Feasibility of Deep Learning-Based Segmentation of the Facial and Vestibulocochlear Nerves on High-Resolution Magnetic Resonance Imaging.

objectiveTo evaluate the feasibility of deep learning-based automated segmentation of the facial and vestibulocochlear nerves within the cisternal and intracanalicular segments on high-resolution magnetic resonance imaging. STUDY DESIGN AND

settingThis was a retrospective imaging study conducted at a tertiary referral center. PATIENTS: Twenty-two adult patients with normal internal auditory canal magnetic resonance imaging and no skull base pathology were included.

methodsManual segmentation of the facial and vestibulocochlear nerves was performed on axial constructive interference in steady state magnetic resonance images using 3D Slicer. The dataset was divided into training, validation, and test subsets. A three-dimensional U-Net convolutional neural network was trained with standard augmentation. Additional Medical Open Network for Artificial Intelligence-based architectures, including Attention U-Net, Dynamic U-Net, and U-Net++, were trained and compared under identical preprocessing and training conditions.

resultsAll models generated anatomically plausible segmentations on qualitative review. The baseline U-Net achieved a test Dice similarity coefficient of 0.6398. Dynamic U-Net demonstrated the highest validation performance (Dice = 0.6236), while U-Net++ achieved the highest test performance (Dice = 0.6716). Attention U-Net demonstrated lower performance in this small-structure segmentation task. Performance trends were consistent with the challenges inherent to segmenting thin neural structures with limited voxel representation.

conclusionsDeep learning-based segmentation of the facial and vestibulocochlear nerves on high-resolution magnetic resonance imaging is feasible within a limited retrospective dataset. Model selection appears important for small-structure segmentation, with Dynamic U-Net and U-Net++ demonstrating relatively higher performance trends within this limited dataset. Although performance metrics were modest and derived from a limited test dataset, automated segmentation showed consistent anatomic overlap with manual labels on qualitative review. These findings provide preliminary technical groundwork for future validation in larger cohorts and extension to clinically relevant applications such as cochlear nerve integrity in implant candidates.