Normalization for patches ome-zarr

[alishbaimran]

We want to compute statistics from the FOV-scale zarr store and store it with patch-scale zarr store, which dataloader will parse.

The process for this is:

• Compute statistics per FOV (not patch) and store these with metadata.
• Normalize using FOV statistics at training and test time - use existing CLI or preprocessing script.

This is current structure of the patches ome_zarr:

=== Summary === Format: omezarr v0.4 Axes: T (time); C (channel); Z (space); Y (space); X (space); Channel names: ['RFP', 'Phase3D'] Row names: ['A', 'B'] Column names: ['3', '4'] Wells: 4 Positions: 2629

This is structure of track_labels zarr where we can store the meta data: (fov information can be stored here) === Summary === Format: omezarr v0.4 Axes: T (time); C (channel); Z (space); Y (space); X (space); Channel names: ['tracking'] Row names: ['A', 'B'] Column names: ['3', '4'] Wells: 4 Positions: 61

Need help on how to integrate metadata into the track ome_zarr, how this will be used to generate the patch ome_zarr and then eventually the dataloader for normalization.

[mattersoflight]

Need help on how to integrate metadata into the track ome_zarr, how this will be used to generate the patch ome_zarr and then eventually the dataloader for normalization.

@alishbaimran our initial approach can be

• compute FOV statistics from track_labels.zarr/row/col/fovN/ array.
• write them to each cropped FOV in patches.zarr in /row/col/fovNcellM/.zattr file.
• use the statistics while loading patches for normalizing the data.