At 15:21 21/01/2005 +0000, Randall Benson wrote:
>Dear SPMers,
>
>I am new at doing efMRI.  I am doing an efMRI study using synthetic speech
>syllables as auditory stimuli.  The phonetic cues are sufficiently subtle
>that I need to interleave the stimuli and image acquisition.  I should
>mention that the subjects are aphasic patients with speech perception
>problems so we don't want acoustic masking by the scanner noise.  There
>are two kinds of stimuli along with silent trials.  So 3 stimulus
>conditions randomly but deterministically presented including: 1) paired
>syllables, 2) paired notes, 3) silence.  Stimuli last two seconds for the
>pair.  Stimuli are triggered by the scanner. I ran a subject using this
>setup:
>
>2 seconds to acquire 27 slices
>2 second of "gap" or delay (to separate the HRFs of the epi and stimulus)
>2 seconds of stimulus
>2 seconds of gap (so as to catch the peak of the HRF to the stimulus)
>
>Task is to decide same or different and press corresponding keypress.
>Accuracy to be used as a factor for the syllables.
>
>Questions are:
>
>1. Do you think this is the optimal (most efficient) design given the
>interleaving constraint?  My concern is that I am not doing a good job of
>sampling the HRF with a single phase.  This may not be important, however.
>I'm also concerned that it's not efficient, given that we are sampling at
>only 1/8 Hz.  The quickest we could sample would be 1/4 Hz with our
>constraint of interleaving stimulus and imaging.

The delay between stimulus onset and the middle of the scan is ~5 seconds.
This is reasonably optimal if you assume that (1) the canonical HRF applies
in the brain areas that you are interested in and (2) you are interested in
a response to stimulus onset. However, since you are only acquiring a
single scan for each event, you will lose power if either of these
assumptions are false.

We have used a more rapid interleaved design (TA=1.1secs, TR =2.5secs) in
auditory studies with shorter stimuli (<1sec duration). By collecting scans
~1, 3.5, 6.0, 8.5 seconds after each stimulus we can acquire several data
points during the expected HRF to each stimulus. This should afford greater
sensitivity.

The equivalent scheme for 2sec stimuli and 2sec TA would be a 4sec TR and
scans ~3, 7, 11 seconds after stimulus onset. This collects more data, but:
(a) misses the predicted peak of the HRF to your stimuli (around 5 secs
post-stimulus), and (b) leaves you more susceptible to influences of
scanner noise (since the BOLD response to the scanner noise will peak
during the subsequent scan).

I would therefore not recommend this method unless you can use shorter
stimuli, and/or a shorter acquisition.

>2. How best to statistically analyze the data.  The goal is to identify
>regions which not only activate for for the syllables than the notes but
>to look at correct and error trials to find the region which is associated
>with correct performance.  If I am sparsely imaging the HRF but am doing
>so at the peak does this mean a t-test is as good as anything?  I don't
>know what the F-test is in the context of the efMRI analysis.  Is it just
>looking at differences in variance or is it a converted wilkes-lambda or
>some such regression analysis?

If you stick with your original design (8sec TR) then all you need to do is
a t-test to compare those scans following each of your different conditions.

The F-test is used in SPM either where you have multiple basis functions
for each scan (e.g. when you are using temporal/dispersion derivatives) or
if you are interested in differences between conditions where you are not
assuming which condition is more active than the other. Since neither of
these cases seem to apply to your design, I would suggest that you stick
with t-contrasts

>3. Is it true that for efMRI the soa files are expressed as V/TR + 1
>because it expects t=0 to be (V/TR) to be =1?

I'm not sure I understand your question. However, event-times in SPM2 can
be specified in seconds or scans. If specified in scans, then the first
scan in your run will be scan 0. In specifying your events, you can either
use a canonical HRF and specify event times as being at an appropriate
position between your scans (e.g. at -0.625, 0.375, 1.375, etc.), or you
can use a single-bin FIR model (equivalent to a simple box-car) and specify
events as being immediately before the corresponding scan (0,1,2 etc). For
your long TR design, I would tend to favour the latter approach. For a
short TR design we have used the canonical HRF with some success.

hope this is helpful,

matt


****************************************************
  Dr Matt Davis
  MRC Cognition and Brain Sciences Unit
  15 Chaucer Road, Cambridge, CB2 2EF

  email: [log in to unmask]
    tel: 01223 273 637  (direct line)
    tel: 01223 355 294  (reception)
    fax: 01223 359 062

****************************************************