Detection of Fascicles in Peripheral Nerves with a Region-based Convolutional Neural Network

Daniel Tovbis (1,2), Anne Agur (3), Jeremy Mogk (3), José Zariffa (1,2)

1. Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, Canada

2. Toronto Rehabilitation Institute, University Health Network, Toronto, Canada

3. Division of Anatomy, Department of Surgery, University of Toronto, Toronto, Canada

Background:

Neuroprostheses are devices that can interact with the nervous system to help restore function after neurological injuries, for example in the form of implanted functional electrical stimulation systems. To enable this communication, a neural interface (NI) can be placed in or around a peripheral nerve. New NI designs are tested in computational models; however, these often use simplified neural anatomies. Prior research has shown that anatomical variations can cause significant differences in recorded nerve signals, and simplified models may lead to inaccurate results.

By gathering structural data on fascicular anatomy and its variability across the population, a statistical model of neural anatomy in target nerves could be formed. This requires the acquisition of a large dataset, making manual processing infeasible. As a result, there is a need for an automated method to generate peripheral nerve models from anatomical data.

A key part of such a method is the ability to identify fascicles in nerve slices. We therefore aim to develop an automated image processing method capable of identifying fascicles in peripheral nerves, and quantify its performance for use in a reconstruction algorithm.

Methods:

Three median nerve segments were collected from human cadaveric specimens and stained with hematoxylin and eosin (H&E). The images were morphologically processed and aligned using intensity-based registration. A region-based convolutional neural network (RCNN) with the VGG-16 network architecture was trained to identify fascicle bounding boxes.

The algorithm’s performance was quantified by comparing the output of the detection to manual labels using F1 score.

Results:

The network was tested on each of the three segments after being trained on the other two. The image sets contained 1625, 763, and 533 fascicles respectively. The network achieved a mean F1 score of 0.9792±0.0043, with 77, 34, and 15 false positives and 1, 2, and 2 false negatives respectively.

Conclusion:

This fascicle detector will serve to construct models that reflect the variability in peripheral nerve anatomy across the population, and form the basis for a database used by engineers and neurosurgeons to inform the design and implantation of NIs. Better designed and implanted NIs will have greater specificity to target muscles, resulting in better surgical outcomes for patients with spinal cord injuries, and thus, improved quality of life.