Mayo Clinic AI spots pancreatic cancer on routine scans 475 days before diagnosis

On a CT scan that a radiologist reads as completely normal, an artificial intelligence model can already see pancreatic cancer coming. That is the central finding of a study published in the journal Gut: a Mayo Clinic-developed AI called REDMOD identified hidden tissue changes associated with pancreatic ductal adenocarcinoma — the most common and lethal form of the disease — an average of 475 days before patients received their clinical diagnoses.

The stakes are unusually high. Pancreatic cancer carries a five-year survival rate of just 13%, the lowest among major cancers, according to figures cited by ibtimes.com.au. The National Cancer Institute, as cited by Mayo Clinic, reports that more than 85% of patients do not receive a diagnosis until after the disease has already spread. Roughly 67,530 Americans are expected to be diagnosed in 2026, with approximately 52,740 projected to die from the disease.

Dr. Ajit Goenka, Mayo Clinic radiologist and senior author of the study, has described the central problem in pancreatic cancer as the inability to see the disease when it is still curable and noted that the AI can identify the signature of cancer from a normal-appearing pancreas, according to ibtimes.com.au.

REDMOD — short for Radiomics-based Early Detection MODel — works by analyzing extremely fine patterns in the texture of pancreatic tissue on standard computed tomography scans, a technique known as radiomics. These patterns are not visible to the human eye. The system also automatically outlines the pancreas, separating it from surrounding organs, removing the inconsistencies that arise when different specialists trace the organ's boundaries by hand.

The researchers tested the model by running it on abdominal CT scans from 219 patients who had been treated across multiple hospitals. Every one of those patients had scans that experienced radiologists had read as showing no evidence of disease — yet every one was subsequently diagnosed with pancreatic cancer. The scans had been taken at different intervals: 40% were taken three to twelve months before diagnosis, 35% between one and two years before, and 25% more than two years before — up to around three years. Nearly two thirds of the cancers developed in the head of the pancreas.

Those scans were compared against images from 1,243 people who remained cancer-free for up to three years afterward, matched by age, sex, and scan date. Patients who later developed cancer had an average age of 69; those in the comparison group averaged 64.

The performance gap between the AI and human readers was wide. Overall, REDMOD correctly identified cancer-related signals in 73% of cases, compared with 39% for experienced radiologists reviewing the same scans. The gap widened further for the hardest cases — scans taken more than two years before diagnosis — where REDMOD achieved a 68% detection rate against radiologists' 23%, nearly three times the human rate. Among an independent group of 539 patients drawn from several hospitals, the AI correctly classified just over 81% of scans as cancer-free; that figure rose to 87.5% when applied to a public dataset of 80 patients maintained by the U.S. National Institutes of Health.

Consistency across time added weight to the findings. Analyzing scans taken months apart from the same patients, the researchers found that REDMOD returned matching results in 90% to 92% of cases — a level of consistency they said points to the system picking up genuine disease-related changes rather than random noise.

This temporal window holds profound significance, as attaining such early detection would substantially augment the probability of cure and improved survival.— The study researchers

The researchers also quantified what earlier detection could mean in practice: modelling studies suggest that raising the share of localized pancreatic cancers from 10% to 50% would more than double survival rates. By comparison, when pancreatic cancer is still localized, the five-year relative survival rate is about 44%; when it has spread to distant parts of the body, that figure falls to just 3%, according to medicalnewstoday.com.

For those at risk, certainly at higher risk, this may actually be a game changer... meaning that we're going to see more and more patients being captured earlier in the game, where we actually can make a bigger difference.— Tanios S. Bekaii-Saab, MD, FACP, division chair of Hematology/Medical Oncology at Mayo Clinic and disease group leader for Gastrointestinal Cancers at the Mayo Clinic Comprehensive Cancer Center

Bekaii-Saab noted that earlier detection does more than open the door to surgery — it also allows clinicians to better leverage emerging treatments and tailor therapies to the tumor's biology, according to onclive.com.

Despite the results, the researchers are careful about what comes next. The study was retrospective — researchers looked back at existing scans rather than testing the model in real time on new patients. The study population also lacked sufficient ethnic diversity, which could affect how well the model performs across different groups. Before REDMOD can enter routine clinical care, it must be validated prospectively, particularly in people at higher risk: those with unexplained weight loss or newly diagnosed diabetes, both of which can be early signs of pancreatic cancer.

Mayo Clinic is moving toward that next step. The institution is preparing to launch a prospective study known as AI-PACED — Artificial Intelligence for Pancreatic Cancer Early Detection — that will assess how AI-guided detection performs in the care of high-risk patients, with researchers monitoring early detection rates, false positives, and outcomes over time, according to ibtimes.com.au.

A separate line of AI research, also from Mayo Clinic, is pursuing an even earlier warning signal. Researchers presented findings at the American College of Surgeons Clinical Congress 2026 in Washington, D.C., showing that a model built on electronic health records — including routine blood tests and clinical history from nearly 40,000 patients — could identify people at elevated risk of pancreatic cancer up to five years before diagnosis. Those findings have not yet been peer-reviewed. Lead author Chris Varghese, MBChB, said the model aims to move detection "upstream — from reacting to symptoms to proactively identifying patients who may benefit from earlier, targeted screening."

We found that the model's predictions were strongly informed by routine blood tests, particularly components of the complete blood count, which are often obtained for other reasons or as part of routine care.— Cornelius Thiels, DO, MBA, surgical oncologist at Mayo Clinic and study co-author

Together, the two approaches point toward a future in which a patient's existing health data — blood tests already drawn, CT scans already taken for other reasons — could flag pancreatic cancer long before any symptom appears. Whether that future arrives depends on the prospective trials now beginning.

Why it matters — Pancreatic cancer kills most of the people it touches because it is almost always caught too late — REDMOD and related AI tools represent the first credible path toward catching it while surgery and treatment can still make a decisive difference.

Reported by sciencedaily.com, mayoclinic.org, eurekalert.org, ibtimes.com.au, onclive.com, medicalnewstoday.com