correct way to store polygon coordinates in spatial_polygon_obsm? about vitessce-python HOT 5 CLOSED

I believe you need to use a list not numpy arrays, but that is based on limited experience: https://github.com/hubmapconsortium/portal-containers/blob/dc568234c76017c7cd9644a4d15ef0f7b9d84e24/containers/sprm-to-anndata/context/main.py#L57-L66
Also, I could be wrong about this, but you probably also need to store the centroid. I'm going to open an issue to look into this so it doesn't fall off my radar.
Also as a heads up vitessce/vitessce#927 is coming.....so if you want to future-proof what you are doing now for that (i.e switch the segmentation visualization to bioformats6 image pyramids of the masks in the future and use those instead of polygons), I'd recommend storing the centroid as well since it will be needed for selection (spatial_centroid_obsm)

In short, I'd say 1) stick to basic python types like lists and 2) also store the centroid.

Thanks for trying this stuff out!!!!

from vitessce-python.

Ok great. Are there example SPRM data files around somewhere I could download? If I load those up I can take a look at all the pieces of data that go into the AnnData based on the link you posted.

It's my pleasure to try vitessce out for this application. I'll definitely let you know about my experience too 👍

from vitessce-python.

@NHPatterson I can message you a test HuBMAP dataset and view config. The IEC is pushing at the moment trying to get it out the door and we only have data on the test instance which is IP whitelisted.

from vitessce-python.

Thanks for this. I now understand the issue. It seems the AnnData store ( and zarr) can't accept ragged arrays.

From the dataset you sent me:

adata_proc.obsm["poly"].shape
Out: (11377, 1, 10, 2)

Essentially each polygon must be represented by 9 points in this case (one redundant vertex). It seems downsample_shape
in the linked script above performs this operation. I think the bit-mask solution will address this issue but may have issues with overlapping regions-of-interest? We sometimes have annotations for both gross (i.e., occupying >1/4 of the image area) and fine morphology.

I don't know what's possible here to address the polygon coordinate issue, but perhaps a string parser could be employed using WKT representation? This would produce one string for every polygon regardless of number of vertices. But this is now likely out of the scope of this issue. 😄

from vitessce-python.

@NHPatterson I agree with you on the last point certainly - I was trying to conform to the old way of doing this in Vitessce in cells.json with the AnnData stuff so I ended up just re-implementing it but there's nothing to stop us from busting out a parser or something. That would certainly be nice. Message me an example of you think could go wrong with the BitmaskLayer, would be very curious to understand.

from vitessce-python.