On a recent day at a hospital in western Beijing, a cancer radiologist named Chongchong Wu loaded a suspicious-looking lung scan into a computer program resembling Photoshop. A neural network trained on thousands of example scans highlighted nodules in red squares, which she examined carefully. She corrected two false positives where the network mistakenly identified blood vessels as potential malignancies. But she also found a nodule that she’d previously overlooked, perhaps indicating an early sign of disease.
China is embarking on a big initiative to add AI to health care with tools like this one. In some ways the trend mirrors what is happening in the US and Europe. In China, however, restrictions on the use of data and new technologies are looser, and the need for automation is more pronounced. There are 1.5 doctors for every 1,000 people in China, compared with 2.5 in the US.
China is moving quickly. Some 131 companies are currently working on applying AI in the country’s health-care sector, according to Yiou Intelligence, a Beijing-based consultancy. Starting next month, a hospital in Beijing will run all its lung scans past an AI algorithm in order to expedite the screening process.
The Chinese government has called for such technology to help with computerized medical diagnosis as part of the first stage of its grand plan to embrace AI by 2020 (see “China’s AI Awakening”). In a report published in February, IDC predicted that China’s market for AI health-care services will reach 5.9 billion yuan ($930 million) in 2022. This market is also being targeted by China’s big tech companies. Both Alibaba and Tencent have research units dedicated to developing AI diagnostic tools.
The way people perceive AI in China may make it easier for the technology to flourish in medicine. In the West, advances in AI have prompted debates about job losses, but most Chinese doctors seem keen to automate away their most repetitive work.
Using AI in medicine comes with challenges, though. The diagnostic tools may reach their conclusions using complicated mathematical processes that defy explanation. So far there is little debate in China about who will be responsible for mistakes when medical diagnoses are outsourced to these algorithms.
Last year, the China Food and Drug Administration incorporated AI diagnostic tools into its list of permitted medical devices, but companies need to apply for accreditation for each product before setting a price.
The AI software Wu used, developed by a Beijing-based startup called PereDoc, has been installed in more than 20 hospitals in China. PereDoc has amassed a network of more than 180 hospitals that serve as research collaborators.
Crafting algorithms that can process medical images, such as CT scans and x-rays, is a particularly hot field for China’s startups. One reason is that image classification plays to the strength of the latest deep-learning algorithms.
But AI is also being used in other areas. Peijun Lv, a prosthodontist in Beijing, is collaborating with Tsinghua University to develop an AI program that can design dentures. A prototype algorithm was trained using rules on denture design, drawn from textbooks, and 30,000 real cases labeled by doctors. “It can replicate the expertise of experienced doctors,” says Lv. He plans to run clinical trials of the algorithm later this year.
And Peng Liu, a lymphoma doctor in Beijing, is working with researchers at Tsinghua to develop a machine-learning algorithm that can use ultrasound data to detect blood clots caused by lymphoma treatment. If caught early, often via an ultrasound scan of the patient’s veins, blood clots can be easily treated. But hospitals often do not have enough resources to screen every patient unless there are specific symptoms.
Other researchers in China are tackling general medical knowledge. iFlytek and Tsinghua University jointly created an AI system that scored higher than over 96 percent of human contestants in last year’s Chinese medical licensing exam. The difficulty of creating a system like this is not incorporating the breadth of existing medical knowledge, but teaching machines to understand the intricate connections between different facts and use these findings to reason and make decisions.
At its core, this is a natural-language-processing system that’s particularly adept at dealing with medical questions. The way it reaches a conclusion on a multiple-choice question is completely different from the way a human chooses the best answer. The algorithm looks for evidence needed to answer a particular question by calculating statistical similarities between words represented mathematically.
A detailed analysis of the exam results shows where machines cannot compete with humans: common sense and ethics. The algorithm scored lower than the national average on the section that tests ability to exercise judgment under stressful situations such as family disputes.
Ji Wu, an associate professor at Tsinghua University who led the project, is exploring ways to put this algorithm to clinical use. But he admits it’s not going to be as simple as installing this software in every doctor’s computer.
Doctors who do use the new tools can find them a big help, though. At Chongchong Wu’s hospital in Beijing, for example, the outpatient department sees about 10,000 people every day, so she doesn’t have enough time to read every image as carefully as she’d like. The scan-processing program, she says, “can relieve my burden.”
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Data analytics reveal real business value
Sophisticated analytics tools mine insights from data, optimizing operational processes across the enterprise.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.