Exporting nnUNet design choices to improve the performance of a custom network.

[kitbransby]

I'm interesting in exporting all the design / hyper-parameters / pre-processing choices of a trained nn-UNet pipeline, and using it in a different model. The new segmentation model will share the same encoder, but have a different decoder (graph network).

My motivation is that nn-UNet has configured a pipeline where more powerful segmentation features are encoded from my image dataset, and I may be able utilise these in a new model. As many of the choices are dataset-dependent I'm hoping that they are transferable to new architectures.

However, FabianIsensee mentioned here that it is very difficult to recreate nnUNet performance without using their code. It is how the 'little' things add up.

Interested to know

1. Is this my approach logical?
2. How would I go about doing this?