Open Collections

Holmes, Nathan

University of British Columbia

Assessing the morphology of myelin in transmission electron micrograph (TEM) images is an important research tool for studying certain neuropathologies and neuroregeneration. Additionally, TEM can be used to validate the accuracy of other methods of measuring the state of myelin as it allows direct visualization of myelin ultra-structure. Often this requires laborious manual quantification of TEM micrographs to extract useful morphological data, especially in cases where the feature of interest is not commonly measured such that software is not commercially available to aid in quantification. A recent study produced by our lab involved comparing measurements of myelin based on magnetic resonance imaging (MRI) T2 relaxation with TEM images in conditions of healthy and degenerating myelin in the rat spinal cord following a dorsal column transection injury. The MRI T2 techniques can quantify the amount of water trapped between layers of the myelin sheaths, as opposed to water in other spaces in healthy tissue. However, what the MRI T2 measurement was actually corresponding to in degenerating tissue was unknown. The TEM portion of this analysis involved manually tracing the area of the myelin sheath, and thresholding the image to separate intact myelin from vacuous spaces occurring in the myelin as it degenerates. This made it possible to determine that the MRI T2 technique only measures to the quantity of intact myelin in degenerating tissue. This thesis uses the TEM images and the manually quantified intact myelin measurements as data-sets for training machine learning approaches to automate this task. Three approaches are trained and evaluated: a k-nearest neighbour (KNN) approach based on natural clustering of image histograms, a standard SegNet deep-learning artificial neural network, and a more generic dense prediction artificial neural network featuring skip-connections to preserve low level data. The KNN approach was found to be completely unsuitable for the task. The SegNet approach had substantial difficulties which indicate that it did not have the flexibility to fully learn the data-set. Finally, the generic dense prediction network performed robustly in both the training and test data-sets. Additional testing is performed to contrast the performance of the two artificial neural networks.

Text

2023-11-23

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/86658

Zoology

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/