DKI Tractography

[Ehsan Misaghi]

Hello everyone,

I'm analyzing a DKI dataset using ExploreDTI and I have a few questions regarding that:

1) Could you point me to a paper/document that states the types of equations and technical steps ExploreDTI uses to analyze the DKI datasets and do DKI tractography later on and how is that different from the ones it uses to process the DTI datasets and DTI tractography?

2) I know that in the initial step of importing data into ExploreDTI, I tell the program that my dataset is a DKI dataset (using NaN as the b-value). So, does the program take that into account when it is actually doing tractography on DKI datasets? More specifically, does ExploreDTI calculate kurtosis stuff based on the 3D dODFs or based on the 2D glyphs? and can it actually delineate the crossing fibers, e.g. tracts like the corticobulbar tract, if I input the DKI dataset?

3) The protocol we acquire our images with is organized such that it takes 10 b0 images at first and then alternates between 30 directions of b=1000 and b=3000. So, the 4D NIfTI file would contain 70 images. We then organize the file this way: average the 10 b0 images and make it one b0 and then change the organization of the other b-values such that it becomes 30 b1s and 30 b2s, basically leaving us with 61 images (and we change the b-matrix accordingly). So, my question is: Are there any differences between using any of these two image organizations?

Thanks a bunch,

Ehsan

[Sjoerd Vos]

Hi Ehsan,

1) That depends also on what tensor fitting option you've chosen. The most recent fitting options are iWLLS (iterative weighted) and REKINDLE (outlier rejection). These work very similarly for DTI and DKI, with the only difference being the signal equation used in the fitting: DTI uses the second order tensor with 6 independent components (PJ Basser et al., 1994); DKI expands this by also fitting a fourth order tensor with 15 independent components (JH Jensen, MRM 2005). For both DTI and DKI, the first eigenvector from the second order tensor is used for tractography, as this describes the principal diffusion direction.

2) For both DTI and DKI the second order tensor is fit, and the first eigenvector is used. Also fitting the kurtosis tensor does affect the diffusion tensor, see Tax et al., ISMRM 2012, p3629. This is not working well with crossing fibres. Although DKI-derived dODFs have been proposed to resolve crossing fibres, the performance of these w.r.t. other crossing fibre models such as CSD haven't been evaluated thoroughly (to my knowledge). Anyway, ExploreDTI does not calculate the DKI-ODFs.

3) I would suggest not reordering the data at all. The averaging of b=0-images does not add anything and might only cause uncertainty if one or more are corrupted in a minor or major way.

Although I realise the following may be difficult if your datasets are already acquired, but I would strongly suggest to scan the multiple b=0-images not all at the beginning but spread out through your acquisition. My most recent paper that's just accepted in MRM demonstrates the presence of signal drift in dMRI data as a decrease in signal intensity as the acquisition progresses. This can be corrected after acquisition when multiple b=0-images are acquired spaced throughout the acquisition. It's a continuation of the ISMRM abstract I presented at the 2014 meeting (Vos et al., ISMRM 2014, p4460), and I'll post the link to the paper here as soon as it's online.

HTH,

Sjoerd

[Ehsan Misaghi]

Hi Sjoerd,

Thank you so much for your answer. I was trying to look into other programs and compare them, that's why it took so long to follow-up with you. So, I actually have new questions now!

1) Bottom line, I have DKI data that I wanna do tractography on and we want to be able to extract tracts that may be crossing (e.g. CBT). Would you still suggest using ExploreDTI? If yes, what would be the best fitting method? and how can I choose that in ExploreDTI?

2) I saw you wrote about CSD. Do you think I should still use the DTI tractography method or the CSD one?

3) In general, what would be your suggestion as a typical pipeline for analyzing DKI data and doing tractography on them?

Thanks a lot,

Ehsan

[Sjoerd Vos]

Hi Ehsan,

1) You can definitely still use ExploreDTI to do crossing fibre tractography. My preferred method is Constrained Spherical Deconvolution (CSD; Tournier et al., NeuroImage 2007), which is also in ExploreDTI. This is currently only the single-shell method, so you'd have to extract the DWIs that you want to use for this plus the b=0-images into a separate mat-file. This is probably easiest to do by selecting only a few of the DWI volumes from the nifti (one of the plugins can help you there) and creating a new mat-file. CSD works best with higher b-values (e.g., 2000 or 3000) so if you've got high-angular sampling at such a b-value I suggest using that shell.

2) I'd suggest CSD, see 1)

3) If sticking to ExploreDTI: I'd suggest getting the quantitative values from DKI (or the WMTI model; Fieremans et al., NeuroImage 2011) through the 'export stuff' plugins. Any tract-based things you want I'd derive from CSD. This can be combined to get tract-based quantitative metrics as well.

HTH,

Cheers,
Sjoerd