Isotope Spillover

[DarioS]

Spillover due to isotopes is a probem for imaging mass cytometry. Has the training data been corrected before modelling?

[rossbar]

This is a good question with a potentially quite complex answer! However, I'll do my best to share a bit of the mental model for the deepcell-types approach (as I understand it).

First, the easy part of the question:

Has the training data been corrected before modelling?

No - or, at least not explicitly. We don't necessarily have control over what processing folks do to data before they publish it, so there's always some inherent variability there based on whether/how data were preprocessed before publishing. The only explicit preprocessing done by the deepcell-types pipeline itself is some combination of:

• precentile thresholding in a rudimentary effort at outlier rejection
• Image rescaling to a universal scale to reduce variability in image feature size across datasets
• Techniques for attempting to reduce variability in contrast distributions per-channel, i.e. histogram normalization. This may or may not be used (IIRC histogram normalization is not currently being used in the development version of the pipeline).

Ultimately, these preprocessing steps are focused only on reducing spatial and contrastive variability across datasets. The goal is to leverage the power of the vision transformer to "learn" the visual features themselves.

What this means in practice is that I'd expect the model to be resilient to spillover if the visual effects of spillover are well-represented in the training data. I don't have a ready answer to this question - I suspect that IMC is among the least-well represented types of image in the training data (see the publication for info on how many training instances are drawn from codex, mibi, imc, etc.). That's just a function of the amount of publicly-avaialble IMC data we've managed to find - any leads on IMC data that could be incorporated for training would be appreciated!

[DarioS]

Have you incorporated TRACERx consortium's forty thousand lung cancer manual segmentations yet?

[rossbar]

Have you incorporated TRACERx consortium's forty thousand lung cancer manual segmentations yet?

I don't believe we have, though some of the datasets mentioned there (the Hickey intestine and barrett's esophagus images e.g.) are part of the deepcell-types training data. I defer to @xuefei-wang for further details!

[xuefei-wang]

Thanks for the question and the pointer to the TRACERx dataset!

To confirm, we do not perform explicit spillover correction. The current approach assumes that by training on a diversity set of data, the model learns to be robust to such artifacts, if it is well represented in the data. And as Ross mentioned, we don't always know the full preprocessing history of the public datasets we aggregate - there might be additional correction steps done by the authors of the data, feel free to check out the original papers, as cited in our preprint!

Regarding TRACERx, we have not included it yet, and we really appreciate you bringing it to our attention. We will definitely look into adding it to our future training sets. Thanks again!