Running tedana ME-ICA after NORDIC

Are there any recommendations about removing thermal noise from images using NORDIC before running tedana with meica? Can these two denoising tools be stacked to produce better results? Would the addition of NORDIC lead to mislabeling problems in tedana’s decision trees?

This remains an open question (actually several) - in theory NORDIC will suppress the noise source that the PCA step tries to separate out, so there is absolutely going to be an interaction. It is currently not clear how much the decision tree will be impacted, nor is it clear how to appropriately handle the change in the degrees of freedom due to NORDIC (which matters heavily depending on your analysis strategy).

In the past, I have had subcortical components seperate out better (where g-factor amplification was the highest) but several other users did not see such an effect, which leads me to believe that the outcomes depend heavily on the level of thermal noise in the data.

I will say, you can stack them - it will “work” but you may need to manually select the amount of variance to retain or the number of components. I’d argue that a time and space consuming comparison needs to be done - and to think about whether thermal noise is really a major issue for your data and subsequent analysis. In general, if you are going to be smoothing your data, or doing group statistics, it generally makes less sense to use NORDIC, though this is not a hard rule by any means.

Thanks so much for your reply!

A quick follow-up question: The tedana documentation stresses that “any step that will alter the relationship of signal magnitudes between echoes should occur after denoising and combining of the echoes.” My understanding is that NORDIC is run on each echo separately. Is this a concern?

Your understanding is correct (though this is also a point of investigation) - and it is mostly not a concern. If NORDIC is only removing thermal noise, then the relationship of the signal should be unchanged.

However, and this is important - for example if the estimate of the thermal noise is too high and NORDIC removes more than it “should”, then signal amplitude may be reduced. (see, and original blog post: NORDIC denoising on VASO data – layer fMRI blog).

Now - given the amount of noise in fMR, how the T2* is calculated (typically one value for the entire time series), and how much signal NORDIC could really remove in typical unlucky usage - this is likely not a major concern. It is possible that that T2* estimates would change (slightly) and that the resulting optimally combined signal could in turn be slightly different, but I wouldn’t expect it to be a huge effect, unless things go very very wrong.

In sum - definitely right to be concerned, looked closely at the resulting data, but I wouldn’t expect it to be a disaster - at least not an obvious one. I think any negative effects would be subtle and while that is in some ways worse (a sneaky/hidden bad thing you don’t know about!) it also probably wouldn’t be an issue for most analyses. That said, this is an area of active interest and investigation, so things are a bit volatile.

1 Like