My apologies for flooding this board with questions. As you may have guessed, I’m in the middle of a few large projects using MRtrix.
To estimate the cross-sectional area for a pathway of interest, I generated 10 million streamlines (whole brain), applied SIFT2, and used tck2connectome to get the summed SIFT2 weights (multiplied by mu coefficient) for a pathway between two Freesurfer labels (R IPS and R Pars Opercularis; part of SLF III). Long story short: I became suspicious/puzzled when this metric turned out to be negatively correlated (across individuals) with a behavioral metric I expected it to be positively correlated with. I decided to check the log scaling coefficients from mtnormalise and found that they were negatively correlated with the cross-sectional area metric for the tract I was looking at (R(40) = -0.32, p=.039). Looking at the scatter plot, it looks like the normalization factor is explicitly placing an upper bound on this tract’s weighting.
What the best way to deal with this? If the intensity normalization is working correctly, it shouldn’t have any statistical relationship with the tract weight, right? If I ignore intensity normalization and perform an “old school” FA based analysis, I see the expected positive relationship, but it doesn’t seem to be driven by global factors because there is no association with the contralateral tract or other branches of the SLF. This all makes me suspicious that the normalization is producing an artificial negative correlation.
I suppose my question is if normalization is really necessary or wise in this case. All of the subjects’ data were collected on the same scanner with the same sequence – what are the expected sources of global signal differences in this case?. Is it legitimate in theory to re-run the SIFT2 analysis without normalization, show correlations of interest, and then show that these don’t reflect global scaling differences because other candidate tracts don’t show the same association (using the appropriate statistical procedures; i.e. not just showing a null effect)?
P.S. I looked more closely and the effect is entirely driven by the correlation between the summed SIFT2 weights and normalization scale. There is no relationship between the mu coefficient and the normalization scale.
@ThijsDhollander might be able to shed more light on this, but looking through the code, it looks like the normalisation scale (
lognorm_scale ?) is the geometric mean of the scaling factor within the mask, and that this is the factor by which the image has been divided. So this means that cases with low normalisation scale values correspond to images whose densities have been increased on average, and vice versa. In which case a negative relationship would be expected: low normalisation scale implies higher normalised density, and hence higher SIFT2 connectivity value – I think…
However, I wouldn’t expect this relationship to be tract-specific –
mtnormalise performs bias field correction via a 3rd order polynomial, which should not introduce local differences. At this point, it all depends on whether you find correlations in your tract of interest only, or whether the effect is more widespread? It also depends on the pathology under investigation, and whether you expect to find global differences, or purely local differences. If global differences are expected, then this will most likely show up in the normalisation scales too; for example, a global increase in MD will mean higher b=0 image intensities, which may lead to comparatively lower DW image intensity values due to the scanner’s internal intensity calibration. These would then be (correctly) up-scaled again by
mtnormalise, introducing a relationship between MD and the normalisation scale coefficient. And if there is a relationship between pathology and normalisation coefficient, and a further relationship between pathology and connectivity in your tract of interest, then that would explain your observation of a correlation between normalisation and SIFT2 weight – but it wouldn’t necessarily mean it was artefactual, far from it.
This is obviously all speculation at this stage… We’d need to drill down further to figure out the source of the relationship, and whether the result you obtained is actually correct or not. I reckon the first thing to do would be to look at whether there are global relationships between the normalisation factor, MD (averaged over the whole brain), and pathology.
Thanks, this is helpful. I’m going to have to look a bit more closely and get back when I have a clearer sense of whats going on.
While I don’t think I can contribute anything constructive regarding the
mtnormalise scaling, the presence of an inverse relationship between connection density and your behavioural measure (as opposed to the relationship being regressed out) is concerning; particularly since this is not the first time I’ve had a report of such an observation (others being entirely independent of
mtnormalise). I do worry slightly that this may be due to a particular streamlines tracking bias, which I can explain to you in more detail if you’re interested in trying to ascertain whether or not it is a contributing factor in your experiment. There are quite a number of other experiments you could potentially use here too. So I for one will be following this closely.
Hi Rob (@rsmith),
I’d be very interested to hear more details about this potential bias. We have actually collected behavioral/psychophysical performance on a number of tasks, and are hoping to find correlations between individual differences in those measures and the connection density of certain pathways. So, ideally, I’d like to be confident that the tractography is right before moving on to these statistical analyses. Any more information would be much appreciated!
Hi Rob (@rsmith),
Happy new year! Would you able to give me any more information about this potential tracking bias? Is it related to https://github.com/MRtrix3/mrtrix3/issues/1204? If so, it sounds like it will be fixed in RC3? The threads I saw were related to SSST; is the story similar for MMST?
Thanks for your help!
Apologies for the silence on this, but I really haven’t been able to sit down and think long and hard about what might be going on. The effect I was thinking of previously has the potential to produce negative correlations where one might expect positive correlations, but this would be a result of the intrinsic design of the tracking algorithm and therefore does not have a fix other than using a different tracking algorithm; and as far as I can reason, this effect should be independent of
So if you’re still stuck as to what experiments you could potentially do in order to figure out what’s happening in your data, I think we would need to take the discussion offline and dig around in greater detail.
The effect I’m thinking of is not related to Issue #1204, and is not restricted to single- or multi-tissue CSD but is intrinsic to iFOD1/2 and how they subsequently interact with SIFT / SIFT2. But it’s only a hypothesis at this point.