Dear Joe,
thank you for a most timely and clear response! This helps me to better
understand how sparse sampling should work in FEAT :). Then, in the
simple case of 1 EPI / 1 stimulus and the Stimulus->EPI latency kept
constant at 4 seconds, I should just use
Slice time correction = None, FILM prewhitening = OFF, Convolution = None, Add temporal derivative = OFF, Paradigm file latencies [unchanged, no need to adjust]
I still do not quite understand why, using the above FEAT settings, I
get different results if I use slice time correction: this should not
make any difference if Convolution = None?
***
However we already have data where we vary the Stimulus-> EPI latency by
jittering the stimulus presentation time (TR is kept constant at 10
seconds to preserve equilibrium), and thus we collect more time samples
of the HDR curve than just one. Then I believe that the slice timing
correction becomes essential. Let's assume that we have three time
samples at 3.2/4.4/5.6 seconds (from stimulus onset to EPI onset). Could
you tell me what would be the answers to the same questions I posted in
the previous mail in this situation?
***
Also we have data with TR=2.5 seconds (EPI full volume acquisition
duration = 1.2 seconds again). This does not allow enough time for BOLD
response to acoustical scanner noise to return to baseline between
stimuli. Currently we present a single stimulus between each full volume
acquisition, but keep the Stimulus-> EPI latency fixed; this is thus an
event-related design that requires deconvolution of the HDR waveforms
(that we sample once every 2.5 seconds), and slice timing is essential.
In this case, what would be the answers to the questions about FEAT
slice timing correction / stimulus paradigm file timing correction setup?
Thank you very much again,
Tommi
Joe Devlin wrote:
> Dear Tommi,
>
> What you describe is a classic "sparse sampling" design in that it meets
> all three criteria:
>
> 1) the stimuli are presented in silence,
> 2) the BOLD response to the stimulus is sampled at, or near, its peak in
> auditory areas, and
> 3) you allow sufficient time for the BOLD response to the scanner
> noise to
> return to baseline levels.
>
> Points 2 and 3 distinguish sparse sampling from interleaved or clustered
> sampling. If you are interested in auditory cortex responses
> uncontaminated by the scanner noise, you need to use sparse sampling.
> The
> important thing about analysing these data is to note that you are not
> acquiring BOLD waveforms -- that would require far more frequent
> sampling. instead you are acquiring categorical data (like PET
> data). In
> other words, each data point represents a condition being ON or OFF -- it
> is not a continuous sampling. So all of the time course issues in
> fMRI go
> out the window. High pass filtering may still be useful but prewhitening
> should be turned off (see point 3).
>
> Because you aren't collecting time series data in the true sense, the TR
> becomes arbitrary and you certainly don't want to use convolution at
> all. Slice timing should also be avoided because it relies on
> interpolating temporal responses. Since you didn't collect a
> waveform, you
> can't interpolate on it. For similar reasons, you can't use a temporal
> derivative either.
>
> So given these constraints, which were imposed by the data collection
> technique used, the TR is meaningless since it is used to handle the fact
> that data in typical fMRI experiments are acquired continuously, but each
> voxel is acquired discretely. The only function it plays in the sparse
> sampling analysis is to specify the event timings. So you might as well
> set the TR=1 and then just number your events sequentially using 3-column
> format (or use 1-column format if your happy putting all those zeros in).
>
> The many steps which you described in your email would be necessary if
> you
> were using a shorter TR which didn't take point 3 into account (so-called
> clustered or inter-leaved acquisition). Then you would be sampling a
> continuous waveform in long discrete intervals and you'd have to be very
> careful with respect to "jittering" the stimulus onset and the TR
> acquisition in order to get an unbiased and sensitive measure. In sparse
> sampling, jittering is unnecessary because you presumably did preliminary
> testing in order to guarantee that point 2 was met -- namely that the
> response in your ROI (auditory cortex in this case) was maximal approx
> 4-5s
> after onset, since that's when you sample the region.
>
> I think Steve is planning to add some of these details to our FAQ
> soon, if
> he hasn't already.
> Hope this helps.
> Joe
>
>
> Joseph Devlin, Ph. D.
> FMRIB, Dept. of Clinical Neurology
> University of Oxford
> John Radcliffe Hospital
> Headley Way, Headington
> Oxford OX3 9DU, U.K.
> Phone: +44 (0)1865 222 494
> Fax: +44 (0)1865 222 717
> Email: [log in to unmask]
|