MIC-DKFZ/nnUNet

Features

• Segmentation Model Training - Provides a comprehensive framework for preparing datasets and executing supervised training for biomedical image segmentation.
• Dataset-Driven Configurations - Automatically selects the optimal network architecture and training hyperparameters based on the specific properties of the input medical dataset.
• Automated Architecture Selection - Automatically adjusts network architectures and hyperparameters based on the specific properties of the image dataset.
• Automated Medical Imaging Pipelines - Functions as an automated medical imaging pipeline that configures preprocessing and network architecture based on dataset properties.
• Image Segmentation - Trains deep learning models to automatically identify and outline anatomical structures in 2D and 3D medical images.
• Automated Configuration Pipelines - Automatically selects the optimal network architecture and hyperparameters based on the specific properties of the medical dataset.
• Sparse Annotation Handling - Allows learning from datasets with incomplete manual segmentations by ignoring unlabeled regions during training.
• Patch-Based Training Strategies - Processes large 3D volumes by extracting small overlapping sub-volumes to enable training within limited GPU memory.
• Inference Execution - Provides a process that applies trained segmentation models to new images to generate predicted masks and labels.
• Supervised Learning Models - Implements a supervised learning system for generating predicted masks and labels from medical images.
• U-Net Architectures - Utilizes a symmetric encoder-decoder U-Net structure with residual connections for multi-scale spatial feature capture.
• PyTorch Semantic Segmentation Libraries - Provides a deep learning framework for semantic segmentation of 2D and 3D biomedical images using PyTorch.
• Dataset-Adaptive - Adapts network architecture, preprocessing, and hyperparameters based on the specific properties of the provided biomedical dataset.
• Pixel Normalizers - Provides utilities for scaling pixel intensity values using dataset-specific statistics to ensure consistent distributions across modalities.
• Sliding-Window Inference - Generates predictions by scanning the image with overlapping patches and averaging results to remove edge artifacts.
• Segmentation Inference - Applies trained segmentation models to new clinical images to generate predicted masks and labels.
• Segmentation Inference Engines - Implements a tool for applying trained deep learning models to new medical images to perform semantic segmentation.
• Cascaded Segmentation Pipelines - Implements a cascaded model pipeline that sequentially applies coarse and fine networks to refine anatomical boundaries.
• Sparse Annotation Training - Develops segmentation models using datasets with incomplete or coarse manual labels by ignoring unlabeled regions.
• Model Performance Benchmarking - Ships an evaluation system for comparing model segmentation accuracy and processing speed using standardized datasets.
• Anatomical Region Training - Provides a training approach specifically designed to improve the segmentation of large anatomical structures.
• Anatomical Region Weighting - Modifies the training objective to prioritize specific anatomical regions or ignore unlabeled areas in sparse datasets.
• Anatomical Region Weighting - Implements specialized loss functions that prioritize targeted anatomical areas to improve segmentation accuracy.
• Medical Dataset Formatting - Provides utilities for formatting raw medical images and labels into standardized structures for training.
• Medical Image Normalization - Converts and normalizes raw medical imaging files into a standardized format for use in supervised learning pipelines.
• Biomedical Image Processing Toolkits - Provides a set of tools for intensity normalization and dataset preparation of medical imaging formats.
• Cell Segmentation - Self-configuring framework for 3D biomedical segmentation.

Star history