Hi Matt Thanks for the input, quite interesting stuff. You hear a lot of argument about doing this or that all the time, and you are right there is no consensus yet. I have not used ICA-FIX so I cannot comment about that, but I will check it out now that you mention it does not need other type of motion correction afterwards. I guess the take home message is that in any case, you need to do more than Realingment and in Monty's case, ICA-FIX that may have been enough. Thanks again Eduardo On Fri, Mar 13, 2015 at 11:03 AM, Matt Glasser <[log in to unmask]> wrote: > Here is my attempt at a rational discussion of these issues based on > similar ones and ongoing work inside the HCP: > > I would say that exactly how best to deal with motion confounds between > two groups remains an open research question and I would be concerned > about any reviewers that make prescriptive statements about these issues > (at the same time, the Power work and that of other groups has shown this > is an absolutely huge problem in neuroimaging, not just in fMRI but also > in diffusion, or even morphemetric structural analyses). In particular > for fMRI, what techniques are needed when using different quality data > (e.g. ³standard² neuroimaging acquisitions with low spatial and temporal > resolution vs ³HCP-style² neuroimaging acquisitions with high spatial > resolution) are not yet clear. A lot of this denoising work has been done > on standard neuroimaging data, but how much applies to HCP-style imaging > data has not yet been proven. > > Particularly with HCP-syle imaging data, one won¹t be going wrong by > performing ICA+FIX on the data (particularly when one¹s runs are long). > The reason is the performance of FIX at classifying components into signal > and noise is very high in this scenario. ICA+FIX, when performing well, > is very good at specifically removing spatially localized artifacts of > many sources (and it includes 24-parameter movement regression) without > removing signal of interest. Thus, things like regressing out white > matter or CSF or variations on these themes aren¹t needed when using > ICA+FIX (if ventricle or white matter signals are strong, they come out as > artifactual components). ICA+FIX removes many spatially localized > nonlinear effects of motion, including interactions with b0 variations in > the magnetic field. > > Notably, ICA+FIX does not specifically clean artifactual parts of the > global signal (the 24-parameter movement regression portion does clean the > global signal, but other nonlinear effects of motion or other global > artifacts‹e.g. CO2 fluctuations‹are not cleaned). At the same time, we > have seen that the global signal variance after ICA+FIX clean up localizes > to primary sensory brain areas in a strikingly spatially specific way. > What that means is that most brain areas are positively correlated with > primary sensory areas, but may be anti-correlated with each other (e.g. > the default mode and the anti-default mode networks). This aspect of the > global signal is probably not artifactual. > > One strategy for dealing with global artifacts is to use partial > correlation as implemented in FSLNets. Another strategy would be to > regress out the global signal (though, as noted above, this will also > remove apparently neurobiologically valid whole brain to primary sensory > areas correlations and may cause apparent spatial shifts in the borders > between areas when one of the areas' time courses is particularly > correlated to the global signal). > > Other lossy strategies such as band pass filtering, scrubbing, etc. have > not yet been shown to be needed after ICA+FIX on high temporal and spatial > resolution fMRI data (i.e. HCP-style data). Nor have partial correlation > and global signal regression been compared and contrasted in this kind of > data. Ideally, what we would want to show is that even after ICA+FIX and > when using either partial correlation or global signal regression, band > pass filtering and scrubbing still substantially reduce between group > motion differences without also reducing differences between an > uncorrelated third variable of interest (e.g. gender, if gender is > uncorrelated with movement). What we wouldn¹t want to do, is just > continue to use these techniques ³just because people have previously used > them,² because these techniques are lossy (treating signal and noise the > same), we need to check to see if they are still helpful/don¹t make any > differences/are actually harmful in reducing between motion group > differences without also reducing non-movement correlated third variable > differences of interest. > > How to properly clean you data has certainly been one of those issues that > generates strong opinions in different investigators, who often then talk > past each other. There has been at temptation of reviewers in one or the > other of the political ³camps² on the various issues to say ³you must do > this² or ³you must not do this,² when some of the various approaches to > clean up probably do similar things (though some may be more efficient > than others in reducing artifacts while not reducing signals of interest). > I¹ve also by no means covered all of the other acquisition and analysis > techniques out there that people have developed to try to address this > issue, just ones that I¹m more familiar with. > > Peace, > > Matt. > > P.S. The above is based on my interpretation of many internal HCP > discussions with investigators in a variety of the ³camps² together my > looking at HCP data in relation to some of these issues. Because of the > outstanding work yet to be done, the HCP cannot offer a consensus position > on these issues (though hopefully consensus will be achievable after the > above analyses have been done, at least for HCP-style data). > > > On 3/13/15, 10:57 AM, "Monty Waite" <[log in to unmask]> wrote: > > >Is that really the case Eduardo?? Can I see what the collective thoughts > >of others are on this as would mean redoing a whole load of analysis..... > > > >(Ps thanks Michael for earlier comments) >