Dear SPM Community,
I would very much appreciate expert input on the following questions:
[1] The scalar DTI derivatives such as the FA, ADC, axial and radial diffusivity have non-normal distributions, that from exploratory data analysis appear to be describable as ~ log-normal. The same is true for the residuals as seen in the SPM file ResMS.hdr, ResMS.img. It would be possible to transform the data by such transforms as a log or square root function or even by a Box-Cox transformation to better achieve a normal distribution prior to applying an SPM GLM analysis such as a one-way analysis of variance. Q-Q plots of LOG or SQRT transformed data still show an S shape but much more substantially overlap the y = x line. Given this and an experimental design that compares patient cases to controls, is such a data manipulation required or is an assumption that such a systematic error will cancel under a case-control design justifiable?
[2] If such a transformation is considered appropriate and justifiable despite the robustness of the GLM, then when should this be applied before or after image registration? A more minor point - is there an easily accessible error metric within SPM8 or SPM5 that indicates the co-registration 'goodness' using mutual information?
My Many Thanks in Advance
David F. Moore, MD, PhD
|