Segmentation of highly crowded images

[SarkkinenJ]

Creating ground truth for highly crowded images

Hi all, and a huge thanks for developers of StarDist! In an other project StarDist ensured the succession of the project, when it made segmentation of rather crowded images possible.

Currently I´m dealing with more crowded images (TMA sections, 4um thick, IF stained with DAPI stain) and the ready made model for fluorescent nuclei does not perform well in the most crowded areas (here is a link to an example image, and StarDist performance mask with rois as zip, images also as jpeg below). As you can notice, StarDist works well but is missing the most crowded areas. I tried to train my own model but the problem is producing accurate ground truth: labeling of the most crowded areas is tedious for sure but also super challenging doing manually.

• Would you guys have some ideas on how to produce ground truth for images like this?

• Or would people at StarDist (or elsewhere) have already some material which would be possible to use to train more efficient algorithm for crowded images?

If someone could help me with the issue I would be obviously happy either to share the model or the nuclei images. RoiSet.zip

Best, Joona

[maweigert]

Hi @SarkkinenJ ,

Hi all, and a huge thanks for developers of StarDist! In an other project StarDist ensured the succession of the project, when it made segmentation of rather crowded images possible.

thanks, great to hear :)

* Would you guys have some ideas on how to produce ground truth for images like this?

Yes, that looks pretty tough and the most crowded regions you mention seem to be almost impossible to annotate. If you do have annotations of a little bit less crowded regions, you could try applying random gaussian blur augmentations during training (e.g. with augmend and see whether that would help for prediction. But then again, if you as a human cannot delineate objects, then it is going to be hard to judge the quality of these predictions.

Hope that helps!

M