Hi
There is no easy answer, generally we recommend only small amounts of smoothing for ICA (in parts because one main reason for smoothing - ensuring validity of GRF-based inference - does not apply to ICA. Personally I set smoothing to about 1-1.5 x voxel size, i.e. for 3mm isotropic voxels set smoothing to ~4mm. If you later on want to compare maps to eg FEAT maps you can add more smoothing retrospectively to the melodic output.
you also might want to look into explicit denoising your data first in order to remove (at least the linear) effect of head motion prior to common pre-processing followed by FEAT&ICA
hth
Christian
On 30 Oct 2013, at 19:12, Paulo Branco <[log in to unmask]> wrote:
> Dear FSL users,
>
> I'm writing this message in hope I can get some issues clarified: I've been reading the FSL jiscmail and found some related posts but I'm still a little confused on what parameters to use. I'm attempting at running single-session ICA for individual subjects, and then compare this with individual-based FEAT-GLM z-maps. I've managed to do so well, but so far it seems the dimension estimate is spliting some components that have very similar spatial and temporal attributes in some subjects, while on others the components are rather unconvincing. I was reading into the issue and found that smoothing can affect the dimension estimate and this can, in turn, be troublesome in cases where there is some head movement which is a problem I'm currently facing (patient data).
> In my case I'm using 6mm smoothing, to be able to then compare to the 6mm smoothing FEAT maps. Some papers in the field fix the number of components (say, 40) to "stabilize" the issue but my intuition is that doing so, based on the previous arguments would actually be even worse.
>
> I'm wondering what you think is a sensible approach to this problem to avoid over fitting/underfitting. So far I see some alternatives:
>
> 1) Calculate the #ICs using 0 smoothing and then fix that number on a new analysis running at 6mm smooth (or smooth after the analysis).
> 2) Let MELODIC estimate the components; In this case some subjects would be OK while others would be bad, depending on the motion artifacts and extent of activation?
> 3) Fix the components in a sensible number (say, the mean components on all subjects);
> 4) Lower the smoothing in both ICA and FEAT and hope for the best (less smoothing, better estimate, better ICs?).
>
> Could you kindly explain me the pros and cons of choosing each option? Or point me to relevant literature on the topic? Maybe I'm being too picky and this is not at all that relevant, but I'd like an expert opinion.
>
> Thank you,
> PB
|