ENH: agenda for beamformer module

[britta-wstnr]

Let's keep record of what issues and enhancements are still open with the beamformer code (not including visualization etc.). Here is a list of open issues and PRs and other things that need to be done:

• [ ] Eigenvalue inversion beamformer ("automatic" regularization) This is in issues #5362 and #5690.
• [ ] Clarify SAM vs LCMV in docstring of LCMV
• [ ] Non-converging test results with reg=0.0 Check out why one of the beamformer tests in test_lcmv does not converge across CIs for reg=0. For discussion and link to specific test, see https://github.com/mne-tools/mne-python/pull/6042#discussion_r265037976
• [x] Address combination of different sensor types in DICS The combination of different sensor types when computing the CSD matrix needs to be addressed, for discussion see #5135, #7021
• [x] Center of head bias even with normalization #5266
• [x] Pass covariance to beamformer to get static image This is issue #5229. It relates _lcmv_source_power, which could need some work / refactoring with make_lcmv.
• [x] Speed up LCMV Speed up the LCMV code and the LCMV tests. (This was taken care of in some PR already. Also in #6603)
  • [x] Have beamformers use mne.minimum_norm._prepare_forward

    5881 (#6042)

  • [x] Port MNE depth bias tests to LCMV

    5984

  • [x] Projections in beamfomer code Still open from #5135: how should DICS handle projections? (Also related to #5342 - is LCMV handling projections the right way?)
  • [x] [BUG] Orienation selection in unit-noise-gain DICS Compare this comment, the normalization done at the beginning of the DICS pipeline results in a wrong orientation selection formula
  • [x] Further refactoring of the DICS/LCMV code Once the DICS bug (see above) is solved, the last steps of refactoring can be done (confirm equal results when switching DICS and LCMV computation blocks), this concerns this code block (review). At the moment, results do not converge, probably due to the DICS orientation selection bug.
• [x] Orienation selection for unit-gain LCMV This is started in #4659 and is waiting for easy access to leadfield normalization, which should be granted after merging the refactoring PR #5135.
• [x] Store orientations in filters object Store the picked source orientation in the spatial filter, e.g., for plotting.
• [x] NAI for DICS beamformer as suggested by @mmagnuski in #5135

And then, there is this older issue about beamformer enhancements: #3853. Some of this is done now, but other things might remain, like the different output types (z-scores etc. - which, however, could also just be computed on the output directly).

Anything missing, @agramfort @larsoner @sarangnemo @wmvanvliet ?

[larsoner]

Sounds reasonable to me

[agramfort]

sounds like a plan !

[wmvanvliet]

Yeah, I agree on all these points.

Regarding projections with DICS: turns out, applying SSPs to a CSD works exactly the same as applying them to a covariance matrix! See this gist: https://gist.github.com/wmvanvliet/fe04c40e27c8f81422596163d41563ca

[larsoner]

I wrote a similar test that failed, but looking back at my code I only left-multiplied by the proj op, grrr. Now it works just fine:

https://gist.github.com/larsoner/5a5f7c53c79416212e28a56292f434d7

[britta-wstnr]

Added storing of source orientations in filters. cc @sarangnemo

[britta-wstnr]

Added speeding up LCMV as brought up by @larsoner in #5135

[britta-wstnr]

@larsoner I added the combination of different sensor types.

[larsoner]

@britta-wstnr the max-power has a sign ambiguity. Any reason not to orient it so that positive is in the direction of the vertex normal? Basically this means taking the max power orientation and multiplying by the sign of the dot product with the vertex normal. It's a principled way to deal with the arbitrary flip.

[britta-wstnr]

@larsoner How would we do that in the volume case, do volume fwd models have that information as well? (Note that beamformers are usually computed for the full source grid, not a surface solution.)

[larsoner]

The normals are defined as up Z in that case IIRC. So it will still work

[larsoner]

Done in https://github.com/wmvanvliet/mne-python/pull/6 which probably makes the most sense to put in with #5447 (which is why I made a PR to that branch)

[larsoner]

@wmvanvliet can you take a look at the list above and update the checkboxes?

[wmvanvliet]

done!

On 7 Nov 2018, at 23:39, Eric Larson notifications@github.com wrote:

@wmvanvliet can you take a look at the list above and update the checkboxes?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[mmagnuski]

(I just moved unchecked checkboxes to the top to make them more visible)

[larsoner]

Clearing 0.19 milestones related to beamformer, @britta-wstnr let me know if there were things you wanted to get into 0.19

[britta-wstnr]

@larsoner I don't think any of the things here are critical. @wmvanvliet and I have some work planned in October (relating to the Spring Coding Sprint comparison between types) and hopefully will close the pending issues along the way!