Comparison of Radiomics-Based Machine-Learning Classifiers in Diagnosis of Glioblastoma From Primary Central Nervous System Lymphoma
Chaoyue Chen, Aiping Zheng, Xuejin Ou, Jian Wang, Xuelei Ma
Abstract
Purpose The purpose of current study was to evaluate the optimal magnetic resonance (MR) radiomic-based machine-learning algorithms in differentiating glioblastoma (GBM) from primary central nervous system lymphoma (PCNSL). Method One-hundred and thirty-eight patients were enrolled in this study. A total number of 43 texture features were extracted from post-contrast T1-weighted image. The optimal features were selected with five selection methods (distance correlation, random forest (RF), least absolute shrinkage and selection operator (LASSO), eXtreme gradient boosting (Xgboost) and Gradient Boosting Decision Tree (GBDT)). Three radiomic-based machine-learning classifiers (linear discriminant analysis (LDA), support vector machine (SVM) and Logistic Regression (LR))were tested to evaluate their diagnostic performance. The diagnostic performances of classifiers were evaluated with receiver operating characteristic (ROC) analysis, with which areas under curves (AUC) of the radiomic classifiers were calculated and compared. Result The optimal model was Distance Correlation as selection method and LDA as classifier with the highest AUC of 0.978. Brilliant diagnostic performance would be observed in each classifier if combined with suitable selection method (Distance Correlaion+SVM=0.959; LASSO+LR=0.966). Conclusion Radiomic-based machine-learning algorithm can be useful in differentiating GBM from PCNSL. Moreover, Distance Correlation as selection method and LDA as classifier are the suitable model.