Methods to stitch a position grid

[ieivanov]

She is ready!

This PR adds methods to stitch a grid of positions. Stitch parameters are computed with >>> mantis estimate-stitch. Images are then stitched with >>> mantis stitch. Image registration parameters are computed on pairs of images using phase cross-correlation; the pipeline can be extended to include more complex transformations. The image mosaic is compiled using "average" blending; other blending modes may be implemented later.

See examples in /hpc/projects/intracellular_dashboard/ops/2024_03_05_registration_test/live/3-stitch/ and /hpc/projects/intracellular_dashboard/ops/2024_04_11_Manual_HELA/1-stitch/v3/

Replaces #135
See #135 (comment)

TODO:

• Convert compute_image_translation to CLI and save in config file
• Test with non-square image and non-square grids

[talonchandler]

Big job! Looks great. I like the slurm job dependencies.

I'm in the middle of recreating your 2024_03_05_registration_test/live/3-stitch/ (~20% through the final step filling the last store) and everything's going smoothly.

Here are the small snags I've run into:

• --slurm parameter tripped me up on the stitch call. Seems required, so I suggest dropping it for now?
• the default --temp-dir wasn't writeable for me. Default to ./?
• When I open the stitched zarr in napari I get GL_MAX_TEXTURE_SIZE 16384 in at least one axis and will be downsampled., and the viewable resolution is noticeably worse than the originals. Do we need pyramids? Another trick? I don't think this needs to block this merge, though.

[ieivanov]

Thanks for your review @talonchandler! I implemented your suggestions for --temp-path and --slurm options. Let's merge this and we'll fix other issues through smaller PRs.

When I open the stitched zarr in napari I get GL_MAX_TEXTURE_SIZE 16384 in at least one axis and will be downsampled., and the viewable resolution is noticeably worse than the originals. Do we need pyramids? Another trick? I don't think this needs to block this merge, though.

That's true - so far this only happens for our largest (OPS) datasets and it hasn't bothered me much. It'll be nice to have a proper fix, which I think is implementing pyramids. This is not a blocker for now - analysis still happens on a FOV basis and detailed inspection can be done on smaller ROIs by loading a selection of the data through python.

[talonchandler]

LGTM! Thanks @ieivanov.