A Mayo Clinic AI re-analysed CT scans that had originally been read as normal — and detected the signature of pancreatic cancer in 73% of patients who later developed it, a median of 16 months before their diagnosis.

A Mayo Clinic team has reported a careful but striking result: an artificial intelligence model picked up subtle signs of pancreatic cancer on CT scans that had originally been interpreted as normal, often more than a year before patients received a clinical diagnosis.

This is not medical advice. I am reading a study and the Mayo Clinic announcement carefully, not suggesting that anyone should seek, avoid, or reinterpret a CT scan on the basis of an article.

The work centers on pancreatic ductal adenocarcinoma, the most common form of pancreatic cancer. It was published in Gut by Sovanlal Mukherjee, Ajit H. Goenka, and colleagues, and described in a Mayo Clinic News Network release. The finding is worth taking seriously, but it should not be read as the final word. This was a retrospective validation study, not proof that the system improves outcomes in real-world screening.

What the AI was trying to see

The model is called REDMOD, short for Radiomics-based Early Detection Model. It is designed to analyze routine contrast-enhanced abdominal CT scans and look for radiomic patterns in the pancreas that are too subtle for ordinary visual interpretation. In plain English, it is not looking for an obvious tumor. It is looking for texture and structural signals in the image that may reflect early biological changes before a mass is visible.

That distinction matters. The scans in question were not scans where a visible tumor was missed in the ordinary sense. The paper describes visually occult pre-diagnostic disease: cases where the cancer was not discernible as a focal mass or suspicious abnormality on the original read, and where expert review still faced the same problem.

Pancreatic cancer is a difficult disease to catch early because symptoms often appear late and early tumors can be hard to see. Mayo’s release, citing National Cancer Institute figures, notes that more than 85% of patients are diagnosed after the disease has spread and that five-year survival for those patients remains below 15%. Those numbers are part of why earlier detection tools attract so much attention.

The 73% figure comes from an independent test set

In the study, REDMOD was trained on a multi-institutional cohort of 969 people, including 156 pre-diagnostic cases and 813 controls. It was then tested on an independent set of 493 people, including 63 pre-diagnostic cases and 430 controls.

On that independent test set, the model identified 46 of the 63 pre-diagnostic cases. That is the 73.0% sensitivity figure in the paper. The median lead time was 475 days before diagnosis, which is roughly 16 months. In the same test set, the model correctly identified 349 of 430 controls, giving a specificity of 81.1%.

The comparison with radiologists is also important, but it needs to be stated precisely. In a head-to-head evaluation, REDMOD’s sensitivity was 73.0%, compared with 38.9% for pooled readings from two board-certified abdominal radiologists. At more than 24 months before diagnosis, the model’s sensitivity was 68.0%, compared with 23.0% for the radiologists.

That does not mean the AI is “better than doctors” in the simple way the internet often likes to say. It means that, in this retrospective test, the model was more sensitive to a particular hidden signal on pre-diagnostic CT scans. The clinical question is what happens when such a signal is used prospectively, in actual care, where false positives, follow-up tests, cost, anxiety, and outcomes all matter.

False positives are not a footnote

The specificity figure is where the practical caution enters. REDMOD correctly classified most controls, but not all of them. With 81.1% specificity in the independent test set, some people without later pancreatic cancer would still be flagged as suspicious. In a low-prevalence disease, even a reasonably specific test can create a meaningful number of false alarms if applied too broadly.

This is one reason the Mayo team frames REDMOD as an investigational tool and not a population-wide screening test ready for clinical use. Mayo’s professional summary says REDMOD is not currently approved by the FDA for clinical use or population screening. The paper’s conclusion also points toward prospective validation in high-risk cohorts as the necessary next step.

That is not a minor caveat. Retrospective validation asks whether a model can find a signal in scans where the future outcome is already known to the researchers. Prospective validation asks a harder question: if the model is used before the diagnosis is known, can it help clinicians find cancer earlier without causing unacceptable harm from false positives and unnecessary workups?

Why routine CT scans are the interesting part

One reason this result matters is that the model was built around standard-of-care CT images, not a new scan created solely for research. Many abdominal CT scans are performed for reasons unrelated to pancreatic cancer. If a reliable AI system could analyze scans that already exist and flag people who deserve closer follow-up, the practical pathway would look different from asking large numbers of people to undergo a new screening procedure.

The study also tried to avoid an overly tidy single-center result. The authors tested the model across scans from multiple institutions, imaging systems, and protocols. They also reported external specificity validation in two independent cohorts, including a public NIH dataset, and found specificity of 81.3% in a multi-institutional cohort and 87.5% in the public dataset.

The paper reports another useful check: longitudinal stability. In patients with multiple scans, REDMOD showed 90% to 92% concordance across serial imaging. In simpler terms, the model’s risk signal did not appear to swing wildly from one scan to the next in the tested setting.

What this does and does not show

The study supports a narrow but important claim: some CT scans that look normal to human readers may contain measurable image patterns associated with later pancreatic cancer, and a radiomics-based AI model can detect those patterns in a controlled retrospective validation.

It does not show that REDMOD should be used for everyone. It does not show that a positive AI flag means a person has pancreatic cancer. It does not show that earlier detection through this system will necessarily improve survival, though that is the hope being tested. And it does not remove the need for radiologists, gastroenterologists, oncologists, surgeons, and other clinicians to interpret findings in context.

Mayo says the next stage is AI-PACED, a prospective clinical study evaluating the model in people at elevated risk, including those with new-onset diabetes and validated risk scores. That is the right direction for a tool like this. The promise is not in replacing the clinical process. It is in adding a signal that may help clinicians decide who needs a closer look.

My read is that the most careful version of the story is also the most interesting one. REDMOD did not magically see cancer where no evidence existed. It found a statistical imaging signature below the threshold of ordinary visual perception. If that signal holds up prospectively, it could become one piece of a much earlier warning system for a cancer that is often found too late.

For now, the result belongs in the space between technical validation and clinical adoption. It is not a screening recommendation. It is a serious sign that ordinary scans may contain more information than human eyes can comfortably extract, and that the hard part now is proving whether reading that extra information helps patients in the real world.