See documentation :-)
Reworked inference code to be more flexible
Minor bug fixes
nnU-Net V2 delivers the same segmentation performance, but has now been re-worked from the ground up to optimize for usability and facilitate modifications for method developers.
nnU-Net plans are now more powerful and simpler at the same time. Making manual adjustments to the U-Net configurations is much easier. Configurations can inherit from each other, allowing rapid experimentation. You can chain configurations in custom cascades.
No longer are you bound to use NIFTI files. Datasets can be provided in a wide range of imaging formats and you can even add your own I/O adapters to make nnU-Net compatible with your favorite format!
Hierarchical labels, such as the tumor regions used in BraTS, are now natively supported!
Trainer classes are unified: no more messing around with special trainers for distributed training, cascaded training and hierarchical labels. nnUNetTrainer is all you need.
As part of our usability upgrade, nnU-Net V2 supports Linux, Windows and macOS. It can be run on CUDA capable GPUs, CPU and Apple MPS devices.
nnU-Net V1 and can V2 be installed at the same time. There stands nothing in your way for trying it out! Existing datasets (not trained model weights!) can be directly transferred from V1 to V2.
v1.7.1