A bit more explanation: smoothing only spreads signal out, diluting it with non-signal or unrelated signal, it can't concentrate it "towards" some other relevant signal in another subject. For any case that has non-negligible overlap of signals to begin with, smoothing basically just dilutes the overlap with other nearby things. Thus, if you have one fixed group ROI to begin with, which nearly all subjects have substantial signal overlap with, smoothing is only going to hurt you (per-subject ROIs can do even better, but generating them well is harder). Our preprint paper that Matt linked to explores this in terms of how much signal from other cortical areas or other tissues gets mixed in by smoothing (and by volume-based cortical analysis), but this particular idea of "increasing overlap" could also be tested by itself without taking crosstalk into account, and would likely show that until initial overlap gets rather small, common amounts of smoothing only make the overlap worse.
By far, the biggest of the useful effects of smoothing is attenuating noise (and for some things, there are other ways to accomplish this goal - averaging data across a sizable ROI will be far more effective at reducing spatial noise, so if that is part of your analysis anyway, you don't get this benefit from adding smoothing, but you do get the detriments of the smoothing). If your noise is largely independent per-voxel, then most of this attenuation occurs at relatively small smoothing amounts anyway. Also, be careful about regressing out WM and CSF, if your masks don't leave sufficient space from cortex, then it can pick up what is basically the mean gray matter signal, and cause you to accidentally do GSR. Tim On Wed, May 30, 2018 at 6:58 AM, Glasser, Matthew <[email protected]> wrote: > Indeed that is one of the major fallacies in brain imaging. Have a look > at this paper in press at PNAS: > > https://www.biorxiv.org/content/early/2018/04/23/255620 > > Peace, > > Matt. > > From: David Hofmann <[email protected]> > Date: Wednesday, May 30, 2018 at 5:14 AM > To: Timothy Coalson <[email protected]>, Matt Glasser <[email protected]> > Cc: hcp-users <[email protected]> > Subject: Re: [HCP-Users] Additional smoothing of FIX extended resting > state data? > > Hi all, > > thanks for the comments! The idea to smooth the data was based on others > papers which did not use HCP data though. I always thought that smoothing > is a "good idea" for group studies in order to account for the > between-subject variability in the ROI based on the different brain sizes > and shapes. > > The analysis I want to run uses the data from ROIs to calculate > connectivity between ROIs (DCM). The ROI extraction in SPM uses the > component that explains the most variance in a PCA. The extraction runs on > the smoothed volumes. The ROIs are based on some probabilistic atlas (e.g. > anatomy toolbox). I have about 300 subjects. > > I thought the results will be relatively robust for different levels of > smoothing. But this is not the case. Since the CSF and WM signals have > been regressed out, I did not assume that this will have influence. > > The unsmoothed results look much better though. > > greetings > > David > > 2018-05-30 0:51 GMT+02:00 Timothy Coalson <[email protected]>: > >> Volumetric smoothing in particular is not advised, as it causes signal >> from one bank of a sulcus to bleed into the opposite bank. Analyses that >> average all signal within an ROI should have no benefits (and will have >> detriments) from smoothing, as the within-ROI averaging itself is a form of >> smoothing (but with a well-chosen ROI, it won't bring in signal from the >> opposite sulcal bank, etc, in theory). Is this the kind of ROI analysis >> you are doing? If so, I wouldn't trust the differences caused by adding >> volume-based smoothing, because they could be from nearby areas instead. >> >> Tim >> >> >> On Tue, May 29, 2018 at 4:24 PM, David Hofmann <[email protected]> >> wrote: >> >>> Dear all, >>> >>> as far as I read in previous posts on the list, spatial smoothing of the >>> volumetric resting state data is not recommended. But this was with >>> regard to group ICA. >>> Would you also recommend not to apply any further spatial smoothing for >>> the (volumetric) resting state data, when running ROI-based (group) >>> analysis on multiple subjects? >>> >>> Comparing the results of smoothed (4,6, FWHM) and unsmoothed data gives >>> highly different results in my case, so I'm a bit confused now. >>> >>> greetings >>> >>> David >>> >>> _______________________________________________ >>> HCP-Users mailing list >>> [email protected] >>> http://lists.humanconnectome.org/mailman/listinfo/hcp-users >>> >> >> > _______________________________________________ HCP-Users mailing list [email protected] http://lists.humanconnectome.org/mailman/listinfo/hcp-users
