基于SWI序列的支持向量机模型鉴别高级别胶质瘤与脑内单发转移瘤

Support Vector Machine Model Based on SWI Sequences to Discriminate High-Grade Glio-ma from Intracerebral Solitary Metastases

摘要: 目的：探讨基于磁敏感加权成像(SWI)序列的支持向量机模型鉴别高级别胶质瘤与脑单发转移瘤的临床价值。方法：回顾性分析经病理确诊的103例高级别胶质瘤及脑单发转移瘤患者的SWI序列图像，在获得的图像上手动勾画感兴趣体积(Region of interest, ROI)并提取影像组学特征，所有病例按照70%:30%分为训练组和验证组，训练组用于筛选特征和建立机器学习模型，特征筛选由独立样本t检验，Mann-Whitney U检验和最小绝对收缩与选择算子(least absolute shrinkage and selection operator, LASSO)完成，特征筛选后的数据建立支持向量机(support vector machine, SVM)模型。应用受试者操作员特征(ROC)曲线评价模型的诊断性能，结果表示为曲线下面积(AUC)，准确度、灵敏度、特异度，阳性预测率和阴性预测率。验证组数据用于进一步验证。结果：基于SWI序列建立了鉴别高级别胶质瘤及脑内单发转移瘤的诊断模型。训练组中，模型曲线下面积为0.951，诊断的特异度、灵敏度，准确度，阳性预测率及阴性预测率分别为0.889，0.800，0.857，0.889，0.800。验证组中模型曲线下面积为0.868。特异度、灵敏度，准确度，阳性预测率及阴性预测率分别为0.875，0.889，0.880，0.933，0.800。结论：基于SWI序列的支持向量机模型能有效提高术前鉴别高级别胶质瘤及脑内单发转移瘤的诊断效能。

Abstract: Objective: To investigate the clinical value of a support vector machine model based on magnetic susceptibility weighted imaging (SWI) sequences to identify high-grade gliomas and solitary brain metastases. Methods: SWI sequences of 103 patients with pathologically confirmed high-grade gliomas and single brain metastases were retrospectively analyzed, and the volume of interest (ROI) was manually outlined on the acquired images and the imaging histological features were extracted, and all cases were divided into a training group and a validation group according to 70%:30%. Feature screening was done by independent sample t-test, Mann-Whitney U-test and least absolute shrinkage and selection operator (LASSO), and support vector machine (SVM) model was built on the data after feature screening. The diagnostic performance of the model was evaluated by apply-ing the subject operator characteristic (ROC) curve, and the results were expressed as area under the curve (AUC), accuracy, sensitivity, specificity, positive prediction rate and negative prediction rate. Validation group data were used for further validation. Results: A diagnostic model based on SWI sequences was established to discriminate high-grade glioma from intracerebral solitary me-tastases. In the training group, the area under the curve of the model was 0.951, and the specificity, sensitivity, accuracy, positive prediction rate and negative prediction rate were 0.889, 0.800, 0.857, 0.889, 0.800, respectively. The area under the model curve in the verification group was 0.868. The specificity, sensitivity, accuracy, positive predictive rate and negative predictive rate were 0.875, 0.889, 0.880, 0.933 and 0.800, respectively. Conclusion: The support vector machine model based on SWI sequences can effectively improve the diagnostic efficacy of preoperative identification of high-grade glioma and intracerebral solitary metastases.