search

MIC-DKFZ / nnUNet

Apache License 2.0
5.86k stars 1.75k forks source link

Customizing training parameters for stronger focus on structural information #2536

Open code8865 opened 3 weeks ago

code8865 commented 3 weeks ago

Hi there,

I’m currently working with “intentional manipulated” CT data. This manipulation can cause voxel regions to appear unique. I would like the network to "reinterpret" these areas based on learned structural information, by assigning the appropriate labels in line with the expected anatomy.

nnUNet’s default setup handles such problems reasonably well. However, I’m looking for ways to optimize the network training so that it focuses even more on structural consistency or spatial relationships.

Does anyone know methods (e.g. another loss function) to focus more on these structural information when dealing with unexpected data deviations?

Thanks a lot in advance!

ykirchhoff commented 2 days ago

Hi @code8865,

I am not sure I understand what you are exactly trying to do here, could you maybe give me some more details?

Best, Yannick