Recently, there has been a great interest in methods that can decompose brain networks into clusters with similar connection patterns. However, most of the currently used clustering methods in neuroimaging are based on the stringent assumption that the cluster structure is modular, that is, the nodes are densely connected within clusters, but sparsely connected between clusters. Furthermore, multi-subject network data is typically fit by several subject-by-subject analyses, which are limited by the fact that there is no obvious way to combine the results for group comparisons, or on a group-averaged network analysis, which does not reflect the variability between subjects. In the first part of this thesis, we consider the analysis of a single binary-valued brain network using the Stochastic Blockmodel (Daudin et al., 2008) and compare it to the widely used clustering methods, Louvain and Spectral algorithms. For this, we use the Caenorhabditis elegans (C. elegans) worm nervous system as a model organism whose wealth of prior biological knowledge can be used to validate the functional relevance of network decompositions. We show that the ‘cores-in-modules’ decomposition of the worm brain network estimated by the Stochastic Blockmodel is more compatible with prior biological knowledge about the C. elegans than the purely modular decompositions found by the Louvain and Spectral algorithms. In the second part of this thesis, we propose three multi-subject extensions of Daudin et al.’s Stochastic Blockmodel that can estimate a common cluster structure across subjects. Two of these (non-trivial) models use subject specific covariates to model variation in connection rates in the data. The first and trivial model assumes no variability between subjects, the second model accounts for a global variability in connections between subjects, and the third model accounts for local variability in connections between subjects that can differ across individual within/between-cluster connectivity elements. In the third part of this thesis, we propose a mixed-effect multi-subject model which can account for the repeated-measures aspects of multi-subject network data by including a random intercept. For the second and third part of the thesis, we use intensive Monte Carlo simulations to investigate the accuracy of the estimated parameters as well as the validity of inference procedures. Furthermore, we illustrate the proposed models on a resting state fMRI dataset with two groups of subjects: healthy volunteers and individuals diagnosed with schizophrenia.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:682913 |
| Date | January 2015 |
| Creators | Pavlovic, Dragana M. |
| Publisher | University of Warwick |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://wrap.warwick.ac.uk/77571/ |
Page generated in 0.0017 seconds