Radiomics MRI Phenotyping with Machine Learning to Predict the Grade of Lower-Grade Gliomas: A Study Focused on Nonenhancing Tumors

Abstract

Objective: To assess whether radiomics features derived from multiparametric MRI can predict the tumor grade of lower-grade gliomas (LGGs; World Health Organization grade II and grade III) and the nonenhancing LGG subgroup.

Materials and methods: Two-hundred four patients with LGGs from our institutional cohort were allocated to training (n = 136) and test (n = 68) sets. Postcontrast T1-weighted images, T2-weighted images, and fluid-attenuated inversion recovery images were analyzed to extract 250 radiomics features. Various machine learning classifiers were trained using the radiomics features to predict the glioma grade. The trained classifiers were internally validated on the institutional test set and externally validated on a separate cohort (n = 99) from The Cancer Genome Atlas (TCGA). Classifier performance was assessed by determining the area under the curve (AUC) from receiver operating characteristic curve analysis. An identical process was performed in the nonenhancing LGG subgroup (institutional training set, n = 73; institutional test set, n = 37; and TCGA cohort, n = 37) to predict the glioma grade.

Results: The performance of the best classifier was good in the internal validation set (AUC, 0.85) and fair in the external validation set (AUC, 0.72) to predict the LGG grade. For the nonenhancing LGG subgroup, the performance of the best classifier was good in the internal validation set (AUC, 0.82), but poor in the external validation set (AUC, 0.68).

Conclusion: Radiomics feature-based classifiers may be useful to predict LGG grades. However, radiomics classifiers may have a limited value when applied to the nonenhancing LGG subgroup in a TCGA cohort.

Keywords: Grade; Lower-grade glioma; Magnetic resonance imaging; Radiomics; The Cancer Genome Atlas.

Fig. 1. Patient enrollment process for entire LGG group and nonenhancing LGG subgroup in (A) institutional cohort and (B) TCGA cohort.

FLAIR = fluid-attenuated inversion recovery, LGG = lower-grade glioma, TCGA = The Cancer Genome Atlas, TCIA = The Cancer Imaging Archive, T1C = contrast-enhanced T1-weighted

Fig. 2. Workflow of image processing, radiomics feature extraction, and machine learning.

GLCM = gray level co-occurrence matrix, GLRLM = gray level run-length matrix, GLSZM = gray level size zone matrix, ROC = receiver operating characteristic

Fig. 3. Heatmap of AUC values.

Heat map of AUC values from machine learning classifier to predict grade (A) in entire LGG group in internal validation for institutional test set (n = 136) after training on institutional training set (n = 68) and entire LGG group in external validation for TCGA validation set (n = 99) after training on entire institutional cohort (n = 204); and (B) in nonenhancing LGG subgroup in internal validation for institutional test set (n = 73) after training on institutional training set (n = 37) and nonenhancing LGG subgroup in external validation on TCGA cohort (n = 37) after training on entire nonenhancing institutional cohort (n = 110). AUC = area under curve, GBM = gradient boosting machine, LDA = linear discriminant analysis, RF = random forest, RFE = recursive feature elimination, ROSE = random over-sampling examples, SMOTE = synthetic minority over-sampling technique