Multiscale Spatial Characterization of Glioma MRI with MUSE and Radiomics

Purpose. Characterizing glioma heterogeneity requires measurements that retain the physical scale and orientation of intratumoral intensity dependence. We evaluate the original 3D variographic field produced by the MUSE tool (Modeling Uncertainty as a Support for the Environment) alongside directional radiomics and independent image-based molecular classification. Methods. We analyzed postcontrast T1-weighted magnetic resonance images and enhancing-tumor segmentations from the public UCSF-PDGM collection. Verified MUSE outputs were available for 416 independent baseline patients, of whom 375 had paired directional gray-level co-occurrence matrix (GLCM) measurements. Each MUSE field retained 328 original experimental directional variograms with physical lag distances, semivariances, orientations, and voxel-pair support. Agreement was assessed at aligned offsets and through between-tumor comparisons. Patient-level permutation tests evaluated associations with isocitrate dehydrogenase (IDH) status on 2,508 native direction-lag coordinates. A standalone 3D residual convolutional neural network (CNN) provided a separate IDH classification benchmark. Results. At 9,371 geometrically aligned offsets, native MUSE semivariance and half GLCM contrast correlated strongly (Pearson r = 0.954). Across 31,599 eligible between-tumor comparisons, representation-level dissimilarities showed Spearman ρ_S = 0.862, with sensitivity to GLCM feature family and long-offset sampling support. In 364 grade-4 tumors, including 24 IDH-mutant cases, the original MUSE field was associated with IDH after conventional adjustment (global permutation p = 0.00050), but the association was attenuated after primary directional GLCM adjustment (p = 0.256). The independent CNN achieved an out-of-fold IDH ROC-AUC of 0.792 and PR-AUC of 0.414 across 416 patients. Native MUSE examples revealed scale- and direction-dependent spatial behavior, including distances beyond those sampled by the GLCM protocol. Conclusion. Variograms by MUSE provide a physically indexed description of tumor spatial heterogeneity that can accompany radiomic analysis and CNN classification. The observed statistical overlap is substantial, and incremental clinical discrimination beyond directional GLCM features remains unestablished.