Definitely look at the paper Christian suggested. In addition, here's my two
cents.
On Tuesday, March 27, 2012 06:05:39 you wrote:
> Dear MELODIC users,
> I am new to resting state analysis and would very much appreciate some
> pointers. We are exploring the effect of a new intervention on resting
> state functional connectivity.We will start by exploring the resting state
> motor network. We have 20 resting scans (1.5T, TR=2420ms, TE=40ms, ~
> 8min) from 10 patients; each patient was scanned twice, once before an
> intervention ("Before") and once after ("After"). I want to do a group ICA
> of my Before scans and a group ICA of my After scans, identify the motor
> RSNs, and compare them. From what I have read – the best way of doing this
> seems to be simply concatenating all my before scans together, and all my
> after scans together and then ‘pressing play’ on MELODIC.
This is adequate for exploratory analysis. Later on, you will probably want to
concatenate ALL scans together and then use MELODIC + dual regression to
compare coactivation between groups.
> My questions;
> Which preprocessing steps are recommended before Melodic analysis; realign,
> normalise to MNI, smooth?
Yes, you should do all of these. Note that the MELODIC GUI can do all of them
for you. Some people also do frequency filtering or removal of nuisance
regressors, but I personally prefer to deal with those after ICA. Also, note
that concatenation is NOT a step you should do before running ICA. MELODIC
will take each subject or run and do this for you.
> How many components should I split my signal
> into?
You can let MELODIC decide for you. If you want to manually choose a number of
components, 50 is a good guess for a large group with 4x4x4 mm voxel size at 3
Tesla. There are some papers by Christian Beckmann, Vince Schmithorst, and
Vince Calhoun that deal with this.
> - the big paper I read from the FSL group use a 20 component ICA.
This also sounds reasonable. Was this before or after they rejected
artifactual components, though?
> What is the best method for differentiating between noisy signals and the
> low frequency fluctuations I am looking for (specifically within the motor
> network for the moment. We may also want to explore other RSNs in the
> future if we are successful)
Look at the literature to see what others have done. I often look at the
spatial extent and spectral characteristics of the data. If there is a task
involved, you can also check for correlation with the task time course.
> What is the best method for comparing the
> functional connectivity before and after an intervention.
Look up dual regression. FSL has a script to do it.
> We have not
> collected control data, I have seen other papers use libraries of data
> available online to compare their results to. Is this a legitimate
> comparison given the differences in acquistion parameter between studies
> and centres etc.? Many thanks
It depends what your hypothesis is. Are you interested in the effect of an
intervention on a patient group? If so, you may want healthy controls who have
undergone the same intervention as a baseline comparison. Otherwise, you can
probably get away with using data acquired with different parameters if you are
careful. Look at Biswal's "Toward discovery science of human brain function"
for an example of this.
|