Hi Jack,

Let me first state plainly that the concern refers to doing a voxelwise comparison, among subjects, of tractography results. This is not at all a critique of tractography (an excellent method), nor of its models (that are necessary).

Tractography is a bit different than other modalities in that it returns, at each voxel, a quantity that refers almost exclusively to the model, as opposed to a quantity (model-based or not) that could more directly represent or be traced back to the underlying biology of the tissue within the voxel, and that could be used for group comparisons that would tell something about how that biological aspect of the voxel varies across subjects.

In addition, the reconstructed tract needs to be seen as a whole: a voxel may have a morphological and functional content that is identical among subjects, yet in some of them this particular voxel may not be included in a tract because the propagation stopped or was rerouted elsewhere. A voxelwise comparison would see none of this.

You can definitely use tractography results to derive masks and compute summary quantities for the tracts such as FA and other parameters. You can make qualitative or semi-quantitative analyses that preserve the global interpretation of the tract. You can also do connectivity analyses as have been done in various amazing papers. These would all be fine.

About TBSS: it ensures that the voxelwise comparisons done inside the bulk of the WM take the areas of highest local FA in each subject, that tend to be at the centres of tracts, thus addressing issues with local misalignments that were creating problems in earlier VBM-style analyses. These are exactly the sort of registration problems I alluded earlier. If there were a TBSS-style tractography, it would be possible to project FA, MD, and other quantities (model-based or not) to these TBSS-style, tractography-derived tracts, and then do a voxelwise analysis of these quantities. I'm unaware of such a method though, and it's clearly not easy to develop a reliable, meaningful one.

All the best,

Anderson

On 4 September 2015 at 00:22, Jack Grinband <[log in to unmask]> wrote:

Hi Anderson,
I'm not sure I'm convinced by your argument. All data analyses make some model assumptions and we do statistical tests given those assumptions. In this particular case, as long as the assumptions of the statistical test are independent of the assumptions of the propagation model used to identify the tracts, there should not be any problem. It is certainly reasonable for you to question the underlying assumptions of the tractography, but I don't see how that invalidates the statistics.

Moreover, it sounds to me that you are arguing that it's not possible to make group comparisons of tractography data and that randomise cannot be used for tractography statistics. This seems contrary to the recommendations in the TBSS user guide (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/TBSS/UserGuide). Perhaps Steve has an opinion on this?

jack

On Wed, 2 Sep 2015 09:19:14 +0100, Anderson M. Winkler <[log in to unmask]> wrote:

>Hi Zachary,
>
>Exactly, I am suggesting that this voxelwise comparison isn't meaningful.
>First, there is the issue of the tracts not being in perfect register, but
>even if we put aside the registration aspect, what would an eventual
>voxelwise significant result mean? It would depend on the model used for
>the propagation, but then the comparison wouldn't be about underlying
>biology, but about the model used, and what the values in each voxel
>represent for that model.
>
>Please note that I am not in any way suggesting that tractography wouldn't
>be good or appropriate. What I am saying is that a comparison across
>subjects isn't quite trivial as running a voxelwise GLM (be it parametric
>or non-parametric, as in randomise).
>
>All the best,
>
>Anderson
>