Mayo Clinic Pancreatic Cancer AI Sees a Risk Signal That Radiologists Miss

Mayo Clinic pancreatic cancer AI detected hidden warning signs on routine CT scans up to three years before patients received a clinical diagnosis. The model found 73% of prediagnostic cancers in an independent test set. Expert radiologists reviewing the same scans found 38.9%.

That gap is the real story. The scans were not obviously abnormal, and the patients had undergone imaging for reasons unrelated to pancreatic cancer. Yet the model found patterns associated with cancers that became clinically visible months or years later.

The system, called the Radiomics-based Early Detection Model, or REDMOD, does not diagnose cancer by itself. It assigns risk using quantitative patterns across pancreatic tissue that appears normal to human readers.

Researchers from Mayo Clinic, MD Anderson Cancer Center, and the University of Washington reported the results in Gut. Their retrospective study gives medical AI a demanding target: detecting a rare cancer before a visible tumor provides a clear diagnostic clue.

However, the model has not shown that its alerts improve survival in clinical practice. Its next test is whether hospitals can use those alerts without creating excessive false positives, invasive procedures, and patient anxiety.

Mayo Clinic Pancreatic Cancer AI Found Signals Years Early

REDMOD changed what researchers could extract from an otherwise normal-looking CT scan, not what a radiologist could visibly see.

The researchers assembled prediagnostic CT scans from 219 people who later developed pancreatic ductal adenocarcinoma. This disease, often shortened to PDAC, is the most common form of pancreatic cancer.

Those scans were taken between three and 36 months before diagnosis. They had originally been interpreted without identifying a suspicious pancreatic mass or another focal abnormality.

The study also included 1,243 control scans from people without pancreatic cancer. The control participants had at least three years of follow-up, reducing the chance that an undiagnosed cancer had been mislabeled as normal.

Researchers divided the data into separate development and testing groups. REDMOD trained on 969 scans, including 156 prediagnostic cases and 813 controls.

The independent test set contained 493 scans. It included 63 prediagnostic cancer cases and 430 controls, none of which were used to train the system.

In that test, REDMOD achieved an area under the receiver operating characteristic curve of 0.82. This measure, known as AUC, reflects how well a model separates positive cases from controls across different thresholds.

The model reached 73% sensitivity and 81.1% specificity at the selected threshold. Sensitivity measures the share of cancer cases detected, while specificity measures the share of controls correctly classified as negative.

The median lead time was 475 days, or about 16 months, before clinical diagnosis. Some evaluated scans came from as early as three years before diagnosis.

These results support the headline claim that Mayo Clinic pancreatic cancer AI can flag risk up to three years early. They do not mean every cancer was detected three years in advance.

That distinction matters. The three-year figure describes the outer boundary of the study window, while 475 days describes the typical lead time in testing.

According to the peer-reviewed REDMOD study, performance changed with the interval before diagnosis. The model’s advantage over radiologists became larger among scans taken more than two years earlier.

REDMOD detected 68% of cases in that subgroup, compared with 23% for the radiologists. The researchers cautioned that longer lead-time subgroups contained fewer cases, making their estimates less certain.

Still, the result challenges a familiar assumption. A CT scan judged normal may contain measurable tissue changes long before a pancreatic mass becomes visible.

Why Earlier Pancreatic Cancer Detection Remains So Difficult

The clinical pressure comes from a narrow treatment window and the absence of a standard screening strategy for average-risk adults.

Pancreatic cancer often produces vague symptoms or no recognizable symptoms during its earliest stages. By the time weight loss, abdominal pain, or jaundice prompts imaging, the cancer may have spread.

Earlier detection matters because surgery offers the clearest path toward long-term survival for eligible patients. Many patients never reach surgery because clinicians find the disease after it has advanced.

Screening the entire population is not a simple answer. Pancreatic cancer is uncommon compared with breast, lung, colorectal, or prostate cancer.

Even a test with apparently strong accuracy can generate many false alarms when deployed against a rare disease. The number of people without cancer is so large that imperfect specificity becomes a major clinical problem.

The National Cancer Institute says there is no standard pancreatic cancer screening test for average-risk people. Its current screening guidance focuses surveillance on people with elevated risk.

