Hi, would you be able to help us with some pointers how to do connectome analysis of two groups with significantly different brain size?
We did seed_random_per_voxel =10 for whole brain tractography followed by tck2connectome. The NBS analysis of the streamlines number, however, detected almost all networks had significantly changed.
Could someone suggest what would be a good normalisation procedure and analysis pipeline (or a paper?) for comparing structural networks of brains with different size?
I commenced writing a manuscript about five years ago describing how this should be done; it’s literally become a self-deprecating meme at this point. But I will endeavour to get @alan-connelly to read it so that I can push it out.
You shouldn’t be performing direct comparisons of streamline counts between subjects in the absence of a method such as SIFT. I can only suggest reading the relevant manuscripts to properly appreciate why this is the case.
While there does exist what I believe to be the “best” method for properly accounting for reconstruction differences between subjects in the quantification of connection density using streamlines tractography, I still nevertheless find it hard to advocate that particular approach at this point in time, as without the relevant article describing why this is the best approach, you would likely run into issues at peer review time. The rationale behind it would not make immediate sense to anyone not well-versed with both AFD and SIFT.
What you could do instead is either:
Instead of using -seed_random_per_voxel, use e.g. -seed_image, and generate an identical number of streamlines for each subject. You could then optionally divide the matrix values by the total number of streamlines generated.
From the matrices you are obtaining via -seed_random_per_voxel, divide by the number of voxels contained in the seed mask for each subject.
These two options vary slightly in how they behave in the presence of inter-individual variability in the likelihood of successful streamline generation from each seed point, but are reasonably comparable in their interpretation: the value contained within each connectome edge is something like “the fraction of the subject’s total white matter fibre connectivity that belongs to this particular pathway”. People have done experiments like this fixing the number of streamlines for each subject, so there’s a precedent for it; but this is the precise interpretation of such.
This raises a higher-level issue. Instead of thinking about “how do I generate connectomes and perform NBS and account for differences in brain size”, I would encourage thinking more carefully about “what am I trying to quantify?”. If one subject has a significantly larger brain than another, but the overall network connection geometry is equivalent, then I would expect the fibre connection densities to be greater in the larger brain biologically, not just in a streamlines reconstruction.
If instead, your question is in fact something like “are there differences in connection density in specific pathways, over and above what would be predicted by brain volume differences alone?”: I would personally advocate including brain volume as a nuisance regressor in your analysis, rather than “scaling” the underlying connectome values by brain volume or the like. The latter destroys the physical interpretation of fibre connectivity afforded by the SIFT model, and makes the actual hypothesis test kinda funky (e.g. something like “the fibre-connectivity-divided-by-brain-volume is different between groups”). There are other statistical caveats in here that I can go into if you’re interested, but I feel like I’ve probably put more than enough on the table for you already…
There’s repeatedly been questions about how to appropriately scale connectome data / reconstructed pathway density data in order to properly facilitate quantitative comparisons between subjects.
While I’ve still not submitted my manuscript where this issue is described more completely, the physically appropriate solution is described and demonstrated in this response to a recent article; this should at least provide adequate evidence for use of this approach for any applications.