>1) my functionnal image voxels are 3,75 x 3,75x 6 mm size. In which
>cases and with what kernel should
>I realign (= apply the realignment matrix to) my images? So far,
>recommendations I got from different users and
>methodologists varied between:
>- do not realign at all if mouvement along timeseries was small, i. e.
>did not exceed 2mm translation and/or 1 degree
>rotation, because realignment may bring undesirable secondary effects.
>If mouvement was larger, use 5mm kernel.
I don't now what "undesirable secondary effects" they allude to. Certainly
due to imperfect interpolation and other error sources realigned data are
never identical to those one would have obtained had the subject remained
absolutely motionless. However, as far as I can see realignment always
With regards to the kernel, I assume you mean the size of the kernel used
for the Sinc interpolation (b.t.w. its unit is voxels, not mm). In
principle there should be a trade off between speed and fidelity, where
larger kernels is slower and give higher accuracy. I tend to use 11x11x11
kernels, but I susect that is overdoing it.
>- realign if mouvement was large; if it was small (see above), realign
>only between the first images of each series
>- spm realignment proceeders are particularily effective if movement was
>small (!!!). So apply them systematically
>precisely in that case. For larger mouvement, use different realignment
>pre-treatment (which one then?).
This would have been the case for pre SPM99 realignment where a small angle
approximation was utilised and a fixed number of iterations were used
rather than a stringent converegence criterion. With SPM99 I certainly
would not have any concern using it even for large movements.
>Can anyone help me to find my way within all these contradictions?
>2) time slicing is to be applied to functionnal images before spatial
>realignment. Naturally , as the
>signal in each voxel has been modified by the time slicing interpolation
>proceedures, we expect the realignment matrix to be different too.
>However, I applied "realign correster" to my data before and after time
>slicing, and the two realignment curves looked the same.
>What is the exact impact of time slicing on realignement matrix?
I think the jury may still be out on the order of realignment and slice
timing. The correct thing would be to use a 4D interpolation kernel, which
is simply inpractical. Therefore the question of the order arises. It
should be noted that the "ordered" approach is always wrong, but one order
may give larger errors than the other.
It would seem to me that interpolation should be done first in the
direction in which the gradients are largest, i.e. the spatial direction.
The issue changes slightly for interleaved acquisition where adjacent
voxels voxels (in the z-direction) will have been acquired 0.5TR apart and
hence where temporal interpolation may be quite seriously affected by
However, the issue of temporal/spatial interpolation will affect mainly the
accuracy of the final interpolated data, and less so the estimation of
movement parameters. I would be surprised if you could find any difference
between these by simply eyeballing your parameter curves (which you can't).
>Thanks in advance
Good luck Jesper
Wellcome Dept. of Cognitive Neurology
12 Queen Square
London WC1N 3BG
phone: 44 171 833 7484
fax: 44 171 813 1420