By the time doctors detect pancreatic cancer, it’s often too late to treat effectively. But a new study suggests that artificial intelligence might be able to find signs of the disease before tumors are visible on a scan.

An AI model developed at the Mayo Clinic in Rochester, Minnesota, detected abnormalities on patients’ CT scans up to three years before they were diagnosed with pancreatic cancer, according to research published this week in the journal Gut.

The scientists behind the model, which is now being evaluated in a clinical trial, trained it by feeding it CT scans from patients who had been screened for other medical conditions then were later diagnosed with pancreatic cancer. The team then had radiologists review the scans and compared their ability to find early signs of cancer to that of the AI model. The model was found to be three times better at identifying the early signs.

“We knew, based on the biology of the disease, that this is not something which is coming all of a sudden in three months. … We knew that the signal was there. We just needed to find a way to be able to detect it,” said Dr. Ajit Goenka, a radiologist at the Mayo Clinic and an author of the study.

With a five-year survival rate of 13%, pancreatic cancer is on track to become the second leading cause of cancer deaths by 2030. Around 80% of patients are diagnosed after the disease has reached an advanced stage.

Unlike colon or breast cancer, there is no routine screening for pancreatic cancer in healthy people. Feeling for a lump is nearly impossible, since the pancreas is buried deep in the abdomen. And typical symptoms like stomach pain and sudden weight loss usually don’t begin until the cancer has spread to other organs.

Early markers of the disease are often too subtle to be seen by the human eye on a scan. In many cases, patients’ scans appear normal as little as six months before they’re diagnosed.

“I analyze these images every day,” said Dr. Daniel Jeong, a diagnostic radiologist at Moffitt Cancer Center in Tampa, Florida, who was not involved in the Mayo Clinic research. “We’re really looking for a measurable mass that could represent the cancer. So these tumors need to grow to a certain level to become visible.”

Goenka said one signature of early cancer that the AI model was able to detect is abnormal cells in the pancreas that shelter and protect cancer from the body’s immune defenses. Scientists have known such cells exist but struggled to find them on scans.

Goenka said the model could someday be useful for people who already have risk factors for the disease — such as a family history or diabetes — but don’t have any symptoms. From there, doctors would likely recommend additional blood work and imaging.

“Unfortunately, if they have symptoms and if it’s truly pancreas cancer, you don’t need AI for that,” he said.

The AI model is one of several recent advancements in pancreatic cancer research. Last month, the results of an early-stage trial showed that an mRNA vaccine prolonged survival in eight patients with pancreatic cancer. An experimental drug called daraxonrasib was also found to double patients’ life expectancy in a later-stage trial. The Food and Drug Administration said Friday that the drug’s developer, Revolution Medicines, could expand access to it in a controlled setting for patients with advanced pancreatic cancer who have already tried other treatments. Daraxonrasib hasn’t been approved by the FDA yet.

Additionally researchers are working on advanced blood tests that could look for hallmarks of the disease.

“We’re making, actually, major strides. It hasn’t turned this disease around,” said Dr. Tamas Gonda, director of the pancreatic disease program at NYU Langone’s Perlmutter Cancer Center.

Other doctors not affiliated with the Mayo Clinic study said that by catching signs of pancreatic cancer before it spreads to major organs or blood vessels, the AI model might increase the number of patients who could be candidates for surgery, chemotherapy or radiation.

“This really could be a game changer for us for early detection,” said Dr. Pam Hodul, a surgical oncologist at Moffitt Cancer Center.

AI has shown increasing promise as a diagnostic tool for a range of other conditions. In a study published this week in the journal Science, researchers found that AI either matched or outperformed human physicians at diagnosing emergency room patients.

Goenka said it could be a while before the Mayo Clinic’s model reaches the public, since the clinical trial needs to follow participants for three to five years to see who develops cancer and who does not.

“In a disease where we have been just wandering in darkness for decades, this is a milestone that shows us the finish line, but we still have to get to the finish line,” he said.