tedana roadmap discussion

[handwerkerd]

Several years ago, a bunch of of us put together a roadmap for tedana. We're actually getting close to achieving the main goals of the roadmap! We're definitely not there yet, but this is a good time to start thinking ahead.

We can host this discussion here. As people put forward ideas, here are a few categories to consider:

• Current stuff better: A functionally that tedana currently has, but you have a way you want to see it improve
• Novel methodology: Ways to improve the underlying methods
• Connections to other tools: Are there parts of tedana that should be more accessible to external programs
• Software Documentation
• Multi-echo fMRI education documentation & resources
• Shared datasets for education, validation and testing

[handwerkerd]

Starting off with some ideas several of us have been discussing over the years.

• Current stuff better
  • More ways to visualize and interact with logged information from a tedana execution
    • Flow chart visualization of an executed decision tree
    • Tools to collect and organized summary measures across runs and populations (i.e variance explained, number of components, variance in accepted vs rejected)
• Novel methodology
  • Find better ways to estimate appropriate variance explained for PCA/ICA (maybe going back to using multi-echo information)
  • Switch away form ICA to a method that can maximize the distance between kappa & rho within each component rather than maximizing statistical independence as ICA does
  • As the core measure of what makes a component good/bad are kappa & rho and the elbow calculations the best we can do?
  • Better integrate non-multi-echo noise removal tools with multi-echo noise removal (parts of which might be beyond the scope of tedana, but making connections could be within the scope)
• Connections to other tools
  • Continue work on modularization code to make it easier to have other programs connect to parts
  • Reword the main tedana.py workflow so that it's easier to access parts and easier for contributors to improve parts without needs to understand how everything fits together
  • Decide of other parts of tedana should become stand-alone packages (maybe the modularized decision tree and metric calculation sections???
• Multi-echo fMRI education documentation & resources
  • Expand the multi-echo analysis Jupyter book:https://me-ica.github.io/multi-echo-data-analysis
• Shared datasets for education, validation and testing
  • Build up validation tools to support comparisons across multiple data sets and decision trees.
  • Decide how we can make sharing data specifically for tedana development & validation easier.

[eurunuela]

@dowdlelt and I brainstormed about the novel methodology that could be added to or supported by tedana. The list of ideas is too unstructured and unpolished to post here (can be accessed here), but I guess we will edit this comment as we write them more clearly.

To improve classification:

• Compare component time series to motion regressors
• Compare time series with physiological regressors
• Examine data for slice-to-slice artifacts (spatial FFT over slice dimension - high-frequency component means movement artifact probably ( high frequency, in this case, being slice to slice)
• Provide slice timing (or JSON) Multiband artifact - are spatial maps of simultaneous slices more similar than actual neighbor slices. Probably Multiband issue, noise
• Symmetrical (left-right) smooth activation (or symmetrical absolute values) should probably be classified as bold-like
• Tissue maps, does it look like an edge effect? Csf? Etc
• Reconstruction time series from unclear components, smooth data, and try again - maybe bold components would be more clear and could be classified
• Provide Field strength - how much does a component overlap with CSF like t2star values, wm t2star values, etc
• Provide Field map, how much does component overlap with high susceptibility regions (identify respirations maybe)
• Provide task if available - try and keep components that look like a task ( controversial)
• Provide Field strength - convert components to percent signal change - if 3t and percent signal change reaches 20, that's noise! (Probably, requires testing)
• Extracts skull from the first echo, if components overlap with that, it's noise!
• Revisiting the smoothing idea - kappa values should remain stable as smoothing increases, perhaps there is a kappa+1 voxel smoothing, kappa+2 voxel smoothing, etc columns
• Multiple runs, cross-validation (hard)
• Multiple runs, concatenation - MELODIC takes multiple runs for ICA, tedana should too. Increase sensitivity! More components to sort, but BOLD should be similar across runs ( and reinforce) whereas noise should vary from run to run
• Actually, there is an idea, that even if you tedana each run independently, you could add another column, does a component have a match in another run. Consider spatial - networks (or task activation) should be similar run to run so one component would be very very similar to one of the components in another run.
  • Movement could be similar, but probably not
  • Physiological would be similar but could be identified through other methods.

Metrics:

• Currently, kappa is based on some math done to the spatial map. What if we instead took a grey matter mask and calculated, within that only. And within WM only. A kappa grey to white ratio. It's another metric
• And can we calculate a kappa-like value based not on the spatial maps of each echo, but on the kappa/rho nature of each component's timecourse?

Rica:

• The SVM (or whatever, could be deep learning too) trained on Rica classifications idea for "automagic" classification of components.
  • Rica would upload the classification file to a server, and these manual classifications would be used to train a classification algorithm.