I’m not responsible for that addition to the docs; and I’m certainly not taking responsibility for the consequences of it. This isn’t generally correct; and relates to a certain kind of pattern and amount of “non-brain” voxels; the likes of which you wouldn’t get e.g. from the current dwi2mask. This in fact by the very definition of the core mechanism of dwi2mask.
In any case, that addition to the docs seems to be aimed at what is maybe in more nuance described here and beyond: Mtnormalise: spurious FODs outside brain mask - #13 by jdtournier . I recommend anyone to read that in great detail to appreciate the kind of voxels this aims to talk about. The problem appears to arise if you actively dilate a mask that is itself already overestimated by a good margin. So please, don’t do that in your pipelines. Due to the nature of bias field correction wrt multi-tissue signal compartments, that is naturally asking for the correction to introduce something that is not the correction of a bias field.
Alright, all that aside, @maxpietsch, what I specifically commented on in your point is more specific than that though. This might reveal that the wording in the docs itself is misleading even. My point related to you using the phrase “brain parenchyma”: i.e. only the brain tissue, excluding all fluid (maybe even up to a degree of partial voluming), i.e. ventricles and all. This would be a bad idea, because the tissue balance factor estimation relies on tissue signal to be robust of course. The most robust amount of signal for each “tissue” (including CSF) is found in the areas where there’s most of it. For CSF, that’s in the ventricles. For an adult brain, if you were to reduce the mask to a very strict definition of brain parenchyma, the balance factor for CSF will become highly sensitised to extremely small effects of any kind, anywhere in the parenchyma, of the CSF compartment (potentially spatially specific).
I don’t have the time to write this up in my former levels of detail, but maybe imagine this one (even though not quite the same, but there’s rough similarities in effects): you might just get the (similar-ish) problems some people get who ran MSMT-CSD with WM-GM-CSF on single-shell data, where in the end GM doesn’t get involved (but is also not always exactly zero, just small traces); and then go on to run mtnormalise on that. Every once in a while, you might even get the “Non-positive tissue balance factor was computed.” error. That itself would flag how bad it can get.
So well, the best you can have is areas of “very much present” summed signal across all tissues involved, and that also per tissue in some areas. To extend this to interesting scenarios I come across at times: if you’ve got e.g. a (surgical) resection of small or large parts of the brain, and the intra-cranial space gets naturally filled in with CSF… well, it’s quite a good idea to keep this in your mask, also for mtnormalise; certainly if this is to be integrated with other data in a quantitative analysis.