Few-Shot Sample-Specific Deep Learning Multiclass Segmentation and Fiber Tractography using 3D-Microscopy with Ultraviolet Surface Excitation (3D-Muse) Images of Human Vagus Nerve

At CWRU, the Reconstructing Vagal Anatomy (REVA) project images the vagal complex from brainstem to abdomen using multiple modalities and resolutions. Here we describe the analysis of 3D Microscopy with Ultraviolet Surface Excitation (3D-MUSE), a section-and-image system that enables microscopic visualization and quantification of the morphology and microanatomy of 15-mm vagus nerve samples. Variations in staining, tissue, and image quality across samples require few-shot training of sample-based neural networks to segment fascicles, perineurium, and epineurium. After pre-training a Vision Transformer on all unlabeled images, we fine-tuned a U-Net Transformer with 1-10% segmented images. Highest performance achieved mean Dice scores of 0.91±0.02 with 1% (10) training images, and 0.94±0.01 with 4% (40) MUSE image segmentations reveal nerve density variations, positive correlation between perineurium thickness and fascicle diameter, and thickness variation due to fascicle splits and merges. Fiber tractography reveals dynamic fascicular organization and requires further investigation into small fascicle (<150µm-diameter) behavior.