[FieldTrip] denoising data

[Todorovic, A.]

Dear 'trippers,

I'm curious to hear your thoughts on how to best denoise data. What I'm doing is working, but I'd like to hear whether it's logical or reasonable. 

In particular, I have two issues. One is that I use ICA to remove blinks, so I have to fit denoising into a pipeline that incorporates ICA (which, I am guessing, works well only without including reference channels). 

The other is that I see there are two functions for denoising, ft_denoise_synthetic and ft_denoise_pca. 

In sum (1) I'm not really sure WHEN the best moment for denoising is, given that I use ICA, and (2) I don't understand when it's better to use ft_denoise_synthetic and when ft_denoise_pca. I do see that ft_denoise_pca has the option of preprocessing reference channels separately, which makes it easier to use after ICA.

I used to know only about the first function, and my solution was to do ft_denoise_synthetic (using the 'G3BR' option) prior to the ICA. Then I would remove components which contain artifacts, and continue with making ERF/TFRs/whatever. This produces data that is somewhat cleaner than when I skip the denoising step.

I was curious about the ft_denoise_pca when I saw it, so I tried running it on filtered, preprocessed data that I had after I rejected ICA components. [In the process of doing ICA, I used the ft_denoise_synthetic option, as above.] This produced a different TFR at the end of the road, which again looked cleaner. Significantly cleaner, actually, but N=1. 

Now I'm not sure if it was a logical step to use both denoising functions, and if it would have been a better idea to do things differently. I'd like to hear both whether something in the logic is wrong, and whether it's inelegant.

Cheers,
Ana