Region-based training using already existing ROIs

[fohofmann]

Hi everyone, Hi Fabian,

I was wondering if there is an efficient (and maybe already implemented) option to use region-based training and region-based inference, if the outermost label is already existing (e.g. from already established models). In other words, is there an option to train a model for a sub-segmentation of an existing segmentation?

Example:
Input: CT Abdomen
Model 1 (already existing and validated): Segmentation of liver
Model 2 (to be trained): Segmentation of Liver tumor using the existing labels (in both, training and inference) from Model 1

Any ideas?
Thank you, and all the best,
Felix