Add Minian segmentation extractor

[alessandratrapani]

Add segmentation extractor for Minian output.
Supported output format: Zarr (for now)
Similarly to Caiman uses CNMF to perform cell identification.

Output structure (example):

minian/
 ├── A.zarr 
 │   ├── A (851, 608, 608) float64
 │   ├── animal () <U7
 │   ├── height (608,) int64
 │   ├── session_id () <U9
 │   ├── unit_id (851,) int64
 │   └── width (608,) int64
 ├── C.zarr
 │   ├── C (851, 8931) float64
 │   ├── animal () <U7
 │   ├── frame (8931,) int64
 │   ├── session () <U14
 │   └── session_id () <U9
 │   └── unit_id (851,) int64
 ├── b.zarr
 │   ├── animal () <U7
 │   ├── b (608, 608) float64
 │   ├── height (608,) int64
 │   ├── session () <U14
 │   ├── session_id () <U9
 │   ├── unit_id () int64
 │   └── width (608,) int64
 ├── b0.zarr
 │   ├── animal () <U7
 │   ├── b0 (851, 8931) float64
 │   ├── frame (8931,) int64
 │   ├── session () <U14
 │   ├── session_id () <U9
 │   └── unit_id (851,) int64
 ├── c0.zarr
 │   ├── animal () <U7
 │   ├── c0 (851, 8931) float64
 │   ├── frame (8931,) int64
 │   └── session () <U14
 ├── f.zarr
 │   ├── animal () <U7
 │   ├── f (8931,) float64
 │   └── frame (8931,) int64
 ├── S.zarr
 │   ├── S (851, 8931) float64
 │   ├── animal () <U7
 │   ├── frame (8931,) int64
 │   ├── session () <U14
 │   └── session_id () <U9
 │   └── unit_id (851,) int64
 ├── max_proj.zarr
 │   ├── animal () <U7
 │   ├── max_proj (608, 608) float64
 │   ├── height (608,) int64
 │   ├── session () <U14
 │   ├── session_id () <U9
 │   └── width (608,) int64

• A: Spatial footprints of cells. Should have dimensions (“unit_id”, “height”, “width”). --> image_masks
• C: Temporal components of cells. Should have dimensions “frame” and “unit_id”. --> roi_response_denoised
• b: Spatial footprint of background. Should have dimensions (“height”, “width”). --> background_image_masks
• f: Temporal dynamic of background. Should have dimension “frame”. --> roi_response_neuropil
• b0: Baseline fluorescence for each cell. Should have dimensions (“frame”, “unit_id”) and same shape as C --> roi_response_baseline
• c0: Initial calcium decay, in theory triggered by calcium events happened before the recording starts. Should have dimensions (“frame”, “unit_id”) and same shape as C
• S: Deconvolved spikes for each cell. Should have dimensions (“frame”, “unit_id”) and same shape as C --> roi_response_deconvolved
• max_proj: the maximum projection --> summary_image

TODOs

• add extractor
• update CHANGELOG
• implement tests

[h-mayorquin]

Do tag me when this is ready for review.

[alessandratrapani]

I am uploading on the s3 bucket the testing folder for the minian segmentation data

[alessandratrapani]

@h-mayorquin ready for review

[h-mayorquin]

I took the first look. Looks good and simple to me.

1. Is the structured documented? (the one that you shared on the PR) I am inferring the meaning of A, C, b, f, etc from the equation of CNMF here:

https://minian.readthedocs.io/en/stable/pipeline/notebook_5.html

But maybe there is other place in the documentation where this is stated more precisely, if so, we probably should link that in the docstring.

2. The example that you uploaded, can you add a readme and share it with me so that I can upload it to gin.

[h-mayorquin]

I am not sure this docstring is very helpful, I think it should be oriented to explain things to final users and not implementation details.

[alessandratrapani]

Ok, sure! I just wanted to be consistent with all the other extractors, but I guess a more detailed docstring won't hurt.

[h-mayorquin]

I don't understand how the "f" (the temporal activities of the background) is the response neuropill. Can you say how those concepts match?

[alessandratrapani]

Let's say first that the background can contain neuropil signal. The problem is that I wanted to map this "f" signal with something already in the SegmentationExtractor.

Now, I am not super happy with this formulation I think we should have two separate variables:

1. _roi_response_neuropil with the respective _neuropil_images_masks. This can be used to store the actual neuropil signals that sometimes refer to the area surrounding the ROIs
2. _roi_response_background with the respective _background_images_masks. This can be used more generally to store the signal in the field of view that excludes the areas in the identified ROIs (as in this case).

I opened an issue on roiextractor to discuss this with the others: #375

[h-mayorquin]

OK, makes sense.

[h-mayorquin]

What's the status of this?