@ARTICLE{Xiao, author = {Wang, Z. and Xiao, X. and He, K. and Wu, D. and Pang, P. and Wu, T. and }, title = {A study of MRI-based machine-learning methods for glioma grading}, volume = {20}, number = {1}, abstract ={Background: Preoperative classification of gliomas is essential to identify the optimal treatment and prognosis. The aim of this study was to identify the optimal machine learning methods that can be used to accurately grade gliomas based on magnetic resonance images (MRI). Materials and Methods: A total of 153 glioma patients from two medical institutions were enrolled. Four methods, namely support vector machine-recursive feature elimination (SVM-RFE), least absolute shrinkage and selection operator (LASSO), max-relevance and min-redundancy (mRMR), and decision trees were used to screen glioma features. Five follow-up classifiers, including decision trees (DT), naive Bayes (NB), K-Nearest Neighbor (KNN), logistic regression (LR), and the support vector machine (SVM), were then used to develop the models. Receiver operating characteristic (ROC) curves were then plotted, and the area under the curve (AUC) was calculated to evaluate the prediction performance of the models. The accuracy, sensitivity, and specificity of the models were also calculated. Results: A total of 1070 predictive features based on image histograms, shape, and texture were extracted from preoperative T1-weighted contrast-enhanced imaging (T1-CE) MRI scans. The SVM-RFE and SVM models yielded the highest prediction performance with an AUC, sensitivity, specificity, and accuracy of 0.985, 94.2%, 89%, and 91.1%, respectively, while LASSO and NB had the lowest accuracy, with an AUC, sensitivity, specificity, and accuracy of 0.854, 97.9%, 72.3% or 85.1%, respectively. The average AUC and accuracy for the four methods were SVM-RFE (0.967, 91.3%), LASSO (0.951, 88.1%), mRMR (0.935, 90%), and DT (0.954, 90.4%). In the validation cohort, the average AUC and accuracy were SVM-RFE (0.837, 80%), LASSO (0.786, 76.6%), mRMR (0.817, 82.2%) and DT (0.70, 71.1%). Conclusion: The radiomics models could yielded a good performance in differentiating LGG from HGG, and the SVM-RFE combined with other machine-learning methods could provide the best average performance. }, URL = {http://ijrr.com/article-1-4081-en.html}, eprint = {http://ijrr.com/article-1-4081-en.pdf}, journal = {International Journal of Radiation Research}, doi = {10.52547/ijrr.20.1.18}, year = {2022} }