next up previous
Next: Results Up: Evaluation data Previous: Artificial activation in real

Artificial signal in synthetic noise

In a similar way to that outlined above, we added various source signals into Gaussian and autoregressive noise. The background noise parameters (i.e. voxel wise mean and std. deviation in the case of Gaussian noise and AR parameters in the case of autoregressive noise) were estimated from the resting state FMRI data. We then added 10 spatial maps and associated time courses taken from ICA decompositions of various other true FMRI data sets. The sources were chosen to represent different source processes that commonly are identified in real FMRI data, e.g. high frequency noise within the ventricular system, fluctuations in the B  $ _0$ field homogeneity spatially located near tissue-air boundraries, activation maps etc. This should not bias the results of the comparison in favor of PICA given that these spatial maps originated from different data sets and as such are not mutually spatially independent. Similarly, the associated time courses are not uncorrelated. This, we belive, is a more faithful representation of FMRI data. In FMRI it is often possible (and almost always advisable) to create experiments under an orthogonal experimental design which hopefully renders the temporal responses to external stimulation to be mutually orthogonal. Any additional source processes, however, can have arbitrary correlation with any column of the design. We therefore did not impose any constraints on the associated time courses.


next up previous
Next: Results Up: Evaluation data Previous: Artificial activation in real
Christian F. Beckmann 2003-08-05