Artificial Intelligence, Diagnostics

Nvidia’s plan to take the ‘anxiety’ out of AI’s use in radiology

A new partnership between Nvidia and the American College of Radiology will give a wider range of clinicians the ability to build, share and adapt their own AI models.

machine learning AI

Much of the media attention around AI in radiology has had a bit of a Terminator tilt to it. Headlines based on the primacy of AI algorithms in detecting potential issues are often paired with the conclusion that radiologists are quickly becoming obsolete.

The truth, of course, is a little less black-and-white.

Development of AI writ large in still in its early stages and there’s a wide chasm between current applications and an algorithm replacing your doctor.

But suspicion over AI technology – as well as the standard issues of advancing innovation in healthcare – has built walls between providers and AI-based tools that can improve patient outcomes and make clinician’s lives easier.

Santa Clara, California-based Nvidia is attempting to change the paradigm by expanding access to AI tools for radiologists in way that engenders trust in the technology.

From its initial roots in gaming, Nvidia has branched out to become one of the preeminent companies in AI through efforts in industries like transportation, energy and healthcare.

Through a partnership with the American College of Radiology (ACR), the company is trying to democratize access to the technology and give clinicians a central role in actually developing and validating the algorithms meant to improve care.

The broad strokes of the collaboration involves the incorporation of Nvidia’s Clara AI toolkit into the ACR AI-LAB software platform, allowing more than 38,000 clinicians to build, share, validate and customize AI algorithms, while keeping constituent patient data housed where it was created.  

“The mission of the AI lab is to help institutions develop AI for their own facilities and for their own patient populations,” Nvidia’s VP of Healthcare Kimberly Powell said in a call with reporters.

By expanding access to this toolkit, a larger swath of radiologists will be given AI-assisted annotation tools to create more robust datasets, as well as the ability to use their own datasets to adapt pre-trained models in a process known as Transfer Learning.

Dushyant Sahani, chairman of the department of radiology at University of Washington’s School of Medicine, said he viewed the collaboration between Nvidia and ACR as a way to alleviate some of the “anxiety” some clinicians feel about AI.

“The ACR has a tremendous database of skilled radiologists around the country and if you engage those radiologists to be part of this AI initiative, not only are they taking ownership of what AI is, but also helping to steer where AI should go,” Sahani said.

To Sahani, making radiologists a key part of the way that these technologies are developed adds credibility to the tools themselves and insures their ability to address real clinical issues.

The collaboration with the ACR was built on an earlier pilot involving multiple partners including The Ohio State University’s Wexner Medical Center and the Massachusetts General Hospital and Brigham and Women’s Hospital’s Center for Clinical Data Science.

In the pilot, OSU successfully imported, customized and validated a cardiac computed tomography angiography model created by the Boston-based Center for Clinical Data Science.

Center for Clinical Data Science Executive Director Mark Michalski underscored the importance of transfer learning in building a level of personal familiarity with the technology.

“When radiologists – and clinicians in general – have a hand in developing these technologies, not only does it give you insight into the model itself but it engages both the patients and the physician in that collaborative process,” Michalski said.

“It makes sense that when people get involved and feel like they’re a part of building something, they’re more likely to trust it and adopt it.”

Transfer learning, according to Michalski, also has the potential to address some of the ethical concerns that come along with AI use. Chief among them being the incorporation of historical bias into models that then go on to perpetuate existing inequities.

The proliferation of AI has led to establishment of special research centers like Stanford’s Institute for Human-Centered Artificial Intelligence and and industry groups like the Alliance for Artificial Intelligence in Healthcare to guide the ethical development and implementation of AI technology.

“If you have a patient population thats underserved by the healthcare machinery and – by extension – underrepresented by the data, they’re underserved by the model,” Michalski said. “One of the things we’re investigating is whether transfer learning has a role in beginning to mitigate and recognize these biases.”

While there’s still plenty of work to be done in clinically validating algorithms, one limiting factor that continues to lag behind the science is the vital (if not particularly sexy) issue of workflow integration. In fact, Michalski said around 80 percent of his organization’s work is in building out that integration process.

A common sentiment is that best algorithm will wither on the vine if a clinician lacks the ability easily use and integrate it into their normal practice.

This is where Nvidia’s collaborations with industry partners like Nuance and GE Healthcare comes in.

Burlington, Massachusetts-based Nuance – known for their speech recognition and transcription services, has been a key bridge in what the company calls the “last mile” of AI implementation.

Karen Holzberger, the head of Nuance’s healthcare diagnostic solutions business, said the company serves nearly 70 percent of the country’s radiologists through its PowerScribe platform, which helps clinicians with reporting and documentation.

The company has no plans to build out their own algorithms. Instead, Nuance is integrating with Nvidia’s AI technology to allow for automatic documentation and easier creation of diagnostic reports, which are then corrected and contextualized by clinicians.

“We are uniquely positioned to wrap up that feedback and take that back to improve the models themselves,” Holzberger said. “This is a way for institutions to start scaling AI into more disease states, while keeping it in existing workflows, so that radiologists can get home at a reasonable hour.” 

GE Healthcare has also integrated its Edison platform with ACR AI-LAB allowing for faster deployment of AI algorithms and the easier incorporation of EHR data into personalized models.

“The first mile of AI development – data aggregation and annotation – requires cutting-edge technology for the collection of data and clinical expertise for analysis of that information. Edison, with its 100 plus developer services and new integration with ACR AI-LAB, will enable the ACR membership to accomplish the first mile more easily and at scale,” said Karley Yoder, GE Healthcare’s Director of Product Management for AI.

While there are still plenty of unanswered questions when it comes to the intersection of radiology and AI, what’s clear is that the resulting technology will be big business. A 2018 report from Signify Research estimated that the market for AI in medical imaging will top $2 billion by 2023.

Which is not to say that Nvidia is stopping there.

Alongside their partnership with the ACR, the company also announced that they were joining forces with the ATOM Consortium to apply data-driven AI-based approaches to accelerate drug discovery.

Still, it is yet to be seen whether these cross-disciplinary collaborations can break down the traditional borders in the way healthcare is delivered and clinical data can be harnessed.

“What I’m most excited to see is the technology of AI landing in the hands of not just a handful of academics, but a wide range of clinicians,” Michalski said. “This can’t be a siloed technology. In order to be successful it kind of needs to be something everyone owns.”

Photo: Hemera Technologies, Getty Images

Shares1
Shares1