Cross-entropy loss weighted by distance from the surface

[chrisrapson]

The ideas in this paper look intuitively promising to me. It has a reasonable number of citations, so I guess I'm not the only one. Has anybody looked at integrating it with nnunet?
Multiclass Weighted Loss for Instance Segmentation of Cluttered Cells by Guerrero-Pena et. al., ICIP 2018

The idea I like best tries weighting the cross-entropy by the distance from the surface of an object, so that the network concentrates more on learning the boundaries between labels. Perhaps nnunet's method of selecting patches for training already takes care of that?

Adding a new class corresponding to borders between classes isn't as intuitive to me. It seems to be motivated by ambiguity of class labels for a voxel on the border. However, if the labels were created or visualised as masks, then there should be no ambiguity and borders run between voxels?