nnU-Net — Self-Configuring Medical Segmentation

nnU-Net — Self-Configuring Medical Image Segmentation

nnU-Net (no-new-net) is a self-configuring deep learning framework for medical image segmentation from the German Cancer Research Center (DKFZ). Feed it any labelled 3D medical dataset and it automatically determines optimal preprocessing, network topology, training scheme, and post-processing — consistently achieving top results in biomedical segmentation challenges (over 30 MICCAI competition wins).

Key Features

• Fully automated configuration: spacing, normalisation, patch size, batch size, architecture
• Three network configurations: 2D U-Net, 3D full-resolution U-Net, 3D U-Net cascade
• Handles CT, MRI, ultrasound, histology, microscopy
• Ensemble inference and test-time augmentation
• Multi-GPU training via PyTorch DDP
• Built-in cross-validation and model selection

Quick Start

pip install nnunetv2
export nnUNet_raw=/data/raw
export nnUNet_preprocessed=/data/preprocessed
export nnUNet_results=/data/results

# Preprocess and train
nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity
nnUNetv2_train DATASET_ID 3d_fullres 0

# Inference
nnUNetv2_predict -d DATASET_ID -i /input -o /output -f 0 -c 3d_fullres

npx ai-supply add nnunet-medical-image-segmentation

Curated mirror of the open-source nnU-Net (Apache-2.0). Get it from the source.