ROCHESTER, Minnesota — A Mayo Clinic-led research team developed an artificial intelligence system that identified subtle pancreatic cancer-associated changes in CT scans a median of 475 days before patients were diagnosed. The results raise the possibility of using existing medical scans to help flag a dangerous cancer earlier, but the model remains investigational and is not approved as a stand-alone screening or diagnostic test.
The study, published in Gut in April 2026, evaluated the Radiomics-based Early Detection Model (REDMOD). In an independent testing group, the model identified 73% of the scans associated with a later pancreatic cancer diagnosis. Expert radiologists reviewing the prediagnostic scans detected 38.9%, according to the peer-reviewed study and Mayo Clinic.
What the AI actually detected
Pancreatic ductal adenocarcinoma can develop without obvious symptoms, and an early tumor may be too small or subtle to recognize on a routine CT scan. REDMOD does not literally see a tumor that is invisible to radiologists. Instead, it automatically identifies the pancreas and measures small differences in tissue patterns associated with a higher likelihood of future cancer diagnosis.
The system uses a 40-feature radiomic signature drawn largely from detailed, filtered image data. Such quantitative patterns can differ even when a scan appears normal to a human observer. The study’s authors describe a potential role for automated analysis of CT scans already performed for other medical reasons, particularly for people whose clinical histories put them at elevated risk.
The numbers behind the 475-day finding
The researchers trained REDMOD on 969 CT scans: 156 from patients who were subsequently diagnosed with pancreatic cancer, and 813 controls. They then assessed it on 493 separate scans, including 63 prediagnostic cases and 430 controls. In that independent test set, REDMOD correctly flagged 46 of the 63 future cases.
- 73% sensitivity: the proportion of future cancer cases the model identified in the independent test set.
- 81.1% specificity: the proportion of controls correctly classified.
- 38.9% sensitivity: the combined rate achieved by radiologists who reviewed the earlier scans without AI assistance.
- 475 days: the median interval from the CT scan carrying an AI-detectable signal to the eventual clinical diagnosis.
Performance was also examined on scans obtained more than two years before diagnosis, where the AI identified a substantially greater share of cases than radiologists. These are retrospective detection results—not evidence yet that the system reduces cancer mortality.
False alarms remain a major hurdle
In the independent test set, the model flagged 46 scans belonging to people who would be diagnosed with cancer, but also flagged 81 controls. Its reported precision was 36.2%: approximately one in three positive results in that particular test group was a true case.
The test group contained an unusually high proportion of people who later developed pancreatic cancer compared with the general population. That means the proportion of false alerts could be higher in routine population-wide use. A false-positive finding can cause distress and lead to unnecessary scans, appointments or procedures. Clinicians will need to establish how to investigate alerts safely and efficiently.
Why early warning could matter for U.S. patients
The National Cancer Institute estimates that 67,530 people in the United States will receive a pancreatic cancer diagnosis in 2026, while 52,740 will die from the disease. The five-year relative survival rate is 13.7% overall. It is 43.6% for disease diagnosed while localized, compared with 3.4% when the cancer has spread to distant sites.
These differences help explain the urgency of early-detection research. But a promising signal on a scan is not the same as a confirmed diagnosis, and the trial did not demonstrate that AI-triggered interventions extend life. Future research must show whether earlier flags lead to effective treatment and better outcomes.
What happens next
Mayo Clinic says its prospective AI-PACED research program is evaluating how clinicians might incorporate these alerts into care for patients at elevated risk, including some people with new-onset diabetes. Researchers will need to assess false-positive rates, consistency across hospitals and imaging machines, how patients are followed, and whether the approach produces measurable clinical benefits.
The takeaway is neither that AI has already cured pancreatic cancer nor that a normal CT scan guarantees safety. The study suggests that quantitative information in existing scans may help specialists identify risk earlier than visual interpretation alone—provided prospective trials establish a safe way to use it.
Research video and related reading
Watch Mayo Clinic’s REDMOD demonstration and read its institutional report. A separate short-form explainer discusses the findings; some claims in the short require the qualifications described above.
