T2WI及对比增强T1WI影像组学模型鉴别纤维型与非纤维型脑膜瘤

目的 观察T2WI及对比增强T1WI （T1C）影像组学模型鉴别纤维型与非纤维型脑膜瘤的价值。方法 回顾性分析423例经病理证实的单发低级别脑膜瘤患者，按7 ∶ 3比例分为训练集（n=296）和验证集（n=127）；提取训练集T2WI和T1C中病灶3 376个影像组学特征，以SelectPercentile单因素分析法及最小绝对收缩和选择算子（LASSO）筛选最优影像组学特征，分别以分类器逻辑回归（LR）、支持向量机（SVM）、随机森林（RF）、线性SVC （LinearSVC）、自适应增强（Adaboost）及决策树（DT）构建鉴别纤维型与非纤维型脑膜瘤的影像组学模型，即模型_LR、模型_SVM、模型_RF、模型_linearSVC、模型_Adaboost及模型_DT，以验证集验证其效能。结果 基于T2WI和T1C共筛出13个最优影像组学特征，以之构建的模型_LR、模型_SVM、模型_RF、模型_linearSVC、模型_Adaboost及模型_DT鉴别训练集纤维型与非纤维型脑膜瘤的AUC分别为0.755、0.739、0.819、0.746、0.990及0.607；在验证集的AUC分别为0.698、0.636、0.752、0.670、0.591及0.609。模型_Adaboost鉴别训练集纤维型与非纤维型脑膜瘤的AUC为0.990，在验证集为0.591，出现过拟合；模型_RF在训练集及验证集中的AUC均高于模型_SVM、模型_linearSVC及模型_DT（Z=2.65~8.25，P均<0.05）；模型_RF在训练集中的AUC高于模型_LR（Z=3.27，P<0.01），在验证集的AUC与模型_LR差异无统计学意义（Z=7.95，P=0.05）。模型_RF诊断效能最佳。结论 术前T2WI及T1C RF影像组学模型可有效鉴别纤维型与非纤维型脑膜瘤。

T2WI and contrast enhanced T1WI radiomics models for distinguishing fibroblastic and non-fibroblastic meningioma

Objective To observe the values of radiomics models based on preoperative T2WI and contrast enhanced T1WI (T1C) for distinguishing fibroblastic and non-fibroblastic meningioma. Methods Data of 423 patients with single low-grade meningioma confirmed by pathology were retrospectively analyzed. The patients were randomly divided into training set (n=296) or validation set (n=127) in the ratio of 7:3. Totally 3 376 radiomics features were extracted based on T2WI and T1C of training set using Shukun technology platform. SelectPercentile univariate analysis, the least absolute shrinkage and selection operator (LASSO) were used to screen the optimal radiomics features. Classifier logistic regression (LR), support vector machine (SVM), random forest (RF), linearSVC, adaptiveboost (Adaboost) and decision tree (DT) were used to construct radiomics models for distinguishing fibroblastic and non-fibroblastic meningiomas, i.e. model_LR, model_SVM, model_RF, model_linearSVC, model_Adaboost and model_DT. The efficiency of the models were verified using validation set. Results Totally 13 optimal radiomics features were screened based on T2WI and T1C. The area under the curve (AUC) of model_LR, model_SVM, model_RF, model_linearSVC, model_Adaboost and model_DT for distinguishing fibroblastic and non-fibroblastic meningiomas in training set was 0.755, 0.739, 0.819, 0.746, 0.990 and 0.607, in validation set was 0.698, 0.636, 0.752, 0.670, 0.591 and 0.609, respectively. AUC of model_Adaboost for distinguishing fibroblastic and non-fibroblastic meningiomas in training set was 0.990, in validation set was 0.591, indicating overfitting. AUC of model_RF was higher than that of model_SVM, model_linearSVC and model_DT in both sets (Z=2.65-8.25, all P<0.05), in training set was higher than that of model_LR (Z=3.27, P<0.01), but being not significantly different with AUC of model_LR in validation set (Z=7.95, P=0.05). Model_RF had the best diagnostic performances. Conclusion RF radiomics model based on preoperative T2WI and T1C could effectively distinguish fibroblastic and non-fibroblastic meningioma.