In China, AI is finding deadly tumours that doctors might miss

By Vivian Wang

BEIJING – Three days after Qiu Sijun, a retired bricklayer in eastern China, went for a routine diabetes checkup, he received a call from a doctor he hadn’t met before. The doctor, the head of the hospital’s pancreatic department, wanted him to come in for a follow-up.

“I knew it couldn’t be anything good,” Qiu, 57, recalled.

He was partly right. The bad news was that Qiu had pancreatic cancer. But there was good news, too: The tumour had been detected early. The doctor, Zhu Kelei, was able to remove it.

This was possible only because of a new artificial-intelligence-powered tool that the hospital was testing, which had flagged Qiu’s routine CT scan before he had any symptoms. The tool is one example of how Chinese tech companies and hospitals are racing to apply AI to some of medicine’s most stubborn problems.

Pancreatic cancer is one of the deadliest forms of cancer, with a five-year survival rate of around 10%, largely because early detection is so hard. Symptoms often do not appear until the cancer has advanced.

Tests used to confirm its presence, such as contrast CT scans, involve large amounts of radiation, so many experts advise against widespread screening. But lower-radiation alternatives, such as noncontrast CTs — in which no contrast dye is injected into patients’ bloodstreams — produce less clearly defined pictures, making it hard for radiologists to identify abnormalities.

Artificial intelligence may change that. The tool at Zhu’s hospital, which was developed by researchers affiliated with Chinese tech giant Alibaba, was trained to look for pancreatic cancer in noncontrast CTs.

The tool is called PANDA, short for ‘pancreatic cancer detection with artificial intelligence’. At the hospital where Zhu works, the Affiliated People’s Hospital of Ningbo University in eastern China, doctors started using it as part of a clinical trial in November 2024.

The tool has since analyzed more than 180,000 abdominal or chest CTs, helping doctors detect about two dozen cases of pancreatic cancer, 14 of which were in the early stage, Zhu said. The tool found 20 cases of intraductal adenocarcinoma, the most common and deadliest type of pancreatic cancer. (Qiu had a neuroendocrine tumour, which is a rarer and less aggressive cancer.)

All of those patients had come to the hospital with complaints like bloating or nausea and had not initially seen a pancreatic specialist, Zhu said. Several of their CT scans had raised no alarms until they were flagged by the AI tool.

“I think you can 100% say AI saved their lives,” he said.

In April, Alibaba said the US Food and Drug Administration had granted PANDA “breakthrough device” status, meaning that its review would be expedited to help it get to market. It is also the subject of several clinical trials in China.

Researchers cautioned that more real-world data was needed to show whether the tool could detect enough early-stage cases to outweigh the risks of false positives and unnecessary testing. Scientists elsewhere are studying other AI-assisted approaches to early pancreatic cancer detection that focus more narrowly on high-risk groups, in large part because the cancer’s prevalence is low.

Several experts not involved with the Chinese research said they were sceptical that noncontrast CTs could offer as much valuable information as other forms of imaging.

Even the engineers behind PANDA initially shared that concern, said Ling Zhang, a senior staff algorithm engineer at Damo Academy, Alibaba’s research arm, who is one of the tool’s creators.

To address it, they asked a radiologist to manually annotate the contrast CTs of more than 2,000 known pancreatic patients with the locations of their lesions. The engineers then algorithmically mapped the highlighted lesions onto the same patients’ noncontrast CTs. Those noncontrast scans were then fed to the AI model, so that it could learn to detect potential cancer even in less detailed images.

When the tool was subsequently tested on more than 20,000 noncontrast CTs, it correctly identified 93% of people who had pancreatic lesions, according to a study published in Nature Medicine in 2023.

“The effectiveness actually surprised us,” Zhang said.

At the Ningbo hospital, the system is being used to analyze scans that doctors had already ordered for other reasons, so there is no additional testing cost to the hospital or patients. (In China, many people routinely undergo noncontrast CTs as part of their annual checkups; at the Ningbo hospital, a noncontrast CT costs about $25, before insurance.)

Zhu and his team review any scans that the system marks as high-risk, and if necessary, they call the patients in for more detailed testing.

The model still can’t compare to a pancreatic specialist, Zhu said.

It sometimes highlights cases of pancreatitis, and it cannot say whether a tumour originated in the pancreas or spread from a different organ. Since its launch, the model has issued alerts for about 1,400 scans, but only about 300 of them needed follow-up, doctors decided.

Dr Ajit Goenka, a radiologist at Mayo Clinic who is researching the early diagnosis of pancreatic cancer, said it was crucial to reduce the number of false alarms. It is possible that hundreds of people in Ningbo “faced the terror of a potential pancreatic cancer diagnosis, underwent unnecessary callbacks, and likely endured expensive, invasive follow-up testing — only to find out they were healthy,” he wrote in an email.

The tool might also be more useful for trainee doctors than for experienced specialists, said Dr Diane Simeone, a pancreatic surgeon at the University of California San Diego. Some of the tumours that the tool caught in the Nature Medicine study should have been “super obvious” to well-trained radiologists even without AI, she said.

But she acknowledged that it could be a valuable backstop for hospitals where specialists are in short supply. (PANDA is also being tested at a clinic in rural Yunnan Province.)

“You’re going to have different skill sets at different centres, depending on where in the world you are or the clinical volume,” Simeone said.

In Ningbo, the technology’s apparent success has brought some new challenges. The hospital currently does not have enough staff to contact all the patients who need follow-up, Zhu said. And its aging hardware has trouble keeping up with the model’s large amounts of data. Several times, when Zhu tried to pull up PANDA on his computer, it froze.

Detecting cancer before patients have symptoms can also create its own problems. In China, widespread medical corruption has eroded public trust in doctors. Some people may refuse to come in for follow-ups, Zhu said, because they worry that the hospital is simply trying to make money.

Qiu was not one of them. He did not hesitate when Zhu advised him to have his tumour removed, though he later said he did not use AI or understand how it worked. During a follow-up visit with the doctor in November, Qiu said he felt perfectly healthy and was busy growing vegetables on his family farm.

“He said I was very lucky,” Qiu said. “So there was nothing else for me to say. I could only be relieved.”

-New York Times