Glioma Versus Metastasis Lesion Identification using Multi-Voxel MRS with Ensemble Random Forest: A Pilot GLIMMER Study

The accurate and early differentiation of new brain lesions is crucial for early treatment planning. Radiologically distinguishing between primary and secondary brain lesions is often difficult. The purpose of this study was to evaluate multi-voxel proton magnetic resonance spectroscopy (¹H-MRS) for the characterization of intracranial lesions and to develop a Random Forest (RF) classifier to differentiate intracranial metastases from diffuse gliomas in a clinical setting. A total of 33 participants were included in this study. Multi-voxel ¹H-MRS was used to quantify metabolites, along with their ratios and gradients between contrast-enhancing tumor core, peritumoral edema, and ipsilateral normal white matter. For each region, a regression model was fitted to estimate the marginal mean metabolite. Then a RF classifier adapted for class imbalance was trained to discriminate between the two groups, using age, sex, metabolite ratios and selected gradients as predictors. The predictors were first reduced via Monte Carlo Feature Selection (MCFS) before training and evaluation of RF on out-of-bag data. A significant difference between metastases and diffuse gliomas was observed in the edema region for the N-acetylaspartate/choline ratio (NAA/Cho) (difference = -0.38, p = 0.031, Cohen’s d = -0.98) and for NAA/Cho in contrast-enhanced tissue (difference = 0.29, p = 0.046, Cohen’s d = -0.81). The RF classifier achieved an out-of-bag AUC of 0.84 (95% CI: 0.70–0.98) in differentiating intracranial metastases from diffuse gliomas. The results suggest that the RF model that integrates multi-voxel 1 H-MRS features demonstrates preliminary preoperative differentiation potential requiring large-scale multicenter prospective validation.