Room: AAPM ePoster Library
Purpose: The aim of this study was to identify malignant glioma grades (grade III and IV) based on imaging biomarkers of a few conventional magnetic resonance imaging (MRI) sequences.
Methods: Preoperative conventional contrast-enhanced T1-weighted MR images (cT1WIs) and T2-weighted MR images (T2WIs) were utilized in this study. Morphology, first-order, second-order (texture), and wavelet features were extracted from tumor regions on the MR images as the imaging biomarkers. Six machine learning (ML) algorithms: logistic regression (LR), support vector machine (SVM), standard neural network (SNN), random forest (RF), and k-nearest neighbor (k-NN) were employed to construct classification models. The imaging biomarkers were selected using statistical and regularization-based methods before constructing the models. Leave-one-out cross validation (LOOCV) in a training dataset were performed as an internal validation. An external validation using the training and test datasets with selected imaging biomarkers for all folds of LOOCV in the internal validation was performed.
Results: The number of cases in the training and test datasets were 157 (III: 55, IV: 102) and 67 (III: 22, IV: 45), respectively. The mean area under the receiver operating characteristic (ROC) curve (AUC) values for all six classification models in the internal and external validation were 0.903±0.022 and 0.751±0.032, respectively. The best performances in the internal and external validations were SVM with AUC value of 0.932, RF with AUC value of 0.800.
Conclusion: The identification of malignant glioma grades using MRI biomarkers has been investigated in this study. It has been suggested that a few conventional MRI sequences could sufficiently predict the malignant glioma grades in the imaging biomarker analysis.
Funding Support, Disclosures, and Conflict of Interest: This study was supported by JSPS Grant-in-Aid for Scientific Research Grant numbers 18J00599 and 18K15625.