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

Area of Science:

• Neuroimaging
• Radiomics
• Machine Learning

Background:

• Preoperative differentiation of meningiomas, brain metastases, and gliomas is challenging due to overlapping MRI features.
• Intratumoral heterogeneity further complicates accurate diagnosis.

Purpose of the Study:

• To evaluate if habitat analysis of multiparametric spin-and-gradient-echo (SAGE) imaging can incorporate spatial heterogeneity for three-class tumor differentiation.
• To assess the performance of different feature sets and classifiers for intracranial tumor classification.

Main Methods:

• Retrospective multicenter study with internal (90 patients) and external (28 patients) cohorts.
• SAGE-derived vascular architecture maps were partitioned using K-means clustering (K=2-6).
• Feature sets included global mean, whole-tumor radiomics, and habitat-derived radiomics, evaluated with 10 classifiers via cross-validation.

Main Results:

• K=3 was selected as the optimal habitat representation.
• The best internal model (Mean + Habitat K3 + RBF SVM) achieved a macro-AUC of 0.700±0.086.
• The best external performance (Mean Only + Linear SVM) achieved a macro-AUC of 0.849, with SHAP analysis highlighting global mean and habitat features.

Conclusions:

• SAGE-based habitat analysis offers spatially resolved features for differentiating three types of intracranial tumors.
• This approach enhances the characterization of tumor heterogeneity for improved classification.