Hi Jack
>
> 1. For cluster based thresholding, is the thresholding applied to
> resel-corrected data or
> uncorrected data? It wasn't clear to me from the manual if cluster-
> and resel-based corrections
> are mutually exclusive.
>
These are different techniques, though the spatial smoothness enters
into the equations in both cases.
> 2. For estimating number of resels, does feat estimate smoothness only
> from the size of the
> spatial filter or does it also try to estimate the smoothness in the
> pre-filtered data?
>
The smoothness is estimated from the 4D residuals (using 'smoothest')
after regressing out the model from filtered_func_data... you certainly
don't want to estimate smoothness from the pre-filtered data as this
will not reflect the smoothness introduced by the convolution with a
Gaussian
> 3. I have consistently come across the following situation: I do a
> third level analysis (Z
> thresh=2.3, clust P thresh=0.01). There is no activation after
> thresholding. When I look at the
> zstat map and threshold at Z=2.3, I get large blobs (for instance, I'm
> looking at one such blob that
> is approximately 5cm x 2cm x 3cm). So, obviously the clustering, at P
> thresh=0.01, is eliminating
> large areas of activation.
>
> But the thing that I find really puzzling is the unthresholded zstat
> map itself. If I set the color look
> up table to render3, I get very meaningful patterns of "activation".
> For example, I get no activation
> to an auditory stimulus when I use the default thresholding, but when
> I look at the render3 zstat
> map, I get the whole superior temporal gyrus turning bright red. My
> point is that by looking at the
> raw zstats, one can easily identify the auditory cortex despite having
> no significant activation after
> thresholding. Any comments on this scenario? Will histogram mixture
> modeling eliminate these
> "false negatives"?
> many thanks,
Oh yes, there's a lot of fun to be had when thresholding stats images
and you're opening a hugh can of worms here - formally, the above
statement of 'clustering... eliminating large areas of activation' is
incorrect - all this stuff is about null-hypothesis testing and
strictly speaking there is no such thing as 'activation' within this
framework - just areas where you're happy to accept the null-hypothesis
and areas where you're not. What your test is telling you is that even
when you (based on your prior knowledge) think that there is a
'significant area of activation' the stats is not significant enough to
fail the null hypothesis test... and there are many potential reasons
why this might happen, the increased smoothness at higher levels (due
to spatial registration) is one possible candidate.
It's very difficult to predict what's going to happen when you change
thresholding parameters or even approches. The next release version
will provide an additional method using spatial mixture model, for now,
you can 'fool' melodic into running the non-spatial Gaussian/Gamma
mixture model on your data, just save all maps to threshold in a single
4d file, create a fake mixing matrix (number of columns = number of
maps to threshold, number of rows > number of columns) and use
melodic -i example_func --ICs=zmaps --mix=tmp_mix -o blah.ica -v
--report --Ostats
melodic will produce probability maps In the blah.ica/stats folder
which you can use to threshold at any level - you can inspect the
goodness of fit by checking the histogram plots in the report folder
hope this makes sense
cheers
christian
>
> jack
>
--
Christian F. Beckmann
Oxford University Centre for Functional
Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
Email: [log in to unmask] - http://www.fmrib.ox.ac.uk/~beckmann/
Phone: +44(0)1865 222782 Fax: +44(0)1865 222717
|