AT WAKE RADIOLOGY in North Carolina, roughly 50 doctors scrutinize x-rays and other images for local medical providers. Within a few weeks, they should start to get help on some lung CT scans from machine-learning algorithms that highlight potentially cancerous tissue nodules. Although Wake is based in a region known as the Research Triangle, for its intensity of high-tech R&D, the lung-reading software hails from elsewhere—China.
Infervision, a four-year-old Beijing startup, has amassed more than a million scans from Chinese hospitals that it’s using to train and test algorithms. Gathering medical data is much easier for Chinese companies than for their US counterparts, because patient populations are larger and the burden of privacy regulations smaller.
“In the US, particularly for big academic hospitals, you have to go through so many processes and it can take a really long time to access data,” says Yufeng Deng, Infervision’s chief scientist. Chinese institutions do take steps to protect patient privacy, such as anonymizing records used in research, he says, but they are not bound by as many rules and regulatory processes. “In China it’s less well-defined—if the hospital’s information people say it’s fine, you can do it,” Deng says.
To create its algorithm to identify lung nodules, Infervision gathered more than 400,000 lung scans from Chinese partners, such as leading Beijing research center Peking Union Medical College Hospital. Over two years, it then sent each image for review by three radiologists at its Beijing office. Their annotations created the feedstock to train and test image-processing algorithms, in the same way internet companies train systems to recognize cats, dogs, and people. Infervision has published peer reviewed studies in Chinese and US journals on its algorithms’ performance. In a pilot at Shanghai Changzheng Hospital, two radiologists found that Infervision’s product could dramatically boost their ability to annotate lung nodules, the company says.
William Way, a radiologist and Wake’s chief medical officer, is hopeful his own doctors will also get a productivity boost. He says the technology looks good enough to be a kind of virtual assistant that improves accuracy and consistency on a tricky task. “The process of looking for lung nodules is pretty tedious,” he says.
Infervision’s software is intended to speed up the work of radiologists, not make diagnoses on its own—Deng says he can’t imagine AI being ready to do that within a decade. The company is in the process of seeking FDA approval to market its product in the US. In the meantime, it will provide software to Wake, Stanford Children’s Hospital, and other US partners interested in testing the tool for free. The startup has also opened offices in Germany and Japan. In China, Infervision is refining its lung-analysis software to look for other things, like bone fractures and emphysema, and testing algorithms that analyze brain scans for signs of stroke.
Infervision’s leaders think their Chinese data trove provides an advantage over competitors working only with hospitals in the US, where data access is harder won. The startup’s strategy helps illustrate the global competition in developing AI technology—and one way that China may have an advantage. Training machine-learning algorithms requires large amounts of example data relevant to the task at hand. China’s huge population and relatively loose privacy rules give the country’s AI developers more to work with than those in the US.
The laser scans collected by prototype self-driving cars on Chinese roads, or billions of Mandarin social media posts, are unlikely to be of much value in Brooklyn or Kalamazoo. But human anatomy makes insights derived from medical scans like those accumulated by Infervision a naturally global currency. “China has a decided advantage with respect to the quantity of data and the nurturing of the government,” says Eric Topol, a professor at Scripps Research and author of a forthcoming book on AI and medicine.
That doesn’t mean Chinese companies will always win. AI-assisted medical imaging is still relatively unproven in the clinic, and more data doesn’t automatically mean better results for an AI program. Topol says it’s an open question how well algorithms trained on Chinese patients and scanning equipment will perform when given data from US patients and imaging technology.
Deng of Infervision says that the startup refined its algorithms with around 2,000 US-sourced images to adapt them to American patients and imaging equipment. Wake and other partners will provide feedback on any errors radiologists spot in its recommendations.
Gathering a big trove of health data in the US typically requires negotiating with multiple partners, all of whom know their data’s value. That can drive up costs beyond the means of startups without large cash reserves. IBM has spent more than $3.5 billion since 2015 acquiring health care software companies and amassing millions of patient records and billions of images of all kinds of medical conditions. The company has published research on medical-image-processing AI software but not yet launched a commercial service. IBM didn’t respond to requests for comment.
Other countries are trying to use easier access to medical data to boost their own AI industries. French president Emmanuel Macron’s AI strategy announced last yearincludes a pledge to make data from France’s universal health care system available for AI research. The Canadian province of Ontario is using its single-payer health system to lure more investment into its already vibrant AI R&D scene.
China’s government has also made health care and AI a priority. Support for development of medical uses of the technology is part of a national AI strategy launched in 2017. The country’s over-stretched hospitals are also generally more open to technology that might augment the work of doctors than their counterparts in the US, says David Yuan, a partner at Redpoint Ventures, which invested in Infervision. The startup has around 300 employees and has raised more than $73 million from investors, including the Chinese arm of storied Silicon Valley venture firm Sequoia and leading Chinese venture firm Qiming Ventures.
In North Carolina, Wake’s chief information officer, Matt Dewey, says AI-enhanced efficiency can help US radiologists and patients, too. In addition to Infervision, Wake is testing algorithms from other startups that estimate the bone age of pediatric patients and measure the volume of different parts of a person’s brain. All were integrated into Wake’s existing systems via radiology platform EnvoyAI. In a few years, Dewey envisions Wake’s embrace of AI helping the company draw more referrals than competitors because it can process cases more quickly and provide richer, more accurate reports. “The thing that we’re really selling is our reports,” he says. “If we do that better because of AI then we should get more patients.”