Machine learning for grading prediction and survival analysis in high grade glioma

Abstract

We developed and validated a magnetic resonance imaging (MRI)-based radiomics model for the classification of high-grade glioma (HGG) and determined the optimal machine learning (ML) approach. This retrospective analysis included 184 patients (59 grade III lesions and 125 grade IV lesions). Radiomics features were extracted from MRI with T1-weighted imaging (T1WI). The least absolute shrinkage and selection operator (LASSO) feature selection method and seven classification methods including logistic regression, XGBoost, Decision Tree, Random Forest (RF), Adaboost, Gradient Boosting Decision Tree, and Stacking fusion model were used to differentiate HGG. Performance was compared on AUC, sensitivity, accuracy, precision and specificity. In the non-fusion models, the best performance was achieved by using the XGBoost classifier, and using SMOTE to deal with the data imbalance to improve the performance of all the classifiers. The Stacking fusion model performed the best, with an AUC = 0.95 (sensitivity of 0.84; accuracy of 0.85; F1 score of 0.85). MRI-based quantitative radiomics features have good performance in identifying the classification of HGG. The XGBoost method outperforms the classifiers in the non-fusion model and the Stacking fusion model outperforms the non-fusion model.

Keywords: Cerebral glioblastoma; Machine learning; Magnetic resonance imaging; Radiomics analysis.

Fig. 1

Patient exclusion criteria flowchart.

Fig. 2

This study uses two subtypes of high-grade glioma as examples. Case 1 (first column): An 11-year-old male with grade III glioma. Case 2 (second column): A 19-year-old female with grade IV glioma.

Fig. 3

Variable selection by the LASSO regression model. (A) Choice of the optimal parameter (λ) in the LASSO regression model with logλ as the horizontal coordinate and regression coefficients as the vertical coordinate; (B) Plot of λ vs. number of variables with logλ as the bottom horizontal coordinate, binomial deviance as the vertical coordinate, and number of variables as the top horizontal coordinate.

Fig. 4

Six model ROC curves for data imbalance not handled with smote.

Fig. 5

Six model ROC curves after using smote to deal with data imbalances.

Fig. 6

ROC curve for the stacking fusion model with an AUC of 0.95.