nnU-Net v2

At a glance

What it does

Train a strong 2D/3D U-Net on your own annotated CT/MRI/PET volumes without architecture engineering; used for tumour, organ-at-risk and lymph-node segmentation in radiotherapy and radiomics research.

Tasks, data types and cancers

Architecture

Training data

No fixed training corpus. nnU-Net ships no pretrained weights: it configures itself and trains from scratch on whatever labelled dataset the user supplies. In the original work it was applied without any manual adjustment between datasets to the Medical Segmentation Decathlon (10 tasks) and to further biomedical segmentation challenges.

Linked datasets

• benchmark BraTS — Brain Tumor Segmentation Challenge Free registration — standard glioma segmentation benchmark where nnU-Net variants placed first repeatedly

Evaluation

How to run

pip install nnunetv2
nnUNetv2_plan_and_preprocess -d 001 --verify_dataset_integrity
nnUNetv2_train 001 3d_fullres 0
nnUNetv2_predict -i imagesTs/ -o preds/ -d 001 -c 3d_fullres

Regulatory status and intended use

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

• Needs voxel-level labels; no pretraining shortcut.
• Long training times; the cascade variant is heavy.

Sources

1. Isensee F et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021
2. Isensee F et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation (preprint)
3. GitHub — MIC-DKFZ/nnUNet