Quantification of unmyelinated nerve fibers in electron microscopy images using deep learning

The vagus nerve is a mixed nerve containing sensory and motor fibers that carry signals between the brain and body. Selective targeting of different fiber subsets is a major goal of neuromodulation strategies, including vagus nerve stimulation (VNS). Improved methods for identifying and quantifying specific fiber types within the vagus nerve (such as A, B, and C-fibers) would provide important insights for selective neuromodulation at the ultrastructural level and enable more realistic in-silico testing of different VNS parameters. While larger myelinated A- and B-fibers are straightforward to identify, most of the fibers in the vagus nerve are unmyelinated sensory C-fibers that are difficult to identify due to their small size and poor contrast in most imaging modalities. Traditional methods have failed to provide quantitative information about unmyelinated fibers, which number in the thousands to tens of thousands, depending on the species studied. In this study, we trained deep learning (DL) based convolutional neural network (CNN) to identify and quantify unmyelinated C-fibers in the mouse cervical vagus nerve. Cross-sections from wildtype C57BL6 mice were processed with glutaraldehyde-paraformaldehyde fixation, followed by 1% osmium tetroxide, then stained in uranyl acetate. Ultra-thin sections were imaged with a transmission electron microscope (FEI) and Ultra Scan 4000 imaging system (GATAN). These TEM images were used as input images for training the model. Our trained DL model has instance segmentation architecture that can provide the spatial distribution (mask) and count of type-C unmyelinated fibers simultaneously for a given test image. We validated the fine-tuning of the DL model with various hyperparameters (learning rate, batch size, number of iterations) for accurate type-C fiber detection. Generation of the target images requires intensive manual annotations; therefore, we limited the training dataset to a subset of our dataset. We incorporated a rapid inference-based image denoising network on the input images that enhanced the contrast information during the pre-processing step. Our trained model detects more than 50% type-C fibers for a given test image. By adding additional target images, we expect our model will identify more than 75% type-C fibers, in addition to providing quantitative measurements such as average area, the radius of fibers, and their spatial distribution within the nerve based on segmentation results. The demonstrated novel method implements an automated pipeline to quantify unmyelinated fibers from mouse vagus nerve TEM images, but it may also be applied to other nerves and species of interest.