A Model Based on Radiomics and Machine Learning in Glioma Grading

Abstract

Radiomics-based researches have shown the predictive abilities of machine learning methods in medical diagnosis. However, different machine learning approaches affect the prediction performance. This paper proposes a method based on Tree-based Pipeline Optimization Tool (TPOT) to find the best classification method in glioma grading. This study utilized the public multimodal Brain Tumor Segmentation Challenge (BraTS) 2019 magnetic resonance imaging (MRI) dataset. 3860 radiomics features were extracted from multi-modal MRI images, including tumor morphological features, first-order gray features, texture features, etc. Then the least absolute shrinkage and selection operator (LASSO) was used to select 88 best radiomics features. Finally, the TPOT was used to construct the brain glioma grade prediction model based on the selected features. The accuracy of the model optimized by TPOT was 100% and the area under the ROC )AUC( was 1 in the training group, and 95.52% and 0.98 in the test group, respectively. Based on machine learning algorithms, brain glioma can be graded automatically by radiomics method.