Comparison of machine learning classifiers for differentiation of grade 1 from higher gradings in meningioma: A multicenter radiomics study

Background and purpose: Advanced imaging analysis for the prediction of tumor biology and modelling of clinically relevant parameters using computed imaging features is part of the emerging field of radiomics research. Here we test the hypothesis that a machine learning approach can distinguish grade 1 from higher gradings in meningioma patients using radiomics features derived from a heterogenous multicenter dataset of multi-paramedic MRI. Methods: A total of 138 patients from 5 international centers that underwent MRI prior to surgical resection of intracranial meningiomas were included. Segmentation was performed manually on co-registered multi-parametric MR images using apparent diffusion coefficient (ADC) maps, T1-weighted (T1), post-contrast T1-weighted (T1c), subtraction maps (Sub, T1c - T1), T2-weighted fluid-attenuated inversion recovery (FLAIR) and T2-weighted (T2) images. Feature selection was performed and using cross-validation to separate training from testing data, four machine learning classifiers were scored on combinations of MRI modalities: random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM) and multilayer perceptron (MLP). Results: The best AUC of 0.97 (1.0 and 0.97 for sensitivity and specificity) was observed for the combination of ADC, ADC of the peritumoral edema, T1, T1c, Sub and FLAIR-derived features using only 16 of the 10,914 possible features and XGBoost. Conclusions: Machine learning using radiomics features derived from multi-parametric MRI is capable of high AUC scores with high sensitivity and specificity in classifying meningiomas between low and higher gradings despite heterogeneous protocols across different centers. Feature selection can be performed effectively even when extracting a large amount of data for radiomics fingerprinting.