tsalo
commented
6 months ago A few things to try varying:
- Number of echoes from 3 to 8.
- Do we expect low numbers of echoes to perform well? We certainly assume as much in Tedana.
- Echo times from just covering the range of T2 values to 1.5x the max T2 value (see Logan's post here) to something far beyond that.
- We could probably disable the adaptive mask to see if there's a problem with including bottomed-out echoes, and specifically to see what the nature of that problem is.
- Lagged BOLD-based global signals (i.e., sLFOs).
- Tedana, as it currently exists, shouldn't be able to do much with this. It should identify a temporally-blurred version of the global signal and accept it.
- Localized task-related signals.
- Spatially correlated BOLD and non-BOLD signals.
- Tedana's spatial ICA shouldn't be able to handle this, so we might expect components that have both high Kappa and high Rho.
- Temporally correlated BOLD and non-BOLD signals.
- Tedana's spatial ICA should handle this successfully.
- Motion-correlated non-BOLD signals.
- Amount of thermal noise.
- We can probably use the amount of variance removed by the PCA for this.
- Patterns of thermal noise.
- Does the scale of the thermal noise vary by echo? I assume so.
- Does the time series of the thermal noise to vary by echo?
I was thinking that we could probably simulate data with specific parameters, using real data as a basis. If we have (1) TE-(in)dependence model fit maps, (2) component weight maps, (3) component time series, and (4) variance explained maps, we could probably predict the multi-echo data for a range of echo times, numbers of echoes, etc. We could then run tedana on the simulated data to see how the various parameters impact the results.