nnU-Net is a semantic segmentation framework that automatically adapts its pipeline to a dataset. It analyzes the training data, creates a dataset fingerprint, configures suitable U-Net variants, and provides an end-to-end workflow from preprocessing to training, model selection, and inference.
It is primarily designed for supervised biomedical image segmentation, but it also works well as a strong baseline and development framework for researchers working on new segmentation methods.
If you are looking for nnU-Net v1, use the v1 branch. If you are migrating from v1, start with the TLDR migration guide.
- First-time setup: Installation and setup
- First run on your own data: Getting Started
- Task-oriented docs: How-to Guides
- Formats, commands, and configuration details: Reference
- Concepts and rationale: Explanation
Install PyTorch for your hardware first, then install nnU-Net:
pip install nnunetv2For the full setup, including nnUNet_raw, nnUNet_preprocessed, and nnUNet_results, see Installation and setup.
Start with the documentation home.
Useful entry points:
- New users: Getting Started
- Dataset preparation: Prepare a dataset
- Training workflow: Train models
- Inference workflow: Run inference
- Recommended residual encoder presets: Residual Encoder Presets in nnU-Net
- Contributing: CONTRIBUTING.md
nnU-Net is built for supervised semantic segmentation. It supports 2D and 3D data, arbitrary channel definitions, multiple image formats, and dataset-specific adaptation of preprocessing and network configuration.
It performs particularly well in training-from-scratch settings such as biomedical datasets, challenge datasets, and non-standard imaging problems where off-the-shelf natural-image pretrained models are often a poor fit.
For a concise overview of the design, see How nnU-Net works.
Please cite the following paper when using nnU-Net:
Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021).
nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation.
Nature Methods, 18(2), 203-211.
Additional recent work on residual encoder presets and benchmarking:
- nnU-Net v2 is a complete reimplementation of the original nnU-Net with improved code structure and extensibility.
- Not every dataset creates every configuration. For example, the cascade is only generated when the dataset characteristics justify it.
- Detailed historical changes are summarized in What is different in v2?.
nnU-Net is developed and maintained by the Applied Computer Vision Lab (ACVL) of Helmholtz Imaging and the Division of Medical Image Computing at the German Cancer Research Center (DKFZ).