Select a calendar:
Filter August Events by Event Type:
Events for August 23, 2019
Fri, Aug 23, 2019 @ 09:30 AM - 11:30 AM
Thomas Lord Department of Computer Science
PhD Candidate: Daniel Moyer
Date: Friday August 23rd, 2019
Time: 9:30 am - 11:30 am
Location: GFS 213
Committee: Greg Ver Steeg, Paul Thompson, Aram Galstyan, Aiichiro Nakano
Representation Problems in Brain Imaging
Diffusion weighted Magnetic Resonance Imaging (dMRI) is a imaging modality that, conditional on direction, queries the propensity for water diffusion in living tissue. These images are desirable for studying tissues with anisotropic diffusion profiles, e.g. the white matter of the human brain. I here describe two problems at the interface between diffusion imaging and computational science, as well as proposed solutions for those problems, reconciling differences between current methodology, existing theory, and practical constraints.
The first part of this thesis describes methodology from the intersection of network science and neuroscience. Previous literature has worked to estimate white matter axon bundles (fasiculi) from dMRI. Combined with segmentations of the grey matter, this leads intuitively to the construction of macro-scale brain networks, where nodes are cortical regions and edges are estimates of axonal connections. The work I present here deconstructs this progression, reconciling the inherently spatially continuous cortex with discrete network theory. The resulting model is a generalization of discrete graphs, using point process theory to create continuous interaction densities (continuum adjacency functions).
The second part of this thesis describes a method for learning representations of data that invariant under changes in specified outside factors. These can be applied to diffusion imaging data, which in particular suffers from instrument/observer biases (\"site/scanner\" biases). Due to the relative complexity of the domain, modelling the particular effects of each instrument may be difficult; moreover, such an approach does not generalize to unseen instruments. Instead, the described method can learn representations that are minimally informed of the imaging site. Subsequent derived data will then be at least as uninformed of the site variable.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon