Semantic segmentation training on sparesly annotated data

[kovecsroli]

Hi all,

I am working on segmenting specific sites on chromatin fibers (see images below). I have sparesaly annotated data, set up, and trained an nnUNet to recognize the interesting areas. However, I get very low accuracy (dice validation score 6%). Did I miss something while creating the masks with the BG being 0 and the foreground masks being 1? I read about the ignore label option and the partial loss, where the network omits the non-annotated sections: https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/ignore_label.md

Am I right that for this to work, I would need to have the non-labelled chromatin parts labelled as a different class, and I could give it the ignore flag? Could it be done without segmenting the chromatin itself and only considering the labelled parts in the loss function?