Developing Brain Connectivity- Effects of Parcellation Scale on Network Analysis in Neonates

Student thesis: Doctoral ThesisDoctor of Philosophy


Diffusion magnetic resonance imaging (dMRI), tractography and the use of
network measures have combined to form an established approach for exploring
brain connectivity. When applied to the human brain, a definition of regions
of interest (ROIs) which act as network nodes is required. In adults, regions
commonly represent brain areas that are assumed to be functionally coherent.
During early development however, a complete set and locations of ROIs in the
brain is yet to be established. This motivates the use of random parcellation
schemes with varying numbers of regions or scales. However, network measures
can be scale dependent, making comparisons across multiple scales challenging
and hindering group comparisons.
To address such scale dependence, network measures are commonly normalised
using random surrogate networks which act as a baseline. In this work,
the efficacy of commonly used normalisation techniques is determined and new
methods for generating randomised surrogate networks are introduced. Furthermore,
a subset of measures is derived by investigating inter-measure correlations
and the framework is then applied to serial dMRI data of a preterm
cohort. It is shown that a new method for generating surrogate networks for
normalisation improves on established approaches and eliminates scale dependencies over a local range, allowing for meaningful group comparison.
While normalisation may be used for group comparison over a local range,
scale dependence can remain over larger ranges. This work shows that the
nature of the scale dependence varies between cohorts, and proposes a multiscale framework for group comparison. Using this framework to characterise the scale dependence, it is possible to differentiate the groups of neonates studied.
This approach, however, requires the calculation of networks at multiple
scales. Therefore the use of a node-merger scheme is also proposed to infer
network properties at a coarse scale from a single network estimated at a fine
scale. This approach allows for multi-scale group comparison based on a single
starting network.
Date of Award2015
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorPaul Aljabar (Supervisor) & Jo Hajnal (Supervisor)

Cite this