Hi Alfonso,

Hope all is well! I analyzed my resting state fMRI scans using the default pre-processing pipeline (output in MNI space) and, at the end, saved the denoised dswau* file. I noticed that compared to the swau* file, dswau* seems to be skull extracted and in a few of my subjects there was a very small number of brain matter voxels that had consequently been removed, particularly in the occipital regions. A few questions that I had were:

1) Since I'm performing some analyses where I'm looking into the BOLD signal along the whole cortical ribbon (including the BOLD values of those removed voxels), I was wondering if there was a way to save the dswau* file but with the skull included (for instance by potentially commenting out the code that applies the skull removal)? Or is the skull removal an essential step for the calculation of the BOLD signal and the creation of the dswau* file?

2) If (1) is possible, should I just re-run the denoising step (which should overwrite everything?), or should I run the pipeline from the beginning?

Thanks in advance for all the help!

Best,
Panos

Just wanted to re-circulate this. Thanks again!

Best,
Panos

Hi Panos,

That is likely the effect of masking with the "analysis mask" (a general brainmask defined in Setup.Options), rather than explicitly a skull removal step. I would perhaps suggest to change the analysis mask option to either no-masking, or some larger / more liberal explicit mask? (this would require to re-run Setup and Denoising steps)

If you prefer to create an alternative denoised dataset just for 'out-of-CONN' analyses, but without changing or having to redo or affect any of your current CONN analyses, you may also do that as:

1) open your CONN project, and manually create a new secondary dataset (in 'Setup.functionals', select 'primary dataset', and then select in the same menu 'new'; if you want, select then the 'label' option in the same menu and give this new dataset a name/label -e.g. "external" without the quotes)

2) run the command:

conn_batch(...
   'Setup.preprocessing.sets', 'external',...                          % change this to the label or number of your new secondary dataset
   'Setup.preprocessing.steps', {'functional_regression', 'functional_bandpass'}, ...
   'Setup.preprocessing.reg_names', {'realignment','scrubbing','White Matter','CSF'}, ...
   'Setup.preprocessing.reg_dimensions',[inf, inf, 5, 5], ...   % change this to modify the options for this denoised dataset
   'Setup.preprocessing.reg_deriv', [1, 0, 0, 0], ...               % change this to modify the options for this denoised dataset
   'Setup.preprocessing.bp_filter', [0.008 inf])                     % change this to modify the options for this denoised dataset

That will create a new set of functional files (named bdswau*.nii) containing the denoised and band-passed functional data. Since these are stored as a secondary dataset they will not affect any of your existing analyses (also since this step is run as a preprocessing step it will not be masked at all). Last, you may also modify the denoising options above if you feel that the 'out-of-CONN' analyses would benefit from different denoising settings.

Hope this helps
Alfonso
Originally posted by Panos Fotiadis:

Hi Alfonso,

Thank you, that is extremely helpful!
I just ran the "no masking" option for right now and noticed that the value at each voxel between the non-masked denoised volume is approximately 3x the value of the equivalent voxel in the masked denoised volume. Is that because each value (at each voxel) represents the percent BOLD signal difference between the raw BOLD value of that voxel and the mean BOLD value of the whole brain (and the mean BOLD signal is now different in each case)? 

Thanks again,
Panos

Hi Panos,

Yes, you are right, that sounds like a product of the percent signal change (PSC) scaling of the BOLD signal, since the estimated global BOLD signal will be considerably smaller when using the 'none' masking option compared to explicit- or implicit- masking (in this case I would probably recommend skipping PSC-scaling entirely).

Best
Alfonso

Originally posted by Panos Fotiadis:

Hi Alfonso, 

Sounds good, that makes sense! A couple last questions:
1.   Since the average global BOLD signal in the denoised output is set to be 100 (at least based on my understanding from what I've read - might be wrong!), would a PSC value of 300 at a given voxel mean that this particular voxel displays a 3% higher BOLD signal than the overall brain?
2.   When using raw BOLD values for the denoised output (instead of PSC), what would the units be?

Thanks again for all your help,
Panos

Hi Panos,

A value of 300 would mean 3 times the average BOLD signal (but since the average BOLD signal has been computed over a region that includes a lot of out-of-brain voxels, it is not all that meaningful).  Typically SPM defines the global BOLD signal through two steps: 1) compute a "preliminary" global BOLD signal by averaging across the entire image and find those voxels with average BOLD signal above 1/8th of this preliminary global BOLD signal value; 2) re-compute the global BOLD signal now averaging only over the voxels found in step (1). This procedure is the one that CONN will use for PSC-scaling if you select the "implicit" masking option (and the implicit mask will be defined as those voxels with average BOLD signal above 80% of the global BOLD signal). When you select the "explicit" masking option in CONN the global BOLD signal will instead be computed as the average BOLD signal across all voxels within the explicit mask (this is similar to the procedure used by FSL for PSC-scaling). The differences between these two procedures is often noticeable, and it creates quite a few complications when attempting to compare effect-sizes across studies. In any way, back to our case here, using PSC-scaling in CONN together with the "none" masking option would use the global BOLD value obtained in step (1) of the SPM procedure described above (simply the average BOLD signal across the entire image), which will typically result in a considerably understimated global BOLD signal value compared to the other two methods (in addition to making the procedure very sensible to other factors such as variations in FOV, etc.), so that is why I would probably recommend to use either implicit masking or no PSC-scaling in this case.  

And regarding (2), if skipping PSC-scaling, the BOLD signal of the denoised data will simply be in the same units as the original functional data (typically raw BOLD signal values are expressed in arbitrary units)

Hope this helps
Alfonso
Originally posted by Panos Fotiadis:

Hi Alfonso,

That is extremely helpful, once again. Thank you!! (also my apologies, I meant to say 3x not 3% in my previous message...).

Best,
Panos