fMRI bandpass filter artifacts about cato HOT 3 CLOSED

A secondary issue was introduced by the mirror filtering. In the current implementation, the signal is mirrored both vertically and horizontally 'around' the first and last point. However, if the signal is highly erratic, there is a high probability this mirroring creates jumps (example on the right in image).

It appears that for our use case only mirroring in the time dimension is better.

from cato.

I compared the current solution as implemented in #28 to Gustafsson's method that is available in scipy.signal.filtfilt. Note that this method is referenced in the MATLAB filtfilt documentation but it appears it is not actually implemented (see also here).

Here are the various options side by side for a few voxels:

If I compute the within-voxel absolute z-scores and then average over all voxels to get a rough metric of how 'exceptional' the timeseries are at the edges, I get for this particular subject:

Based on this I don't have a preference for either method. However, I couldn't find a clean implementation of Gustafsson's method for MATLAB, and I would rather avoid implementing our own version, so based on this I suggest we keep the fix as in #28.

from cato.

Thank you for analyzing and comparing the different methods! I agree with the proposed solution of padding the signal using the the time-mirrored timeseries. This method is might not be perfect, as indicated by the small bumps still visible in the mean absolute z-score plot, but it provides a good balance between retaining as many time points in the analysis and having a well-filtered signal. I will merge #28 into the main branch and release a new update for CATO.

from cato.