Differentiation of Pathologically Distinct Intracranial Tumors Using SAGE-Based Habitat Analysis: A Multicenter Study

Preoperative differentiation of meningiomas, brain metastases, and gliomas remains challenging because of overlapping magnetic resonance imaging features and intratumoral heterogeneity. We evaluated whether habitat analysis of multiparametric spin-and-gradient-echo (SAGE) imaging could incorporate spatial heterogeneity into three-class tumor differentiation.

This retrospective multicenter study included an internal development cohort of 90 patients scanned at 3.0 T and 5.0 T and an external cohort of 28 patients scanned at 3.0 T. SAGE-derived vascular architecture maps were partitioned by K-means clustering across candidate K values from 2 to 6. Three feature categories were extracted from SAGE parameter maps: global mean parameters, whole-tumor radiomics, and habitat-derived radiomics; their combinations yielded seven feature sets. After ComBat harmonization, 10 classifiers were evaluated using 30-repetition stratified Monte Carlo cross-validation. The model with the highest mean internal validation macro-AUC was evaluated externally. Shapley additive explanations (SHAP) analysis of a prespecified secondary tree-based model examined feature contributions.

K=3 was selected as the most parsimonious and reproducible habitat representation. Mean + Habitat K3 + radial-basis-function support vector machine achieved the highest mean internal validation macro-AUC (0.700±0.086). Its external macro-AUC was 0.629. In a post hoc comparison, Mean Only + Linear support vector machine achieved an external macro-AUC of 0.849. SHAP analysis indicated contributions from global mean and habitat-derived features.

SAGE-based habitat analysis provided spatially resolved features for three-class intracranial tumor differentiation.