Application of MRI radiomics-based machine learning model in prediction of tumor mutational burden and classification of molecular subtypes for low-grade glioma patients

Since 2016, the World Health Organization (WHO) has updated glioma classification by incorporating molecular biology parameters, including lowgrade glioma (LGG). In the new scheme, LGGs have three molecular subtypes: IDH-mutated 1p/19q-codeleted, IDH-mutated 1p/19q-noncodeleted, and IDH-wildtype 1p/19q-non-codeleted entities. The first component in this study proposes a model prediction of LGG molecular subtypes using Magnetic Resonance Imaging (MRI). MRIs were segmented and converted into radiomics features, thereby providing predictive information about the brain tumor classification. With 726 raw features obtained from the feature extraction procedure, this study developed a hybrid machine learning-based radiomics by incorporating a genetic algorithm and eXtreme Gradient Boosting (XGBoost) classifier, to find out 12 optimal features for the tumor classification. To resolve imbalance data, the Synthetic Minority Oversampling Technique (SMOTE) was applied. The XGBoost algorithm outperformed the other algorithms on the training dataset by accuracy value of 0.885. The author continued evaluating the XGBoost model, then achieved an overall accuracy of 0.6905 for the three-subtype classification of LGG on external validation dataset. The proposed model is among the few ones that resolved the three-subtype classification challenge of LGG with a high result compared to previous studies that did the same work.
Besides LGG classification in genotype, by discovering the suppression of negative immune regulation, immunotherapy is promising as an effective treatment method for low-grade glioma patients. However, the therapy could not apply well to all types of LGGs, and tumor mutational burden (TMB) has been shown to be a potential biomarker for the susceptibility and prognosis of immunotherapy in low-grade glioma patients. Hence, predicting TMB benefits brain cancer patients. The second component of this study investigated the correlation of MRI (Magnetic Resonance Imaging)-based radiomic features and tumor mutational burden in low-grade gliomas by applying machine learning methods. Six machine learning - genetic algorithm incorporated models were built to find the most effective features and compare the models to choose the best performance. The Light Gradient Boosting Machine- genetic algorithm incorporation succeeded to select 11-radiomics signature for TMB classification. The imbalanced data problem was solved by using the SVMSMOTE technique in Python coding and got an impressive result. The LightGBM model resulted in high accuracy of 0.7936, and reached a balance between sensitivity and specificity, achieving 0.76 and 0.8107, respectively. To the author’s knowledge, the proposed study represented the best model for classification of TMB in LGG patients at present.