Spatial pattern-driven interpretable model and biological correlates in brain glioblastoma-lymphoma differentiation

Abstract

Summary Glioblastoma (GBM) and primary central nervous system lymphoma (PCNSL) often exhibit overlapping appearances on routine MRI, complicating pre-treatment diagnosis. In 1,109 patients from five centers, we constructed standard-space tumor probabilistic maps and derived atlas-anchored spatial features to augment conventional radiomics. The spatial radiomics classifier outperformed radiomics alone (external test area under the ROC curve [AUC], 0.98) with acceptable calibration and decision curve benefit, and SHapley Additive exPlanations (SHAP)-enabled anatomy-grounded interpretation. Aligning tumor localization with the Allen Human Brain Atlas and a normative functional connectome linked GBM-enriched territories to developmental-oncogenic programs and network hubness, whereas PCNSL-enriched territories showed immune-inflammatory/proliferative programs, and associations with network hubness did not survive spatial-autocorrelation correction. These results provide shareable reference maps and an interpretable, multicenter-generalizing tool for GBM-PCNSL differentiation, while offering biological context for diagnosis-specific location susceptibility.