Evidence-Bound Reasoning: Neuro-Semantic Verification of Biomedical AI in Glioblastoma Radiogenomics

Radiomic biomarkers derived from magnetic resonance imaging (MRI) have been widely investigated as non-invasive tools for tumor characterization and prognostic modeling in glioblastoma (GBM). However, their clinical translation remains limited, in part due to sensitivity to tumor segmentation variability. In this study, we systematically investigate the relationship between feature robustness and predictive utility in GBM survival modeling using the University of Pennsylvania Glioblastoma Imaging, Genomics, and Radiomics (UPENN-GBM) cohort. A total of 4,752 radiomic features were obtained from multiparametric MRI across three tumor subregions: enhancing tumor (ET), peritumoral edema (ED), and necrotic core (NC). Feature robustness was quantified using the intraclass correlation coefficient (ICC) based on the automatic and expert-refined segmentation versions. Among features with valid ICC estimates, 48.1% were classified as robust. Survival prediction was evaluated using cross-validation with Coxnet, Random Survival Forest, and Gradient Boosting Survival Analysis models. In this cohort, radiomic feature inclusion showed no consistent improvement over the clinical baseline, and robustness filtering produced no detectable performance gain. Model-selected features were less robust than the overall feature pool, indicating a lack of enrichment for robustness. These findings suggest that robustness alone is not a reliable criterion for feature selection in radiomics-based survival modelling.

Background: Radiogenomics allows identification of radiological biomarkers for genomic phenotypes. In glioblastoma, these biomarkers could potentially complement patient stratification strategies. We aim to develop and analytically validate radiological biomarkers that capture immune cell signatures within IDH-wildtype glioblastoma microenvironment using radiogenomic analysis. Methods: This was a retrospective multicenter study using curated open-access anonymized imaging and genomic data from TCGA-GBM, CPTAC, IvyGAP, REMBRANDT and CGGA datasets. Imaging data consisted of MRI-based radiomic features extracted from necrotic core, enhancing and edema regions of deep learning-based auto-segmented tumors. Radiomic feature selections were performed using nested cross-validated LASSO. Support vector machine and ensemble models were trained using seventeen immune and cell-specific score labels extracted from deconvoluted transcriptomic data using pan-cancer and glioblastoma immune signature matrices as reference standards. Seventeen classifier models trained in three cross-cohort strategies were validated on three held-out datasets assessing stability and generalizability. Results: One-hundred-and-seventy-six patients were included in the study. The immune-related radiomic signatures obtained after feature selection were shape, first order and higher order radiomic features. Models predicting macrophage subtype immune signature showed stable mean performance on balanced accuracy (0.67) and precision (0.89) metrics for three independent holdout datasets with ensemble model outperforming support vector machine model. Conclusion: Radiogenomic models non-invasively predicted the macrophage subtype M0 immune signature in IDH-wildtype glioblastoma. These biomarkers have the potential to stratify patients for immunotherapy within prospective glioblastoma clinical trials.

Multimodal medical large language models remain structurally weak for neuro-oncology because volumetric evidence is compressed into generic visual tokens and diagnostic conclusions often lack an auditable link to MRI regions. We present NeuroMosaic, a 3D multimodal language model that converts multi-sequence brain MRI into anatomy-indexed regional tokens, aligns them with clinical narrative and molecular concepts, and generates evidence-linked outputs. The architecture combines a multi-resolution volumetric tokenizer, a neuroanatomical graph router, a molecular concept memory, and selective risk control. Across four glioma cohorts, NeuroMosaic achieved an internal subtype macro-F1 of 0.827 and external macro-F1 values of 0.784, 0.761, and 0.742. On UPenn-GBM, it improved over the strongest matched-input baseline by 3.6 percentage points (95% CI: 1.8 to 5.4, adjusted p = 0.0018), with IDH, 1p/19q, and MGMT AUROCs of 0.918, 0.861, and 0.781. Evidence pointing accuracy reached 0.703, and targeted evidence deletion reduced correct-answer probability by 0.187, compared with 0.046 for random deletion. These results establish anatomy-indexed routing as a measurable mechanism for accurate, grounded, and calibrated volumetric medical-language reasoning.