Ivadomed: Providing support for EEG modality in IVADOMED

[mpompolas]

@jcohenadad
This is to start the conversation for opening up the toolbox to EEG.
As mentioned to Guillaume (GitHub handle?) on Slack, I imagine the ivadomed toolbox to be accessed at a later stage of the analysis pipeline, since in electrophysiology there are multiple pre-processing steps needed for making comparisons across conditions.

I attach a diagram to make this a bit more clear. Ivadomed comes into play after the last step:

Assuming that we want to perform classification among conditions, the easiest way would be to create an importer for different types of analysis features (for our purposes we can just think of them just as 2D or 3D matrices).
For example, to compare conditions in:

1. The time domain (nTrials x Channels x TimeSamples)
2. The frequency domain (nTrials x Channels x FFTbins)
3. Time-frequency decompositions (nTrials x Channels x FFTbins x TimeSamples)

Any suggestions are welcome

[jcohenadad]

My first reaction would be: can the described pre-processing pipeline in https://github.com/ivadomed/ivadomed/issues/525#issue-742138726 follow the BIDS convention? I.e. processing outputs would be placed under the derivatives folder, and be used as input in ivadomed. If we go with a BIDS folder as input, this will minimize the amount of refactoring needed on the ivadomed side.

So, if we can "bidsify" the 2D/3D matrices mentioned in https://github.com/ivadomed/ivadomed/issues/525#issue-742138726, that would be great.

[mpompolas]

Yes, I agree. That would be the easiest first step.
I'll reach out to Guillaume to get an example dataset to create the BIDS folders based on the way he stores his metadata.

If we are satisfied with the way these example files handle metadata we can use them as a guide for whoever wants to use the software.
If not, we do one more iteration on this.

Regarding applications, classification is a standard approach (comparison of patients/controls for example).
What would also be very interesting (I'm not sure how we can do this yet with ivadomed - I only used statistics for that), would be to also find the exact time or/and frequency that differentiates conditions

[mpompolas]

I made a BIDSified version for mne epochs:

This is on my computer right now. Will create a repo

[mpompolas]

Notes:

https://docs.google.com/presentation/d/16iArDrH5Lqm3rL-MTBkrHQ9eF0Qy328UkIpmKMJ0-MY/edit?usp=sharing