EquispacedMaskFunc doesn't generate equidistant masks about fastmri HOT 4 CLOSED

This is an interesting issue. Certainly from the image-domain SENSE algorithm you'd want the exact spacing, but the inexact spacing in the repo at the moment might actually help things from a compressed sensing/incoherence point of view.

Ultimately though what's in the repo should be targeted most to matching what's in the public test and challenge data sets - those cannot be changed. If they have the same staggered offset characteristics, then we will have to leave this.

from fastmri.

Yes I was actually thinking that from a GRAPPA point of view (since it was also mentioned in the paper).

Currently GRAPPA approaches will fail unless using a tailored implementation with twice (more or less) as many geometries.

I also think in any case it might be nice to clarify that point somewhere. I understand in this case it's a bit more tricky to change the test and challenge set since the values could be directly used (as opposed to the sign issue for the v1 of the knee dataset), but it could be so benefitial for the comparison to other approaches. I have typically been asked a lot about the comparison with GRAPPA and I was hoping to be able to make a public point with the brain dataset (as opposed to using my homemade masks).

I think if you don't plan to take any action though, we can close this.

from fastmri.

Yeah - of course people on the fastMRI team have also used GRAPPA.

I just opened file_brain_AXT2_210_6001948.h5 - also from the test set - and I'm also observing this behavior there.

I will add a note to the function that due to its implementation some customization of GRAPPA methods will be necessary for the challenge set. I'll also relay this to the team about the challenge data in case we want to make any changes.

from fastmri.

Yes I actually read that paper and it added to the confusion when seeing this "problem". I actually sent an e-mail to one of the authors, @tullie , who forwarded it to @anuroopsriram , to understand more about what they did to tackle this issue of custom GRAPPA.

Closing this, since I don't think any more action is required.

from fastmri.