Human microscopic vagus nerve anatomy using deep learning on 3D-MUSE images.

Area of Science:

• Neuroscience
• Medical Imaging
• Biomedical Engineering

Background:

• Current imaging techniques like micro-CT and MRI lack the resolution to detail human vagus nerve (VN) fascicles and boundaries.
• Accurate VN anatomy mapping is crucial for developing effective neuromodulation therapies.

Purpose of the Study:

• To develop a high-resolution imaging method for detailed VN anatomy analysis.
• To create the first comprehensive VN connectome for computational modeling.
• To automate the segmentation of VN structures using deep learning.

Main Methods:

• Developed 3D serial block-face Microscopy with Ultraviolet Surface Excitation (3D-MUSE) for high-resolution VN imaging (0.9-μm in-plane, 3-μm thickness).
• Trained multiple deep learning models (2D U-Net, Attention U-Net, Vision Transformer) for automatic segmentation of VN fascicles, perineurium, and epineurium.
• Utilized pseudo-3D image generation and self-supervised learning for model training.

Main Results:

• 3D-MUSE successfully captured fine details of VN anatomy, including myelinated fibers and connective sheaths.
• A 2D U-Net model trained on pseudo-3D images achieved the highest segmentation accuracy (Dice score of 0.936).
• The developed models show promise for automated segmentation and nerve fiber tractography estimation.

Conclusions:

• 3D-MUSE provides unprecedented resolution for human vagus nerve imaging.
• Deep learning models, particularly 2D U-Net with pseudo-3D data, effectively automate VN structure segmentation.
• This work lays the foundation for a detailed VN connectome to advance neuromodulation therapy modeling.