多序列MR影像组学融合模型预测胶质瘤分级的价值

目的:探讨多序列MR影像组学特征及融合模型预测胶质瘤分级的价值。方法:回顾性分析2016年1月~2021年6月间梅州市人民医院经病理证实且行MR检查的胶质瘤病人资料。将多序列MR DICOM格式的原始图像导入ITK-SNAP软件进行VOI勾画，采用GE A.K分析软件进行影像组学特征提取，用ANOVA+Mann-Whiney、Spearman相关分析和LASSO模型进行特征筛选。选择Logistic回归（LR）算法构建单序列模型，选择LR、线性判别分析（LDA）、支持向量机（SVM）3种机器学习算法构建融合序列模型。通过受试者操作特征曲线评估不同模型的预测效能，并计算其曲线下面积（AUC）。结果:共有150例患者纳入本研究，其中低级别胶质瘤（LGG）41例，高级别胶质瘤（HGG）组109例。LGG组与HGG组病人间的性别、年龄差异无统计学意义（P>0.05）。共获得5个最优特征集，其中T1WI特征集包含15个特征，T2WI特征集包含31个特征，Flair特征集包含25个特征，DWI特征集包含12个特征，T1C包含4个特征，融合序列集包含44个特征。基于T1WI、T2WI、Flair、DWI及T1C各序列MR影像组学预测模型在训练集和测试集中的AUC分别为0.719 0、0.651 4，0.769 5、0.711 4，0.7410、0.610 2，0.721 9、0.747 0，0.815 7、0.754 5。基于LR、LDA、SVM算法的融合模型在训练集和测试集中的AUC分别为0.952 4、0.767 0，0.894 8、0.688 1，0.928 6、0.704 5。结论:基于多序列MRI影像组学特征的单序列模型中，T1C单序列预测模型对胶质瘤分级诊断效能最高；相较于单序列预测模型，多序列融合预测模型具有更高的诊断效能，且基于LR算法构建的融合模型较LDA算法及SVM算法构建的融合模型显示出更高的预测效能。

Predictive value of multi-sequence MRI radiomics fusion model for glioma grading

glioma grading. Methods The data of glioma patients who were pathologically confirmed and underwent MR examination in Meizhou Peoples Hospital from January 2016 to June 2021 were retrospectively analyzed. The original multi-sequence MR images in DICOM format were imported into ITK-SNAP software for VOI delineation. After extracting radiomics signatures with GE A.K analysis software, ANOVA+Mann Whitney, Spearman correlation analysis and LASSO model were used for feature screening. Logistic regression (LR) algorithm was selected to build a single sequence model, while LR, LDA and SVM were adopted to establish fusion sequence models. The prediction performances of different models were evaluated using the receiver operating characteristic curve, and the area under the curve (AUC) was calculated. Results A total of 150 patients were enrolled in the study, including 41 in low-grade glioma group and 109 in high-grade glioma group. There was no significant difference in gender and age between LGG group and HGG group (P>0.05). Five optimal feature sets were obtained, and there were 15, 31, 25, 12, 4 and 44 features in T1WI feature set, T2WI feature set, Flair feature set, DWI feature set, T1C, and fusion sequence set. The AUC of MRI radiomics-based prediction models of T1WI, T2WI, Flair, DWI and T1C was 0.719 0, 0.769 5, 0.741 0, 0.721 9 and 0.815 7 on the training set, and 0.651 4, 0.711 4, 0.610 2, 0.747 0 and 0.754 5 on the test set, respectively. The AUC of the fusion models established with LR, LDA and SVM was 0.952 4, 0.894 8 and 0.928 6 on the training set, and 0.767 0, 0.688 1 and 0.704 5 on the test set, respectively. Conclusion Among the single sequence models based on multi-sequence MRI radiomics, T1C single sequence prediction model has the highest efficiency in glioma grading. Compared with single sequence prediction model, multi-sequence fusion prediction model has higher diagnostic efficiency, and the fusion model established with LR shows higher prediction efficiency than the fusion models established with LDA and SVM.