|
|
Exploring Brain Network Connectivity through Hemodynamic Modeling
Havlíček, Martin ; Hluštík, Petr (oponent) ; Šmídl,, Václav (oponent) ; Jan, Jiří (vedoucí práce)
Functional magnetic resonance imaging (fMRI) utilizing the blood-oxygen-level-dependent (BOLD) effect as an indicator of local activity is a very useful technique to identify brain regions that are active during perception, cognition, action, and also during rest. Currently, there is a growing interest to study connectivity between different brain regions, particularly in the resting-state. This thesis introduces a new and original approach to problem of indirect relationship between observed hemodynamic response and its cause represented by neuronal signal, as this indirect relationship complicates the estimation of effective connectivity (causal influence) between different brain regions from fMRI data. The novelty of this approach is in (generalized nonlinear) blind-deconvolution technique that allows estimation of the endogenous neuronal signals (system inputs) from measured hemodynamic responses (system outputs). Thus, it enables a fully data-driven evaluation of effective connectivity on neuronal level, even though only fMRI hemodynamic responses are observed. The solution to this difficult deconvolution (model inversion) problem is obtained through a nonlinear recursive Bayesian estimation framework for joint estimation of hidden model states and parameters. This thesis is divided into three main parts. The first part proposes a method to solve the above mentioned inversion problem. The method uses a square-root form of a nonlinear cubature Kalman filtering and cubature Rauch-Tung-Striebel smoothing extended to a joint estimation problem defined as a simultaneous estimation of states and parameters in a sequential manner. The method is designed particularly for continuous-discrete systems and obtains an accurate and stable solution to model discretization by combining nonlinear (cubature) filtering with local linearization. Moreover, the inversion method is equipped with the adaptive estimation of measurement, state, and parameter noise statistics. The first part of the thesis is focused only on the single time course model inversion; i.e. estimation of neuronal signal from fMRI signal. The second part generalizes the proposed approach and applies it to multiple fMRI time courses in order to enable the estimation of coupling parameters of a neuronal interaction model; i.e. estimation of effective connectivity. This method represents a novel stochastic treatment of dynamic causal modeling, which makes it distinct from any previously introduced approach. The second part also deals with methods for Bayesian model selection and proposes a technique for detection of irrelevant connectivity parameters to achieve improved performance of parameter estimation. Finally, the third part provides a validation of the proposed approach by using both simulated and empirical fMRI data, and demonstrates robust and very good performance.
|
|
|
Exploring Brain Network Connectivity through Hemodynamic Modeling
Havlíček, Martin ; Hluštík, Petr (oponent) ; Šmídl,, Václav (oponent) ; Jan, Jiří (vedoucí práce)
Functional magnetic resonance imaging (fMRI) utilizing the blood-oxygen-level-dependent (BOLD) effect as an indicator of local activity is a very useful technique to identify brain regions that are active during perception, cognition, action, and also during rest. Currently, there is a growing interest to study connectivity between different brain regions, particularly in the resting-state. This thesis introduces a new and original approach to problem of indirect relationship between observed hemodynamic response and its cause represented by neuronal signal, as this indirect relationship complicates the estimation of effective connectivity (causal influence) between different brain regions from fMRI data. The novelty of this approach is in (generalized nonlinear) blind-deconvolution technique that allows estimation of the endogenous neuronal signals (system inputs) from measured hemodynamic responses (system outputs). Thus, it enables a fully data-driven evaluation of effective connectivity on neuronal level, even though only fMRI hemodynamic responses are observed. The solution to this difficult deconvolution (model inversion) problem is obtained through a nonlinear recursive Bayesian estimation framework for joint estimation of hidden model states and parameters. This thesis is divided into three main parts. The first part proposes a method to solve the above mentioned inversion problem. The method uses a square-root form of a nonlinear cubature Kalman filtering and cubature Rauch-Tung-Striebel smoothing extended to a joint estimation problem defined as a simultaneous estimation of states and parameters in a sequential manner. The method is designed particularly for continuous-discrete systems and obtains an accurate and stable solution to model discretization by combining nonlinear (cubature) filtering with local linearization. Moreover, the inversion method is equipped with the adaptive estimation of measurement, state, and parameter noise statistics. The first part of the thesis is focused only on the single time course model inversion; i.e. estimation of neuronal signal from fMRI signal. The second part generalizes the proposed approach and applies it to multiple fMRI time courses in order to enable the estimation of coupling parameters of a neuronal interaction model; i.e. estimation of effective connectivity. This method represents a novel stochastic treatment of dynamic causal modeling, which makes it distinct from any previously introduced approach. The second part also deals with methods for Bayesian model selection and proposes a technique for detection of irrelevant connectivity parameters to achieve improved performance of parameter estimation. Finally, the third part provides a validation of the proposed approach by using both simulated and empirical fMRI data, and demonstrates robust and very good performance.
|
host ::
RSS.