Fix OME-Zarr chunk size issue for stitching module

[miketaormina]

For discussion, is this the proposed solution outline:

Solution outline

1. Pystripe outputs single plane Tiffs (maybe zarr in the future) that conform to TeraStitcher directory structure.
2. TeraStitcher ingests a specific (parameterized) channel from a multi-channel data set and calculates offsets.
3. Found merging parameters are broadcast to the other spectral channels in the data set.
4. Once the Merge step of TeraStitcher workflow, --sliceheight, --slicewidth, and slicedepth are set to the Y, X, and Z chunk size desired and --volout_plugin is set to TiledXY|3Dseries (is there a TiledXYZ|3Dseries?).
   • maybe zarr in the future.
5. code.src.zarr_converter module ingests the resulting folder and writes a Zarr array with a 1:1 correspondence between Zarr chunks and Tiff files.
   • Do we need to edit the AICS-imageio package to allow this, or perhaps not use that package?

Separately, it seems like step 4 and 5 are a convenient place to create and store the image pyramid levels using TeraStitcher's --resolutions flag. Do we think this should be concurrent work, or a separate issue?

Is there a convenient size of data set that is useful for testing? I can have Erica acquire a 2 x 2 tiled data set in the middle of a brain, which should cover several chunks in x, y, and z. Alternatively, we could make a synthetic data set for that.

[camilolaiton]

Hello Mike,

I agree with the solution outline. In regards to the open questions/suggestions:

• Is there a TiledXYZ|3Dseries? R/ I haven't seen a TiledXYZ|3Dseries but there is TIFF (tiled, 3D) in paraconverter which outputs x-y-z tiles. Right now, we're using TIFF (tiled, 4D) which allows us to output tiles c-x-y-z and the folder structure facilitates us to read data per channel using Dask. documentation. Originally, parastitcher did not allow the TIFF (Tiled, 4D) option. Therefore, I included it in the following commit. However, I just tested it in the console, it's not in this repository's source code since I'm launching paraconverter. The main idea is that we only had to worry about a single script. I will update this in the future...
• About the pyramids, we're using xarray-multiscale to generate a pyramid representation of the original image. I believe this can give us more options in terms of image pyramid algorithms (in the case we want to change it in the future). Also, we provide this algorithm as metadata in the OMEZarr .zattrs. However, we can switch to TeraStitcher's --resolutions option.
• Is there a convenient size of data set that is useful for testing? R/ I'm using a multichannel toy dataset that TeraStitcher has in the documentation, but it would be extremely useful to have our own toy dataset based on our actual requirements.

[miketaormina]

I agree with you point about using the xarray-multiscale -- we should keep as much control over algorithms as we can.

John is acquiring a bunch of 1 x 2 tiled data sets right now as part of a photo-bleaching experiment. Maybe one of these would be good enough? They might be kind of large in the z direction though, so maybe we should acquire one specifically for this purpose.

[sharmishtaa]

Switch stitching parameters, increase search region and try other algorithms in terastitcher

[camilolaiton]

Closing issue with #9