Dear FSL list,
I have been using the TBSS framework on a dataset without problems
initially. However, when I started to use tbss_sym (FSL 4.1.5) to look
at left-right asymmetries, I observed a systematic bias. In all
subjects, both healthy controls and the patient group, FA was generally
greater on the right side compared to the left with the opposite
observed with MD/axial/perpendicular diffusivity. Thus when using
randomise to compare left-right asymmetry, most of the white matter
skeleton shows a significant difference, including in areas such as the
SLF where one would generally expect FA to be greater on the left (the
opposite to what I observed).
I have performed detailed experiments with the acquisition including
looking for left/right bias in diffusion measurements on phantom data
and the pre-processing steps which seem to exclude these as the source
of the bias. I have however noticed something unusual with the
symmetric skeleton generated by tbss_sym. The MNI152 and FMRIB58
templates both have a x-dimension of 182, so I expected the midline to
be between voxels x=90 and x=91 (counting from zero). However it
appears that the midline is offset by one voxel and lies between x=89
and x=90. Thus when tbss_sym flips the white matter skeleton in order
to generate a symmetric skeleton, the skeleton generated does not lie
symmetrically on the template to which data is normalised by TBSS.
The mean_FA_skeleton appears in the correct location on the
FMRIB58_FA_1mm template (and the normalised data itself), whilst the
mean_FA_symmetrised_skeleton is offset to the left. Has anybody else
observed a similar effect and could this be a source of the bias observed?
Many thanks,
Gavin
|