Biopsy-Validated Deep Learning Mapping of Glioma Infiltration Beyond Contrast Enhancement Using Multiparametric MRI

Background: Glioblastoma (GBM) exhibits diffuse tumor infiltration beyond contrast-enhancing margins into FLAIR-hyperintense brain tissue, where infiltrative tumor is not visible on conventional MRI. We developed and validated the infiltrative Tumor Burden (iTB) map, a deep learning biomarker that uses multiparametric MRI (mpMRI) to identify infiltrative tumor within these non-enhancing regions. Methods: Preoperative mpMRI datasets from 80 patients with WHO grade II-IV gliomas and 24 patients with brain metastases were retrospectively analyzed. Imaging included pre- and post-contrast T1-weighted, T2-weighted, FLAIR, dynamic susceptibility contrast perfusion and diffusion MRI-derived maps (rCBV, rCBF, ADC). Spatially localized stereotactic biopsies from non-enhancing glioma regions provided histopathologic ground truth while peritumoral brain metastasis regions served as biologically grounded non-tumor controls. A 3D convolutional neural network was trained on voxel-centered mpMRI patches using patient-level cross-validation and an independent test set, with data augmentation optimized using a greedy local search strategy. Results: Model development included 121 biopsy-localized samples from 56 glioma patients and 104 non-tumor samples from 26 brain metastasis patients. Augmentation optimization significantly improved independent test-set performance, increasing accuracy from 74.5% to 85.5% and AUC from 0.859 to 0.942 (p = 0.008). Ensemble-derived voxel-wise iTB maps demonstrated low infiltration probabilities in brain metastases and extensive regions of elevated infiltration probability in glioblastoma, consistent with expected biology. Conclusions: The iTB biomarker generates voxel-wise maps of tumor infiltration within non-enhancing FLAIR abnormalities. Validated against image-localized histopathology and independent testing, it achieved high diagnostic performance and may improve surgical planning, radiotherapy targeting, and assessment of glioma burden.