Hi,
fsl_regfilt is specifically tailored at removing MELODIC noise components, using a full melodic_mix matrix to generate spatial map estimates, allowing a selection of components and subtracting the relevant outer product of the melodic_mix columns with the spatial map estimates from the data. fsl_regfilt returns as output a file fit for further data processing (with the voxel-wise mean intact etc...
fsl_glm implements a OLS procedure and you need to as it to output the residuals to get back your data after regressing out the 'noise'. You'd also need to reintroduce the mean yourself (if you think you need it later).
The 'optimal' procedure will possibly vary depending on the structure of the noise in relation to the signal of interest and your personal preference as to how aggressive you want to denoise the data. I'd go with not denoising per se but adding motion and CSF signal to further regression analysis. Also, I recommend _not_ to remove the global signal as that procedure is the spawn of the devil... if you find your results to 'improve' after GSR then that is likely because you're not modelling the signals correctly in the first place earlier on...
hth
Christian
On 7 Aug 2013, at 18:19, Yael Grossman <[log in to unmask]> wrote:
> Hello all,
> I just read a paper published by Kelly et al in J Neuro Methods 2010 about artifact removal. I'm still new to FSL and fMRI analysis in general, so I've been using fsl_glm to regress out artifacts like motion and global signal. However, this paper mentions using MELODIC to get a series of ICs, then choosing the noise-ICs and regressing them out using fsl_regfilt. I was wondering which would be a better method to use for denoising data, what I've been doing so far or the procedure outlined in this paper.
>
> Thank you!
> Yael
|