Christian,
Thanks for your reply. So my understanding is, I take both sets of one-way
t-maps from randomise (tstat1 and tstat2 for an upaired two group ttest),
convert them to z-maps. Then run melodic--like shown below--to test group
differences. The reason I am doing this, is to try and match what has
already been published in the Filippini paper. I suspect the result of this
will be similar to, but slightly less conservative than the TFCE output of
randomise?
We would also like to run some correlations on the dual regression output
with behavioral variables, can this all be done with mixture modeling as
well, or best to just stick with randomise? Basically, can any design.mat
and design.con be used with GGMM? Thanks!
melodic -i 4d_zmaps_from_randomise_tstat1_or_tstat2 --ICs=(dummy option)
--mix=(dummy mixing matrix) --Sdes design.mat (same .mat from randomise
run) --Scon design.con (same .con from randomise run)
Chris Bell
University of Minnesota
On Sep 13 2010, Christian F. Beckmann wrote:
>Hi
>
>> I have performed a dual regression analysis including running the final
>> randomise step. I would now like to threshold the maps using the
>> gaussian/gamma mixture model
>> as was done in the Filippini paper. Is there a way to implement this
>> using FSL tools? My best guess is that I need to convert the tstat
>> images output from randomise to zstat images, and then somehow implement
>> melodic using a design.mat and a design.con file to show areas of
>> significant group differences controlling for the local FDR?
>
> Yes, see the second type of usage you get from just typing 'melodic' at
> the command line. You will need to create a fake mixing matrix, just use
> matlab to write out a N+1 X N ASCII matrix of random values.
>
>
>> Alternatively, I could use TFCE from the standard output of randomise.
>> Or another recent paper (Napadow, "Intrinsic Brain Connectivity in
>> fibromyalgia is associated with chronic pain intensity") used a mixed
>> effects model implemented in FLAME. I appreciate any advice, on the best
>> way to handle the dual regression output. Thanks!
>
>What's wrong with using the p-images from randomise?
>
>>
>> One more question. In the context of group ICA, the gaussian/gamma
>> mixture model delineates between significantly activated and
>> non-activated voxels. Is there even a way to use the GGMM to test group
>> differences, or am I misinterpreting the paper?
>>
>
> If the stats image feeding into the GGMM reflects group differences then
> that should work fine.
>hth
>Christian
>
>
>> Filippini et. al.
>> "(ii) using these time-course matrices in a linear model fit (temporal
>> regression) against the associated fMRI data set to estimate
>> subject-specific spatial maps. Finally, the different component maps are
>> collected across subjects into single 4D files (1 per original ICA map,
>> with the fourth dimension being subject identification) and tested
>> voxel-wise for statistically significant differences between groups
>> using nonparametric permutation testing (5,000 permutations) (55). This
>> results in spatial maps characterizing the between-subject/group
>> differences.
>>
>> These maps were thresholded using an alternative hypothesis test based
>> on fitting a Gaussian/gamma mixture model to the distribution of voxel
>> intensities within spatial maps (see ref. 31 for further details) and
>> controlling the local false-discovery rate at P < 0.05."
>>
>> Chris Bell
>> University of Minnesota
>
|