Quantification of Axonal Regeneration via Machine Learning in a Rat Sciatic Nerve Defect Model
Benjamin R Mintz, PhD1, Arun Nemani, PhD2, Andrew Mesaris, B.S.1, Paul Bonvallet, PhD1 and Ankur Gandhi, PhD1, (1)Integra LifeSciences, Princeton, NJ, (2)Draycon Labs, Chicago, IL
Introduction
Axon counting of histological sections remains the standard for assessing the reinnervation potential of experimental technologies and procedures in peripheral nerves even though histomorphology techniques require training, time, and result in limited datasets. Advances in machine learning can streamline this process, yield supplemental data, and have the potential to reduce variability, but challenges remain in adapting to standard histological techniques. Here, we demonstrate results of a new computational method based on the U-net image segmentation model applied towards automated analysis of peripheral nerve regeneration from histological cross-sections.
Methods
A 15mm sciatic nerve defect was generated in athymic rats. Treatments included allograft, collagen conduit, and autograft. Samples were fixed in situ and stained with Toluidine Blue. Manual axon counting of digital slide scans was performed by region using a modified cell counting plugin in ImageJ (FIJI) to set baseline measurements. Three histological images were selected as independent training, validation, and test images. Ground truth labels were created based on manual count and class selection for training and validation images to assess model performance and generalizability. Modeling used a U-Net based (Zaimi 2018) deep learning model to train features pertaining to axons and myelin structures. Morphological metrics including G-ratio, axon and myelin diameter, thickness, density, eccentricity, and solidity were calculated for each prediction image.
Results
Segmentation performance metrics indicate high performance on a holdout image where the axon and myelin pixelwise f1 score is 0.9005, and 0.8229 respectively, and the pixelwise accuracy is 0.9772, 0.9409. We also predict another holdout sample and show a high accuracy with regards to axon counts (prediction count = 369, ground truth = 374). Figure 1 shows a sample of this image. Average morphological feature statistics for this prediction image can be seen in Table 1.
Conclusions
The application of machine learning demonstrated here provides proof of concept for an automated approach robust enough to generate reliable axon-based metrics from standard Toluidine Blue histological cross-sections while reducing potential for human error and variability. Further, the additional data generated by this method as opposed to counting alone further enables the performance assessment of peripheral nerve interventions.
Figure 1
Table 1:
| Axon area (um^2) | Axon diam (um) | Axon and myelin area (um^2) | eccentricity | gratio | Myelin area (um^2) | Myelin thickness (um) | solidity |
mean | 2.514 | 1.578 | 6.499 | 0.752 | 0.557 | 3.985 | 0.559 | 0.926 |
std | 2.945 | 0.844 | 5.269 | 0.160 | 0.112 | 2.504 | 0.137 | 0.061 |
Back to 2021 Abstracts