Nvidia has introduced a new healthcare artificial intelligence (AI) platform and released version 1.0 of its Medical Open Network for AI (MONAI) at this week's GPU technology conference (GTC) 2022 conference in San Francisco.
IGX is a hardware and software platform designed to support the use of real-time AI at the edge for imaging applications such as minimally invasive surgery and robotic surgery, said Kimberly Powell, vice president of healthcare at Nvidia.
"IGX is the ideal computing platform for the imaging robotics pipeline that consists of streaming sensor processing, reconstruction, detection, segmentation, and real-time visualization," she said. "Now all future medical devices can take advantage of the same breakthroughs that enable self-driving cars, to be powered by AI and to become software-defined."
Three robotic surgery startup companies -- Activ Surgical, Moon Surgical, and Proximie -- have already signed on to use IGX to run Nvidia's Clara Holoscan platform for their surgical robotics systems, according to the company.
Nvidia's IGX healthcare AI platform is being used by three robotic surgery startups. Image courtesy of Nvidia.Nvidia has also released version 1.0 of Medical Open Network for AI (MONAI), an open-source framework for development of medical imaging AI applications. Developed in collaboration with King's College London in the U.K., MONAI 1.0 features Active Learning, Nvidia's interactive and "intelligent" approach to labeling of training data. The company estimates Active Learning can reduce the amount of labeling necessary for training by 75%.
The new release also includes Auto3D, a framework for selecting and training a 3D segmentation model.
"It automatically goes through state-of-the-art 3D AI architectures and chooses the one best fit for your application," Powell said.
In addition, MONAI now supports Nvidia's Federated Learning Application Runtime Environment (FLARE) to facilitate federated learning for algorithm training. Furthermore, the framework's Model Zoo now includes over 15 pretrained models for applications ranging from 3D segmentation, detection, and classification for multiple different modalities, according to Powell.
Algorithms can also now be developed for streaming imaging applications such as endoscopy, ultrasound, and surgical video. Native, end-to-end support for these applications includes labeling tools, the right model architectures for training, and the capability to deploy the algorithms, Powell said.
Since its initial release in 2020, MONAI has been downloaded over 600,000 times, and half of the downloads have taken place over the last six months, according to Powell.
