Normalization of imaging based transcriptomic data

[Pointillomic]

Hello!

This might be a naive question, but why is it that most single cell methods perform a size factor/library size scaling during the normalization step, even for imaging based methods?

I can understand that with sequencing based methods, better covered cells might take up reads and out compete the more poorly covered cell types, but scaling them all based on coverage seems to be making the assumption that gene expression is roughly the same for every cell. This is particularly confusing for imaging based methods as they won't have issues with better covered cells taking up more signal from less well covered cells. Furthermore, this cell level normalization would likely distort visualization of expression across the tissue, which I imagine should be more similar to the raw count plots (similar to regular in-situ hybridization experiments). For the purposes of dealing with some technical variability while getting expression values, I was considering pooling cells and normalizing only within pools (without scaling the pool back to the average cell). However, I was worried that this would cause other issues. For clustering, I would of course use variance stabilizing normalization like the pearson residuals.

I have been racking my brain about different normalization methods and what they assume/want to achieve, so it would be great to get more input! Also, please let me know if I seem to be misunderstanding something - any thoughts on this would be greatly appreciated!

[Pointillomic]

Also, for clarification, I am looking at this in the context of the CosMX platform, since there seem to be quite a bit of variability even within image based platforms