Can region-based training support partly overlapping regions

[CaB15]

I have made several segmentation modes using nnUnet and am incredibly pleased with how they work.

I have only been using label-based training, not region-based training (see https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/region_based_training.md). I am slightly doubtful about the aim of region-based training. Will region-based training and inference support delineations that are partly overlapping or only delineations that are subsets of each other? My current understanding is that it aims at subsets, although I initially believed it was aimed at partly overlapping structures.

In my case, I have two regions, A and B, which can overlap but are not necessarily subsets of each other. Thus, I assume that the network in nnUNet will need at least three label levels (e.g. 1=(A not B), 2= (B not A), and 3=(A and B)) for training. During the data preparation, I can change the A and B delineations to the three levels, but will region-based training support the process in any way? Finally, I would like to make an inference that results in delineations matching A and B, which might overlap in certain regions. But am I right that the inference will "only" output the three levels (thus nnUnet has a single label output per voxel), and then I need to combine them in a local post-processing, or can regions training support this in any way?

Best,
Carsten