MONAILabel Endoscopy App – Multilabel Segmentation Inference from CVAT Not Working

[bilalcodehub]

Dear Team,

I’ve configured MONAILabel with the endoscopy app for multilabel segmentation. The server reads labels from a text file and creates CVAT tasks with custom labels. After annotation, active learning runs, trains a new model, and updates CVAT by adding a new task and marking the previous one as completed.

However, when using model-assisted labeling on a new task, it only segments one label. This seems to be because the default tool segmentation model in the endoscopy app is designed for a single class. I’ve modified the app to read custom labels with adjusted colors but haven’t changed the network’s output shape.

My questions:

1. Does MONAILabel automatically adjust the model’s output shape based on the defined labels, or do I need to change network configurations manually?
2. If manual adjustment is required, which file should I modify as currently I don't see the model config anywhere?
3. Can I switch the endoscopy app’s model from a simple U-Net to another model supported by MONAILabel’s core library?

Any guidance would be appreciated!

Many thanks and
Kind regards,
Bilal