantage over CT images with regards to AR performance of the 15 selected texture features, while 3T MRI is better than 1.5T MRI to classify liver fibrosis. Finally, the texture analysis is more effective during equilibrium phase than in any of the other phased images.
肝脏纤维化的纹理模式是在 CT 或磁共振 (MR) 图像上从初始阶段到终末期诊断和分类慢性肝病的重要生物标志物之一。使用纹理特征的计算机辅助诊断 (CAD) 在最近的研究进展中已经很流行。然而，到目前为止，如何正确选择有效的纹理特征和图像参数在很大程度上仍未确定且不明确。在这项研究中，我们研究了来自 CT 和 MR 图像的不同类型的数据集，以选择最佳的参数和特征，以便对纤维化进行适当的分类。
共有 149 名患者接受了多探测器 CT (MDCT) 扫描，218 名患者接受了 1.5T 和 3T 超导磁共振扫描仪进行腹部检查。所有病例均通过活检作为实验的金标准进行验证，范围从 0（无纤维化）到 5（肝硬化）。对于每个病例，通过 CT 或 MR 扫描仪获取至少四个连续相位图像：预对比、动脉、门静脉和平衡相。对于这两种成像方式，在肝脏 ROI 内提取 15 个基于灰度共生矩阵 (GLCM) 的纹理特征作为一组输入向量。使用支持向量机 (SVM) 和留一法 (leave-one-case-out method) 检查这些输入子集的每种组合，以将纤维化分为两组：非肝硬化或肝硬化。此外，由经验丰富的放射科医生从每个序列图像中以分散的方式手动选择 10 个肝脏 ROI，并对每个病例的 10 个 ROI 中的每个特征进行平均，以减少验证时间。输入项的数量从 15 个特征的各种组合中选择，根据每个组合的特征的准确率 (AR) 进行计数，以确定肝脏阶段评分，然后与金标准进行比较。
根据从每种组合中计算出的准确率 (AR)，无论使用哪种模态，分类肝脏纤维化程度的最佳纹理特征数量范围为 4 到 7。通过计算所有 15 个特征的最大 AR 值的平均值之和得出的整体性能，在 CT 图像中为 66.83%，而在 MR 图像中分别为 68.14%和 71.98%；在 15 个纹理特征中，均值灰度值和熵是所有 3 个成像数据集最常用的特征。相关特征的 AR 值最低，因此在所有数据集均被剔除为有效特征。AR 值随着造影剂的注入而增加，CT 和 MR 图像在平衡期均达到了最高的 AR 性能。
与两种成像方式的分类准确率进行比较，在所选 15 个纹理特征的 AR 性能方面，MR 图像优于 CT 图像，而 3T MRI 比 1.5T MRI 更有利于肝脏纤维化的分类。最后，纹理分析在平衡期比任何其他相的图像都更有效。
Zotero 插件