mic-dkfz/nnunet

★ 8,416PythonAudience · researcherComplexity · 4/5Setup · hard

Mindmap

mindmap
  root((nnunet))
    What it does
      Auto configuration
      Medical segmentation
      2D and 3D support
    Workflow
      Dataset preparation
      Auto preprocessing
      Model training
      Inference on new data
    Use cases
      Tumor segmentation
      Organ boundary labeling
      Competition baseline
    Setup
      PyTorch first
      pip install
      Env path config

mindmap root((nnunet)) What it does Auto configuration Medical segmentation 2D and 3D support Workflow Dataset preparation Auto preprocessing Model training Inference on new data Use cases Tumor segmentation Organ boundary labeling Competition baseline Setup PyTorch first pip install Env path config

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Train an AI model to segment tumors or organ boundaries in MRI or CT scans without manually choosing a network architecture.

USE CASE 2

Use nnU-Net as a strong out-of-the-box baseline when entering a biomedical image segmentation competition.

USE CASE 3

Run inference on a new set of medical images using a trained nnU-Net model.

USE CASE 4

Benchmark a new segmentation method against nnU-Net to demonstrate improvement over the self-configuring baseline.

Tech stack

PythonPyTorchpip

Getting it running

Difficulty · hard Time to first run · 1day+

Requires a GPU, PyTorch installed for your specific hardware, and three environment variables pointing to raw data, preprocessed data, and results directories.

In plain English

nnU-Net is a Python framework for automatically training AI models that identify and label regions within medical images. The task it is designed for is called semantic segmentation: given an image, the model labels each pixel or voxel according to what it depicts, such as identifying a tumor region, an organ boundary, or a specific tissue type. What makes nnU-Net different from most AI frameworks is that it configures itself. You give it a labeled training dataset, and it analyzes the data to understand its characteristics, including image dimensions, spacing, and size. Based on that analysis it automatically decides what kind of network architecture to use, how to preprocess the data, and how to train the model. This removes a large amount of manual tuning that is usually required when applying deep learning to a new imaging problem. The framework supports both 2D and 3D image data and handles a variety of medical image formats. It works well for biomedical research problems, competition datasets, and other specialized imaging tasks where general-purpose models trained on natural photographs are not a good fit. Researchers also use it as a baseline when developing new segmentation methods, because it provides a strong default result without customization. Installation requires installing PyTorch first for your specific hardware, then installing nnU-Net via pip. The full setup involves configuring a few environment paths for raw data, preprocessed data, and saved model results. Detailed documentation covers dataset preparation, training, running inference on new images, and contributing to the project. nnU-Net is developed at the German Cancer Research Center and Helmholtz Imaging. The original paper describing the method was published in Nature Methods in 2021. Users are asked to cite that paper when publishing work that uses nnU-Net.

Copy-paste prompts

Prompt 1

I have a labeled MRI dataset in NIfTI format. Walk me through the complete nnU-Net workflow from dataset preparation to training a model to running inference on new scans.

Prompt 2

How do I configure the three environment paths nnU-Net requires for raw data, preprocessed data, and saved model results?

Prompt 3

nnU-Net automatically chooses between 2D, 3D full-resolution, and 3D low-resolution architectures. How does it make that decision and how do I override it?

Prompt 4

I want to cite nnU-Net in a medical imaging paper. What is the correct citation and which benchmark configurations should I report?

Open on GitHub → Explain another repo

← mic-dkfz on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.