Machine learning–based quantification of neurovascular compression for correlation with trigeminal neuralgia pain outcomes

Abstract Machine learning–generated segmentations of the trigeminal nerve and surrounding vasculature can quantitatively assess the magnitude of neurovascular compression (NVC) in patients with trigeminal neuralgia (TN). Using the magnetic resonance imaging (MRI) of 183 patients undergoing microvascular decompression (MVD) for TN, this study trains and evaluates the nnU-Net machine learning method to generate 3-dimensional segmentations of the trigeminal nerve region, from which quantitative metrics such as surface area of NVC can be extracted and correlated with postoperative pain outcomes. The accuracy of nnU-Net–generated segmentations was determined by comparison with manually labeled ground-truth (GT) segmentations: resulting model F1 and IoU scores were 0.820 ± 0.012 and 0.743 ± 0.011, respectively, suggesting nnU-Net can generate segmentations with high fidelity. For contextualization, an SE-ResNet152–based U-Net model was trained using the same patient MRIs and was outperformed by the nnU-Net model based on F1 and IoU scores. Predicted nnU-Net segmentations in the inference dataset of 100 additional patients were then correlated with post-MVD pain recurrence rates. Higher NVC surface area was observed in patients without post-MVD pain recurrence than in patients with pain recurrence ( P = 0.004). Furthermore, higher surface area of NVC (HR 0.884 per mm 2 , 95% CI 0.798-0.979, P = 0.018) and presence of NVC (HR 0.421 relative to absent NVC, 95% CI 0.189-0.934, P = 0.033) were each associated with significantly decreased risks of pain recurrence. Given its ability to yield high-fidelity segmentations whose quantitative metrics correlate with clinical outcomes, our nnU-Net model is a proof-of-concept automated approach that can evaluate patients seeking TN treatment.