A Comprehensive Evaluation of Non-Vascular Prepontine Cistern Anatomy Influencing Trigeminal Nerve Vulnerability Using Machine Learning–Based Morphometric Analysis

Abstract

Abstract Background: Trigeminal neuralgia (TN) is a severe neuropathic pain disorder traditionally attributed to neurovascular compression. However, emerging evidence suggests that non-vascular anatomical variations of the prepontine cistern may significantly contribute to disease susceptibility. Objective: To quantify non-vascular morphometric features of the trigeminal nerve and adjacent cistern and evaluate their discriminative value using a leakage-free, machine-learning–based MRI pipeline. Methods: We retrospectively analyzed 131 participants (71 with idiopathic TN (iTN) and 60 controls) who were imaged with temporal MRI. Two neuroradiologists independently assessed the neurovascular conflict status, achieving inter-rater agreement of 97% (κ = 0.91). Measured parameters included trigeminal nerve thickness (root and porus trigeminus level), Meckel cave area (axial and coronal plane) and height (sagittal plane), cisternal length (Mean), cisternopontine angle, sagittal angle, and trigeminoclival angle. Model selection employed nested, paired splits across 20 outer repetitions with Optuna-based tuning; Average Precision (PR-AUC) was the optimization target. Six classifier families (Random Forest, SVM, MLP, XGBoost, KNN, Bagging) were evaluated; SHAP and LIME were applied post hoc for interpretability. Results: TN showed thinner nerve diameters (particularly at the porus), larger Meckel cave areas (axial and coronal) and height, smaller sagittal angles, and shorter cisternal length; several of these differences remained significant after multiple-comparison control (e.g.,porus diameters and Meckel cave areas, Holm-adjusted p < 0.01; sagittal angle, Holm p = 0.0092). On held-out test sets, discrimination was consistently high: for SVM, PR-AUC was 86.16 ± 4.39% and ROC-AUC was 87.40 ± 4.52%; the other models clustered closely around ROC-AUC (≈0.85–0.87). Friedman testing demonstrated a global difference on F1 across models; post-hoc Wilcoxon–Holm confirmed that only Random Forest exceeded KNN, while RF, SVM, and XGBoost did not differ pairwise on F1 or ROC AUC. SHAP/LIME prioritized porus-level diameters and Meckel cave measures as leading contributors, aligning with groupwise morphometric shifts