Those groups include people with substantial family histories, inherited cancer-associated variants, and certain pancreatic lesions. Some clinical programs also investigate new-onset diabetes as a risk-enrichment signal.

Risk enrichment means selecting a smaller population whose cancer rate is higher than the general population’s. That makes a secondary imaging or diagnostic test more practical.

REDMOD fits this targeted approach better than universal screening. The model is designed to analyze CT scans already obtained during routine care, especially among patients with established warning factors.

This design avoids exposing everyone to new CT radiation solely for pancreatic screening. It also turns previously acquired imaging into another source of longitudinal risk information.

Yet it creates a coverage limitation. People without an eligible abdominal CT scan cannot benefit from the model, regardless of their underlying risk.

The training population also does not reproduce true population screening conditions. The study used about one prediagnostic cancer case for every six controls.

That ratio is considerably more cancer-enriched than an unselected clinical population. The researchers used it as a demanding low-prevalence design relative to balanced case-control experiments, not as a copy of population incidence.

A lower real-world prevalence can sharply reduce positive predictive value. Positive predictive value measures how often a positive alert corresponds to actual disease.

A 2024 systematic review found wide variation among electronic health record models for pancreatic cancer risk. It also identified incomplete reporting and limited real-world validation across the field.

A later meta-analysis found that positive predictive values were often below 1% for EHR-based approaches. The analysis also reported extreme variation among studies, despite encouraging average discrimination.

These findings do not invalidate REDMOD, which analyzes imaging rather than structured records. They show why strong retrospective accuracy is only the opening requirement for a screening system.

How REDMOD Reads a Normal-Looking Pancreas

The model converts pancreatic tissue into measurable texture patterns, then combines three classifiers to estimate risk.

REDMOD begins with automated pancreas segmentation. Segmentation means identifying the three-dimensional boundary of the organ across individual CT image slices.

The researchers used a deep learning architecture called 3D nnU-Net for this step. Automated segmentation reduces the manual labor that would otherwise make large-scale analysis difficult.

After isolating the pancreas, the system calculates radiomic features. Radiomics converts medical images into numerical measurements describing intensity, texture, shape, and spatial relationships.

These measurements can capture pixel-level variation that does not form a visible tumor. A radiologist sees anatomy, while radiomics also measures how image values vary within that anatomy.

The system initially examined hundreds of measurements. Its final signature used 40 selected features, including four first-order features and 36 texture features.

First-order features summarize the distribution of image intensities without considering their location. Texture features measure spatial patterns and relationships among neighboring image values.

Thirty-six of the 40 selected features came from filtered images. Filters expose patterns at different scales, including variations that may disappear in the original CT image.

Wavelet-filtered texture features accounted for 90% of the final signature. These features produced an AUC of 0.82, compared with 0.74 for unfiltered features.

The difference was statistically significant in the study. It suggests that the model relies on distributed structural changes, not simply on a hidden mass overlooked by a reader.

REDMOD then sends the selected measurements into three conventional machine-learning methods. These are logistic regression, random forest, and extreme gradient boosting.

Each method calculates a probability. The system averages those outputs through soft voting and converts the combined probability into a classification.

This ensemble matters because the three methods capture different relationships. Logistic regression favors simpler boundaries, while the tree-based models can represent nonlinear interactions.

The training process also used Synthetic Minority Over-sampling Technique, known as SMOTE. This method creates synthetic examples within the minority class to reduce imbalance during training.

SMOTE does not add real patients. It helps the classifier learn a decision boundary without allowing the larger control group to dominate every prediction.

The model includes an adjustable classification threshold. A hospital could tune that threshold toward greater sensitivity or greater specificity without retraining the underlying algorithms.

That flexibility is useful, but it transfers an important decision to clinical deployment. A lower threshold can catch more cancers while producing more false-positive alerts.

A higher threshold can reduce unnecessary follow-up but miss more early disease. The technically best threshold is therefore not automatically the clinically safest choice.

Mayo Clinic says the intended output is an additional risk signal, not an automated cancer diagnosis. Its clinical summary describes REDMOD as a complement to radiologists and gastroenterologists.

That role preserves human review around any subsequent decision. It also recognizes that a statistical texture signature cannot establish malignancy without further clinical assessment.

