Multicenter MRI Radiomics Features to Predict IHD1 Gene Mutation Status of Low-Grade Glioma

Abstract

Background: IDH mutation has been incorporated into the World Health Organization classification of gliomas, and its role in treatment recommendations is under development. Purpose: We aim to predict IDH1 mutation status from T1, T1-Gd, T2, and T2-fluid-attenuated inversion recovery (FLAIR) MRI sequences. Material and method: We used 119 patients' data from the cancer genome atlas low-grade glioma (based on histopathologic criteria) (TCGA-LGG) public database. We extracted 103 image biomarker standardization initiative-compliant radiomics features from whole tumors of all MRI sequences, including shape, histogram, and texture features. An extra tree classifier was used to select A subset of features to maximize the prediction model performance and minimize the size of the feature space. A support vector machine (SVM) classifier tuned with a Bayesian optimizer was employed to construct the classifier. Results: The extra tree classifier selected about one-third of the features for each MRI sequence. The Bayesian optimizer selected radial kernel for all sequences and its corresponding hyper-parameters including γ, \mathcal{C} for each sequence. The AUC-ROC curve values were above 0.96 ± 0.01) for all MRI sequences validation dataset, and the lowest and highest values of AUC for test data were 0.97 and 0.98 obtained from T2/T2-FLAIR and T1-Gd, respectively. The minimum test accuracy was just above 92% for T2-FLAIR and the highest value was just under 94% for T1. Conclusion: Radiomics biomarkers from MRI sequences, including T1, T1-Gd, T2, and T2-FLAIR, could predict the IDH1 mutation status with a clinically acceptable performance after tuning an SVM classifier.