Multisite Data

Hi all,

Is there any documentation on using multisite diffusion data with MRtrix3?


Hey Claude,

Depends in part on whether or not “multi-site” is a Trojan horse for some other difference :upside_down_face:

If you have identical scanner hardware, and an identical sequence protocol, then ideally inter-site differences would be indistinguishable from intra-site differences. The smallest differences that one may envisage creeping in would be in e.g. B0 shimming, B1- and B1+ fields, ADC gain, that sort of thing; but there are explicit processing steps already in place to deal with all of these, even if imperfectly. If you were doing a quantitative analysis you would likely include site of acquisition as a nuisance regressor, and you’d be pretty safe.

Where things get more complex is if it’s not just the site that’s varying, but also the acquisition protocol. E.g. If you have different b-values between sites, that’s clearly going to be problematic. But rather than writing an essay trying to enumerate and disentangle all possible differences here, I might defer and wait for more specific site differences that you may be concerned about.


Hi Rob,

I fear that you are correct regarding the Trojan horse here :slight_smile: The scanner manufacturer is the same but the models are different. To the extent possible the sequences are kept the same (eg b-values should be the same but the TE would be different).

I could elaborate but perhaps it will be too much to dump on the public forum.


Hi @rsmith

as an added question to this, if you were doing tractography on two different scanners and trying to compare … would you go for an RF per scanner or a single RF for the study?

I would advise using a per-scanner RF. And to take into account the different sites in your subsequent statistical analysis as much as possible.

1 Like

For the sake of tractography only, it probably doesn’t matter a great deal, the data variance from the tracking process is going to outweigh any subtle differences in FOD amplitude or shape. Per-scanner RF is safer, though I would personally be concerned about the quality of the hardware if DWI data of the same nominal b-value differs noticeably in shape between scanner models. It’d be worth generating the per-scanner RFs and then simply comparing them subjectively.

Note also that if you were to quantify Fibre Bundle Capacity (FBC) based on e.g. SIFT2, there’s a normalisation term in there that disappears with the use of a common response function but may need to be more carefully considered if different RFs are used between sites. But as previously, effect would probably be small relative to the data variance, plus if you were to use site of scan as a nuisance regressor the effect would disappear in the GLM anyway.

1 Like