Using Cellpose for IMC data

[antopHU]

Hi,

I'm writing to ask for help regarding the use of Cellpose on Imaging Mass Cytometry (IMC) data. I have approximately 22 channels attributed to the cytoplasm and 2 nuclear markers. I'm using the model pretrained on the TissueNet dataset. However, I'm not obtaining meaningful results because the cytoplasmic channels seem to be under-evaluated by the program; only extreme values of cell probability expand the mask beyond the nucleus.

I was wondering if there are specific settings or code modifications that could improve the results. Additionally, I'm working with pancreatic ductal adenocarcinoma (PDAC) tissues, and compared to images processed with Ilastik and CellProfiler, my results tend to predict overly round cells. Is there a way to adjust the program to produce more irregularly shaped masks?

Thank you in advance

[carsen-stringer]

the issue is that the tissuenet model is trained on cells that are mostly round - you can try other built-in models or try training your own model. if training your own model doesn't work, please post an issue here. alternatively/additionally this might be a good question for https://forum.image.sc where there might be forum posters working with the same data

[antopHU]

Hi,

Thank you for your response.

I'm currently working on collecting data as I plan to train my own model. Following your advice, I’ve also posted on the forum at image.sc.

I have one more question: I’m wondering how to extract the underlying data from the "cell pose" plot generated by the plot.show_segmentation function. Sometimes, I notice significant differences between the segmentation displayed in that plot and the actual mask that’s produced. Could you explain why this discrepancy occurs?