Hi Genevieve,

The orthoplane consensus algorithm works on run length encoded segmentations, which makes it very fast but requires conversion from the regular labelmap. Normally the run length encodings are generated on 2D slices immediately after model inference and then discarded when the consensus algorithm finishes (there's no reason I couldn't add an option to save them though).

Before getting started, I recommend upgrading to the newest version of empanada:

pip install --upgrade empanada-napari==0.2.3

Here are two scripts that should work for you assuming you used one of the MitoNet models and stored the results in zarr.

This one converts a labelmap stored in zarr to a run length encoding json file:
https://gist.github.com/conradry/7376e7504456e9c194638c78d009523e

The only parameter to watch out for is the label_divisor. This is the value of "Max objects per class in 3D" from the napari plugin. Probably this was left as default so either 1,000,000 or 100,000 depending on your version (you'll get an error if you guess incorrectly).

This one accepts a list of json files and runs the consensus algorithm.
https://gist.github.com/conradry/67691f226093539edf60f0e4265191b2

A sneaky secret about the consensus algorithm is that it works for as many segmentations as you give it. (E.g. you can mix-and-match segmentations from different models, downsampling factors, confidence thresholds, etc.)

For help with the parameters see:
https://github.com/volume-em/empanada/blob/01c6e7aa3ad0e3c3334df8b129b0122724b6ad2e/empanada/consensus.py#L348-L377

Let me know if anything is unclear or doesn't work.

Best,
Ryan

Thanks for the suggestions Ryan!