Medical AI finds deadly pancreatic cancers 3 years earlier

Pancreatic cancer has long been one of medicine's most formidable adversaries. By the time most patients receive a diagnosis, the disease has already spread beyond the pancreas, leaving doctors with few curative options and patients with devastating odds. A five-year survival rate below 14 percent tells the story in stark terms: the problem isn't that treatments don't exist, it's that we almost always find the cancer too late.

A new artificial intelligence system developed by researchers at the Mayo Clinic may fundamentally shift that equation. The model, called the Radiomix-based Early Detection Model, can identify subtle signatures of pancreatic cancer on CT scans that appear completely normal to human radiologists—sometimes up to three years before a traditional diagnosis would occur. The findings were published this week in the journal Gut, capping a multi-year research effort by the Mayo team.

The scale of the problem cannot be overstated. Pancreatic cancer accounts for just 3 percent of all cancer cases in the United States, with roughly 67,000 new diagnoses annually, yet the National Cancer Institute projects it will become the second-leading cause of cancer death by 2030. More than 85 percent of patients are diagnosed only after metastasis has already occurred. The window for curative surgery—a Whipple procedure or distal pancreatectomy—closes rapidly once the disease spreads. Any technology that can reopen that window, even by months, represents a potential turning point.

The Mayo Clinic team trained their AI on nearly 2,000 abdominal CT scans, a portion of which came from patients who were later diagnosed with pancreatic cancer. Critically, every one of those scans had originally been read as normal. The AI identified 73 percent of the future cancer patients at an average of 16 months before their actual diagnosis. In cases where scans were available two or more years before diagnosis, the detection rate jumped even higher—the model caught nearly three times as many early cancers as conventional screening would have flagged.

"The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable," said Dr. Ajit Goenka, a Mayo Clinic radiologist and the study's senior author. "This AI can now identify the signature of cancer from a normal-appearing pancreas, and it can do so reliably over time and across diverse clinical settings."

What makes this model particularly promising for real-world deployment is its design philosophy. Rather than requiring dedicated screening protocols or specialized imaging sequences, the AI is built to run passively on any abdominal CT scan ordered for any reason—a kidney stone workup, a trauma evaluation, routine follow-up. If the scan includes the pancreas, the model analyzes it. If the AI detects elevated risk, it flags the patient for closer follow-up, potentially triggering earlier MRI or endoscopic ultrasound before a tumor is ever visible to the human eye.

The stability of the results is also noteworthy. Researchers reported that the model produces consistent risk scores on repeat scans months apart, which means it could be used not just for one-time detection but for longitudinal tracking—watching how a patient's risk profile evolves over time. That kind of serial monitoring is essentially impossible with current clinical workflows, where a "normal" scan is filed away and forgotten.

There are important caveats. The study is retrospective, meaning the AI was tested on scans where the outcome was already known. Prospective clinical trials—where the model's flags are acted upon in real time—will be needed before this technology can be integrated into standard care. False positives remain a concern; flagging too many patients for invasive follow-up could cause harm and erode physician trust. And the model's performance across different scanner manufacturers, hospital systems, and patient demographics still needs broader validation.

But the potential upside is enormous. If even a fraction of the 85 percent who are currently diagnosed late could be caught earlier, the survival statistics for pancreatic cancer could look dramatically different within a decade. A disease that kills roughly 50,000 Americans a year might finally meet a detection tool worthy of the challenge.

## What This Means For You

If you or a family member has ever had an abdominal CT scan for any reason, this research suggests that in the near future, a second look by an AI system could detect cancer risk that your radiologist never saw. While this technology isn't available in clinics yet, it's moving toward clinical trials—and when it arrives, it could transform routine scans into powerful early-detection tools. For patients with family histories of pancreatic cancer or genetic risk factors like BRCA mutations, this could eventually mean far more effective surveillance without additional invasive procedures. Ask your doctor about AI-augmented screening options, especially if you're at elevated risk. The earlier we find this disease, the more lives we save.