Prepare a Dataset

This guide covers the dataset preparation step before experiment planning and training.

Required format

nnU-Net expects datasets in the nnU-Net dataset format. Start with the concise reference here:

• Dataset and input format reference

The key points are:

• each dataset lives in nnUNet_raw/DatasetXXX_Name
• training images go into imagesTr
• training labels go into labelsTr
• optional test images go into imagesTs
• dataset.json describes modalities and labels

Existing data sources

If you already have data in a different layout:

• Medical Segmentation Decathlon: Convert MSD datasets
• nnU-Net v1 datasets: use nnUNetv2_convert_old_nnUNet_dataset

Input formats

nnU-Net v2 supports multiple file formats. The exact supported formats and image I/O details are documented in:

• nnU-Net dataset format

Inference inputs

Inference input folders follow the training dataset's naming and file-ending conventions. See:

• Dataset and input format reference

Next step

Once your dataset is in place, continue to Plan and preprocess.