基于MRI影像组学在脑胶质瘤分级中的研究

Classification of glioma based on MRI radiomics

摘要:

目的 构建基于MRI影像组学模型，以鉴别高级别胶质瘤（HGG）与低级别脑胶质瘤（LGG）。方法 回顾性收集经手术病理证实的120例脑胶质瘤患者病例，按照7∶3的比例随机分为训练集（n=84）和测试集（n=36）。入组患者均于术前行对比增强T₁WI（CE-T₁WI）和T₂液体衰减反转恢复（T₂ FLAIR）序列扫描。使用ITK-SNAP软件手动勾画肿瘤，形成感兴趣区（ROI）。采用最小绝对收缩与选择算子法（LASSO）筛选影像组学特征，通过支持向量机（SVM）学习算法建立影像组学预测模型。绘制受试者工作特征（ROC）曲线评估预测模型的性能，得到曲线下面积（AUC）、敏感度、特异度及准确率。结果 从CE-T₁WI和T₂ FLAIR序列中分别提取影像组学特征1 835个，经去冗除杂处理后，两个序列各筛选出10个、6个特征。在SVM构建的单一序列模型中，CE-T₁WI模型预测性能较佳，其训练集与测试集的AUC值分别为0.898和0.845。联合模型CE-T₁WI+T₂ FLAIR的性能优于单一序列模型，在训练集和测试集的AUC分别为0.952和0.818。结论 基于MRI影像组学所构建的影像组学诊断模型在预测脑胶质瘤高低分级中具有一定的诊断效能，其中联合模型CE-T₁WI+T₂ FLAIR的预测诊断价值最高。

Abstract:

Objective To construct an MRI-based radiomics model to differentiate high-grade gliomas (HGG) from low-grade gliomas (LGG).Methods Retrospectively, 120 patients with surgically pathologically confirmed gliomas were collected and randomly divided into a training set (n=84) and a test set (n=36) according to a 7:3 ratio. Enrolled patients underwent preoperative CE-T1WI and T2 FLAIR serial scans. Tumors were manually segmented using ITK-SNAP software to form regions of interest (ROIs). Radiomics features were screened using least absolute shrinkage and selection operator (LASSO), and the radiomics prediction model was built by support vector machine (SVM) learning algorithm. The performance of the prediction model was evaluated by plotting receiver operating characteristic (ROC) curve to obtain area under the curve (AUC), sensitivity, specificity and accuracy.Results From the CE-T1WI and T2 FLAIR sequences, 1835 radiomics features were extracted respectively, and 10 and 6 features were screened from each sequence after the redundancy and decontamination process. Among the single sequence models constructed by SVM, the CE-T1WI model has better prediction performance. It has AUC values of 0.898 and 0.845 for the training and test sets, respectively.The combined model CE-T1WI+T2 FLAIR outperforms the single sequence model with AUCs of 0.952 and 0.818 in the training and test sets, respectively.Conclusion Radiomics diagnostic models constructed based on MRI radiomics have certain diagnostic efficacy in predicting the high and low grades of gliomas, among which the combined model CE-T₁WI+T₂ FLAIR has the highest predictive diagnostic value.