Zhang Zhiwei, Xiao Jingjing, Wu Shandong, Lv Fajin, Gong Junwei, Jiang Lin, Yu Renqiang, Luo Tianyou
Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, 400016, China.
Department of Medical Engineering, Xinqiao Hospital, Third Military Medical University (Army Medical University), Chongqing, 400037, China.
J Digit Imaging. 2020 Aug;33(4):826-837. doi: 10.1007/s10278-020-00322-4.
The grading of glioma has clinical significance in determining a treatment strategy and evaluating prognosis to investigate a novel set of radiomic features extracted from the fractional anisotropy (FA) and mean diffusivity (MD) maps of brain diffusion tensor imaging (DTI) sequences for computer-aided grading of gliomas. This retrospective study included 108 patients who had pathologically confirmed brain gliomas and DTI scanned during 2012-2018. This cohort included 43 low-grade gliomas (LGGs; all grade II) and 65 high-grade gliomas (HGGs; grade III or IV). We extracted a set of radiomic features, including traditional texture, morphological, and novel deep features derived from pre-trained convolutional neural network models, in the manually-delineated tumor regions. We employed support vector machine and these radiomic features for two classification tasks: LGGs vs HGGs, and grade III vs IV. The area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, and specificity was reported as the performance metrics using the leave-one-out cross-validation method. When combining FA+MD, AUC = 0.93, accuracy = 0.94, sensitivity = 0.98, and specificity = 0.86 in classifying LGGs from HGGs, while AUC = 0.99, accuracy = 0.98, sensitivity = 0.98, and specificity = 1.00 in classifying grade III from IV. The AUC and accuracy remain close when features were extracted from only the solid tumor or additionally including necrosis, cyst, and peritumoral edema. Still, the effects in terms of sensitivity and specificity are mixed. Deep radiomic features derived from pre-trained convolutional neural networks showed higher prediction ability than the traditional texture and shape features in both classification experiments. Radiomic features extracted on the FA and MD maps of brain DTI images are useful for noninvasively classification/grading of LGGs vs HGGs, and grade III vs IV.
胶质瘤的分级在确定治疗策略和评估预后方面具有临床意义。为了研究从脑扩散张量成像（DTI）序列的分数各向异性（FA）和平均扩散率（MD）图中提取的一组新的放射组学特征用于胶质瘤的计算机辅助分级，本研究进行了回顾性研究，纳入了108例在2012年至2018年期间经病理证实患有脑胶质瘤且进行了DTI扫描的患者。该队列包括43例低级别胶质瘤（LGGs；均为二级）和65例高级别胶质瘤（HGGs；三级或四级）。我们在手动勾勒的肿瘤区域中提取了一组放射组学特征，包括传统纹理、形态学特征以及从预训练卷积神经网络模型中衍生出的新的深度特征。我们使用支持向量机和这些放射组学特征进行两项分类任务：LGGs与HGGs的分类，以及三级与四级的分类。使用留一法交叉验证方法，将受试者工作特征（ROC）曲线下面积（AUC）、准确率、敏感性和特异性作为性能指标进行报告。在区分LGGs和HGGs时，当结合FA+MD时，AUC = 0.93，准确率 = 0.94，敏感性 = 0.98，特异性 = 0.86；而在区分三级和四级时，AUC = 0.99，准确率 = 0.98，敏感性 = 0.98，特异性 = 1.00。当仅从实体瘤中提取特征或额外包括坏死、囊肿和瘤周水肿时，AUC和准确率仍然相近。不过，在敏感性和特异性方面的效果则好坏参半。在两个分类实验中，从预训练卷积神经网络中衍生出的深度放射组学特征显示出比传统纹理和形状特征更高的预测能力。从脑DTI图像的FA和MD图中提取的放射组学特征对于LGGs与HGGs以及三级与四级的非侵入性分类/分级是有用的。