REDMOD Versus Radiologists Is the Wrong Final Contest

The study shows that AI found more prediagnostic signals, but the practical system still depends on radiologists interpreting those alerts.

Two board-certified abdominal radiologists reviewed the independent test set without knowing the outcomes. They assessed whether each pancreas appeared normal or showed prediagnostic disease.

Their combined sensitivity was 38.9%, compared with REDMOD’s 73%. The difference was statistically significant.

This result does not show that the radiologists performed poorly at their normal jobs. They were reviewing scans without a visible mass, which the original clinical readers had also interpreted as negative.

REDMOD addressed a different information layer. It measured mathematical patterns that human vision does not routinely quantify.

The radiologists maintained higher specificity in some comparisons. That advantage matters because false positives can trigger repeat imaging, endoscopic ultrasound, biopsy, surgery, or prolonged surveillance.

Ajit Goenka, the study’s senior author, described the model as a complementary system in clinical commentary. The radiologist would produce the standard report, while REDMOD would add a separate risk score.

That workflow reframes the opponent. The useful comparison is not machine versus doctor, but AI-assisted review versus conventional review alone.

An AI alert could prompt a radiologist to inspect the pancreas again. It could also send the case to a multidisciplinary team combining radiology, gastroenterology, oncology, and risk genetics.

For a patient with new-onset diabetes or unexplained weight loss, an elevated score might support closer surveillance. Clinicians could consider MRI, endoscopic ultrasound, molecular imaging, or a shorter follow-up interval.

For an average-risk patient with no symptoms, the same score could have a different meaning. The chance of a false alarm would be higher because the underlying probability of cancer is lower.

Hospitals therefore need more than one sensitivity figure. They need performance separated by age, sex, race, diabetes status, scanner manufacturer, imaging protocol, and clinical indication.

The study included scans from multiple institutions, systems, and protocols. That diversity strengthens the evidence that REDMOD did not memorize one hospital’s imaging setup.

The researchers also tested specificity in two additional control groups. It reached 81.3% in a multi-institutional cohort of 539 scans and 87.5% in a public dataset of 80 scans.

These checks are helpful, but they do not replace prospective deployment. Retrospective researchers already know which patients eventually developed cancer and can construct carefully defined comparison groups.

Real clinical queues are messier. Patients have motion artifacts, altered anatomy, pancreatitis, cysts, prior surgery, incomplete histories, and scans created for many unrelated conditions.

A successful system must handle those cases without silently failing. It must also communicate uncertainty in a form clinicians can understand during a busy reading session.

An unexplained red warning would invite automation bias. A calibrated risk score with supporting image regions and trend information would make review more defensible.

Longitudinal stability may help. REDMOD produced consistent classifications across serial scans in 90% to 92% of evaluated cases.

Consistency can show that an alert represents a persistent signal rather than random image noise. However, consistent error remains possible, so stability is not proof of biological truth.

What the REDMOD Numbers Do Not Prove

Retrospective detection is not yet evidence that REDMOD reduces deaths, avoids unnecessary procedures, or works safely in routine screening.

The study’s largest uncertainty is clinical utility. Clinical utility asks whether using a model changes care in ways that produce better patient outcomes.

REDMOD correctly distinguished many future cancer cases from controls. The researchers did not randomly assign patients to AI-assisted care and conventional care.

As a result, the study cannot determine how many cancers would have been treated earlier. It also cannot show whether earlier intervention would improve survival for flagged patients.

Lead-time bias complicates any early-detection claim. Finding a cancer sooner can increase the measured time from diagnosis to death without extending the patient’s life.

A prospective trial must therefore examine disease stage, resection eligibility, treatment, complications, quality of life, and cancer-specific mortality. Detection time alone is insufficient.

False-positive consequences require equal attention. At 81.1% specificity, roughly 19% of controls in the test setting received a positive classification at the chosen threshold.

That percentage cannot be directly converted into a population screening outcome because the study used a selected cohort. It still shows why the downstream protocol matters.

Pancreatic biopsy and surgery are not minor interventions. Even additional imaging can bring radiation exposure, incidental findings, cost, and anxiety.

The NCI lists false positives, overdiagnosis, overtreatment, and false negatives among the potential screening harms. Any REDMOD workflow must measure each category.

