I have been using Tedana for multi-echo denoising but I am finding that it is finding far too few components with
--tedpca mdl (around 20 components for TR=1, volumes=700).Moreover, it was only turning up around 64% of the explained variance. A colleague of mine experienced similar issues and began using the
--tedpca kundu to better results but when I try and make this adjustment, I am getting errors that indicate it is failing to converge.
Any idea why this might be occurring on my end?
--tedpca kundu is removing too few components, so the dimensionality of the data hasn’t been reduced enough for the ICA to converge. This is the main reason we added the MAPCA options to tedana.
You can try
kundu-stabilize, which is more aggressive than
kundu, but less aggressive than the MAPCA methods, or one of the other MAPCA methods (
kic), which are both less aggressive than
Thanks for the response! I have given the kundu-stabilize a go on my data and while I am getting nearly double the components now, I have noticed the variability explained has significantly dropped (to between roughly 35-50% on average). Is there some reason that this is occurring?
I’ll ping @eurunuela and @handwerkerd on this, since they know more about the PCA step than I do, but my guess is that kundu-stabilize is classifying at least one high-variance PCA component as noise, so even though more components end up moving on to the ICA step, they explain less variance. If you look at the PCA metrics file and compare variance explained grouped by classification, do the rejected components explain more variance than you’d expect?
I answered similar questions a week ago or so. I suggest you have a look to better understand how maPCA works: Tedana mask and components - #29 by e.urunuela
I’m hoping to add new changes to tedana that provide more info about the PCA step asap.
Edit: you can actually try these changes using the branch in this pull request https://github.com/ME-ICA/tedana/pull/839