Functional ROI for targeted tracking

I would like to make use of fMRI data to define ROI that can be used for targeted tracking. What’s the best approach in your opinion? What about creating a functional group ROI which can be warped to native space (and possibly intersected with the corresponding anatomical freesurfer parcellation) and used as a seed region, perhaps in conjunction with ACT (e.g. cropping at the gmwmi)? An alternative could be to simply use the functional ROI as a reference and draw the ROI on the group FA map (and then warp…). I’m currently working with data based on 45 directions and b1000, which might set some limits…

Hi Orjan,

I’m no expert on the relative merits of different registration strategies, but I can comment on the use of ACT in this context.

Generally in this scenario I suggest running ACT with the -seed_gmwmi option. This will enforce all streamline seeds to lie within the ROI, propagating from the isocontour between WM and GM. However, the mechanism by which this seeder operates is important. Seeds are drawn at random within the ROI, and each is then optimised toward the GM-WM interface based on trilinear interpolation. Therefore, seeds that are right in the middle of the cortical ribbon, nearer the outer layers, may in fact fail to find this interface, as there will be no underlying tissue gradient, and therefore no streamline will be propagated. This means that your seed ROI needs to be large enough such that it includes the GM-WM interface of your grey matter of interest (preferably including the adjacent superficial WM as well: the seed point optimisation occurs from both directions).

Although 45 directions and b=1000 may not be ideal for tracking, that’s actually more-or-less independent of the mechanism you use to define your streamline seeds.


Thank you for the quick reply!

I could imagine that when warping GM to native space (I will do a 2 group comparison) I might get a small but systematic difference in ROI size between the groups (I haven’t checked yet) - in my case due to unqual fine motor skills. This would make no difference for what I want to represent, but could the systematic unequal ROI size in itself bias the tractography?

I guessed that the 45 directions and b1000 would have been an issue if the suggestion was to warp fod images to the group template (or extract AFD). Is there any general advice or are there “minimum requirements” on data quality below which it would be less than sensible to try such functions?

Thanks again,

If the grey matter region responsible for that function were genuinely smaller in your group of interest, and if your registration / segmentation were able to capture that accurately, then I would suspect that the white matter pathways connecting to that region would also be smaller in cross-section in that group. If this were reconstructed as such, I wouldn’t call it a ‘bias’, but an accurate reflection of the reduced connection density of that reduced-size region to the rest of the brain. That’s a lot of ‘if’s’, though.

I certainly wouldn’t be trying to fix the ROI size in subject space; that would just result in tracking of adjacent pathways in those subjects where the region of interest is genuinely smaller. Whether the region size should be used as a covariate in your statistics is a different question. This sort of thing really requires that you go all the way back to the core hypothesis that you’re testing; specify precisely what you’re trying to quantify, and then see whether any of the processes you’re applying to your data appear to invalidate the interpretation of your data for testing that hypothesis.

Lower b-value data is troublesome for these types of analyses for two reasons.

  1. Incomplete suppression of the extra-cellular water signal means that the AFD interpretation breaks down a bit; your FOD amplitudes are proportional to the raw DWI signal intensity, which will be a combination of intra-cellular and extra-cellular contributions. So although you may generate a result using e.g. a CFE analysis, interpretation of results would be more troublesome.
  2. If crossing fibres are not properly resolved due to the reduced angular contrast, then streamlines will not follow biologically accurate trajectories, which casts doubt on any results that are generated based on those trajectories.

I think we tend to avoid specifying “minimum requirements”, as then people may opt for those ‘requirements’ rather than the recommendations. Whether it’s “less than sensible” to apply such methods on non-optimal data… is probably up to individual interpretation. So I don’t think I can answer that one.