Non-invasive identification of mesenchymal glioblastoma using quantitative radiomic features from advanced diffusion...

Glioblastoma (GBM) is no longer regarded as a single disease, as distinct molecular subgroups exist, with the mesenchymal (MES) having the worst prognosis. As such, there is a critical need for noninvasive methods to determine GBM molecular status. Although conventional magnetic resonance imaging (MRI)-based radiomics showed promise for predicting GBM characteristics, few studies evaluated pipelines that leverage advanced diffusion MRI (dMRI) techniques, such as diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI), enabling characterization and quantification of tumor microstructure.

To identify advanced dMRI radiomic features specific to MES GBM, we enrolled 36 GBM patients (4 mesenchymal, 32 non-mesenchymal), who underwent presurgical DTI and NODDI protocols. Post-surgery samples were processed to establish subgroup-specific GBM sphere-forming cell (GSC) lines, generating 21 xenografts (12 non-mesenchymal, 9 mesenchymal) that were subjected to the same dMRI protocols.

By leveraging a preclinical-to-clinical transfer learning approach, a machine learning classification algorithm was developed to generalize between preclinical and clinical contexts. Models were trained on xenograft-derived data and validated using an independent patient test set. Using bootstrap resampling to estimate confidence intervals, the XGBoost model achieved an area under the receiver operating characteristic curve of 0.93 (95% confidence interval (CI): 0.79-1.00) and a balanced accuracy of 0.86 (0.64-1.00) for MES prediction. A subset of 9 selected features was sufficient to build a model that accurately predicted MES affiliation.

DTI and NODDI radiomics revealed key features that predict MES GBM and correlate with biological and clinical characteristics.

A DTI and NODDI-based model trained on preclinical xenograft-derived data can be validated in a human patient cohort, demonstrating cross-species generalizability of radiomic biomarkers. This approach provides a noninvasive means to molecularly stratify GBM patients, enabling the potential to inform tailored treatment.

We defined a machine learning algorithm that, starting from subgroup-specific glioblastoma xenografts, reliably identifies the mesenchymal affiliation of glioblastoma patients. The specific dMRI features selected from experimental preclinical models of glioblastoma hold a remarkable predictive value. The same features provide insights into subgroup-restricted tumor tissue microstructure and its relationship with the malignant behavior of mesenchymal glioblastomas.