We introduce a new challenging task: for unseen biomedical domains, generate label maps from multiple protocols that are consistent across subjects.

Pancakes addresses it through a protocol-sampling mechanism and outperforms foundation segmentation models.

There are many ways (protocols) to segment a biomedical image. The region you want to segment often depends on the downstream tasks.

Therefore, it is useful to produce multiple plausible segmentation maps. What we care three aspects:

Accuracy: What is the overlap between the ground truth and the prediction?

Consistency: Are label maps from the same protocols yielding the same structures with the same labels for different images?

Diversity: Can the model generate a wide choice of different label maps?

We introduce a new challenging task: for unseen biomedical domains, generate label maps from multiple protocols that are consistent across subjects. Pancakes addresses it through a protocol-sampling mechanism and outperforms foundation segmentation models. Check our paper if you are curious :)

Pancakes produces consistent label maps. Each color corresponds to a different label ID.