Mayo Clinic's AI Spotted Pancreatic Cancer 3 Years Early on Routine CT Scans — Here's How It Works
Mayo Clinic's model finds pre-tumor signs on routine CT scans up to 3 years before diagnosis. Here's the back-testing methodology and what it means clinically.
A Routine CT Scan Taken Three Years Ago Already Knew
Mayo Clinic’s AI model detects pancreatic cancer on routine abdominal CT scans up to 3 years before a clinical diagnosis — and it does this on scans that weren’t taken to look for cancer at all. The model identifies subtle signs of disease before tumors are even visible. That’s the finding, and it’s worth sitting with for a moment before moving on to how it works.
This is the kind of result that sounds like marketing until you understand the methodology. Mayo Clinic back-tested the model on pre-diagnosis scans from confirmed pancreatic cancer patients — meaning they took scans from people who were later diagnosed, fed those earlier scans into the model, and asked: could the model have flagged something? The answer was yes, up to three years out.
You don’t need a background in oncology to see why that matters. But you do need some context to understand why this is genuinely hard, and why the back-testing approach is the right way to evaluate it.
Why Pancreatic Cancer Is the Worst-Case Scenario for Late Detection
Pancreatic cancer has one of the lowest five-year survival rates of any major cancer — around 12% overall. The reason isn’t that it’s uniquely aggressive compared to all other cancers. It’s that it’s almost always caught late.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The pancreas sits deep in the abdomen, behind the stomach. It doesn’t produce symptoms that are easy to notice early. By the time someone has jaundice, back pain, or unexplained weight loss — the classic warning signs — the cancer has usually spread. At that point, surgery is often no longer an option, and treatment shifts from curative to palliative.
Early detection changes this math dramatically. Patients caught at stage I, before the cancer has spread beyond the pancreas, have five-year survival rates closer to 20–30%. That’s still not good, but it’s a different conversation than the 3% survival rate for stage IV disease. The problem is that stage I pancreatic cancer almost never gets caught, because there’s no standard screening protocol for it the way there is for breast cancer or colon cancer.
This is the clinical gap Mayo Clinic’s model is trying to close.
What “Routine Abdominal CT Scan” Actually Means Here
This is the part that’s easy to gloss over but is actually the whole point.
A routine abdominal CT scan isn’t a pancreatic cancer screening. It’s the kind of scan you get when a doctor is investigating kidney stones, abdominal pain, a liver issue, or any number of other things. Millions of these scans are performed every year. The pancreas shows up in the image because it’s in the abdomen — but nobody’s looking at it specifically for early cancer signs.
The Mayo Clinic model changes what you can extract from that existing scan. Instead of a radiologist glancing at the pancreas and seeing nothing obviously wrong, the model analyzes the image for subtle structural or textural changes that precede visible tumor formation. These are changes that a human expert, looking at the scan for its original purpose, would not flag — and arguably couldn’t flag, because the signal is too subtle and too diffuse to catch without computational analysis across thousands of similar cases.
This is what “identifies subtle signs of disease before tumors are visible” means in practice. It’s not that the tumor is there and small. It’s that the tissue is changing in ways that predict tumor development, and those changes are legible to the model even when they’re invisible to the eye.
The Back-Testing Methodology
The validation approach Mayo Clinic used is worth understanding in detail, because it’s the thing that makes this result credible rather than speculative.
They took a cohort of patients who had been diagnosed with pancreatic cancer. Then they went back and found CT scans those same patients had received earlier — scans taken before their diagnosis, when no one knew they had cancer. They ran those historical scans through the model and measured how often it flagged the right patients.
This is called a retrospective or back-tested validation. It’s a standard approach in medical AI research, and it has a specific advantage: you already know the ground truth. You know which patients eventually developed cancer, so you can measure the model’s sensitivity (how often it correctly identifies future cancer patients) and specificity (how often it correctly clears patients who didn’t develop cancer) against a real outcome.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
The finding that the model can detect signs up to three years before clinical diagnosis means that in the back-tested cohort, there were patients whose pre-diagnosis scans — taken 36 months before anyone knew they were sick — were flagged by the model as high-risk. That’s not a small margin. Three years is enough time for meaningful intervention.
The obvious next question is: what’s the false positive rate? A model that flags everyone would technically catch all the cancer patients. The clinical value depends on specificity — how often it’s wrong in the other direction. Mayo Clinic’s published work addresses this, though the full details are in their research rather than in the summary that’s been circulating. The model is designed for specialist use, which implies it’s meant to be one signal among several rather than a standalone diagnostic.
What the Model Is Actually Doing
The model isn’t doing anything exotic in terms of its architecture. It’s a deep learning model trained on CT imaging data — the same general class of approach that’s been applied to chest X-rays for pneumonia detection, retinal scans for diabetic retinopathy, and mammograms for breast cancer.
What’s specific to this case is the training data and the label. The model was trained to predict future pancreatic cancer diagnosis from current imaging, not to detect an existing visible tumor. That’s a harder problem. You’re not training the model to find a thing that’s there; you’re training it to find a pattern that predicts a thing that will be there.
This requires a large dataset of longitudinal patient records — patients with known outcomes, matched to their historical imaging. That kind of dataset is hard to assemble. It requires years of follow-up data, careful record linkage, and enough cancer cases to train on. Mayo Clinic, as one of the largest integrated health systems in the world, is one of the few institutions with the infrastructure to build it.
The model’s output is presumably a risk score — a probability that this patient will develop pancreatic cancer within some time window — rather than a binary yes/no. That score then gets routed to a specialist who decides what to do with it: additional imaging, biopsy, watchful waiting, or nothing.
