Emerging technologies like machine learning, deep learning, and natural language processing can promise significant improvements in an array of industries, including the healthcare field. However, it’s rarely fast or easy to make the jump from research to practical use. The supporting technologies and algorithms need time to catch up to the promise.
As Nvidia pointed out just last month, in the healthcare field, AI has made strides, but most of that has been in the area of medical imaging. At the time, Nvidia talked about expanding the reach of AI and deep learning further into other parts of field, in this case in genomics and digital health sensor data.
However, there is still a lot that can be done to move AI from the research space into clinical practices in the medical imaging space, which Nvidia is laying out this week at a convention in Chicago of the Radiological Society of North America (RSNA), an organization of about 50,000 radiologists The GPU maker is rolling out tools and development kits designed to help programmers and healthcare facilities develop AI algorithms that can be used in medical imaging efforts as well as partnerships with Ohio State University, the National Institutes of Health (NIH), and others bridge the gap between research and clinical use of the technology.
“There’s a lot of innovation, there’s a lot of AI, there are a lot of use cases for GPUs in general,” Abdul Hamid Halabi, global health care lead at Nvidia, said during a press conference before the RSNA show began, adding that radiologists are working with imaging machines whose images often need fast processing. “Back-end [operations also have] use cases of imaging, like population counts, like archival and iteration, perhaps, so they can research data lakes. The question is, how do you bring all this innovation and all of this AI that we’re looking at into this imaging workstation, to this imaging chain of improving quality, access and cost?”
That is where Nvidia is stepping in. For the past several years, Nvidia has made AI, machine learning and other such technologies a core part of its business going forward. The company has done it through the development of products like the DGX supercomputers as well as through a wide range of partnerships, which in the healthcare field includes the likes of GE and, more recently, Scripps Research Translational Institute. In addition, the company says it is working with 75 companies to help expand the use of AI in the radiology field.
At Nvidia’s GPU Technical Conference in Japan in September, the company announced its Clara platform, a hardware and software offering that leverages a computing architecture – the Clara AGX – based on the Nvidia Xavier AI computing module and Turing GPUs to enable organizations to more quickly bring AI capabilities to their medical devices. Clara leverages the GPUs and the CUDA platform to give companies a way to more quickly crunch through the massive amounts of data generated by the medical devices so they can be more easily interpreted by physicians and scientists. At RSNA, Nvidia is making the Clara SDK generally available.
“Recognizing that medical imaging, in general, is really computational and the fact that adoption of AI is really skyrocketing, Nvidia now has the Clara platform,” Halabi said. “This is an open platform to enable the medical imaging industry to [develop] and deploy algorithms and breakthroughs in the instruments and network. It takes advantage of Nvidia hardware, whatever hardware platform you have available to you, although we are making specialized hardware for intelligent instruments. The platform can take advantage of any Nvidia hardware. On top of it is the Nvidia Clara SDK.”
The SDK includes accelerated libraries and engines that use the libraries, taking advantage of the acceleration capabilities of the GPUs and CUDA. It also enables individual radiology practices to develop the algorithms that best suit their needs, similar to the need to localize features in autonomous vehicles, he said. A car in the United States needs some different features than those in the United Kingdom.
“You don’t need to retrain the car from scratch, but you do need to teach it about the local environment that it’s in,” Halabi said. “Every radiology practice is unique. It has its own instruments. It has its own patients, demographics, and its own way to practice. So although we’re seeing the flood of algorithms comes through, we do believe that we need to provide radiology with a tool to take that AI and localize it, allowing the physicians to local the algorithms for their own patients.”
The Massachusetts General Hospital (MGH) and the Brigham and Women’s Hospital (BWH) Center for Clinical Data Science, both in the Boston area, have used the Clara SDK to develop a model for detecting abdominal aortic aneurysms and is deploying its on the Nuance AI Marketplace, a site that launched a year ago and is powered by the Clara platform.
Other new tools are the Transfer Learning Toolkit for Medical Imaging and the AI Assisted Annotation SDK. Through the toolkit, radiology departments will be able to combine existing algorithms with a small subset of private data from the clinic to create a new model that is customized for its patients. The SDK “will enable, instead of actually having to go through the tedious process of annotating the image yourself and marking all the edges of every point and making sure that very pixel is correct, you can with a few clicks – no more than six or seven clicks – tell the AI where the organ of interest is, and that will use existing algorithms in order to actually annotate the image for you. Then you’re able, with just a few minor modifications, to update for the annotation department [with] your knowledge as a radiologist. This SDK will be key because it’s going to help most radiologists [make sense] of the data they’re sitting on today and allow them to contribute to the process of deployment of AI as well as the creation of AI for other partners’ benefit.”
Nvidia also is partnering with Ohio State to build an in-house AI marketplace for clinical imaging similar to the Nuance marketplace. It will be based on the Clara platform as a way to ensure easy access to the various AI algorithms that different players in a medical facility need.
“The question that they have is, how to you actually deploy all these different algorithms to all the potential [needs] that the radiologist can consume or the physician can consume,” Halabi said. “It was really clear to them that they had to develop an AI marketplace or apps store locally within the hospital. We’re partnering together to develop an AI marketplace within Ohio State where all of these algorithms will be hosted and then they will be integrated into physicians’ viewers.”
In addition, the NIH partnership will involve creating projects to bring AI technologies to clinical trials as well as developing AI tools. NIH conducts more than 1,600 trials a year, and the first projects between the two organizations will focus on AI tools to streamline clinical trials for brain and liver cancer. Nvidia will place researchers and engineers with clinicians at NIH as part of the partnership.