Nuance Communications, a Microsoft company, and Nvidia today announced a partnership they say will open the pipeline between AI technology and real-world applications. By putting AI-based diagnostic tools directly into the hands of clinicians at scale, the pair hopes to decrease health inequities, cost of care and eliminate the distance from “bench to bedside.”
Nvidia’s open-source and domain-specialized medical imaging AI framework, called MONAI, united with the Nuance Precision Imaging Network is already being offered to clinicians in major medical centers. Nuance offers an AI-powered cloud platform that uses diagnostic imaging to create patient insights.
Mass General Brigham is among the first major medical centers to use the medical imaging AI models, streamlining workflow for its 80,000 employees and speeding up care time for its annual 1.5 million patients.
“Prior to this activity, it was difficult for us to take the assets of works that happened through research and be able to deploy those into clinical purposes,” Keith Dreyer, Mass General Brigham's chief data science officer, told reporters during a press conference. “Now we can set up a pipeline that can allow those researchers as they're completing their products or developing their products, they can mindfully and purpose-built, deliver those into a framework that we can then deploy inside the clinical environment.”
Dreyer said the new workflow allows for model adaptation time to decrease from years down to months and weeks. By decreasing time, the cost goes down as well.
The new workflow allows MGB to develop medical imaging models, application packages and clinical feedback for model refinement. The medical center currently has $2.3 billion in annual research spending.
Nuance and Nvidia’s technology is now being used in MGB’s breast density AI model to reduce patient wait time for scan results down to 15 minutes so patients may speak to clinicians about results before leaving the facility. Previously, patients would spend 24 to 72 hours waiting for potentially life-altering results. Dreyer said this speed is reflected in the speed of model refinement.
“We've seen quite a bit of acceleration and deployment of this,” Dreyer said. “Whereas before, it's difficult to appreciate how challenging it is to be able to have a disconnected developer or researcher creating algorithms, maybe at another facility, and then be able to push those over to us and not be able to deploy them clinically.”
Lung scans of COVID-recovery patients are also being assessed with Nuance and Nvidia’s technology, according to executives. MGB used AI models to evaluate the scores of over 200,000 patient lungs to predict whether lung function would degrade over time, better determining if a patient needed ICU admittance or oxygen.
By using AI algorithms to continuously monitor CT abdomen scans, the variations of the adrenal gland have been assessed in order to predict cancer, executives said.
Nuance’s PowerScribe is already used by 80% of radiologists to interpret medical imagines which are then shared through Nuance’s PowerShare network of 12,000 facilities. MONAI is then integrated directly into clinical practice using Microsoft Azure-powered Nuance Precision Imaging Network, an open ecosystem supporting imaging AI models built by MONAI users.
Nvidia's MONAI Deploy technology then functions as an accelerated processing pipeline delivering MONAI Application Packages into healthcare systems, utilizing interoperability standards including DICOM.
As new images are taken from patients, they flow through the precision imaging network, the AI model that's been developed in MONAI is run against that, and then outputs of those models are driven into real-world clinical workflows, the organizations said.
“So it's sort of this end-to-end factory from the design time of a model into clinical practice and rapidly collapsing the amount of time and efforts required to drive this into clinical practice,” Peter Durlach, chief strategy officer at Nuance Communications, told reporters. “Because from the point of view of organizations deploying these models, they don't have to build any custom infrastructure or do any custom work effectively to make this connection to the clinical workflows happen.”
With 80% of medical conditions starting with an image as a key part of the diagnostic journey and diagnostic imaging playing a role in 23,000 clinical conditions, Durlach emphasizes that the application will directly affect patient outcomes.
Medical images additionally require a fine-tuned eye that may miss subtle nuances which AI could pick up. “Imaging AI is the farthest along of any of the different modalities for AI in healthcare today by leaps and bounds,” Durlach said.
“The key piece here is, 'How do you actually standardize this application packaging to make it deployable across a wide variety of healthcare IT infrastructure?' That's really this MONAI application-specific specification format,” David Niewolny, Nvidia’s director of business development in healthcare, told reporters. “So the great news is MONAI is being widely adopted in the industry. And I think that is one of the key pieces here is what we need to start creating is de facto standards in the industry.”
MONAI currently has 650,000 downloads, 450 GitHub projects and 160 papers published, according to Niewolny.
Starting in radiology, the technology has migrated into pathology, and Niewolny now sees it moving into surgical applications via streaming videos. The advancement “tackles the entire AI life cycle” from AI and pretrained models to AI-assisted labeling tools and training techniques like federated learning models to ensure the safety of private health information.
Historically, there have been challenges to deploying AI algorithms in clinical practice, Niewolny said, which constrains the adoption of radiology and AI at scale.
“Ninety-nine percent of the medical imaging AI applications never ended up making their way to patients; they're never actually able to deploy at scale,” Niewolny said. “So really, we're at the proverbial chasm in AI and medical imaging. What you are hearing here is that chasm being closed.”
By standardizing specifications, models can be reproduced, and the speed of the application feedback loop can accelerate the honing of applicable AI in healthcare, executives said.