Bayesian spatiotemporal modelling of fmri data

Resumen

This thesis is concerned with modelling and analysing functional magnetic resonance ¡maging data. fMRI ¡s a non-invasive technique forobtaining images of brain activity in response to specified periods of stimulation, actions or cognitive operations. A series of images is obtained over time under at least two different conditions and regions of activity are detected by observing differences in blood magnetisation due to the hemodynamic response. The analysis of fMRI data considere two main problems; the detection of brain activity and the estimation of the hemodynamic response curve. In this thesis we propose three different models to address these issues: (i) a fully Bayesian temporal model by using an scaled Poisson probability density function shape for the estimation of the HRF; (ii) a spatiotemporal model in which our prior beliefs about activity characteristics parameters (location and magnitude) are formulated in terms of GMRF models; (iii) a new spatiotemporal model which includes a more flexible estimation of the HRF by using a transfer function model. Although the three models are capable of addressing both issues, the first spatiotemporal model has proven to be more accurate in detecting áreas of activity meanwhile the second spatiotemporal model demonstrates a better performance ¡n the estimation of the HRF. There is a trade-off between sensitive detection of brain activity and accurate estimation of the hemodynamic response.