Mayo Clinic: AI learns to spot pancreatic cancer years before clinical diagnosis

Mayo Clinic has unveiled an AI system that finds signs of pancreatic cancer on routine CT scans well before clinical diagnosis. In a validation study, the REDMOD model spotted the disease an average of 475 days before diagnosis, and noticeably more often than specialists working without an AI hint.

Why it matters

Pancreatic cancer is one of the deadliest cancers not because it cannot be seen at all, but because it is usually noticed too late. In its early stages the tumor causes almost no symptoms, and on scans it often looks like normal tissue. According to Mayo Clinic, more than 85% of patients learn of their diagnosis only after the disease has spread, and five-year survival remains below 15%.

That is why the idea of catching the disease on scans taken for other reasons is so compelling. CT is already widely used in clinics, and if AI can reliably pick out faint precursors of the disease on routine CT scans, doctors gain extra time for follow-up testing, monitoring and, in some cases, treatment at a stage when surgery is still possible.

What the validation showed

The Mayo Clinic team tested the system on data closer to real practice than typical lab demos. REDMOD analyzed almost 2,000 CT scans from different clinics, different scanners and different protocols. The main sample included 219 patients whose scans radiologists had earlier judged normal, although these people were later diagnosed with pancreatic cancer, plus 1,243 people in a control group who received no such diagnosis in the following three years.

• The model recognized 73% of pre-diagnostic cases, at a median of about 16 months before diagnosis
• The average lead time was about 475 days
• On scans taken more than two years before diagnosis, the AI showed roughly a threefold advantage over specialists without the system's support
• REDMOD's sensitivity was 73% versus 39% for radiologists, and 68% versus 23% in the subgroup "more than two years before diagnosis"
• On repeat scans of the same patients, the model's outputs stayed consistent in 90–92% of cases

The approach does not rely on the algorithm "seeing the invisible" by some magic. REDMOD measures hundreds of quantitative tissue features: texture, structure and faint changes that the human eye does not usually pick out as an obvious tumor. According to the authors, the system works automatically and does not require labor-intensive manual preparation of scans before analysis.

"The main barrier to saving lives in pancreatic cancer is our

inability to see the disease while it is still curable."

Where the method's limits lie

Despite the strong numbers, this is not a story about ready-made mass screening tomorrow. The study was a validation study and largely retrospective: the model was tested on existing scans, not in a real patient flow, where routing, follow-up examinations, the cost of errors and the workload on doctors all have to be taken into account. The authors state plainly that prospective validation is needed to confirm the approach's clinical usefulness.

There are also more practical questions. Even a good AI in oncology must not only find suspicious cases but also avoid creating too many false alarms. In addition, the sample was not ideally diverse in ethnic makeup, so how well the results transfer to all populations remains to be proven.

The next step is already under way: in the AI-PACED study, doctors are testing how to build such prompts into the care of higher-risk patients, for example those with new-onset diabetes.

What this means

For AI in medicine, this is a rare piece of news where the value is obvious at once: not a chat interface and not administrative automation, but a chance to move the diagnosis of one of the deadliest cancers months or even years earlier. If prospective trials confirm REDMOD's results, hospitals will get a "second look" tool for CT scans already taken, and early diagnosis will become a working process rather than a theory.