Comparison of Normalization Techniques for Radiomics Features From Magnetic Resonance Imaging in Predicting Histologic Grade of Meningiomas

Purpose: This study aimed to compare the effects of different normalization methods on radiomics features extracted from magnetic resonance imaging (MRI). Materials and Methods: Preoperative T1-contrast enhanced MRI data from 212 patients with meningiomas were obtained from two university hospitals. The tumors were segmented using 3D Slicer software, and the PyRadiomics framework was used to extract radiomics features. We developed four experiments to predict the histological grade of meningiomas prior to surgery. The first experiment was performed without normalization. The next three experiments used the StandardScaler, MinMaxScaler, and RobustScaler to normalize radiomics features. The PyCaret framework was used for feature selection and to explore an optimized machine learning model for predicting meningioma grades. The prediction models were trained and validated using data from the first hospital. External test data from the second hospital were used to test the performance of the final models. Results: Our testing results demonstrated that meningioma grade prediction performance depends highly on the choice of the normalization method. The RobustScaler demonstrated a higher level of accuracy and sensitivity than the other normalization methods. The area under the receiver operating characteristic curve and specificity of the RobustScaler method were comparable to those of no-normalization but higher than those of the Standard and MinMaxScaler methods. Conclusion: The results of our study suggest that careful consideration of the normalization method may provide a way to optimize the experimental results.