nnUNet

nnUNet performs self-configuring deep learning-based biomedical image segmentation by automatically adapting preprocessing, network architecture, training schedules, and postprocessing to new datasets.

Key Features:

• Self-configuration: Automatically adapts preprocessing, network architecture, training schedules, and postprocessing to each dataset.
• Network architecture: Constructs task-specific 2D/3D U-Net variants.
• Memory-aware optimization: Adjusts patch size and batch size to match memory limits.
• Preprocessing strategies: Selects intensity normalization and resampling strategies.
• Training configuration: Configures loss functions and augmentation policies automatically.
• Design rules and heuristics: Formalizes design decisions as fixed rules and interdependent heuristics derived from extensive empirical benchmarking.
• Dataset property adaptation: Accounts for modality, resolution, class imbalance, and hardware constraints when configuring models.

Scientific Applications:

• Biomedical image segmentation: Automated segmentation across imaging modalities and resolutions for clinical and research datasets.
• Benchmark evaluation: Demonstrated performance outperforming specialized models across 23 public challenge datasets.

Methodology:

Automatically adapts preprocessing, network architecture, training schedules, and postprocessing; constructs task-specific 2D/3D U-Net variants; adjusts patch size and batch size for memory limits; selects intensity normalization and resampling strategies; and configures loss functions and augmentation policies, using fixed rules and interdependent heuristics derived from empirical benchmarking.