Hello fellow CONNers,

I was hoping to get some feedback on possible analysis and/or visualization avenues for a particularly interesting pre/post case.  Ideally, I would love to show post-intervention changes in static and dynamic connectivity using CONN, but I'm not certain what would be possible given the inability to use a SPM model .mat file.  Any thoughts?

Regards,
Jeff

Dear Jeff,

For such single case studies, I have some scripts that can be helpful for static analyses. If you run CONN up to the 1st level analyses, you will get seed-to-voxel results for each seed in your setup. You can then use the following script to perform a t-test contrast with FDR correction:

https://github.com/lrq3000/csg_mri_pipel...

With this, you can for example test post - pre ([1, -1]), to highlight the differences in the subject's connectivity post operatively for the seeds you select.

After generating these results, you can visualize them using CONN surface or volumetric renderers as you prefer, by using this script:

https://github.com/lrq3000/csg_mri_pipel...

This will allow you to select any nifti map of your choice to visualize using the CONN visualizing GUI we all love.

About dynamic connectivity analyses, I too rarely use it for the moment to provide any guidance unfortunately.

Hope this helps,
Best regards,
Stephen Karl Larroque
GIGA-Consciousness, University of Liège & FRS-FNRS

Thank you so much, Stephen.  These scripts should certainly come in handy!  

I may circle back with questions if I run into trouble, but that said the scripts are commented beautifully and shouldn't be an issue.

Warm regards,
Jeff

Stephen,

When you were setting up your single-subject pre/post in CONN did you set it as one subject with two sessions or two subjects with one session each?

Thanks,
Jeff

> Dear Jeff,
>
> I set it up as one subject with two sessions, as this feels for me the
> most natural as well as ensuring that both sessions are totally comparable
> statistically (same denoising etc).
>
> Hope this helps,
> Best regards,
> Stephen

Thanks again, Stephen.  The script worked flawlessly, but I forgot to address the most important question.  What exactly is being analyzed to generate the p-corrected source images?  For instance, are the [-1 1] condition difference x source results generated through parametric or nonparametric analyses?

Warm regards,
Jeff

Dear Jeff,

The script is using BETA maps and calculates the thresholds and contrasts from these maps. The significance thresholding is using a parametric voxel-wise FDR approach as supplied by CONN (which slightly differs from how SPM computes FDR values), as there is unfortunately no other way to calculate either cluster-wise thresholding nor non-parametric thresholding at the 1st-level, as both are technically ill-defined at the moment.

Here is an excerpt:
«Thresholding techniques for single subject fMRI are more complicated than for multi subject fMRI, as the fMRI time series contain auto correlation (Woolrich et al., 2001). To be able to perform a permutation test on single subject fMRI data, the auto correlations have to be removed prior to the resampling (Locascio et al., 1997; Bullmore et al., 2001; Friman and Westin, 2005), in order to not violate the exchangeability criterion. Single subject fMRI is further complicated by the fact that the spatial smoothing changes the auto correlation structure of the data. This problem is more obvious for CCA based fMRI analysis, where several filters are applied to the fMRI volumes (Friman et al., 2003). The only solution to always have null data with the same properties, is to perform the spatial smoothing in each permutation, which significantly increases the processing time. This problem was recently solved, by doing random permutation tests on the GPU (Eklund et al., 2011a, 2012).»
From: Anders Eklund, Mats Andersson, Camilla Josephson, Magnus Johannesson and Hans Knutsson, Does Parametric fMRI Analysis with SPM Yield Valid Results? - An Empirical Study of 1484 Rest Datasets, 2012, NeuroImage.

However, Adolf et al devised a simple algorithmic scheme to calculate non-parametric permutations on 1st-level fMRI, accounting for the temporal autocorrelation. The strategy is to do a blockwise permutation of temporal volumes, in other words simply split the resting-state scans and then permute to derive the null distribution. This strategy is implemented in the SPM toolbox StabMultip (and maybe also in FSL PALM?). For more info, see: Adolf, D., Weston, S., Baecke, S., Luchtmann, M., Bernarding, J., & Kropf, S. (2014). Increasing the reliability of data analysis of functional magnetic resonance imaging by applying a new blockwise permutation method. Frontiers in neuroinformatics, 8.

However, I never tried this approach, and I am not sure how it would work in practice. To try this approach, you would need to enable saving the denoised bold timeseries in CONN, to then use them as the input for StabMultip. If you try this approach, I would be very interested in hearing how it went :-)

Hope this helps,
Best regards,
Stephen Karl Larroque
GIGA-Consciousness, University of Liège, F.R.S.-F.N.R.S.

Thanks, Stephen.  

Your response anticipated my next question, which would be the possibility of some sort of cluster threshold.  As you have already figured out, the results from the script allow for even the tiniest clusters.  I wonder if there's a way to just have the results thresholded for an arbitrary cluster (k) size?  While not empirically pleasing, one could reasonably set the cluster extent to a level that would preclude most of the likely noise effects.

