Hide/Show Apps

Comparison of partial directed coherence and dynamic bayesian network approach for brain effective connectivity modeling using fMRI

Öğe, Oğuzhan Can
Two of the approaches attempting to model brain effective connectivity are compared. These methods are Partial Directed Coherence (PDC) and Dynamic Bayesian Network (DBN). PDC is based on linear and deterministic signal modelling. It is derived from the Granger Causality approach and underpinned by the Multivariate Auto Regressive (MVAR) model. On the other hand, DBN is based on probabilistic signal modelling, which gives DBN the ability of detecting nonlinear interactions between signals unlike all the other estimator methods. In order to compare these two approaches, linear and nonlinear multivariate synthetic fMRI data whose connectivity is known beforehand is generated. In the generation process Hemodynamic Response Function (HRF) is applied after the generation of data by MVAR model. During data generation, the length of the signals, signal-to-noise ratio of the HRF, and complexity of the network (number of channels) are chosen as variables. All in all, these two methods are compared in terms of these parameters. After the comparison, it can be deduced that PDC performs better on linear signals, while it fails on nonlinear signals completely. DBN performs better on nonlinear signals and gives a satisfactory result for linear ones. Since connections in the brain are highly nonlinear and Dynamic Bayesian Network is the only brain effective connectivity estimator method that can differentiate nonlinear signals, it is certain to say that DBN is a more appropriate approach for connectivity modelling than PDC. This conclusion is supported by applying two methods to real fMRI collections of dyscalculia patients at the end.