Another uncertainty involves causation. Radiomic patterns may reflect early cancer biology, inflammation, metabolic change, aging, scanner behavior, or several effects combined.

The study’s wavelet analysis supports a tissue-based signal. It does not fully explain which biological processes generate that signal.

A related concern is label construction. Every prediagnostic scan came from a person later diagnosed with cancer, so the analysis looked backward from a known outcome.

A prospective system must flag cancer among people whose futures are unknown. That difference can expose biases hidden by retrospective case selection.

The case-control ratio also remains far above real population incidence. Even stable specificity can yield a large number of false alerts when millions of low-risk scans are processed.

Deployment should therefore start with risk-enriched groups. New-onset diabetes, inherited susceptibility, pancreatic cysts, chronic pancreatitis, and strong family history offer plausible entry points.

Each group has a different baseline risk. A threshold suitable for one group may perform poorly in another.

Researchers must also test whether performance remains equitable across patient populations. Multicenter validation does not automatically guarantee demographic fairness.

CT protocols can differ by contrast phase, slice thickness, reconstruction method, and scanner hardware. Those variations can change the texture measurements on which REDMOD depends.

The external specificity results are encouraging, but sensitivity still needs broad external testing. A model that avoids false positives across sites might still miss cancer patterns in an underrepresented population.

Regulatory status and workflow integration present additional barriers. Hospitals need validated software, quality controls, version tracking, cybersecurity safeguards, and clear responsibility for missed alerts.

Clinicians also need guidance about what an elevated score means. Without an agreed response, a warning can create uncertainty instead of useful action.

The appropriate conclusion is measured. Mayo Clinic pancreatic cancer AI found a meaningful signal in retrospective CT scans, but it has not established a population screening program.

Three Signals Will Determine Whether REDMOD Reaches Clinical Care

Prospective performance, the false-positive burden, and changes in patient outcomes will decide whether REDMOD becomes a clinical tool.

The first signal is the result of prospective validation. Mayo Clinic is advancing the approach through AI-PACED, short for Artificial Intelligence for Pancreatic Cancer Early Detection.

This study is designed to evaluate AI-guided detection among patients at elevated risk. It will observe how model alerts interact with routine imaging, follow-up, and clinical judgment.

Prospective testing removes a major advantage available to retrospective research. Investigators cannot select cases based on knowing who eventually received a cancer diagnosis.

If sensitivity and specificity remain close to the retrospective results, confidence in REDMOD will rise. A substantial performance decline would weaken the case for deployment.

The second signal is the model’s positive predictive value within each intended patient group. This number reveals how many alerts correspond to actual cancer.

Researchers should report it alongside sensitivity and specificity. They should also count additional scans, specialist referrals, endoscopic procedures, biopsies, and operations triggered per detected cancer.

A tolerable false-positive rate in a high-risk clinic might be unacceptable across all abdominal CT scans. Deployment claims should therefore name the exact population and threshold.

The third signal is a shift toward clinically meaningful outcomes. Earlier stage at diagnosis would be useful, but greater surgical eligibility and improved cancer-specific survival would be stronger evidence.

Researchers should also measure harms. A system cannot claim clinical success if it finds more early cancers while causing substantial unnecessary intervention.

The most credible deployment path begins narrowly. Hospitals could analyze existing scans from well-defined high-risk groups, then expand only after prospective evidence supports each new population.

REDMOD also belongs within a broader detection stack. Clinical history, blood biomarkers, diabetes patterns, genetics, and imaging may work better together than any single signal.

Combining tools can create an efficient sequence. A low-cost risk model might identify candidates, REDMOD could analyze existing CT imaging, and targeted procedures could confirm suspicious findings.

That approach would preserve REDMOD’s strongest advantage. It extracts more information from scans that already exist without asking radiologists to see patterns below normal visual perception.

It would also respect the model’s current limits. REDMOD provides a risk estimate that requires clinical context, not a definitive diagnosis.

For clinicians and medical AI teams, the next question is concrete: can prospective studies retain REDMOD’s detection advantage while keeping the follow-up burden acceptable?

For patients, the message requires similar care. An AI flag is not proof of pancreatic cancer, and a normal result cannot rule it out.

The study has shown that routine CT data can contain an early signal. The decisive test is whether health systems can turn that signal into safer care and better outcomes.