NVIDIA, the American tech company, has collaborated with King’s College London and open-sourced MONAI (Medical Open Network for AI), a framework developed for healthcare professionals using best practices from existing NVIDIA tools such as NVIDIA Clara, NiftyNet, DLTK and DeepNuero. Using PyTorch resources, MONAI provides domain-optimised foundation capabilities for developing healthcare imaging training in a standardised way to create and evaluate deep learning models. 

According to the NVIDIA Blogs, MONAI is user-friendly, delivers reproducible results and is domain-utilised for the remains of healthcare data - equipped to handle the unique format, resolution and specialised meta-information of medical images. The first public release provides domain-specific data transforms, neural network architectures, and evaluation methods to measure the quality of medical imaging models. 

Available on GitHub, the open-source code is based on the Ignite and PyTorch deep learning frameworks and brings together state-of-the-art libraries for data processing, 2D classification, 3D segmentation and more. Modular, open-source solutions provide researchers with the flexibility to customise their deep learning development, without needing to replace their existing workflows with an end-to-end system. For example, a researcher can use MONAI code for data pre-processing and transformations and then switch over to an existing AI pipeline for training. 

“In partnership with NVIDIA, Project MONAI is following industry standards for open-source development and building a global community across academia and industry to establish a high-quality framework supporting scientific development in medical imaging AI,” said Seb Ourselin, head of the School of Biomedical Engineering & Imaging Sciences at King’s College London.

The MONAI features flexible pre-processing of multi-dimensional medical imaging data, provides and supports compositional and portable APIs for ease of integration in existing workflows, provides domain-specific implementations for networks, losses, evaluation metrics through customisable design for varying user expertise by supporting multi-GPU data. 

“Reproducibility of scientific research is of paramount importance, especially when we are talking about the application of AI in medicine,” said Jayashree Kalpathy-Cramer, scientific director at the MGH & BWH Center for Clinical Data Science, and associate professor of radiology at MGH/Harvard Medical School. “Project MONAI is providing a framework by which AI development for medical imaging can be validated and refined by the community with data and techniques from the world over.”

Future releases of NVIDIA Clara will also leverage the MONAI framework. We plan to bring together development efforts for NVIDIA Clara medical imaging tools, and MONAI to continue delivering domain-optimised, robust software tools for researchers in healthcare imaging.