search

BodenmillerGroup / steinbock

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

Update the CellProfiler pipeline for pixel classification based segmentation #191

Open nilseling opened 1 year ago

nilseling commented 1 year ago

The CellProfiler pipeline used in steinbock does not match the current pipeline in the ImcSegmentationPipeline.

jwindhager commented 1 year ago

Out of curiosity: in what way?

nilseling commented 1 year ago

So steinbock is using the old pipeline where the probability masks were first segmented and then the detected objects were downscaled. This leads to issues that cells are lost if they are smaller than 4 pixels or so.

The pipeline in the newest version of the ImcSegmentationPipeline (version 3) first downscales the pixel probabilities and then performs segmentation. This leads to continuous numbering of cell IDs.

nilseling commented 1 year ago

Hmm, I can't reproduce this issue anymore. For the test data the cell IDs seem to be continous with the current pipeline. But I remember that there was an issue in the past when reading in segmentation masks that did not contain continous cell IDs. In any case, it would be good to update the pipeline to what the ImcSegmentationPipeline is doing.

Milad4849 commented 1 year ago

It appears to me that the only relevant difference between the two pipeline is the Typical artifact diameter parameter, set to 4 in steinbock and 2 in ImcSegmentationPipeline. Ca you possibly confirm that @nilseling

nilseling commented 1 year ago

No, the two pipelines are conceptually quite different. We can have a look at it together tomorrow