Quantitative Description Of Microtubule Disorganisation in Neurodegenerative Diseases: Software Development and Image Analysis

  • Beatriz Costa Gomes

Student thesis: Phd

Abstract

Axons are the long, cable-like processes of neurons which form the nerves that wire our bodies. The maintenance of these delicate structures requires continuous parallel bundles of filamentous polymers called microtubules, which run all along the axonal core. In ageing and neurodegeneration, axonal microtubules often become curled and disorganised, causing detrimental accumulations of organelles and vesicles. The hypothesis of Local Axon Homeostasis suggests that the disorganisation of axon bundles is caused by the constant mechanical stress from the activity of motor proteins performing life-sustaining axonal cargo transport. In healthy axons, microtubule-regulating proteins prevent this from happening. If such proteins are dysfunctional, microtubules have a heightened probability of becoming disorganised. In this scenario, we would expect that microtubule phenotypes should appear similar, regardless of which microtubule regulatory proteins are absent; their images seem to suggest this. Therefore, my task was to develop objective parametric analyses to describe and compare microtubule phenotypes of different genetic conditions. First, I tested a number of existing software packages (Imaris, CellProfiler, NeuronJ, FiberApp), none of which proved suitable for a number of reasons. Especially curvature descriptions of the curled microtubules were impossible with any of these software packages. I therefore investigated different algorithms that would allow such analyses, and chose the Smoothing Spline Fitting method, which I turned into computational code in MATLAB. Tests of these algorithms using spirals and ellipses of defined parameters revealed a fairly stable reproduction of their curvatures. In addition, I used Hough Transforms to establish tools to identify and analyse straight lines. In my validation tests using sample images, these performed better than classic metrics. In the next step, I designed ALFRED (Advanced Labelling, Fitting, Recognition and Enhancement of Data) as user-friendly software in which the above algorithms can be applied. The development of ALFRED included strategies for image processing (from reduction of background noise to skeletonisation), ROI selection (with the help of MATLAB user interface tools) and path recognition (skeletons transformed into graphs and the shortest path algorithm is applied). When using ALFRED to analyse biological images of neurons, length and area measurements were matching closely with manual analyses. Curvature and straightness analyses of neurons of different genotype revealed a clear distinction between normal and mutant neurons, and different mutant conditions revealed similar valued, as is consistent with the hypothesis. The software is now available for wider application. As a complementary approach, I attempted to compare different phenotypes using machine learning. Data clean-up of neuron images was performed with Ilastik, but no accurate outcomes were achieved in the time available. Two pre-trained classification networks (VGG16 and ResNet18) were used to classify different genotypes, but results were inconclusive, likely also due to the insufficient data clean-up of images.
Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorKarl Kadler (Supervisor), Matthias Heil (Supervisor), Andreas Prokop (Supervisor) & Simon Pearce (Supervisor)

Cite this

'