Application of SIFT on brainstem acquisition

Hello,

We are doing dMRI acquisitions with a SSFP sequence on a post-mortem human brainstem at 11.7T and we want to study the connectivity between different nucleus of this region.
We are going to generate a tractogram using the brainstem as seed (we can’t do whole-brain tractography because we only have the brainstem).

Is it coherent to apply SIFT on this tractogram? I think the estimated density of each fibre population in every voxel of the image will be proportionally reconstructed by streamlines and so, SIFT can be applicable?

Thank you,
Dianna

Hi @Dianna_Le_Coz ,

SIFT in an of itself should definitely still be applicable; however, there’s 2 other concerns for this particular region, that might subsequently render your connectome less meaningful (even when SIFT is used):

  1. I suppose using anatomically constrained tractography (ACT) will be difficult, if not impossible, for this purpose? The problem will be that, even of your SIFTed tractogram, (large) amounts of tracks will be unassigned to your connectome’s nodes. So the actual tracks that will make up the connectome, will be featuring an incomplete subset of the SIFTed tracks; i.e., they themselves will not constitute a SIFTed set. This is tricky though: do you expect tracks actually ending in regions in the brainstem (rather than just passing through?). If they just pass through, then you could use ACT with a “GM” region at the top and bottom of the brainstem, and all the rest simply being white matter. If that makes sense, biologically, then this point might not be an issue after all…

  2. Tractography in general in this region, for the accuracy you require (connectivity between nuclei), might be suffering from a lot of false positives; I can image this region is a (almost the ultimate) challenge due to many, many, “kissing” tracts. In the “kissing” scenario, there’s a lot of ambiguity still remaining in the diffusion data; often too much for post-processing to reliably undo, without making very particular/severe assumptions…

My anatomical knowledge of the brainstem is slightly limited though, so it’d be interesting to hear other peoples’ opinions on this.

Hi Dianna,

A few scattered points:

  • There isn’t any problem with running SIFT on what is effectively a ‘reduced field of view’. SIFT is advertised as ‘only applicable to whole-brain fibre-tracking data’, but it would be more accurate to say ‘not applicable to targeted tracking data’. This is because you can’t have an FOD that arises from some set of fibres, but only a subset of those fibres are reconstructed in the tractogram: the model would be unable to correctly assess where the reconstruction has too much or too little streamlines density, as it has no information about what fraction of the fibres in that voxel are part of the reconstruction. But for a reduced FoV, this isn’t a problem: it can still do what it can to obtain the correct streamlines densities within the image volume.

  • SIFT works best when combined with ACT (even if it’s not explicitly a ‘connectome’ being generated). This is primarily about the derivation of the SIFT ‘processing mask’, rather than the streamlines terminations: it reduces the influence of GM voxels (where the FODs are large in amplitude, but the signal is from isotropically-oriented neurites) on the model fit. Without this, particularly for single-shell acquisitions, streamlines that project deep into GM are more likely to be retained than those terminating near the GM-WM interface (as they help to ‘explain’ the non-zero GM FODs). Whether or not you can apply ACT on your data will depend on how well you can resolve your nuclei of interest.

  • SSFP sequences introduce a reasonably strong T1 dependence on the signal amplitude, differences in which will be interpreted as differences in AFD if using spherical deconvolution. It’s been a while since I read up on this stuff so I can’t comment directly on the magnitude of the effect; but you would want to be confident that any biases introduced by this T1 dependence are outweighed by the reconstruction biases that SIFT is reducing.