Hi Herberto

Not sure about this one. If you don't mind helping me debug this, the only way I imagine filtering may totally flatten the timeseries is if the filter window is beyond the Nyquist frequency. Perhaps this could happen for example if the data TR value is misspecified? Could you perhaps double-check in that same line the values for "tr" and "bp_filter" to see what those are?

Thanks

Alfonso

Originally posted by Herberto Dhanis:

Hi there,

I think I found a bug in the denoising module of CONN. Starting from fully preprocessed images I applied the aCompCor pipeline of CONN followed by the bandpass filter. After doing so I need to estimate SPM.mat files and I noticed that they could not be estimated due to 'no significant voxels'. The most likely way for this to happen is if there is literally on variation in the scans, so I checked both the 'd' and 'b' nifti files (using the module 'd' is right after regression and 'bd' is after the bandpass has been applied on top of the regression). While there was variation in the 'd' files, the 'b' files had no variation between the scans, meaning that the scan at t=1 was the same as the one at t=2,3,...,N.

I am not sure what is going on but I did some debugging. In the screenshot you can see that the timecourse of a voxel becomes completely flat after applying the bandpass filter in line 1207. There is variation prior to that but it complety becomes flat after applying the filter (the filter I applied is the classic [0.008, 0.09] basically copying the example you give at the bottom of the denoise pipeline page).

I used the GUI and curiously this does not happen with the GUI. In case it helps, while the module calls conn_process with option = 35, the GUI does regression and denoising using option = 6. I noticed that the way of regressing and filtering is quite different between both, but I am a bit out of my depth on that code.

Best,

Herberto