Longitudinal MRI habitat changes for predicting postoperative recurrence in glioblastoma

This study aimed to delineate spatially distinct, imaging-defined habitat subregions within the perisurgical FLAIR-hyperintense region using multi-parametric MRI and to evaluate the prognostic value of longitudinal voxel-wise habitat dynamics in predicting tumor recurrence and progression-free survival (PFS) in patients with glioblastoma (GBM). Based on the Response Assessment in Neuro-Oncology (RANO) criteria, GBM patients were categorized into recurrence and non-recurrence groups. T1CE and ADC images were coregistered to the corresponding FLAIR images and resampled onto the FLAIR grid using rigid transformation. The FLAIR hyperintensity areas surrounding the postoperative cavity were delineated. Unsupervised K-means clustering was applied to multiparametric MRI sequences to partition regions of interest into distinct spatial habitats. Longitudinal voxel-wise changes in each habitat were quantified across consecutive scans. Predictive performance was evaluated using receiver operating characteristic (ROC) curve analysis. Patient survival in different risk groups was assessed using Kaplan-Meier curves. K-means clustering determined optimal habitat number as K = 3. Multivariable logistic regression showed that increases in voxel number in Habitat 1 (high T1CE enhancement, moderately hyperintense FLAIR, reduced ADC) and Habitat 2 (non-enhancing T1CE, moderately hyperintense FLAIR, elevated ADC) were independently associated with postoperative GBM recurrence ( P < 0.001). The combined model (AUC = 0.909) outperformed each individual habitat model; bootstrap internal validation indicated an optimism-corrected AUC of 0.857. Survival analysis showed that increased voxels in Habitat 1 and Habitat 2 were closely related to shorter PFS within two years ( P < 0.001). In this retrospective single-center cohort, increases in the voxel counts of imaging-defined perisurgical Habitat 1 and Habitat 2 subregions on multiparametric MRI were independently associated with postoperative GBM recurrence and with shorter progression-free survival, within the covariates available. Because the model has not been externally validated, these findings should be regarded as hypothesis-generating and require prospective multicenter validation before clinical application.