A new artificial intelligence model developed at the Mayo Clinic can detect signs of pancreatic cancer up to three years before a patient ever receives a clinical diagnosis.
That is not a small upgrade in medical technology.
That is, for many patients, the difference between having a fighting chance and running out of time.
The findings, published in the journal Gut in April 2026, mark what researchers are calling a landmark moment in cancer detection.
The AI model, named REDMOD (Radiomics-based Early Detection Model), analyzed routine abdominal CT scans and identified cancer signatures in patients whose scans had been reviewed by radiologists and cleared as completely normal.
It found what trained human eyes missed.
And it did so at a scale and consistency that has the medical community paying very close attention.
“The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable,” said Dr. Ajit Goenka, senior author of the study and a radiologist at Mayo Clinic in Minnesota.
“This AI can now identify the signature of cancer from a normal-appearing pancreas, and it can do so reliably over time and across diverse clinical settings.”
That quote deserves a moment to sink in.
A pancreas that looks completely normal to a specialist.
Cancer already growing inside it.
And an AI that can tell the difference.
Why Pancreatic Cancer Is So Difficult to Beat
To understand why this research matters so much, you need to understand what makes pancreatic cancer so uniquely brutal.
According to the National Cancer Institute, more than 85% of patients are diagnosed after the disease has already spread to other parts of the body.
By that point, surgery, which is the only treatment capable of curing the disease, is no longer a realistic option.
The five-year survival rate sits below 13%.
For context, that means roughly 87 out of every 100 people diagnosed with pancreatic cancer will not be alive five years later.
The Pancreatic Cancer Action Network reported in January 2025 that survival rates have stalled at 13%, despite advances in cancer treatment across the board.
In 2025 alone, an estimated 67,440 Americans were expected to be diagnosed, and 51,980 were expected to die from the disease.
Projections suggest pancreatic cancer will become the second-leading cause of cancer-related death in the United States by 2030.
The core problem has always been timing.
Pancreatic cancer grows silently.
It produces few symptoms in its early stages.
By the time a patient feels something is wrong and visits a doctor, the cancer has typically already spread beyond the point of surgical removal.
Catching it early has been, for decades, an almost impossible task.
That is what REDMOD is now beginning to change.
How the Study Was Conducted
The researchers at Mayo Clinic built and trained REDMOD using CT scan data drawn from multiple hospitals, imaging systems, and clinical settings.
This was a deliberate choice.
Most AI models in medicine are trained and tested on data from a single institution, which means their performance often drops when exposed to real-world variation.
The Mayo team wanted REDMOD to work in the kinds of conditions doctors actually face every day, not just in controlled laboratory environments.
According to the Mayo Clinic News Network, the study analyzed nearly 2,000 CT scans in total, including scans from patients who were later diagnosed with pancreatic cancer, all of which had originally been reviewed by radiologists and deemed normal.
The validation phase tested REDMOD on 63 pre-diagnostic CT scans and 430 control scans from patients who had not developed cancer.
For the pre-diagnostic group, the scans had been taken at various points before diagnosis: some three to twelve months out, others twelve to twenty-four months out, and some more than two years before the eventual diagnosis.
REDMOD ran automatically on each scan without requiring any time-intensive manual preparation from technicians or radiologists.
Its results were consistent across different machines, different hospitals, and different imaging protocols.
That consistency was one of the study’s most significant findings.
Findings From the Study
The results were striking.
REDMOD correctly identified 73% of patients who would later be diagnosed with pancreatic cancer, based on scans that radiologists had already reviewed and cleared as normal.
By comparison, radiologists reviewing those same scans, without AI assistance, identified only 39% of cases.
That is nearly double the detection rate.
The gap widened even further when looking at scans taken more than two years before diagnosis.
As reported by Live Science, in that two-year-plus window, REDMOD identified nearly three times as many early cancers that would otherwise have gone undetected.
On average, the AI flagged cancer signatures a median of 475 days before the clinical diagnosis, which works out to just over fifteen months.
The model also performed well in ruling out disease in healthy patients, correctly identifying 88% of people who did not have pancreatic cancer, keeping false positives to a manageable level.
The model’s accuracy (measured as AUC, a standard metric in medical diagnostics) came in at 82%, with a sensitivity of 73% and specificity of 81%.
For a disease this difficult to catch early, those numbers represent a meaningful step forward.
What the AI Is Actually Seeing
Here is where the science gets genuinely fascinating.
When a radiologist looks at a CT scan and says it looks normal, they are making a judgment call based on visible anatomy.
If there is no obvious lesion, no visible tumor, no clear structural abnormality, the scan gets cleared.
REDMOD does not work that way.
According to reporting by Medical Xpress, the model converts the CT image into a mathematical analysis, first building a three-dimensional model of the pancreas from the two-dimensional scan images, then evaluating the structure pixel by pixel.
What it is looking for are subtle changes in tissue texture and structure, tiny statistical irregularities in how the pancreatic tissue is arranged at a microscopic level, changes that are far too faint and diffuse for the human eye to register as suspicious.
These are called radiomic features, and REDMOD measures hundreds of them simultaneously.
The model captures biological changes that are beginning to happen at a cellular level, long before they organize into a visible mass that any radiologist could identify.
Think of it like this: a human radiologist looks at a field and asks, “Is there a fire?”
REDMOD asks, “Has the chemistry of the grass changed in a way that suggests a fire is coming?”
The distinction is everything.
The Number That Changes the Conversation
There is one statistic buried in this research that deserves far more attention than it typically receives.
When pancreatic cancer is caught while still confined to the pancreas, before it has spread, the five-year survival rate jumps from approximately 13% to 44%.
