++nnU-Net boosts medical segmentation by 22% with registration-based augmentation

Medical image segmentation models like nnU-Net depend heavily on large, diverse annotated datasets, but privacy regulations and high annotation costs often limit data availability. To address this, a team of researchers from Portugal, Germany, and Austria have introduced ++nnU-Net, a novel data augmentation module that operates before preprocessing and training. Instead of standard random augmentations, ++nnU-Net applies a two-stage image registration process to existing samples, generating realistic warped images and corresponding segmentation masks. The pipeline also automatically computes available disk space, generates supplementary binary synthetic masks, and creates checkpoints to ensure robustness.

Evaluated on five distinct 2D medical imaging datasets, ++nnU-Net consistently outperformed the standard nnU-Net baseline. In the most prominent cases, Dice Similarity Coefficient scores improved by approximately 22%. These results demonstrate that registration-based augmentation preserves anatomical plausibility while effectively increasing training data diversity. The method is particularly valuable for data-limited settings, offering a practical and scalable approach to enhance segmentation performance. The full source code is available on GitHub, enabling easy adoption by the medical imaging community.