Hi Jon,
In terms of what the software does, you can refer to @bjeurissen’s answer. As to what’s the right thing to do, well… You will find a variety of opinions on the topic!
In general, most experts will agree that the diffusion tensor is only ‘valid’ in the limit of low b. By ‘valid’ here, we mean that higher order terms become negligible, and the tensor model fits the signal well. So most experts will strongly advise against using b-values much larger than ~1,000 s/mm² in the tensor fit.
The other argument is that the overwhelming majority of the literature provides values derived from regular single-shell b=1,000s/mm² data. If you need your tensor metrics to be comparable to literature values, you’d be well advised to use the b=1,000 s/mm² shell only.
In practice, the impact of including higher b-values in the fit depends strongly on the fitting strategy used. Many implementations (including previous versions of MRtrix) use a straight least-squares fit to the log-signal – sometimes referred to as an ordinary least-squares (OLS) fit. But that tends to dramatically inflate the noise contributions from the low signals (as would typically be observed in the high b-value shells), and that can lead to instabilities and biased estimates.
More modern fitting strategies employ a weighted least-squares strategy – in our case, we use an iteratively re-weighted least-squares strategy. These will appropriately down-weight the low-signal contributions to ensure the noise in those values does not affect the estimated metrics more than would be expected. I suspect the values derived using these fitting strategies should be much more stable and closer to the single-shell values – but bear in mind I don’t have data to back that up, I’ve not looked into the problem in any great detail. And either way, I think the first two points above should already settle the issue…
Hope that helps!
Donald