Also, did you or your group figure out how to run single-subject ROI-to-ROI, rather than the seed-to-voxel approach?  Seems like this too might aid in diminishing the effect of the very tiny cluster effects.

Warm regards,
Jeff

Dear Jeff,

For manual cluster extent threshold, you can try to use one of these softwares (though I never used them): https://www.nitrc.org/projects/cluster_c... (Rubes_cluster_correct.m) or https://www.nitrc.org/projects/peak_nii or FIVE by Aaron Schultz (http://mrtools.mgh.harvard.edu/index.php...) or https://www.nitrc.org/projects/cluster_report/ (an extension of cluster_correct).

Be warned however that Eklund et al disadvise such approach as the extent thus set is arbitrary: as long as a cluster passes the significance threshold bar, it should be presented.

However, I backtrack on what I said earlier: Scott D. Slotnick, who is also the author of the paper "Cluster Success", a reply to Eklund's "Cluster Failure" paper, developed an alternative approach to corrected cluster correction which claims that it can estimate smoothness directly without needing an SPM.mat file, so it might work for 1st-level too if you set the right parameters (the degrees of freedom being the number of BOLD volumes). You can find the scripts and more information here: https://www2.bc.edu/sd-slotnick/scripts....

Hope this helps,
Best regards,
Stephen

Thanks again, Stephen!  

Warm regards,
Jeff

Dear Stephen,

Thanks a lot for sharing the scripts.

I'm a bit puzzles though about how to interpret the outcomes. If I understood correctly the scripts compare the BETA maps per source, between two different conditions (here pre and post treatment). If outputs the pvalues maps, the FDR pvalue corrected maps and thresholds the BETA difference maps based on the FDR corr pvalue (and even splits these into positive and negative difference maps).

Let me update you on my situation here.

I'm investigating a single case study by exploring a patient's functional connectivity patterns before and after treatment (i have two pre and two post treatment resting state scans).

I ran part of the preprocessing outside conn (yet using SPM tools) as I wanted to include denoising using ICA-AROMA, and prewhitening of the data (removing the autocorrelation between the scans). I then denoised the data further using the regression steps (for motion and white/CSF matter) in CONN and then ran the first level analysis (correlation).

Now I implemented your script for comparing (for example) the BETAs of pre 1 and pre 2. Inspection of the pFDR_thresholded_corr_* maps (doesn't matter which source) shows significant differences basically everywhere in the brain (see also Figure 1 attached). Indeed, when inspecting BETA differences maps thresholded based on these pvalues, I see significant differences in BETA patched all over the brain (see also Figure 2). I'm very new to this, but to me this seems very very unlikely, or??

Yet, if these maps are indeed correct, how should I interpret them? It hard to imagine that for each source I find differences between the pre1 and pre2 condition? Can you advice on some strategy to adopt to get from these patched difference maps to something more inclusive?

In case these maps are incorrect, any idea what could have gone wrong???

Thanks a lot in advance!

You are observing the same thing that I saw in my data, and I believe the issue is the lack of any cluster-based threshold.  I believe the pFDR is voxel-wise only.  

To address this, I found Simon Eckhoff's SPM ANATOMY toolbox helpful.  With this toolbox, one can select the pFDR thresholded "neg" or "pos" beta difference maps and then apply a cluster extent.  Albeit this cluster extent is arbitrary and limited to voxel count alone, you can at least make a large cluster threshold to minimize Type II error (e.g., I went with voxel pFDR<0.001 and k > 200 in my data).  Note, if you use the Anatomy toolbox, you will need to apply a "-1" multiplicative transform to visualize the "neg" beta difference map blobs.

Hope this helps,
Jeff

BTW - I'm curious.  How or what did you use to pre-whiten your data?

Thx,
Jeff

Hello,

First thank you very much Stephan for these helpful scripts.

I created my first level maps but I would like to have some clarification please.

The final results can be found in the 'Beta_thresholded' and 'Beta_thresholdedpos/neg', right? These maps correspond to the functional connectivity differences between the 2 selected conditions for a particular seed and subject at a voxel defining threshold of p-unc < 0.001 and cluster at p-FDR < 0.05, right ?

If I visualize these maps using conn_3dVis.m, why I cannot optain the colorbar representing the functional connectivity strength when using the volume visualization mode ? it asks me for a brain activation surface map and when I enter the same thresholded map, it seems that they do not overlap perfectly (please see attached), why?

If I want to visualize these maps using the SPM Anatomy toolbox like suggested by Jeff, I can only open the 'pos' and 'neg' maps, one at a time ? I let then the 'hight threshold' at 0 and put no value in the 'extent threshold' value ? 

Thank you very much in advance for your answers,
Regards,
Nabila

Dear Stephen,

Thank you kindly for the very useful visualization scripts.
I not so sure how to interpret this visualization (attached), which is a comparison between pre/post condition in a single subject.
Also, is it possible to add more information to the visualization such as labels/ROIs and values? If so, how?
Thank you.

Sincerely,
Batiah
Originally posted by Stephen L.: