DMIST-nnUNet Integration

This repository is a fork of the nnU-Net framework. Our goal is to integrate nnU-Net’s powerful automated segmentation capabilities within DMIST (Deep-learning Medical Image Segmentation Toolbox).

• dmist-train: toolbox for training
• dmist-deploy: toolbox for pipeline implementation

Key Objectives

• Integrate custom data preprocessing scripts to create our own nnUNet-compatible IRF imaging datasets.
• Deploy nnUNet models trained on IRF imaging datasets for segmentation tasks.

About nnU-Net

nnU-Net is an automated deep learning-based segmentation framework that adapts itself to the characteristics of any given dataset. Originally developed for the biomedical domain, nnU-Net’s flexibility and out-of-the-box performance make it an ideal starting point for custom applications.

Key Features of nnU-Net

• Automated Configuration: Automatically configures segmentation pipelines tailored to the dataset, with minimal need for expert intervention.
• Versatility: Handles 2D and 3D data, with support for various input modalities and voxel spacings.
• Proven Performance: Recognized as a top performer in numerous medical imaging challenges, consistently delivering state-of-the-art results.

For more details on the original nnU-Net, please refer to the official documentation.

Custom Modifications

1. Custom Scripts

We have developed a series of custom scripts to extend nnU-Net’s functionality for infectious disease research. These include:

• Preprocessing scripts tailored to specific imaging modalities and datasets commonly used in our research.
• SLURM job file creation and management to perform inference, ensembling, and postprocessing tasks on Skyline HPC.

2. New Pipelines and Configurations

Integrate our custom scripts within the existing dmist-deploy framework.

3. Integration with Organizational Infrastructure

Integrate these tools within NIH/NIAID/IRF organizational infrastructure.

Getting Started

To use our custom version of nnU-Net, follow these steps:

1. Clone the Repository (with --recurse-submodules flag to include our private IRF submodule):

   git clone --recurse-submodules [email protected]:mark-rustad/dmist-nnunet-fork.git
   cd nnUNet

2. Create a conda environment (using conda/mamba/micromamba) using the environment specification file nnunet_env_v1.yml:

   micromamba create -n <env_name> -f ./nnunet_env_v1.yml

   Refer to nnUNet's installation instructions in the documentation/installation_instructions.md file for additional environment setup information.

3. Set Up Your Dataset: Convert your dataset into the nnU-Net format as described in the Dataset Conversion Guide.

4. Run Custom Pipelines: Use our custom scripts and configurations to train and evaluate models. Refer to our documentation for specific command-line instructions tailored to infectious disease research workflows.

5. Updating the Submodule: To update the submodule to the latest commit from its branch, run:

   git submodule update --remote

Documentation

• Installation Instructions
• Custom Scripts and Pipelines - TODO:Detailed instructions on using our custom scripts and configurations.
• Usage Instructions
• Extending nnU-Net

Acknowledgements

This work builds upon the original nnU-Net framework developed by the Applied Computer Vision Lab (ACVL) and the Division of Medical Image Computing at the German Cancer Research Center (DKFZ).

We extend our thanks to the original authors and the broader community for their contributions to this exceptional framework.

Citation

If you use this fork in your research, please also cite the original nnU-Net paper:

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.

---

For questions, contributions, or feedback, please contact Mark Rustad at [email protected].