Distinguishing PsP from TP in Glioblastoma: A Multiparametric MRI-Based Radiomics and Machine Learning Approach

Glioblastoma multiforme (GBM), designated as WHO Grade IV IDH-wildtype adult-type diffuse astrocytic tumour, represents the most prevalent and therapeutically challenging primary brain malignancy in adults, with a median survival of 14–16 months following diagnosis. A critical and unresolved clinical challenge in post-treatment GBM management is the reliable differentiation of true tumour progression (TP) from pseudoprogression (PsP), a transient treatment-induced inflammatory phenomenon that produces contrast-enhancing lesions on conventional MRI that are radiographically indistinguishable from genuine tumour recurrence. Misclassification of TP as PsP, or vice versa, carries significant negative consequences for patient outcomes, including delayed initiation of salvage therapy and unnecessary treatment escalation. This study presents a comprehensive radiomics and machine learning framework leveraging dynamic contrast-enhanced MRI (DCE-MRI) and pharmacokinetic modelling to address this diagnostic ambiguity. Radiomic features were extracted using the PyRadiomics toolbox (v3.1.0) from DCE-MRI-derived pharmacokinetic parameter maps, specifically Ktrans, Ve, Vp, Fp, Taui, derived from a novel parsimonious modelling framework that selects the optimal pharmacokinetic model voxel-wise from among five competing formulations (non-linear Tofts, Extended Tofts, Shutter-Speed, 2CXM, and 3S2X) via Akaike Information Criterion minimisation. A cohort of 82 histologically confirmed GBM patients (University of Pennsylvania dataset) was evaluated, with MGMT promoter methylation status incorporated as a molecular covariate. Multiple machine learning classifiers, including Random Forest (RF), Support Vector Machine, Logistic Regression, K-Nearest Neighbours, and XGBoost, were systematically compared across four experimental configurations of increasing pharmacokinetic and molecular complexity. Feature selection was performed via a two-stage pipeline combining Mann-Whitney U univariate filtering and Elastic Net regularisation, reducing the initial 1,073-feature space to a biologically interpretable discriminative subset. Class imbalance was addressed using Adaptive Synthetic Sampling (ADASYN) strictly within training partitions. The Random Forest classifier integrating parsimonious DCE-derived radiomic features with MGMT promoter methylation status achieved the highest discriminative performance, attaining a test AUC of 0.89 ± 0.07, sensitivity of 0.93 ± 0.09, specificity of 0.76 ± 0.19, and F1-score of 0.90 ± 0.07. These findings demonstrate that parsimonious pharmacokinetic model selection represents a methodologically critical step, improving AUC by 26 percentage points over the conventional Extended Tofts-only approach, and that integration of molecular biomarkers produces a synergistic improvement in both sensitivity and specificity simultaneously. The proposed framework demonstrates the feasibility of non-invasive, data-driven TP-PsP discrimination and represents a clinically meaningful advance toward precision oncology strategies for post-operative GBM assessment.