back to Projects List
We have designed, developed, and validated deep learning methods for mediastinal lymph node segmentation. We want to use active learning and MONAI Label to transfer this to additional areas of the body (e.g. abdominal or retroperitoneal lymph nodes). We want to use MONAI Deploy SDK as the framework for pre/post-processing in our production inference pipeline. It’s currently unclear how a MONAI Label App relates to a MONAI Deploy SDK Application. They feel largely disconnected at the moment, despite sharing the same project name. It’s not possible to build a repository of custom Operators in MONAI Deploy that are shared across Applications without publishing a public Python package first.
To work with the MONAI Deploy SDK and MONAI Label teams to identify how to structure a repository for use with both toolkits.
We will use the mediastinal subset of TCIA CT Lymph Node as data for development and performing experiments. Below is our plan during the course of project week:
Not much progress, got pulled into other things all week.
The end goal is a directory structure something like this, but it’s still a bit unclear to us how it will work.
/src /shared python code - custom operators or pre/postprocessing steps - custom models /monai-label-apps (These should be able to use MONAI Bundles. Each Application may have custom input from the user which may be different than e.g. deep edit or deep grow. The MONAI Label applications are expected to be applicable to already-curated (e.g. cropped) datasets.) - /app1 (e.g. https://github.com/Project-MONAI/MONAILabel/blob/main/sample-apps/radiology/main.py, which runs in MONAI Label Server) - /app2 - /app3 /monai-deploy-apps (These set up the pre/postprocessing pipelines prior to running the model which are required for integration at the application level (e.g. cropping to lungs). Each App should be able to be packaged into Docker containers individually. - /app1 (e.g. https://github.com/Project-MONAI/monai-deploy-app-sdk/blob/main/examples/apps/ai_unetr_seg_app/app.py) - /app2 /monai-bundles (These include the actual model, the network definition, the training and inference pre/postprocessing steps (e.g. NormalizeIntensityd). Unclear how this would pull in private pre/postprocessing steps or models. - model bundle 1 (e.g. https://github.com/Project-MONAI/model-zoo/tree/dev/models/brats_mri_segmentation/configs) - model bundle 2