Evaluating the Performance of Segmentation Algorithms on Peripheral Nerve Histological Cross-Sections

Tovbis, Daniel1, 2; Agur, Anne3; Mogk, Jeremy1; Zariffa, José1, 2

1. Institute of Biomaterials and Biomedical Engineering, University of Toronto; 2. Toronto Rehabilitation Institute; 3. Division of Anatomy, Department of Surgery, University of Toronto

Neuroprostheses are devices which can help treat various injuries associated with the nervous system. Understanding the internal fascicular anatomy of peripheral nerves is important to the development of neuroprosthetic systems that can selectively interact with specific neural pathways. However, fascicular anatomy is complex and poorly studied, and there are no standard locations for Neural Interface (NI) implantation when attempting to interface with specific functional targets.

By gathering structural data on fascicular anatomy and its variability across the population, such a standard could be formed. This would necessitate a large data set, such that manual processing would become infeasible. As a result, there is a need for an automated process to generate peripheral nerve models from anatomical data. This algorithm could be used to create models that facilitate the development of novel NI designs, and inform neurosurgeons during NI implantation.

Since in vivo imaging methods cannot resolve fascicles at high levels of detail, the proposed algorithm will base the generated 3D models on histological images of nerve cross-sections. The most crucial step in creating these models is the segmentation of the fascicles in each successive cross-section. Previous studies on fascicular reconstruction have employed semi-automatic segmentation methods, which included user initialization as well as regular expert verification. This novel algorithm is expected to operate autonomously.

Objective:

To characterize the ability of several candidate approaches (active contours, K-means clustering, and watershed-based segmentation algorithms), running in both fully automatic and semi-automatic modes, to segment fascicles in histological cross-sections of a rat sciatic nerve.

Methods:

All algorithms have been implemented in MATLAB. Chan-Vese Active Contours run semi-automatically will have user-defined initial masks. Since fascicles are roughly circular, the automatic mode will first identify circles in the image to generate a mask. Semi-automatic K-means will have user-defined centroid locations, while the automatic version shall iteratively determine centroid locations. Semi-automatic watershed segmentation will have user-defined regions of interest, while the automatic version will attempt to use image intensity to create watershed zones.

Each algorithm will be tested on three sets of nerve images; one with very clearly defined fascicles, one with poorly-defined fascicles, and one average case. The outputs will be compared to a manually-segmented ground truth and quantified using intersection-over-union.

Anticipated Results:

We expect that fully automatic implementations will have worse performance than semi-automatic ones, the degree of which will determine whether or not the time/quality trade-off is acceptable. The results of this research will help determine which segmentation method is ultimately used in the reconstruction algorithm.