About the FA computation for multishell HARDI data

Dear MRtrix experts:
I’m new to the MRtrix. I have a multishell HARDI dataset (10 volumes for b=0, 68 volume for b=1000, and 68 volumes for b=2000). The preprocessing has been finished by using MRtrix. Now I want to obtain the FA map. Some researchers suggest to use only the volumes of b=1000 for the FA estimation. But I would like to obtained FA map using volumes of b=1000 and b=2000 together.
I have no idea how to work it out. Any suggestions will be appreciated.

Zhiliang Long

You should just be able to feed the entire DWI series to dwi2tensor, it should just work. The commands should do the computation, whether the input dwi.mif contains one or more shells:

dwi2tensor dwi.mif dt.mif
tensor2metric dt.mif fa.mif

(you may want to add the -mask option to one or the other command to remove the noisy background, but it’s not a hard requirement).

Just a quick question about this, since it is known that data with shells of B>1000 have non-gaussian diffusion effects, does it make sense to include those shells in the tensor computation?


Dear Claude,

Applying b-values of >2000 are generally used for kurtosis imaging, and incorporating DWI data in tensor analysis, would give you in essence more reliable estimates. According to this topic, you can calculate kurtosis based tensors with dwi2tensor as well: https://mrtrix.readthedocs.io/en/latest/reference/commands/dwi2tensor.html

However, I am not sure about the accuracy of this, as I did not test it yet myself.


Good question…

First off, ‘does it make sense’ to do it? Well, the tensor model extends over b-values, so at face value, it makes complete sense.

In practice, things aren’t so simple: we know the answer depends on the b-values included in the fit, and/or the specific tensor fitting algorithm (see Figure 1 and the associated appendix in our recent paper). So if you’re after results that are comparable with the vast majority of the literature, then you’ll probably want to stick with b=1000 s/mm² data – but note your results will likely still depend to some extent on the fitting procedure.

@ChrisvdW also makes a good point, that multi-shell allows you to fit the full kurtosis terms, even if you’re only interested in the regular tensor components. I’m not sure it removes either of the above dependencies though…