Back to
Projects List
Investigate MONAI generative modeling for Imaging Data Commons
Key Investigators
- Steve Pieper (Isomics, Inc. USA)
- Mikael Brudfors (NVIDIA, UK)
- Andres Diaz-Pinto (NVIDIA, UK)
- Andrey Fedorov (BWH, US)
- Birgitt Peeters (BIDMC, US)
- Umang Pandey (UC3M, Spain)
Presenter location: In-person
Project Description
Generative learning refers to a class of techniques that process large amounts of training data into models that can be used for a variety of tasks such as synthetic data generation, image compression, enhancing resolution, classifying images, and content based retrieval. Recently a generative package has been added to the open source MONAI software.
This project will explore the application of MONAI generative tools to data on the NCI Imaging Data Commons.
Objective
- Study the existing material and collect information from other interested parties
- Make plans about what experiments would be interesting
- If possible do some small experiments to better understand what’s possible and what effort and resources would be required to scale up
Approach and Plan
- Explore creating an
IDCDataset
compatible with MONAI Datasets using idc-index to fetch data
- Investigate adapting tutorial code to work with IDC data
- Try running some small tests, such as running the superresolution tutorials on IDC data
- Document how IDC can be used with MONAI for research
Progress and Next Steps
- Discussed the project with people at project week for feedback
- Contacted Mark Graham of KCL, a MONAI generative researcher/developer for advice
- Implemented a first pass combination of IDC data with MONAI generative notebook
- Ran tests on colab and workstations
- Adapted example data (8-bit) to dicom (16-bit) data to accomodate dynamic range differences
- Explored parallel and federated approaches
Illustrations
No response
Background and References