Machine learning-based calculation of neurovascular compression surface area correlates with post-microvascular decompression pain outcomes for trigeminal neuralgia

BackgroundMachine learning-generated segmentations of the trigeminal nerve and nearby blood vessels have the potential to quantify the magnitude of neurovascular compression (NVC) in patients with trigeminal neuralgia (TN). This study applies the nnU-Net machine learning method to create segmentations of the trigeminal nerve region from patient MRIs and correlate resulting quantitative NVC metrics with postoperative TN outcomes. MethodsMRIs from patients undergoing microvascular decompression (MVD) for TN from 2019 to 2022 at a single tertiary care facility were split into training, testing, and inference datasets. The trigeminal nerve and surrounding vasculature were manually labeled (i.e., segmented) in the training and testing datasets to create ground truth (GT) segmentations. nnU-Net was trained on GT segmentations in the training dataset, and predicted segmentations were evaluated using the testing dataset via the F1 score, IoU score, and paired comparison of resulting metrics. To contextualize nnU-Net performance, a manual SE-ResNet152 model was trained and deployed using the same datasets. Predicted nnU-Net segmentations in the inference dataset were then correlated with the rate of post-MVD pain recurrence. ResultsOf 366 total GT segmentations, 302 (82.5%) trained the nnU-Net model and 64 (17.5%) validated the predicted segmentations. The nnU-Net models F1 and IoU scores on the testing dataset were 0.797{+/-}0.011 and 0.714{+/-}0.011, respectively, which were higher than those for SE-ResNet152. The sensitivity and specificity of nnU-Nets ability to detect NVC were 91.3% and 66.7%, respectively. Surface area of NVC calculated from nnU-Net and GT segmentations were statistically similar. Deployed on the inference dataset (n=100), higher surface area of NVC was observed in patients without pain recurrence following MVD than patients with pain recurrence (p=0.008). Finally, higher surface area of NVC (hazards ratio [HR] 0.914 per mm2, 95% confidence interval [CI] 0.848-0.985, p=0.019) and presence of NVC (HR 0.369 relative to absent NVC, 95% CI 0.156-0.876, p=0.024) were both associated with a significantly decreased risk of pain recurrence in Cox proportional hazards models. ConclusionnnU-Net can generate high-fidelity segmentations of the trigeminal nerve region, and the resulting NVC surface area metric is significantly associated with post-MVD pain recurrence. nnU-Net can be a standardized tool to evaluate NVC severity for patients seeking TN treatment.