Voxel‐Wise Radiomics Habitat Analysis of Posttreatment Gliomas for Noninvasive Differentiation of True Progression and Pseudoprogression

BACKGROUND: Differentiating true progression (TP) from pseudoprogression (PsP) after chemoradiotherapy in gliomas remains challenging because conventional MRI findings overlap. PURPOSE: To assess whether voxel-level radiomics habitat metrics improve TP/PsP classification and complement clinical and molecular features. STUDY TYPE: Retrospective, multi-center. SUBJECTS: 193 glioma patients after treatment: 121 in the training set (54. 2 ± 11. 3 years; 74 men; 85 TP/36 PsP) and 72 in the external testing set (51. 4 ± 10. 4 years; 44 men; 52 TP/20 PsP). FIELD STRENGTH/SEQUENCE: 3 and 1. 5 T; contrast-enhanced T1-weighted spin-echo imaging (T1CE), T2-weighted fast/turbo spin-echo imaging (T2WI), T2-weighted fluid-attenuated inversion recovery fast/turbo spin-echo imaging (T2-FLAIR), diffusion-weighted echo-planar imaging (DWI), and arterial spin labeling (ASL). ASSESSMENT: Voxel-wise radiomics features were extracted from contrast-enhancing tumor ROIs. Gaussian mixture models generated soft habitats, from which voxel-level metrics were calculated. Three models were constructed. The Clinical Model included tumor grade, isocitrate dehydrogenase (IDH) status, O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status, and time interval. Feature selection used least absolute shrinkage and selection operator (LASSO) ; classifier optimization employed Optuna-based Bayesian methods. STATISTICAL TESTS: Receiver operating characteristic (ROC) curve analysis, calibration curves, decision curve analysis (DCA), and SHAP were used. p < 0. 05 indicated significance. RESULTS: In the testing set, the Clinical, Voxel-wise Habitat, and Combined Models achieved areas under the curve (AUCs) of 0. 701 (95% confidence interval CI: 0. 597-0. 806), 0. 832 (95% CI: 0. 736-0. 917), and 0. 890 (95% CI: 0. 819-0. 958), respectively. The Combined Model significantly outperformed the clinical model (difference, 0. 189). Voxel-wise Habitat versus Clinical (difference, 0. 131; p = 0. 164) and Combined versus Voxel-wise Habitat comparisons (difference, 0. 058; p = 0. 105) were not significant. SHAP ranked CBFₕabitatₑdgeₛtandarddeviation, CBFₕabitatₑntropyₘean, and T1CEₕabitatₑdgeₛtandarddeviation as leading contributors. DATA CONCLUSION: Voxel-wise habitat analysis, combined with clinical and molecular features, improved TP/PsP discrimination with interpretable heterogeneity metrics. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 3.