What This Looks Like in Practice
Imagine you’re 58 years old. You have a kidney stone. Your doctor orders an abdominal CT scan. The radiologist reads it, confirms the kidney stone, notes nothing else of concern, and sends you home.
Under the current standard of care, that’s the end of the story for your pancreas.
Under a workflow that includes the Mayo Clinic model, that same scan gets processed by the AI system. The model returns a high-risk flag. A specialist reviews the flag and the scan. They order a follow-up MRI or endoscopic ultrasound. They find early-stage pancreatic cancer. You have surgery. You have a real chance.
That’s the scenario this model is designed to enable. It doesn’t require a new type of scan, a new clinical visit, or a new screening program. It extracts additional signal from imaging that’s already happening.
How Remy works. You talk. Remy ships.
The practical deployment question — how does this get integrated into radiology workflows at scale, who reviews the flags, how do you handle the follow-up load — is a real one. But it’s a logistics problem, not a scientific one. The scientific question of whether the signal is there has been answered.
This is also where the broader infrastructure question becomes interesting. Building a system that routes imaging results through an AI model, generates risk flags, and surfaces them to the right specialist is an orchestration problem as much as a modeling problem. MindStudio handles this kind of multi-step AI workflow — connecting models, routing outputs, and integrating with existing clinical or enterprise tools across 200+ models and 1,000+ integrations — which is part of why medical AI deployment is becoming more tractable even for teams without large engineering organizations.
The Broader Pattern: AI Finding Signal in Existing Data
The Mayo Clinic result is part of a larger pattern that’s been building for several years. AI models are consistently finding predictive signal in data that was collected for other purposes.
Google’s DeepMind published work showing that an AI model could predict acute kidney injury 48 hours before it occurred, using routine electronic health record data. A model trained on ECG data can predict atrial fibrillation years before clinical diagnosis. Models trained on retinal photographs can predict cardiovascular risk factors that have nothing to do with the eye.
The common thread is that biological systems produce correlated signals across many different measurements, and those correlations are often too subtle or too distributed for human pattern recognition to catch. Machine learning models, trained on large enough datasets with known outcomes, can find those correlations.
This doesn’t mean the models are always right, or that they’re ready to replace clinical judgment. It means there’s a class of predictive information sitting in existing clinical data that we haven’t been able to access before. The Mayo Clinic pancreatic cancer model is one of the cleaner examples of this because the outcome is so binary and so consequential: you either get cancer or you don’t, and catching it early has a large and measurable effect on survival.
For AI builders thinking about what this means for their own work, the methodological lesson is worth extracting. The back-testing approach — finding historical data with known outcomes, running your model against it, measuring performance — is applicable far outside oncology. If you’re building AI agents for research and analysis, the same principle applies: validate against cases where you already know the answer before deploying against cases where you don’t. The same logic applies when you’re evaluating any model for a high-stakes task — understanding token-based pricing for AI models matters less than understanding whether the model’s outputs are actually correct on your specific problem.
The Gap Between Research and Deployment
There’s a version of this story that ends with “and now every hospital uses it.” That’s not where we are.
Medical AI has a well-documented deployment gap. Models that perform well in research settings often perform worse in real-world clinical environments, for reasons that range from differences in imaging equipment and protocols to differences in patient populations to the simple fact that integrating a new system into clinical workflow is hard.
One coffee. One working app.
You bring the idea. Remy manages the project.
The FDA clearance process for AI-based medical devices adds another layer. A model that’s been back-tested on historical data needs prospective validation — testing on new patients in real time — before it can be deployed clinically. That process takes time and resources.
None of this diminishes the Mayo Clinic result. It means the result is a milestone, not a finish line. The model has demonstrated that the signal exists and is detectable. The work of getting it into routine clinical use is a separate project.
For engineers thinking about how to build systems that could support this kind of deployment, the data pipeline is often the hardest part. You need to ingest imaging data, run it through the model, generate a structured output, route that output to the right person, and log everything for audit purposes. Remy takes a similar approach to complex system assembly: you write a spec — annotated markdown describing the data flows, rules, and edge cases — and it compiles a complete TypeScript application from it, including backend, database, auth, and deployment. The analogy isn’t perfect, but the underlying idea — that the hard part is specifying what you want precisely enough that a system can execute it reliably — applies in both contexts. A clinical AI deployment has the same core requirement: every handoff, routing rule, and edge case needs to be defined before the system can be trusted.
Thinking about how AI models compare on tasks like this is also worth doing carefully. The gap between a model that performs well on a benchmark and one that performs well on your specific data distribution is often large — a point that comes through clearly when you look at how frontier models compare on real workflow tasks.
What Three Years Actually Buys You
Three years is a long time in cancer biology.
A pancreatic tumor that’s detectable on imaging is typically at least 1 centimeter in diameter. At that size, it’s been growing for years — estimates suggest pancreatic tumors grow slowly in their early stages, with the transition from a single mutated cell to a detectable mass taking a decade or more.
The window the Mayo Clinic model is opening isn’t the window between “no cancer” and “cancer.” It’s the window between “cancer that’s too early to see” and “cancer that’s visible.” Within that window, the tumor is still small, still localized, and still potentially resectable.
Surgical resection — removing the tumor — is the only curative treatment for pancreatic cancer. It’s only possible when the cancer hasn’t spread. The three-year detection window is valuable precisely because it’s enough time to catch patients before spread occurs, refer them to a surgical center, and operate.
This is why the clinical framing matters. The model isn’t just an interesting technical result. It’s potentially moving patients from the “incurable” category to the “curable” category, by finding them earlier in a disease that almost never gets found early.
The signal was always there in those CT scans. We just needed a model trained to read it.
