Hi Christian,
Thanks for your explanations.
So, If I got it right, it is better to use a complete modell for
filtering with fsl_regfilt to avoid an overfitting of the noise. On the
other side FSL offers the unconfound tool, which filters the entire
(noise) design matrix, and thus would give different results.
Dave Flitney allready had a look into my data and pointed out, that
indeed there was a fluctuation of the voxel timeseries, however, this
was below 1% of the mean signal and therefore not visible by eye.
The strange thing is, that using fsl_regfilt somehow increases the mean
signal intensity, but keeps or reduced the amplitude of the
fluctuations. Doing the same filtering with unconfound, using only the
noise part of the design matrix, gives reasonable results. So it seems
that somehow fsl_regfilt does not like my complete design matrix and
does something nasty to my data :-) .
Maybe Dave Flitney still has got the submitted files, otherwise I can
again send them.
Thanks for all the support,
wolf
Christian F. Beckmann wrote:
> Hi
>
>> Hello all,
>>
>> I've got two (and a half) questions about fsl_regfilt.
>>
>> 1) When regressing out part of a design matrix, does it matter
>> whether this design matrix consists only of all possible noise time
>> courses, or whether it also includes the events of interest? To be
>> more specific, I want to filter motion related and created a noise
>> only design matrix that I used for filtering. However, in my study,
>> motion is correlated with the events of interest. When including the
>> noise regressors in the design for a feat analysis the influence of
>> the noise gets nicely reduced, giving more pronounced activation
>> patches. However, I am concerned that my noise only model would
>> reduce effects, when always full model will be fitted with
>> fsl_regfilt. So, does fsl_regfilt fit only the selected regressors,
>> or will the full modell be fitted befor the selected regressors are
>> filtered out?
>
> No, what happens is this:
>
> -you specify a full design matrix (X, say)and using -f) the set of
> regressors you would want to be removed. This effectively partitions
> the matrix X into a signal and a noise part. fsl_regfilt then projects
> the data onto all of X to generate parameter estimate maps S. Using
> the information provided with the -f option these maps are then also
> sub-divided into noise maps and signal maps. The filtered data is
>
> data - X_{noise} * S_{noise}
>
> i.e. the new data is obtained by regressing the full model and then
> removing the noise terms.
>
>
>> 2) In the FAQ melodic is suggested for filtering noise related ICs,
>> but the melodic webpage links to a page, where fsl_regfilt is used.
>> Are this two options of doing exactly the same thing, or is there a
>> difference? If they are different, what is the recommended method?
>>
>
> fsl_regfilt exists because melodic ended up being loaded with too much
> different functionality - the filtering will at some point be entirely
> removed from melodic and fsl_regfilt will be the only tool provided
> for this purpose.
>
>
>>
>> .5) After filtering the data, There is no motion visible when using
>> the movie loop with fslview. I am pretty impressed by this effect,
>> since I have data dominated by artefacts. Is fsl_regfilt really that
>> effective or is this just a bug in fslview causing only one volume
>> being displayed?
>>
>
> not sure what's going on - maybe it is that good ;) If you upload the
> data we can have a look.
> cheers
> christian
>
>>
>> Thanks for your help,
>> wolf
|