Abstract

Brain tumors often feature the genetic biomarker O6-Methylguanine-DNA-Methyltransferase (MGMT) associated with a favorable response to chemotherapy and an improved prognosis. Currently, detecting MGMT presence relies on invasive brain biopsy procedures. Thus, this study aims to develop a data mining-based radiomics methodology for non-invasive identifying and evaluating brain tumor genetic biomarkers using radiomics in magnetic resonance images (MRIs). MRIs with segmentation masks were used to extract variables and employ feature selection techniques. Several machine learning models were trained, where Logistic Regression, employing LASSO selection, emerged as the best-performing model, achieving 61% accuracy. Additionally, an explainability module utilizing Shap values identified three significant variables: a T1CE sequence variable related to texture, a FLAIR sequence variable of first-order statistics, and a T1 sequence variable of first-order statistics. This radiomic methodology, with its performance and explainable nature, could offer diagnostic support to clinicians in tumor management.