So using 7T shouldn’t change how tedana performs. I do wonder if the transformation to MNI space is affecting your results. Have you tried to run tedana in the subject space prior to registering to MNI on one of these subjects that give you no BOLD components?
@e.urunuela I did try tedana in the subject space for a few of those participants. Still the same “No BOLD components” errors. Shall I try increasing the number of components?
I’m late to the party here - but I would try increasing components a bit more (2/3rds? 80 some?). Its really hard to say what is happening. I don’t want to advocate for an exhaustive search…
The automated methods are an attempt to solve an impossible problem. it would be nice if they worked everytime for every dataset - but that has never really been the case (hence there is more than one option). I’m not up to date on how the robust options are suppose to work, but in some cases, a rule of thumb for a given dataset is the best we can do. This is annoying, I know.
As for the MNI transform, we talked about this in the past, and it is cleanest to not do it, in some sense - but on the other hand, this is likely to, if anything, reduce noise in the data as it will interpolate and stretch and squish areas. I would expect it to improve the data if anything from a “its noisy” point of view. If I click around with InstaCorr, I see a lot of structure in the data you originally shared - nice bilateral networks, etc - so i don’t think you are overwhelmed with noise. I’m really at a loss - but maybe you need more components to really pull apart the signal - 1.5mm is a pretty high increase compared to the og data that went into MEICA.
One issue could be the relatively late echoes - We would really want the first echo to be around 12 or so - can’t be helped at 7T I know, but it could be contributing here to the weird behavior. That said, you should still get the benefits of some BOLD fitting, increase CNR and some denoising.
How many comps for the “good” subjects? Is it finding 100s? or are they settling for something low, and it just happens to find a BOLD one in there. The warning wouldn’t show up if it found 1, but you shouldn’t be happy about that either!
This is maybe orthogonal to the exact question at hand, but I have heard various things about whether transforming to standard space really has any affect, particularly a negative one, on the fitting. It would be nice to have some solid examples about this for clarity, to really know what is happening in modern code. If there is a clear answer about this for all data, great; if there isn’t, it would be good to know what the considerations are. It would be helpful to see this working/failing, and how, and what to look out for, rather than always be worried about an order of processing steps that might not actually have a negative effect.
–pt
@dowdlelt thank you for your input. Indeed it makes sense to check other subjects who don’t have the “No BOLD” error to determine the number of BOLD components that are identified on average.
I am attaching a spreadsheet with the total number of components identified through rICA (when “-tedpca 60” was specified) and the total count of accepted/rejected/BOLD. On average, I think around ~45% of the components that are accepted are classified also as likely BOLD, excluding those participants who have the “No BOLD” error. What do you think?
tedana_metrics.txt (2.3 KB)
update: @dowdlelt @e.urunuela increasing to “-tedpca 110” removes the error for some runs, but others still have the error “No BOLD components”
Incredible - Its really hard to know the right range - in some cases it needs to be the typical 1/3rd (ie lower is better), in others more can help.
When I look through the tedana_metrics briefly, I am struck by how low the number of comps is, even for the ones that “worked”. I don’t know what to think - my initial impression is that tedana is too strict on how it determines what is BOLD versus not, but thats just a complaint/concern, not a solution.
I’m a bit of a loss for how to move forward without this turning into a tedana investigation, rather than whatever task you were interested in. Optimal combination will still help you, but then you of course, don’t reap the benefits of multiecho denoising.
While I agree that 7T shouldn’t have a huge affect, in principle, you are facing higher levels of thermal noise (1.5mm voxels), a stronger spatial variance of that noise (total accel 9, g-factor varies substantially over the brain), longer echo times than would be desirable (ideally first echo is half of typical, and middle echo is “typical” for field strength ie 22ms or so), and the late last echo means a lot of areas have lost a lot of signal, while those that do have it would have very large percent signal change. Which of these factors matter(s) is well beyond the scope of a neurostars post!
The only other idea I have is likely to just waste your time:
If thermal noise is the issue, perhaps smoothing, running tedana, and if it identifies noise regressors, using those in a GLM with unsmoothed data.
Otherwise, I’m not sure of how to proceed. The one example dataset I looked at suggested you had plenty of signal (and you should, at 1.5 in 7T) - InstaCorr gave nice correlations across the brain, bilateral - in expected areas, etc etc.
Thanks @dowdlelt for the insights. I have checked in 3T datasets in our lab, and none had the “No BOLD components” error (although those have slightly different lengths of data).
@e.urunuela @handwerkerd I am still unsure why tedana finds zero BOLD components. I could increase the number of PCA components to keep, but if BOLD only appears in low-variance components, does that suggest poor signal quality? Relatedly, what BOLD component threshold is acceptable—if only 1-2 BOLD components are found from 100, is that enough, or does it indicate problems with the data or classification?
I’m afraid we don’t really have an answer at the moment. It may be worth trying this adjustment @handwerkerd is testing in this pull request: Only mean centering PCA components by handwerkerd · Pull Request #1271 · ME-ICA/tedana · GitHub
Sorry, how do I try this new feature?