MRI-derived habitat heterogeneity for overall survival risk stratification in IDH-wildtype, CNS WHO grade 4 glioblastoma.

BACKGROUND: Glioblastoma (GBM), IDH-wildtype, CNS WHO grade 4 has marked spatial heterogeneity, yet routine MRI-based prognostic assessment often relies on whole-tumor summaries. We developed a preoperative MRI habitat-analysis framework to quantify intratumoral and peritumoral spatial phenotypes and to evaluate their value for overall survival (OS). METHODS: This retrospective multicohort study included a development cohort of 473 patients assembled from the UCSF-PDGM-v5 dataset (n = 354) and Yantai Yuhuangding Hospital (n = 119), and an independent external testing cohort from the First Affiliated Hospital of Ningbo University (n = 82). All habitat generation, selection of the optimal habitat number, feature selection, and model development were performed using the pooled development cohort. Multiparametric preoperative MRI was partitioned into voxel-wise habitats using k-means clustering. Habitat-derived variables were selected and combined into a risk score using LASSO-Cox modeling. Baseline, habitat, and combined prognostic models were constructed using Cox proportional hazards regression and evaluated using the C-index, time-dependent AUC, calibration, prediction-error analysis, decision curve analysis, and Kaplan-Meier risk stratification. RESULTS: A four-habitat solution was stable and biologically interpretable. Necrotic-like and edema-dominant peripheral habitats showed the strongest adverse associations with OS. The habitat risk score remained independently prognostic after adjustment for the consistently available baseline variables of age, sex, and MGMT promoter methylation. In external testing, the combined model improved the external C-index from 0.604 to 0.701 and the 12-month AUC from 0.626 to 0.724, with reduced prediction error, improved calibration, and separated high- and low-risk groups. CONCLUSION: MRI-derived habitat analysis captures survival-relevant spatial heterogeneity in GBM and provides an interpretable, noninvasive approach for risk stratification warranting prospective validation.