Abstract
Analysis of structural connections between brain regions enables us to gain insight into
the structural architecture of the human brain and into how connections are
affected by age or pathology. Tractography is the standard tool for automatic
delineation of structural connections or tracts. Post-processing of
tractography results using expert prior knowledge is often performed to ensure
a robust delineation. In this thesis, I present a shortest-path tractography
(SPT) framework that can automatically incorporate any prior knowledge about
the location of a tract. Furthermore, I show how such a prior can be learned
from previous tractography results.
A confound common to all SPT methods is their sensitivity to finding many
false-positive connections, since a path between two locations in the brain is
always found. To address this issue I present two approaches to measure the
statistical significance of a connection and demonstrate their application in
connectivity-based parcellation.
Network models are a common way to represent structural connections of the
whole brain. With supervised learning methods, features are extracted from
these networks and are associated with a parameter of interest. Dimensionality
reduction is often performed as pre-processing, since network analysis
typically suffers from high-dimensionality low-sample-size problems.
Preceding the supervised analysis with unsupervised dimensionality reduction can,
however, smooth the discriminative signals, degrading predictive performance.
In this thesis I present a novel supervised dimensionality reduction algorithm
that clusters network nodes into hubs, which reflect common connectivity
structures in the population, and that retains predictive performance of the
lower dimensional features.
the structural architecture of the human brain and into how connections are
affected by age or pathology. Tractography is the standard tool for automatic
delineation of structural connections or tracts. Post-processing of
tractography results using expert prior knowledge is often performed to ensure
a robust delineation. In this thesis, I present a shortest-path tractography
(SPT) framework that can automatically incorporate any prior knowledge about
the location of a tract. Furthermore, I show how such a prior can be learned
from previous tractography results.
A confound common to all SPT methods is their sensitivity to finding many
false-positive connections, since a path between two locations in the brain is
always found. To address this issue I present two approaches to measure the
statistical significance of a connection and demonstrate their application in
connectivity-based parcellation.
Network models are a common way to represent structural connections of the
whole brain. With supervised learning methods, features are extracted from
these networks and are associated with a parameter of interest. Dimensionality
reduction is often performed as pre-processing, since network analysis
typically suffers from high-dimensionality low-sample-size problems.
Preceding the supervised analysis with unsupervised dimensionality reduction can,
however, smooth the discriminative signals, degrading predictive performance.
In this thesis I present a novel supervised dimensionality reduction algorithm
that clusters network nodes into hubs, which reflect common connectivity
structures in the population, and that retains predictive performance of the
lower dimensional features.
| Original language | English |
|---|
| Publisher | Department of Computer Science, Faculty of Science, University of Copenhagen |
|---|---|
| Publication status | Published - 2016 |