nnU-Net

nnU-Net automates deep learning-based 3D biomedical image segmentation by configuring preprocessing, network architecture selection, training, and post-processing to adapt to diverse imaging datasets.

Key Features:

• Deep learning segmentation: Provides 3D semantic segmentation of biomedical images using deep learning.
• Automated preprocessing: Automatically configures preprocessing tailored to dataset properties such as image sizes and voxel spacings.
• Automated architecture selection: Selects and configures network architectures based on dataset characteristics.
• Automated training: Automates training procedures for the configured networks.
• Automated post-processing: Applies dataset-adaptive post-processing steps to segmentation outputs.
• Rule-based configuration: Uses fixed parameters, interdependent rules, and empirical decisions to adapt configurations to new tasks.
• Dataset variability handling: Designed to handle variability in imaging modality, image sizes, voxel spacings, and class ratios.
• Benchmark performance: Demonstrated superior performance across 23 public biomedical segmentation datasets from international competitions.

Scientific Applications:

• Biomedical image segmentation: Semantic segmentation of 3D biomedical images across modalities and variable dataset properties.
• Research and clinical imaging analysis: Applied to image analysis tasks relevant to scientific discovery and medical care.
• Benchmarking and competitions: Serves as a baseline and state-of-the-art method in international biomedical segmentation competitions.

Methodology:

Performs deep learning–based semantic 3D image segmentation with automated preprocessing, network architecture selection, training, and post-processing governed by fixed parameters, interdependent rules, and empirical decisions.