Creating a Segmentation App Consuming a MONAI Bundle

This tutorial shows how to create an organ segmentation application for a PyTorch model that has been trained with MONAI and packaged in the MONAI Bundle format.

Setup

# Create a virtual environment with Python 3.8.
# Skip if you are already in a virtual environment.
conda create -n monai python=3.8 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
conda activate monai

# Launch JupyterLab if you want to work on Jupyter Notebook
jupyter-lab

Executing from Jupyter Notebook

• Creating a Deploy App with MONAI Deploy App SDK and MONAI Bundle
  • Creating Operators and connecting them in Application class
    • Setup environment
    • Download/Extract input and model/bundle files from Google Drive
    • Set up environment variables
    • Set up imports
    • Determining the Input and Output for the Model Bundle Inference Operator
    • Creating Application class
  • Executing app locally
    • app.py
    • __main__.py
  • Packaging app
  • Executing packaged app locally

Video may show the use of previous SDK verson.

Download 04_monai_bundle_app.ipynb

Executing from Shell

# Clone the github project (the latest version of main branch only)
git clone --branch main --depth 1 https://github.com/Project-MONAI/monai-deploy-app-sdk.git

cd monai-deploy-app-sdk

# Install monai-deploy-app-sdk package
pip install --upgrade monai-deploy-app-sdk

# Download/Extract ai_spleen_bundle_data zip file from https://drive.google.com/file/d/1cJq0iQh_yzYIxVElSlVa141aEmHZADJh/view?usp=sharing

# Download the zip file containing both the model and test data
pip install gdown
gdown https://drive.google.com/uc?id=1Uds8mEvdGNYUuvFpTtCQ8gNU97bAPCaQ

# After downloading it using gdown, unzip the zip file saved by gdown and
# copy the model file into a folder structure that is required by CLI Packager
rm -rf dcm
unzip -o ai_spleen_seg_bundle_data.zip
rm -rf spleen_model && mkdir -p spleen_model && mv model.ts spleen_model && ls spleen_model

# Install necessary packages from the app; note that numpy-stl and trimesh are only
# needed if the application uses the STL Conversion Operator
pip install monai torch pydicom highdicom SimpleITK Pillow nibabel scikit-image numpy-stl trimesh

# Use env variables for input, output, and model paths for local running of Python application
export HOLOSCAN_INPUT_PATH=dcm
export HOLOSCAN_MODEL_PATH=spleen_model/model.ts
export HOLOSCAN_OUTPUT_PATH="output"
export HOLOSCAN_LOG_LEVEL=TRACE

# Local execution of the app directly or using MONAI Deploy CLI
python examples/apps/ai_spleen_seg_app/app.py

# Package app (creating MAP docker image) using `-l DEBUG` option to see progress.
# This assumes that nvidia docker is installed in the local machine.
# Please see https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker to install nvidia-docker2.
monai-deploy package examples/apps/ai_spleen_seg_app \
    --config examples/apps/ai_spleen_seg_app/app.yaml \
    --tag seg_app:latest \
    --models spleen_model/model.ts \
    --platform x64-workstation \
    -l DEBUG

# Note: for AMD GPUs, nvidia-docker is not required, but the dependency of the App SDK, namely Holoscan SDK
#       has not been tested to work with a ROCM base image.

# Run the app with docker image and input file locally
rm -rf output
monai-deploy run seg_app-x64-workstation-dgpu-linux-amd64:latest -i dcm -o output
ls -R output

Deploying a MedNIST Classifier App with MONAI Deploy App SDK (Prebuilt Model) Creating a Deploy App with MONAI Deploy App SDK and MONAI Bundle