Presumably the issue with the sample data was that some volumes were negative, and therefore the mean was misleadingly close to zero?

Yes, that's exactly what the problem was. I think it would have been fine if the mean had actually been below zero, as that would have given rise to non-finite values, which I'm pretty sure would have been ignored in the subsequent mrthreshold operation.

So you're right that this 'fix' is certainly not guaranteed to fix all cases. But I'm pretty sure that if no values were positive, that wouldn't actually be a problem, by the same rationale as above (I'd need to run tests to make sure).

But realistically, this is only supposed to catch the rare cases where regridding operations in e.g. dwifslpreproc produce a lot of negative values in problematic low signal areas (particularly CSF, as was the case here), leading to unrealistically high ADC values, which is what was throwing off the automatic thresholding. I'm not sure how else we could guard against that, other than implementing some form of outlier rejection in mrthreshold...

In any case, I'll merge now, but I agree this doesn't preclude adding more robust safeguards down the track if necessary.