Beta 1


Title Modelling Functional Neuroscience Data
Author Bjerre, Troels (Cognitive Systems, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Supervisor Hansen, Lars Kai (Cognitive Systems, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Institution Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
Thesis level Master's thesis
Year 2009
Abstract This thesis focus on classification of functional magnetic resonance imaging (fMRI) data using spatio-temporal pattern recognition methods based on probabilistic Bayesian principles. Materials/methods: The principles behind fMRI, fMRI pre-processing methods, hemodynamic response (HR) modelling are reviewed, as are the foundations of kernel methods (KMs), sparse kernel based classifiers, and relevant kernel functions. In-depth descriptions of the relevance vector machine (RVM) and spatio-temporal pattern recognition techniques are given. The 'fast' sequential sparse Bayesian learning (SBL) algorithm for RVM training is described, and a direct search method for adaptation of the Gaussian kernel width parameter is modified for use with the fast training algorithm. Additionally, a novel spatio-temporal kernel function is presented, in which the scans within the temporal range of the HR response elicited from the current neuronal activation, are considered for classifying the current brain state. Results: It is shown that the RVM, using Gaussian kernel basis functions, is able to classify linearly non-separable data with low error rate. The generalisability and sparsity of the resulting statistical models are very good when a reasonable value is selected for the Gaussian kernel width. HR optimisation using the suggested spatio-temporal kernel function effectively reduced the error rate, and prevented abrupt fluctuations in the estimated posterior probability. The kernel width adaptation scheme made the RVM virtually non-parametric, when initialising the width parameter within a sensible range. Conclusion: For the fMRI data sets used in this thesis, employing RVM classification, resulted in very sparse models with low classification error rates. Compared to the support vector machine (SVM), which is the current state-of-the art sparse classifier, the RVM provides a usable posterior probability output. Use of the HR optimised spatio-temporal kernel showed promising results.
Imprint Technical University of Denmark (DTU) : Kgs. Lyngby, Denmark
Series IMM-M.Sc.-2009-38
Fulltext
Original PDF ep09_38_net.pdf (2.48 MB)
Admin Creation date: 2009-07-01    Update date: 2010-10-28    Source: dtu    ID: 246053    Original MXD