Dear all (HJJ?),

: Yes, here I am.

I have been going through the ANATICOR approach (with some modifications) quite painlessly thanks to the clear explanations in the Neuroimage paper and the excellent notes on the AFNI site (modifications: I am not using FreeSurfer masks but tissue probability maps from the SPM segmentation routines; also, with a 3mm isotropic EPI, I used a 30mm radius for the computation of the WMeLocal regressor). I have some doubts, though, about how to proceed with regards to the smoothing and the subsequent group analysis.

As I understood it, the suggested/favored approach is to blur within the gray matter (GM) mask only (3dBlurInMask) in native space, and then Talairach the GM-blurred dataset with NN interpolation. This seems to imply that we are restricting our functional connectivity analysis to the GM only.

Now, how are we to combine such GM-restricted data across subjects? The local coverage of the GM mask is quite idiosyncratic even after Talairach normalization, so I am assuming that the common approach of using an intersection mask for the group analysis won't quite work for GM-restricted data. Should we use the *union* of all the Talairach'ed GM masks for the group analysis?

: You can make a probabilistic map with overlapping ratio of GM masks across subjects in the standard space, and then should threshold it by visual inspection as common FMRI studies have been doing. You can map GM masks' EPI time series using 3dcalc (residual time series * GM mask), transform to the standard space, and smooth them with isotropic gaussian smoothing kernel in 3dmerge. Just get the trade-off point between those (localization accuracy vs. correspondence existence ratio) for your data set.

The latter point is further complicated (at least in my head!) by the following consideration. In using the tissue probability maps from SPM (GM, WM, and CSF=ventricles+pial fluid; more on the CSF vs LV mask later...) -- after thresholding them at p=0.5, resampling them to the EPI resolution, and eroding them (only the WM and CSF mask) -- I have intersected them with the EPI mask (from 3dAutomask). Why? Because I think that the nuisance signal to be regressed out from the EPI time series should not be contaminated by sources outside the extent of "good" EPI signal (that is, 3dAutomask-"good"). Is this assumption sensible in your opinion? Likewise, would you do the 3dBlurInMask within the original GM mask (the one with T1-like spatial extent) or within the GM mask intersected with the EPI automask? I would favor the latter but would appreciate your opinion on this.

: Sulcal CSF contains lots of GM-involved signals and it is not a nuisance signal in ANATICOR but the signal of interest. Very small portion of GM signal can spoil the overall characteristics of a certain ROI's signal when you average them, and it can cause the correlation bias because the variances of GM-related signals are much larger than the others. I don't know whether a signal is good or bad. To confirm it, there should be something additional to be a gold standard, but it's extremely hard to estimate. About smoothing, as I wrote above, both have each reason in the two different points of view - localization accuracy versus statistical power. If your sample size is not so large, you should consider a way to secure better statistical powers, which is getting EPI time series in GM tissues only, transforming them to the standard space, and smoothing them with isotropic gaussian kernel - of course, statistical tests should be restricted in the thresholded GM map. That's my one suggestion. In the other case (having large number of subjects), you can just use the 3dBlurInMask in individual native spaces, and then Talairach-ing.

One last point. I am using a CSF mask that contains both ventricles and fluid close to the pial surface, instead of a FreeSurfer LV mask. Since, as discussed in the ANATICOR paper, the pial portion of the CSF mask may be still contaminated (even after erosion) by GM signal because of partial volume effects, I have eroded the mask by one voxel. Now, I can see at least two possible strategies of using this CSFe mask:

a) (semi-)manually remove from the CSFe mask anything but the lateral ventricles, and then extract the average EPI time course within this mask; this corresponds to the LVe regressor of the standard ANATICOR approach

b) use the entire CSFe (eroded) mask, but instead of extracting its average from the EPI data, extract a set of principal components of the EPI signal within that mask (with 3dmaskSVD) and use that multi-column nuisance regressor in 3dFitter instead of the single-column LVe regressor. A choice of 5 principal component was recently shown to work quite well for removing nuisance signal for functional connectivity in a paper by Chai et. al (Chai, Castañón, Ongür, Whitfield-Gabrieli, "Anticorrelations in resting state networks without global signal regression". Neuroimage, in press). I wonder what is your take on this alternative approach.

: "a" looks quite simple and certain. A manual work is little painful, I think. You can also use 3dclust to get the large ventricles masks with less efforts from the CSF map since large ventricles or lateral ventricles are the largest CSF structures in CSF map.

- HJJ