next up previous
Next: Signal Up: Conclusions and Discussion Previous: Conclusions and Discussion

Noise

The noise model consists of a space-time simultaneously specified autoregressive model. We used a model comparison technique in the form of DIC, which balanced model complexity with goodness of fit, to deduce which was the best model out of the ones we considered. This turned out to be a spatially non-stationary, but temporally stationary temporal AR(3) model with a within-matter-type edge-preserving MRF prior on the temporal AR coefficients, combined with a temporally stationary, but spatially non-stationary spatial AR(1) model with a within-matter-type edge-preserving MRF prior on the spatial AR coefficients. We observed matter-type dependence of the spatial and temporal autocorrelation of the noise on three different real FMRI datasets. It was clear from the DIC model comparisons that use of MRF prior and ARD prior were both beneficial. However, it is currently not clear how to implement both within the same model. Further improvement might include attempting to model the large scale temporal variations $ l_{it}$ as in (26,21). In this work we remove the worst of these by high-pass filtering as a preprocessing step. If this could be sensibly incorporated into the model then we could be less conservative with the effective high-pass cut-off point and the uncertainty associated with estimating these large scale variations could be taken into account.
next up previous
Next: Signal Up: Conclusions and Discussion Previous: Conclusions and Discussion