Brain as Language: Neuro-Semantic Modeling of Glioblastoma Radiogenomics

<title>Abstract</title> <p>Glioblastoma (GBM) is an adult-type diffuse glioma, isocitrate dehydrogenase (IDH)-wildtype, central nervous system (CNS) World Health Organization (WHO) grade 4, characterized by pronounced spatial, vascular, and molecular heterogeneity. Radiogenomics aims to associate quantitative magnetic resonance imaging (MRI) phenotypes with molecular tumor characteristics, yet conventional radiomic pipelines usually encode MRI-derived information as high-dimensional numerical feature matrices with limited semantic interpretability. We propose a neuro-semantic radiogenomic language framework that translates multiparametric MRI (mpMRI) radiomics into controlled, evidence-linked patient-level token sequences. Using the University of Pennsylvania Glioblastoma (UPenn-GBM) dataset, radiomic features from 11 MRI sequences or derived imaging measures and three tumor subregions were translated into semantic phenotype domains covering vascular, diffusion, enhancement, infiltration-associated, morphological, textural, and intensity-related tumor characteristics. The framework generated 652 radiogenomic sequence records, of which 611 were linked to clinical metadata and 278 had definitive binary O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation labels. In repeated stratified nested cross-validation, hierarchical semantic modeling achieved a ROC AUC of 0.80 (95% CI, 0.75–0.85), while hybrid radiomic-semantic modeling reached 0.83 (95% CI, 0.78–0.88). The main contribution of this work is a structured representation layer that makes radiomic evidence interpretable, modelable, and auditable, connecting quantitative imaging phenotypes with molecular prediction and verifiable biomedical reasoning.</p>