ngreenwald
commented
8 months ago @JLrumberger, I talked with @raymond301 at the conference I presented at a couple months ago.
My sense is that right now, all of the I/O is baked around the directory structure, but the actual model inference doesn't depend on this at all. The simplest would be to just call the single-channel inference directly. However, the normalization computation also depends on the folder structure, right?
Moving forward, it's definitely the case that people with OME-TIFF or other TIFF-based formats will want to use it. What would be the easiest way to abstract the data storage format so this isn't an issue?
JLrumberger
raymond301
At my institution we run multiple platforms and have developed a centralized infrastructure to support multiple investigators by employing an internal, institutional instance of OMERO. https://www.openmicroscopy.org/omero/
As such we have chosen to standardize out multiplex imaging into OME.TIFF format. I have a ETL processes that converts acquired images, from which ever image acquisition machine used into standardized OME.TIFF format. 1 channel per marker. Including structured metadata. Here is an example of a typical image's format.
As such I have modify the workflow to generate DeepCell Mesmer segmentation masks on those files by giving the memser application the same file for nuclear and membrane input, and selecting my chosen index separately, like so:
My directory structure looks like this:
Can you help me to find away to leverage my image data as input to Nimbus, in it's current OME.TIFF format? I'd rather not write an ELT to 'unpack' a single file OME into 64 individual files for each and every FOV if I can avoid it.