Dear Dr Weber,
Here are some comments from a relative novice; I hope that you will receive
some answers from the SPM experts in the near future.
One brief question about your design: are you expecting that each stimulus
will produce a burst of neural activity which will last for 3 seconds? As
you are no doubt aware, the canonical HRF within SPM is not ideal in such a
situation: a better fit would probably be obtained by convolving a time
series with a 3-second-wide pulse for each event convolved with the HRF. I
am not experienced with using the Fourier basis functions offered by SPM,
so I am not sure if they are of low enough frequency to accommodate such a
long 'event'. Obviously if you only expect a burst of activity at the
start of the event then this is not a problem.
>first, what are your experiences regarding correcting the different
>aquisition times of the slices. We have tried the method of Mark
>D'Exposito and Geoffrey Aguirre (c.f., spm mailbase, July 10 1998, Re:
>Time lag and hrf-modeling, by Darren Gitelman). Are there any better
>solutions available or being developed?
At the moment no direct solution has been implemented within SPM (other
than, as you mention, modelling the events with some basis function plus
its first derivative, or a series of Fourier basis functions, to allow
lee-way of a second or so in specifying the timings)
>
> Second, is it correct that any interpolation across slices is
>problematic (e.g. realignment, normalization) in multislice recordings
>with TR>2sec in an event related study?
If slices are acquired in one direction (ie. 1, 2, 3 ...29, 30) then the
shifting of a voxel from one slice to another slice two or three slices
away will only introduce a timing error of less than 300 ms, which is
negligible on the time-scale of the haemodynamic response, and will in any
case be absorbed in the temporal smoothing. If, however, the slices are
acquired in an interleaved fashion as you describe (1, 3, 5 ...28, 30),
then the timing error even between adjacent slices is half of your TR.
Clearly 'spatially' interpolating between voxels which were acquired about
two seconds apart is not ideal in the context of an event-related analysis.
>As a simple solution one might shift the time vector of the events to
>the acquisition time of the individual slice. Has anyone implemented
>such a solution in SPM?
My understanding is that there are theoretical problems associated with
having a slightly different design matrix for each slice (so that every
slice has the timings of events correctly specified), to do with violating
the assumptions within the Theory of Gaussian Fields. At the moment,
therefore, this approach has not been implemented within SPM. However, if
there is a particular area of brain that is of interest, one can specify
the timings optimally for the slice which contains the centre of this area.
I will assume that the whole of the TR is occupied by scanning, ie that
there is a negligible length of time between the acquisition of successive
whole-brain images.
If brain slices had been acquired from the top of the brain to the bottom,
the optimal way of specifying events so as to demonstrate a possible
activation at the top of the brain (ie. in the first slice) would have been
to specify events using (V/TR) + 1 where V is the vector specifying the
timings of events in seconds and TR is the repeat time in seconds. The
optimal way of specifying events so as to demonstrate a possible activation
at the bottom of the brain (ie. in the last slice) would have been to use
V/TR. The optimal way of specifying events in any brain area in an
intervening slice 'n', when the brain is acquired in 30 slices, would have
been to use (V/TR) + {(30-n)/30}, the small discrepancy between this
formula and that in point 1 being explained by the finite duration of each
slice.
The situation is complicated by the fact that you have interleaved slices.
Optimising the specification of timings for slice 30 would require you to
use V/TR, but optimising the specification of timings for slice 29 would
require you to use (V/TR) - 0.5, since this image is acquired 'half a TR'
earlier than slice 30.
Thus although interleaving the scans reduces the overall error in
specifying timings in any given SPM, it does make it harder (indeed, once
spatial smoothing between neighbouring slices has been used, perhaps
pointless) to try to optimise the SPM for a given slice, and it does
introduce problems with spatial interpolations and smoothing which
constitute pre-processing of the data.
Obviously if the total time of a brain acquisition is much shorter than the
TR, then the above formulae would need to be corrected appropriately.
>
>2.) in relation to question 1.) a fourier set as basis function would to
>our knowledge eliminate the need for correcting for different slice
>acquisition times (at least if no interpolation across slices is
>performed before). Unfortunately, we encountered problems in
>understanding the design matrix and setting the contrasts.
>Furthermore, how many functions would be appropriate and how long
>should the window length be?
With my relative inexperience in using more than one function to model the
events, I share your problems with setting appropriate contrasts. However,
my understanding is that people in the department who have used only the
primary basis function as the effect of interest, and treated any
derivatives as effects of no interest, have obtained plausible results.
>
>3.) we treated our event type 2 as confound using the following spm
>options:
>
> - "make inferences about some events (treat others as confounds)"
> - "events of interest": 1 3
This is correct.
All of the above may already be obvious to you. I am sorry that my level
of expertise is not sufficient to have completely answered all of your
questions. I hope that the real SPM experts write soon!
Best wishes,
Richard Perry.
from: Dr Richard Perry BM BCh MA PhD MRCP(UK),
Clinical Research Fellow, Wellcome Department of Cognitive Neurology,
Darwin Building, University College London, Gower Street, London WC1E 6BT.
Tel: 0171 504 2187; e mail: [log in to unmask]
Pager: 04325 253 566.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|