Hi @mblesac , thanks for your response on this thread. I’d like to clarify something regarding “you need to multiply each connectome by their mu coefficient, obtained from tcksift2
”, since this wasn’t explicitly mentioned on this related thread: Group analysis - #3 by mblesac
Is the “mu” coefficient the one that is output by sift2 using the -out_mu
option? If so, I am guessing this is the same “mu” that is listed in equation 4 in the SIFT2 paper - but this “mu” appears in the equation for the weighting coefficients “f” (equation 5). So I guess my question is: how come using the SIFT2 weights to construct the connectome doesn’t already incorporate the effect of the scaling coefficient “mu”?
The reason I raise this point is that I have a dataset of PKU (phenylketonuria) patients vs. controls, and one unexpected result of the group comparison was that most of the significantly-different edges are stronger in the PKU patients than in controls; I had expected the opposite result given that PKU causes demylenation (so, for example, a consistent result in previous studies is that the mean diffusivity decreases while the fractional anisotropy is preserved).
As I am new to neuroimaging, I have been struggling to understand why I am getting stronger edges among the PKU cases (the edges are connected to regions that make biological sense; it’s just the direction of the change that was surprising). I used all the steps recommended in the other thread: mtnormalise, a common response function, and SIFT2 weights to contstruct the connectome. I didn’t multiply by the “mu” coefficients, but I just checked and the “mu” coefficients are higher in the PKU patients, so multiplying would have just made the edges even stronger in the PKU patients.
I found a discussion of the problem of “normalizing connection density between subjects” on the mrtrix docs that seems to have been written for SIFT (not SIFT2). They mention that “a subject may have decreased fibre density throughout the brain, but be morphologically normal; if the same number of streamlines are generated, this difference won’t be reflected in the tractogram post-SIFT”, and I thought this might have explained my result since the decrease in mean diffusivity in PKU patients seems to be throughout the brain…but then I wasn’t sure if this “should” have been accounted for by the use of mtnormalise and a common response function, because on that page, they also write:
“An alternative approach is to try to achieve normalisation of FOD amplitudes across subjects, as is done using AFD. This requires a couple of extra processing steps, namely inter-subject intensity normalisation and use of a group average response function, which are also far from error-free. But if this can be achieved, it means that a fixed density of streamlines should be used to reconstruct a given FOD amplitude between subjects, and then the cross-sectional area of fibres represented by each streamline is also identical between subjects; this can be achieved by terminating SIFT at a given value of the proportionality coefficient using the -term_mu
option.”
The -term_mu
option doesn’t seem to exist for SIFT2…but maybe this recommendation to use the -term_mu
option is related to your recommendation to multiply the connectomes by the “mu” output by SIFT2??
Another explanation I was considering was that my result was caused by the downsides of having used a common response function. My professor told me that the response function should not vary much since all the subjects were scanned with the same magnet, and he recommended averaging only “control” subjects to get the response function, saying that if there was something abnormal about the PKU cases, that could influence the response function in undesirable ways…but now I am wondering if this created a bias in the response function that causes it to be insensitive to the fact that the PKU cases have overall lower mean diffusivity? The SIFT page linked above does say “group average response function instead of the individual subject response may result in spurious peaks or incorrect relative volume fractions in the FODs, which could influence the tracking results.”
Would it be possible to explain the reason the response function might vary from one subject to another? Given my present results, I am not confident about whether I have “correct proportionality between different connection pathways within a subject” - is that something I would only be able to say if I had used a subject-specific response function?
Any help is appreciated!