Comparison of different radiomic models based on enhanced T1-weighted images to predict the meningioma grade

Abstract

To compare different models in predicting meningioma grade based on enhanced T1-weighted images. One hundred and eighty-eight patients with meningioma were analysed retrospectively. There were 94 high-grade meningiomas which formed the high-grade group comprising 68 World Health Organization (WHO) grade II meningiomas and 26 WHO grade III meningiomas. Ninety-four low-grade meningiomas were selected randomly to form the low-grade group. Least absolute shrinkage and selection operator (LASSO) regression was used to reduce the dimensions of the texture parameters. Support vector machine (SVM), decision tree (DT), conditional inference trees (CIT), random forest (RF), k-nearest neighbours (KNN), back-propagation neural network (BPNet), and Bayes were used to construct models. Receiver operating characteristic (ROC) analysis and decision curve analysis (DCA) was applied and compared among different models. Every model performed well and had a high area under the ROC curve (AUC; all >0.80). In the seven models, the highest accuracy was obtained with SVM and KNN (0.79), the highest sensitivity was obtained with DT and Bayes (0.85), and the highest specificity was obtained with SVM and CIT (0.83). SVM and RF had the highest AUC (0.884). KNN had the largest net benefit when the threshold probability was <0.50, whereas SVM had the largest net benefit when the threshold probability was >0.50. Different radiomic models based on enhanced T1-weighted images can be used to predict meningioma grade. The model of SVM and KNN performed better than other models with a larger net benefit.