Background and purpose: In gliomas, characterization of the molecular landscape plays a critical role in determining prognosis and guiding treatment regimens. Imaging biomarker models hold promise for non-invasive characterization of glioma subtypes. We, comprehensively, assessed the potential of magnetic resonance imaging (MRI) data for predicting the survival status and molecular subtypes of glioma.

Methods: We introduce a novel method for quantifying the spatial distribution of gliomas within brain anatomy. The method measures the volumetric ratio of 32 brain anatomical structures affected by the tumor. This novel feature set was combined with established radiomics to build models for predicting O6-methylguanine-DNA methyltransferase (MGMT) methylation status, isocitrate dehydrogenase (IDH) mutation status, and Overall Survival (OS) time of glioma patients. The performance of these models was evaluated on preoperative MRIs of 1788 subjects from four independent datasets, employing both cross-validation (CV) and cross-dataset evaluation strategies.

Results: The proposed feature set revealed no regular patterns in tumor locations across the brain. Integration of these features with radiomics improved model performance for the three tasks. The best performance, in terms of AUROC, respectively, for CV and cross-data tests were: 0.685 and 0.628 for MGMT status, 0.972 and 0.764 for IDH status, and 0.748 and 0.719 for OS time status.

Conclusions: Our experiments demonstrate the potential of imaging biomarkers for IDH prediction, highlighting the challenges associated with predicting MGMT and OS only from image data. This underscores the need for additional information beyond MRI, for accurate prediction of these prognostic markers.