Mayo Clinic AI spotted pancreatic cancer signals over a year before diagnosis in study

Abdominal CT scans judged free of pancreatic cancer contained warning patterns an AI system could recognize long before patients were diagnosed. In a retrospective study published in Gut, Mayo Clinic’s REDMOD identified cancer-associated imaging signatures an average of 475 days before clinical diagnosis. The finding opens a possible earlier detection window, but its usefulness in patient care remains unproven.

Reading texture, not a visible tumor

Mayo Clinic’s REDMOD, short for Radiomics-based Early Detection Model, analyzes subtle texture changes in pancreatic tissue on routine abdominal CT scans. Radiomics means extracting fine-grained image patterns that are not readily visible to someone inspecting the scan. The target is an imaging signature associated with pancreatic ductal adenocarcinoma, the most common type of pancreatic cancer.

The system also finds and outlines the pancreas automatically, separating it from surrounding organs and tissue. That removes the need for specialists to trace its boundaries by hand, a step that can introduce differences between operators. Automation is therefore part of the framework’s design, not just the final cancer-risk assessment.

Researchers examined scans from 219 patients at multiple hospitals who were later diagnosed with pancreatic cancer. Experienced radiologists had found no evidence of the disease when those scans were originally assessed. The comparison group contained 1,243 people who remained free of pancreatic cancer for up to three years after imaging.

The cancer group covered several lead times: 87 patients had scans three to 12 months before diagnosis, 76 had scans 12 to 24 months earlier, and 56 had imaging more than two years earlier, extending to around three years. This let researchers examine whether the signal persisted well before the eventual diagnosis, rather than only immediately beforehand.

An advantage even on older scans

Sensitivity measures how often a method identifies the people who have the condition being assessed, rather than missing them. REDMOD’s advantage was larger on scans taken more than two years before diagnosis: it detected 68% of cases, compared with 23% for radiologists. Those results concern recognition of early signals in historical images, not diagnoses delivered to patients at that time.

Researchers also checked repeat scans from the same patients taken several months apart. REDMOD produced matching results in 90% to 92% of cases. That supports the stability of its classifications over time, although consistency alone does not establish whether a prediction is correct.

Finding cancer is only half the test

Recognizing scans without cancer is a separate requirement. In an independent group of 539 patients from several hospitals, REDMOD correctly classified just over 81% of scans as cancer-free. On the public NIH-PCT dataset of 80 patients, that figure was 87.5%. These results show that the system did not correctly clear every cancer-free scan.

The researchers also identified limited ethnic diversity in the study population, leaving uncertainty about performance across different populations. They called for prospective testing—following patients forward rather than looking back at existing records—particularly among people at increased risk. Examples include people with unexplained weight loss or recently diagnosed diabetes.

That next step must establish clinical usefulness. The researchers’ goal is to move detection away from late-stage, symptomatic disease toward an earlier opportunity to intervene. Their retrospective findings support testing that approach; they do not yet establish that deploying REDMOD in routine care improves survival.