Hi SCALLOPS Team! I have a use case that I think is slightly different from your standard experiment. I suspect some code changes will be needed to support my use case, but I would appreciate if you have any guidance. Also please let me know if these changes are something you would be interested in having in your code base.
My sequencing data matches your standard assumptions and are collected as overlapping tiles that cover the entire well. My phenotyping data is collected as live cell data. As a result, it is a limited set of FOV that do not cover the entire well and the data has a t dimension.
Illumination correction
The first issue I encountered is during illumination correction where the t dimension is squeezed out. If I understand correctly, typically at this point the data would have a t dimension corresponding to a single round of sequencing and therefore that dimension could be squeezed out in this line.
|
image = image.squeeze(d for d in ("t", "z") if d in image.dims) |
My first thought is that there are two ways to get around this. 1) Create an option to recast the t dimension from live imaging as a z dimension which preserves the t dimension for sequencing rounds. 2) Create a new third dimension for the t that comes from live imaging. However I suspect this will be difficult to make compatible with the OME Zarr specification. What do you think would work best?
Stitching
I haven't had a chance to test this step out yet. However, I'm curious if you think the stitching pipeline can handle sparse tiles that are all non-overlapping. I could imagine an implementation that would place them in real world space and leave zeros where there is no data. Do you have a sense of how far the current implementation is from supporting this case?
Registration
My expectation for registration is that if I can get a stitched image prepared it should fit into the registration workflow regardless of if there are empty tiles. Is that correct?
Hi SCALLOPS Team! I have a use case that I think is slightly different from your standard experiment. I suspect some code changes will be needed to support my use case, but I would appreciate if you have any guidance. Also please let me know if these changes are something you would be interested in having in your code base.
My sequencing data matches your standard assumptions and are collected as overlapping tiles that cover the entire well. My phenotyping data is collected as live cell data. As a result, it is a limited set of FOV that do not cover the entire well and the data has a t dimension.
Illumination correction
The first issue I encountered is during illumination correction where the t dimension is squeezed out. If I understand correctly, typically at this point the data would have a t dimension corresponding to a single round of sequencing and therefore that dimension could be squeezed out in this line.
scallops/scallops/illumination_correction.py
Line 168 in 41b13e9
My first thought is that there are two ways to get around this. 1) Create an option to recast the t dimension from live imaging as a z dimension which preserves the t dimension for sequencing rounds. 2) Create a new third dimension for the t that comes from live imaging. However I suspect this will be difficult to make compatible with the OME Zarr specification. What do you think would work best?
Stitching
I haven't had a chance to test this step out yet. However, I'm curious if you think the stitching pipeline can handle sparse tiles that are all non-overlapping. I could imagine an implementation that would place them in real world space and leave zeros where there is no data. Do you have a sense of how far the current implementation is from supporting this case?
Registration
My expectation for registration is that if I can get a stitched image prepared it should fit into the registration workflow regardless of if there are empty tiles. Is that correct?