search

BodenmillerGroup / steinbock

A toolkit for processing multiplexed tissue images
https://bodenmillergroup.github.io/steinbock
MIT License
49 stars 14 forks source link

Lineage-aware ground-truth segmentation using deepcell within steinbock #193

Closed megleila closed 1 year ago

megleila commented 1 year ago

I am working with kidney tissue, and trying to adopt a way of segmenting cells according to cell type. Tubular cells are quite different from infiltrating immune cells, and this heterogeneity of cell characteristics across the tissue sample makes segmentation challenging. The DeepCell literature reports that lineage-aware ground-truth segmentation is possible for improved accuracy in tissues with complex morphologies that cannot be accurately captured with only two channels of imaging data, ie using a single membrane channel (2) and nuclear channel (1) in the panel.csv file under "deepcell" after the preprocessing step. How would I go about including multiple markers that are cell-type specific to provide differential information about cell morphology during model training that is lost when they are combined into a single channel?

https://www.nature.com/articles/s41587-021-01094-0#Sec9

Milad4849 commented 1 year ago

Hi @megleila, Steinbock incorporates the standard release of deepcell which does not come with the version of the model that takes 6 separate input channels, as mentioned in the paper (I don't know where to find that version of the model). A relevant package offering a "cell type-optimised segmentation strategy" might be useful.

Milad4849 commented 1 year ago

Just a follow-up, I contacted the author of deepcell and he confirmed that the lineage-aware model mentioned int he paper is not available. "It wouldn't work as a general purpose model to generate segmentations the way Mesmer does, since it's not nearly robust enough".

megleila commented 1 year ago

Thank you so much for this follow up, and for the reference to the paper/package using sequential segmentation.

Sent from my iPhone

On Jun 13, 2023, at 3:41 AM, Milad4849 @.***> wrote:



Just a follow-up, I contacted the author of deepcell and he confirmed that the lineage-aware model mentioned int he paper is not available. "It wouldn't work as a general purpose model to generate segmentations the way Mesmer does, since it's not nearly robust enough".

— Reply to this email directly, view it on GitHubhttps://github.com/BodenmillerGroup/steinbock/issues/193#issuecomment-1588725447, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BAH2R6ISBIU3HGZ52I743DDXLAKTBANCNFSM6AAAAAAYXBSQZU. You are receiving this because you were mentioned.Message ID: @.***>