Create a set of proof of concept projects on top of Sonador that shows it can be used for ML and image processing. Create data loaders and other primitives for consuming imaging data from Orthanc from within PyTorch and TensorFlow.

Project 1: Classification of Chest X-Rays Using PyTorch

Show how an image classification model can be implemented to detect pneumonia from data stored in Sonador/Orthanc.

• Introduce the "Oak-Tree imaging environment" and discuss how the components can be used for prototyping machine learning applications on the Sonador stack
  • Determine what needs to be added to the Oak-Tree imaging environment so that it is possible to use a docker-compose ... up command that will launch the components needed for medical imaging ML POC work
  • Configure an entrypoint command that will configure a known imaging server (Orthanc) and set identifiers and API/Access key
  • Deploy static assets to object storage: python manage.py collectstatic
• Write a tutorial that describes how to download the NIH Chest X-Ray Dataset and import it into Sonador
  • (Blog) Incorporate fundamentals for retrieving data using the Sonador CLI tool and client library into the tutorial.
  • (Development) Implement support in the Sonador CLI tool the ability to annotate a dataset from a CSV or other structured data file.
  • (Blog) Add details to tutorial about how to annotate data in Orthanc by using a CSV or other structured data source
• Write a second tutorial that shows how to implement a PyTorch classification model using SonadorAI.
  • (Development) Implement a "data loader" for SonadorAI that provides method to connect to Sonador, execute a query, create an image pipeline, and return Tensor data that is appropriate for use with PyTorch
  • (Blog) Follow Ethan's Chest X-Ray classification blog post to create a classifier to differentiate between pneumonia and normal. An alternative to this might be to look for COVID-19.
  • (Blog) Demonstrate use of the SHAP GradientExplainer

Project 2: Auto-segmentation of LV for Heart Failure

• Demonstrate all components of the Root/Sonador platform
• Orthanc: Store cardiac MRI/ultrasound
• Jupyter/Scripts: Create notebooks to create and train a transfer learning model capable of determining the left ventricle size and ejection ratio
• AirFlow: Transform scripts into a pipeline that can train and apply the model
• Summary: Create blog post describing the use-case