That figure comes from Mayo Clinic’s own reporting, citing the American Cancer Society’s Cancer Facts and Figures.
A nearly fourfold increase in survival, simply by catching the disease at the right time.
The challenge has never been the treatment options.
Surgery, chemotherapy, and combination therapies exist.
The challenge has always been knowing that the disease is there when those treatments can still make a difference.
REDMOD’s ability to detect cancer signatures more than a year before diagnosis represents a window that could, for many patients, make all the difference in whether surgery is even possible.
But Here’s What Most People Misunderstand About This Breakthrough
The conversation around AI in medicine often swings between two extremes.
On one side, breathless headlines declare that AI is about to replace doctors.
On the other, skeptics dismiss every new model as overhyped until it has decades of clinical validation.
Both reactions miss what is actually happening here.
REDMOD is not designed to replace radiologists.
It is designed to work alongside them, flagging patterns in data that human perception simply cannot reliably detect at scale.
There is an important nuance in the study data that illustrates this.
REDMOD outperformed radiologists significantly in detecting pre-diagnostic cancer.
But in the same study, radiologists were only shown the scans without any AI assistance.
The real-world application is different.
A radiologist with REDMOD’s output in front of them is not competing with the AI.
They are incorporating it into their clinical judgment, the same way a doctor uses blood test results or any other diagnostic tool.
The AI does not make the decision.
It surfaces information that would otherwise stay invisible.
There is also an important caveat worth being honest about: the model has not yet been tested in large-scale prospective clinical trials.
The research validated REDMOD using retrospective data, meaning scans from patients already known to have developed cancer later.
Confirming its performance in real-time clinical settings, across far larger and more diverse patient populations, is the necessary next step.
That prospective trial is already underway, through a study called AI-PACED (Artificial Intelligence for Pancreatic Cancer Early Detection), which is currently enrolling patients considered to be at elevated risk for the disease.
How This Research Applies to Real Life
For most people reading this, the practical question is: what does this actually mean for me or someone I care about?
The honest answer right now is that REDMOD is not yet available as a standard diagnostic tool in hospitals and clinics.
It is moving toward clinical implementation, and the AI-PACED trial is a critical step in that process.
But the research opens a door that has been closed for a long time.
For the first time, there is a plausible pathway toward routine screening for pancreatic cancer using a tool that most people already encounter when they have abdominal imaging done for any reason.
No new machine.
No invasive procedure.
No specialized test that requires referral to a major cancer center.
Just an AI model running alongside the CT scan you might already be getting.
That is an enormous practical advantage.
Other approaches to early pancreatic cancer detection, including liquid biopsies that look for tumor DNA in the blood and biomarker panels that screen for cancer-related proteins, are promising but still require separate tests and additional clinical infrastructure.
REDMOD integrates into imaging that is already being done.
This research is part of Mayo Clinic’s broader Precure initiative, which focuses on identifying the earliest possible biological changes in the body before symptoms begin, with the goal of predicting and preventing disease rather than simply treating it after the fact.
What Comes Next
The immediate priority is the AI-PACED clinical trial.
Researchers at Mayo Clinic are now working to understand exactly how REDMOD performs in prospective real-world conditions, how often it generates false positives that lead to unnecessary follow-up, and how clinicians can most effectively incorporate its findings into patient care.
Those are not small questions.
A tool that generates too many false positives creates its own set of problems: unnecessary anxiety for patients, additional invasive testing, and additional costs to healthcare systems.
Striking the right balance between sensitivity (catching real cancers) and specificity (avoiding false alarms) is the work that clinical trials are designed to measure.
The current data is encouraging.
The model’s false positive rate was well-controlled in the validation study, and its performance held up consistently across scans from different hospitals and imaging systems, which is a strong indicator that it will generalize reasonably well to clinical practice.
Researchers have also noted that REDMOD’s predictions remained stable over time when the same patient had multiple scans taken months apart.
That consistency matters enormously for a surveillance tool, which would ideally track changes over time rather than giving different readings from one appointment to the next.
The Bigger Picture
Pancreatic cancer has been one of medicine’s most stubborn problems for decades.
Unlike breast cancer, colorectal cancer, and cervical cancer, all of which now have established screening protocols that have meaningfully reduced mortality, pancreatic cancer has never had a reliable early detection method.
The biology of the disease has made it extraordinarily difficult.
The pancreas sits deep within the abdomen, surrounded by other organs.
It produces no biomarkers in early-stage disease that are distinctive enough to screen for reliably.
And by the time symptoms appear, the window for curative treatment has almost always already closed.
What REDMOD represents is not a cure.
It is not even a treatment.
It is something arguably more important at this stage: a way to see the disease at a point when treatment still has a real chance of working.
The gap between a 13% survival rate and a 44% survival rate is not explained by better drugs or more advanced surgery.
It is explained by timing.
If this AI model can give patients and their doctors more time, consistently and reliably, at scale, the downstream impact on survival numbers could be significant.
The researchers at Mayo Clinic are appropriately measured in how they describe what comes next.
Prospective validation is essential.
Clinical integration takes time.
But the direction is clear, and the evidence behind it is stronger than anything that has come before in this particular corner of oncology.
For a disease that has taken far too many lives far too early, that clarity matters a great deal.
If you found this article useful, consider sharing it with someone who might benefit from knowing about early detection research. The more people understand what is becoming possible in cancer detection, the more conversations happen that might one day save a life.
References and Further Reading
Mayo Clinic News Network: Mayo Clinic AI Detects Pancreatic Cancer Up to 3 Years Before Diagnosis
Pancreatic Cancer Action Network: 2025 Survival Rate Report
Mayo Clinic Magazine: Using AI for Early Detection of Pancreatic Cancer
