Dear Anderson and Paulo,

Sorry to jump in but I have a similar problem and want to make sure that I understand what you mean.

1. By "adding", did you mean that we use "fslmaths --add" ?

2. The ICs used in the "adding" is raw IC (split from melodic_IC by fslsplit) or thresholded IC (thresh_zstat) ?

3. After "adding", should I merge the result IC (by adding above) with other IC (split from melodic_IC) to perform the dual_regression or I just need to perform dual_regression on the result IC since I only care about it.

Best regards,

Khoi 

On Sun, Jun 19, 2016 at 2:11 PM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Paulo,

You are probably better off by simply adding up the ICs that are split apart before doing the dual regression.

About the mixture modelling: yes, that can be used after having summed the ICs (in fact, there will then be no need for thresholding, and thus, no need for mixture modelling).

All the best,

Anderson

On 18 June 2016 at 14:21, Paulo Branco <[log in to unmask]> wrote:

Dear FSL members

We're trying to use resting-state fMRI for single-subject mapping. We're discussing clean ways to solve the problem of IC splitting. For one particular network (sensory-motor) we tend to consistently get 2-3 ICs when using the built-in dimensionality estimation from FSL. We can solve this by lowering dimensionality or simply adding the ICs together, but this is a bit too subjective. We're considering doing a "dual_regression" without group maps:

In short, we run ICA, we perform a template-matching procedure, we select the best-fit candidate (so, one of the "pieces" of the network), we extract its timecourse and run fsl_glm on the filtered_func data, with the timecourse as design and the --demean option. This is similar to a SBCA analysis, but the ROI is selected based on the initial ICA. From visual inspection, the output masks show the whole-network (and some noise added), but we were wondering if we're missing something obvious. This is what the dual_regression script does, correct?

Also, we output the results as z values with the --out_z option. Can we perform mixture-modelling on these images to get a consistent thresholding value across subjects (eg. 50% false-positive rate?). If so, would the following command (as seen on the MELODIC user guide) do the trick?

melodic -i myZstat --ICs=myZstat --mix=grot.txt -o myZstatMM --Oall --report -v --mmthresh=0.5

Thanks in advance!
Paulo