Adaptation to EM samples with extracellular space

[erjel]

Hi,

I wonder whether there are ideas around to deal with 'non-dense' EM datasets. In the ISBI dataset (left) the objects are directly touching each other. In my dataset (right) there is quite a lot of extracellular space which separates the single objects.

In the first (prototype) pipeline I could not see any option to mark a certain label (let's say 0) as background. This leads to a mws segmentation in which the background is split into multiple objects. The only way (I can think of) would require post-processing to discriminate between foreground and background. Anything (even just a pointer) how I can deal with such (partly) separated objects would be super helpful ...

Best wishes,
Eric

[constantinpape]

Hey,
yes we have a bit newer version of the pipeline that can also deal with foreground objects.
The idea is to also predict another channel in the network that predicts a foreground / background mask.
I have implemented this in a different repository: https://github.com/constantinpape/torch-em
You can find an example for mitochondria segmentation, that should be pretty close to your use-case, here:
https://github.com/constantinpape/torch-em/tree/main/experiments/platynereis/mitochondria

The installation instructions are in the top-level readme. Let me know if you run into any issues.