Tedana manual classification

1

Hi all,

I’m going through tedana (v0.0.10) component reports from a large data set, and performing manual re-classification of some components. I’ve come across a few components that I’m not sure how to classify, and thought I would ask for others’ perspectives.

The hard-to-classify components virtually always lie around the middle of the kappa rho plot, with relatively high kappa and rho values. When they are rejected, the stated reason is usually I007, 009, or 010. Here are examples of two such components (4 and 6) from one dataset:

ident05_tomloc2-1_comp4
ident05_tomloc2-1_comp42692×1318 389 KB

ident05_tomloc2-1_comp6
ident05_tomloc2-1_comp62692×1318 399 KB

These components are similar overall, but tedana rejects one and keeps the other. Both components have positive weights across most of gray matter, and their time series have some variance at drift-like frequencies (<.01Hz), but also some variance in frequencies typical of BOLD-driven variation (.01-.1Hz).

The rationale for rejecting such components, from Kundu et al. 2013, is that they localize to “they localize to the sagittal sinus and draining veins indicating pCO2 fluctuations.” However, the weights in these components are not localized to draining veins, but extend across gray matter. Other components have a spatial profile that seems more consistent with draining veins, e.g. 10:

ident05_tomloc2-1_comp10
ident05_tomloc2-1_comp102692×1318 343 KB

This component was accepted, although components with similar spatial profiles in different tasks/runs from this participant were typically rejected.

What do folks think about these components? Would you classify them as BOLD or non-BOLD? Is there any principled basis for differently classifying components 4 and 6? Thanks,

Ben

2

Brief answer: If you are unsure whether to accept or reject a component, accept it. The damage from keeping noise is slightly noisier data. The damage from rejecting neural signal is messed up results.

Longer answer: What is likely happening here is an ICA component contains both noise and T2*-weighted signal. The selection process is deciding whether or not the component has enough useful signal to keep and, given it’s on a borderline, sometimes decides one way or the other. It looks like you have around 35 components, which probably means your time series aren’t super-long. That should be enough to have tedana work, but it also means you’re more likely to get this mixed, higher variance components that aren’t purely signal or noise.

The specific examples you show are all a bit mixed. The spatial maps look heavily weighted in gray matter, but there might be some weighting around edges or CSF. The frequency peaks are less than <0.1Hz, but sometimes a bit higher than what we might see in a normal BOLD-weighted signal. The time series look BOLD-ish, but several have a singled, high variance peak, that is likely not BOLD. It looks like all of them are above the kappa elbow (maybe we should mark the kappa elbow on those plots) but some are above or below the rho elbow (a possibly cause of the variable classifications).

Hope this helps

Dan

3

Makes sense, thanks.

1 Like