Dear FSL list,
I have worked with fMRI data for a number of
years, and I'm now also starting to analyze DTI data. I have a number of
questions that I haven't been able to find the answers to.
1) In
the function dtifit in FSL, a general linear model is setup (Y = XB +
e, just like in fMRI analysis) to estimate the intercept and the 6
tensor components. The "raw" measurements are not used, but instead the
logarithm of the measurements are used in the vector Y. I would prefer
to setup a generalized linear model, instead of a general linear model,
and perform the regression with a logarithmic link function. However,
when doing so it seems like there is a problem with colinearity between
the three of the tensor regressors and the regressor for the intercept.
The reason for this seems to be that the diffusion gradients always have
unit length, and therefore sum to 1 (thereby being correlated with the
intercept regressor). I cannot find any paper that discusses this
problem, is it because everyone uses the general linear model with the
logarithm of the measurements? It will of course work, but I suspect
that the tensor estimates will have a high variance, especially if the
number of measurements collected with a b-value of 0 is very low.
One
solution could be to separately estimate the intercept, and then
proceed with a generalized linear model with 6 regressors instead of 7.
2)
Is there any work done on using linear regression methods without first
taking the logarithm of the measurement values? I'm currently not
interested in non-linear methods, only linear.
3) Is there any
work done on actually using the variance of the tensor components / FA
values for each subject? In fMRI it is standard to calculate t-scores,
where the variance is needed. In the FSL function FLAME, for example,
the variance of each subject can then be used in the group analysis of
brain activity, to for example down weight subjects with a high
variance. When doing group inference using FA estimates, is there a
similar FSL function that uses the variance of the FA estimate for each
subject?
4) In fMRI it is common to include the estimated head
motion parameters in the design matrix, to further suppress effects of
head motion. Is this also common to include the head motion parameters
when estimating the diffusion tensor?
Regards,
Linda