Beta 1


Title Real-time Analysis of Brain Imaging Data : A High-performance Distributed Pipeline for Multivariate Spatio-temporal Brain State Classification
Author Hansen, Toke Jansen (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 2010
Abstract Can we build machines that in real-time can decode intentions in the human mind? This thesis addresses this particular question by combining modalities such as functional magnetic resonance imaging (fMRI), digital signal processing, statistical machine learning and high-performance computing. The thesis introduces machine learning from the generic Bayesian viewpoint, as well as the principles behind fMRI and available pre-processing, feature selection and modelling approaches. As the rst we apply online learning techniques in the context of real-time fMRI, and to understand the applied online support vector machine (SVM), an in-depth analytical description of the classical SVM is given, together with a description of kernel methods and constrained optimization techniques. We develop a real-time scanner module for the Siemens Trio system, as well as a brain-computer interface (BCI) game paradigm, in which a human subject competes against an online multivariate model, that is trained on the subjects brain patterns during the entire game. For efficient sequential pre-processing of acquired fMRI data and the respective multivariate analysis, a generic pipeline framework is developed, allowing the composition of both parallel and distributed classification pipelines. To investigate the temporal buildup of the human readiness potential, we devise a novel nonlinear kernel based searchlight heuristic, yielding similar results as resampling based searchlight approaches, but with a vastly reduced computational complexity, applicable for real-time scenarios. Finally, ensemble methods are used in simulations to construct a spatio-temporal model for brain state classification, based on a pipeline instantiation with more than 200 different parameterized parallel processing online SVMs, all while maintaining a latency applicable for real-time fMRI.
Imprint Technical University of Denmark (DTU) : Kgs. Lyngby, Denmark
Series IMM-M.Sc.-2010-19
Fulltext
Original PDF ep10_19_net.pdf (12.56 MB)
Admin Creation date: 2010-04-08    Update date: 2010-10-28    Source: dtu    ID: 259704    Original MXD