hi,
>are these options equivalent between the two programs
most certainly not, although there may be a good correspondence in some situations. since melodic is data-driven whereas feat is model-driven, the two analyses methods are not equivalent.
for thresholding, melodic's alternat(iv)e hypothesis tests fit gaussian/gamma mixture models to intensity value histograms of the extracted independent component maps divided by the standard deviation of the residual noise. this should be more sensitive but less specific than GLM / RFT-based maps at cluster-corr. p=0.01. however, you may want to try spatial mixture modelling on your FEAT statistic image.
>also, is it appropriate to run MELODIC from the resulting filtered_func_data volume created by FEAT
yes. however, you don't need to impose a temporal filter twice on the data.
best regards-
andreas
-----Ursprüngliche Nachricht-----
Von: Bettyann Chodkowski [mailto:[log in to unmask]]
Gesendet: So 28.11.2004 23:28
An: [log in to unmask]
Cc:
Betreff: Re: [FSL] GLM (film_gls) and avwmaths -mas
i forgot to mention that i am running FSL v3.1.
> The "zeroing" of the zstat image isn't happening
> at the point with avwmaths that you suggest; it is happening
> in the cluster (thresholding) command slightly later,
oh, yes, i see that now. i did not expect the volume thresh_zstat1
to be both input and output to program cluster. it is good to know
that avwmaths -mas is not doing anything unexpected.
> it's just because this person isn't activating sufficiently
> strongly according to your design.
i see. thank you. in fact, if i change the p-threshold from
0.01 to 0.1 i start to see some activation.
> However, MELODIC does give plenty of visual activation,
yes, thank you for pointing this out -- an example of how ICA can
be of help.
i see the default threshold value for mixture model inference is 0.5,
ie, --mmthresh="0.5", whereas the p-threshold value for the cluster
program is -p 0.01. are these options equivalent between the two
programs? in fact, if i run the cluster program with
-p 0.5 --connectivity=26, i see activation.
also, is it appropriate to run MELODIC from the resulting
filtered_func_data volume created by FEAT while simultaneously
turning off the [Pre-stats] analysis? i turned off slice timing
correction, motion correction, BET, and spatial smoothing as i
did this in the FEAT analysis. i did leave on temporal filtering,
however, even though this was done in FEAT, too. i am a bit
inconsistent here so perhaps i should turn off temporal filtering,
too.
> so it may just be that your model isn't a great one?
the timing file is based on the ePrime file that presents the
stimuli. looking at the responses of this subject, she responds
correctly 60% of the time so she is participating. stimuli are
presented for 1500ms and her mean response time is 880ms
(stdev 172ms). i computed the correlation between the ePrime
paradigm timing and each of the components found by MELODIC.
the correlations are very poor, [-0.1, 0.1]!!, which agrees with
no activation via GLM using the same ePrime paradigm timing file.
thank you for all the help and reality-checks,
- bettyann
_____________________________________
BettyAnn Chodkowski F.M. Kirby Center
e: [log in to unmask] Kennedy Krieger Institute
v: 443.923.9524 707 N. Broadway
f: 443.923.9279 Baltimore, MD, USA 21205
|