monet-bundle 1.0b0

A Python package providing the MONet Bundle for nnUNet, extending it with MONAI functionalities and access to the MAIA Segmentation Portal.

MONet Bundle

This repository contains the implementation of the MONet Bundle, with some instructions on how to use it and how to convert a generic nnUNet model to MONAI Bundle format.

For more details about the MONet Bundle, please refer to the Jupyter notebook MONet_Bundle.ipynb.

2025-06-25 UPDATE: Check out the MONet Bundle for FedBraTS and FedLymphoma!

The MONet Bundle has been used in the Federated Brain Tumor Segmentation FedBraTS and Federated Lymphoma Segmentation FedLymphoma projects, which are described in the following paper:

• MONet-FL: Extending nnU-Net with MONAI for Clinical Federated Learning

Download the MONet Bundle

You can download the MONet Bundle from the following link: MONet Bundle ALternatively, you can use the following command to download the MONet Bundle:

wget https://raw.githubusercontent.com/SimoneBendazzoli93/MONet-Bundle/main/MONetBundle.zip

or, through the Python Script:

MONet_fetch_bundle.py --bundle_path <FOLDER_PATH>

Convert a trained nnUNet model to MONAI Bundle

To convert a trained nnUNet model to MONAI Bundle format, you can start with exporting a nnUNet trained model with the nnUNetv2_export_model_to_zip command. This command will export the model to a zip file that can be used in the conversion process.

nnUNetv2_export_model_to_zip -d 009 -o Task09_Spleen.zip -c 3d_fullres -tr nnUNetTrainer -p nnUNetPlans -chk checkpoint_final.pth checkpoint_best.pth --not_strict

For testing purposes, you can use the Task09_Spleen.zip file provided in this repository: https://github.com/SimoneBendazzoli93/nnUNet-MONAI-Bundle/releases/download/v1.0/Task09_Spleen.zip. This file contains a trained nnUNet model for the Spleen segmentation task, for only the 3d_fullres configuration and the fold 0.

Next, you can build the provided Docker image to convert the model to MONAI Bundle format. The Dockerfile is provided in this repository, and you can build the image with the following command:

docker build -t nnunet-monai-bundle-converter .

The converter will first convert the nnUNet model to MONAI Bundle format, and then create the corresponding TorchScript model, which can be used for inference with MONAI Deploy. For testing purposes, you can use the Task09_Spleen.zip file provided in this repository.

To run the conversion, you can use the following command:

wget https://github.com/SimoneBendazzoli93/MONet-Bundle/releases/download/v1.0/Task09_Spleen.zip
python MONet_run_conversion.py --bundle_path <MONAI_BUNDLE_PATH> --nnunet_model <NNUNET_CHECKPOINT_PATH>.zip

Package the MONet Bundle with MONAI Deploy

To package the MONet Bundle with MONAI Deploy, you can use the monai-deploy package command. This command will create a deployable bundle that can be used for inference with MONAI Deploy.

monai-deploy package examples/apps/spleen_nnunet_seg_app -c examples/apps/spleen_nnunet_seg_app.yaml -t spleen:1.0 --platform x86_64

Run inference with MONAI Deploy

The resulting Docker context can be found in the deploy/spleen-x64-workstation-dgpu-linux-amd64:1.0 directory. You can use this context to build a Docker image that can be used for inference with MONAI Deploy:

# Copy the TorchScript model to the Docker context
cp nnUNetBundle/models/fold_0/model.ts deploy/spleen-x64-workstation-dgpu-linux-amd64:1.0/models/model/

docker build deploy/spleen-x64-workstation-dgpu-linux-amd64:1.0 --build-arg UID=1000 --build-arg GID=1000 --build-arg UNAME=holoscan -f deploy/spleen-x64-workstation-dgpu-linux-amd64:1.0/Dockerfile -t spleen-x64-workstation-dgpu-linux-amd64:1.0

To test the resulting Docker image, you can run:

MONet_inference_dicom.py --dicom_study_folder <INPUT_FOLDER> --prediction_output_folder <OUTPUT_DIR> --docker-image maiacloud/spleen-x64-workstation-dgpu-linux-amd64:1.0

Specifying the input and output folders, together with the TorchScript model path. The input folder should contain all the DICOM files of the study you want to process, and the output folder will contain the predictions in DICOM SEG format, and an additional STL file with the 3D mesh of the segmentation.

To create the same Docker image running inference on NIFTI images, you can use the provided Dockerfile in the deploy/spleen-x64-workstation-dgpu-linux-amd64:1.0-nifti directory. The Dockerfile is already set up to run inference on NIfTI images, and it includes the necessary dependencies. To test the resulting Docker image, you can run:

MONet_inference_nifti.py --study_folder <INPUT_FOLDER> --prediction_output_folder <OUTPUT_DIR> --docker-image maiacloud/spleen-x64-workstation-dgpu-linux-amd64:1.0-nifti

Specifying the input and output folders, together with the TorchScript model path. The input folder should contain all the NIfTI files of the study you want to process (one per modality, with the given suffix identifier), and the output folder will contain the predictions in NIfTI format.