nnU-Net is a semantic segmentation framework that automatically adapts its pipeline to a dataset.
The tool addresses the challenge of configuring segmentation models by analyzing training data to create a dataset fingerprint, then automatically selecting and configuring appropriate U-Net variants. It handles the complete workflow from preprocessing through training, model selection, and inference. Rather than requiring manual tuning for each new dataset, nnU-Net inspects the characteristics of your data and adjusts its architecture, preprocessing steps, and training strategy accordingly. It supports both 2D and 3D data with arbitrary channel definitions and multiple image formats.
nnU-Net is designed primarily for supervised biomedical image segmentation, where it performs particularly well in training-from-scratch scenarios such as challenge datasets and non-standard imaging problems where pretrained models from natural image datasets are often ineffective. It also serves as a strong baseline and development framework for researchers exploring new segmentation methods. The tool is built for datasets where automatic configuration provides an advantage over manual hyperparameter tuning, and it works across diverse biomedical imaging modalities. Not every dataset generates every possible configuration; for instance, cascade configurations are only created when dataset characteristics justify them.
The project maintains comprehensive documentation organized by user need, including getting started guides, how-to instructions for dataset preparation and training workflows, and detailed reference materials. A migration guide is provided for users transitioning from the earlier version. The codebase represents a complete reimplementation of the original framework with improved code structure and extensibility, and the project actively incorporates recent research findings, including work on residual encoder presets and validation methodologies for medical image segmentation.
