Hi Ben,
I am responding to your question about how to
choose MELODIC components to remove for
denoising. This question is a commonly recurring
one (for example, in the FSL archives). However,
I have not been able to find detailed guidelines
for how to correctly label components
representing artifacts vs. neural signals of
interest. I understand that it is generally
assumed that "experts" know how to do this
correctly, but I haven't seen reports concerning
how well experts agree with one another, or what
qualifies someone as an expert. The issue may
also be complicated by the possibility that some
components might represent a synthesis of neural
and artifactual signal sources.
Some examples of suggestions for what should be
considered noise are found in Christian Beckman's
"the little FMRI shop of horror" webpage
(http://users.fmrib.ox.ac.uk/~beckmann/homepage/academic/littleshop/toc.html)
and in McKeown et al. (1998). Recently, in Kelly
Jr., et al. (2010), our group provided (1) a
review of previously reported automated and
manual methods for ICA-based denoising, (2) a
suggestion for a manual, operationalized
procedure for denoising fMRI data though removing
ICs labelled as representing noise; and (3) a
suggestion for an approach to enable comparison
of results from various ICA-based denoising
methods, allowing us to determine which are the most effective methods.
Hope this helps,
Robert
References:
Kelly Jr. RE, Alexopoulos GS, Wang Z, Gunning FM,
Murphy CF, Morimoto SS, Kanellopoulos D, Jia Z,
Lim KO, Hoptman MJ. Visual inspection of
independent components: Defining a procedure for
artifact removal from fMRI data. J Neurosci
Methods (2010), doi:10.1016/j.jneumeth.2010.03.028.
McKeown MJ, Makeig S, Brown GG, Jung TP,
Kindermann SS, Bell AJ, Sejnowski TJ. Analysis of
fMRI data by blind separation into independent
spatial components. Hum Brain Mapp, 1998; 6: 160-88.
At 06:16 PM 4/30/2010, you wrote:
>Hi Ben,
>
> > Firstly which is the better method to use - Full perfusion modelling or
> > simple subtraction ?
>
>Generally we recommend using full perfusion modelling because it's a
>more accurate model of the whole ASL signal, and this lets you get
>both the BOLD and ASL signals as different contrasts in a single
>analysis, rather than having to do two separate analyses.
>
> > In the simple subtraction case does the phase need to be TR/2 ?
>
>The comment in the balloon help about having to advance your phase by
>TR/2 is just a minor correction (1-1.5 sec probably) and it's much
>more important for rapid event-related designs than for boxcar
>designs. Don't worry about it -- especially because you should be
>using full perfusion modelling instead.
>
> > If I need ON OFF ON OFF instead of OFF ON OFF ON would I need to set the
> > phase equal to rest period or activity period ( 29.3 secs in my experiment)
> > in all the EVs to shift the model forwards ?
>
>Yes -- the default boxcar design is to start with an OFF period, so if
>you want to start with an ON period just set the phase equal to the
>length of the OFF period. If you only have a single boxcar stimulus
>and you use the Model Setup Wizard for your perfusion model, then the
>phase is only set in one place (EV2).
>
> > How would this change if I need to include TR/2 in the simple subtraction
> > method ? Would it just be rest period + TR/2
> Â ? Â Â So say the rest period is
> > 29.3 secs , would the phase need to be set 29.3 +14.65 ~ 44 secs ?
>
>TR is not the period of your boxcar -- it is the time between volumes
>in your FMRI scan, probably 2-3 seconds. So the TR/2 adjustment is
>just a minor correction, to account for the fact that when you do
>subtraction, the new timepoint #1 is actually the difference between
>the original timepoints #1 and #2. Generally don't worry about this
>too much unless you have a very fast design.
>
> > Is there any way to view the subtracted images to see what they look like ?
>
>I think you'll see this in the filtered_func_data.nii.gz in the output
>directory. They will probably be very noisy, with lots of
>blood-vessel "activity".
>
> > I have used the Full perusion model approach and do not seem to get any
> > activation when i view the zstat1
> image(perfusion) but get activation in the
> > ztat3 image (BOLD) . Strangely, when I used a wrong TR (3 secs instead of
> > 2.93) I got a single cluster in the left motor area.  I have run  MELODIC
> > and there seems to be a lot of noise. The Simple subtraction model gives me
> > a good result in one case but random blobs in the other
>
>It shouldn't make any difference if you change the TR slightly like
>that. However, you should look at your design carefully (with View
>Design) and make sure that EV1 is alternating in every scan. If you
>use TR=3 in one place and TR=2.93 in another, then your perfusion
>signal will flip signs about every 20 timepoints! That would
>definitely ruin your ASL results.
>
> > I was wondering how I should choose the components to remove when using the
> > MELODIC denoising option. I have around 35 components and in most of them
> > the activations seem like artifacts except for one or two where the
> > activations are in the left motor cortex(with a lot fo noise). Is there a
> > 'rule' for selecting components to discard
> and  retain ? Is there a limit to
> > the number of components that can be discraded without causing any
> > additional problems in the results ?
>
>I'm not sure about using Melodic on ASL data. I imagine it will pick
>out the tag-control contrast very strongly, and if you remove part of
>this it will cause you problems in the GLM analysis. It takes a
>trained eye to know what's an artifact and what is signal -- don't
>remove too much!
>
>Cheers,
>Adrian
|