Differentiation of Intracranial Dural Metastases and Meningiomas Using DSC Perfusion MRI and Machine Learning

Background/Objectives: To assess the diagnostic performance of dynamic susceptibility contrast (DSC) perfusion MRI parameters and machine learning methods for differentiating intracranial dural metastases (IDMs) from meningiomas. Methods: This retrospective diagnostic accuracy study included 56 patients (mean age: 57.6 ± 11.2 years; 20 men) with dural-based intracranial lesions (65 lesions): 18 patients with IDM (27 lesions) and 38 patients with meningiomas (38 lesions). All patients underwent DSC perfusion MRI. Relative cerebral blood volume (rCBV), relative cerebral blood flow (rCBF), diffusion metrics, and dynamic time-signal intensity curve parameters were extracted. Group comparisons were performed using nonparametric statistical tests. Machine learning models, including linear discriminant analysis (LDA), were developed using patient-level grouped nested cross-validation to avoid data leakage. Diagnostic performance was evaluated using out-of-fold receiver operating characteristic (ROC) analysis, calibration assessment, and clinically oriented thresholds prioritizing metastasis sensitivity. Results: rCBV_mean and rCBF_mean were significantly higher in meningiomas than in dural metastases (median rCBV_mean: 4.71 vs. 2.95; median rCBF_mean: 3.44 vs. 2.02; both p < 0.001). Diffusion metrics and dynamic perfusion parameters, including wash-in time, percentage signal recovery, and wash-out slope, did not differ significantly between groups (p > 0.05). Univariate ROC analysis demonstrated strong discrimination for both rCBF_mean (AUC: 0.82; 95% CI: 0.72, 0.90) and rCBV_mean (AUC: 0.82; 95% CI: 0.72, 0.91). An LDA model integrating rCBF_mean and rCBV_mean achieved an out-of-fold AUC of 0.81 (95% CI: 0.72, 0.89) and improved specificity (85%) at a fixed metastasis sensitivity of 85%. Conclusions: DSC perfusion MRI-derived rCBF and rCBV are robust biomarkers for differentiating IDMs from meningiomas. An interpretable machine learning model integrating these parameters improves diagnostic specificity while maintaining high sensitivity.