|dc.description.abstract||Diffusion magnetic resonance imaging (dmri) is a technique that can be used to examine the
diffusion characteristics of water in the living brain. A recently developed application
of this technique is tractography, in which information from brain images obtained using
dmri is used to reconstruct the pathways which connect regions of the brain together. Proxy
measures for the integrity, or coherence, of these pathways have also been defined using
The disconnection hypothesis suggests that specific neurological impairments can arise
from damage to these pathways as a consequence of the resulting interruption of information
flow between relevant areas of cortex. The development of dmri and tractography have
generated a considerable amount of renewed interest in the disconnectionist thesis, since they
promise a means for testing the hypothesis in vivo in any number of pathological scenarios.
However, in order to investigate the effects of pathology on particular pathways, it is necessary
to be able to reliably locate them in three-dimensional dmri images.
The aim of the work described in this thesis is to improve upon the robustness of existing
methods for segmenting specific white matter tracts from image data, using tractography,
and to demonstrate the utility of the novel methods for the comparative analysis of white
matter integrity in groups of subjects.
The thesis begins with an overview of probability theory, which will be a recurring theme
throughout what follows, and its application to machine learning. After reviewing the principles
of magnetic resonance in general, and dmri and tractography in particular, we then
describe existing methods for segmenting particular tracts from group data, and introduce a
novel approach. Our innovation is to use a reference tract to define the topological characteristics
of the tract of interest, and then search a group of candidate tracts in the target brain
volume for the best match to this reference. In order to assess how well two tracts match we
define a heuristic but quantitative tract similarity measure.
In later chapters we demonstrate that this method is capable of successfully segmenting
tracts of interest in both young and old, healthy and unhealthy brains; and then describe
a formalised version of the approach which uses machine learning methods to match tracts
from different subjects. In this case the similarity between tracts is represented as a matching
probability under an explicit model of topological variability between equivalent tracts in
different brains. Finally, we examine the possibility of comparing the integrity of groups of
white matter structures at a level more fine-grained than a whole tract.||en