- Mayo Clinic’s REDMOD model, published April 29, 2026 in Gut, detected pancreatic cancer on routine CT scans up to three years before clinical diagnosis, identifying 73% of prediagnostic cancers, nearly double the rate achieved by specialists reviewing the same scans unaided.
- A 2026 Radiology study of a large Swedish screening population found AI cancer scores were elevated years before conventional diagnosis, flagging early signals in 39.3% of individuals two years before their recorded diagnosis.
Mayo Clinic’s REDMOD model found pancreatic cancer in CT scans taken up to three years before patients were ever diagnosed, spotting nearly twice as many cases as specialists reviewing the same images. For one of medicine’s hardest-to-catch cancers, that gap in detection time can determine whether surgery remains an option.
Pancreatic Cancer’s Narrow Window
The REDMOD (Radiomics-based Early Detection Model) study, published in Gut on April 29, 2026, is built on a consequential idea: pancreatic cancer leaves faint traces in tissue long before it becomes visible to a clinician. The model measures hundreds of quantitative imaging features from routine abdominal CT scans, picking up biological changes in tissue that appears normal to the human eye. In its validation study, REDMOD identified 73% of prediagnostic cancers at a median of roughly 16 months before diagnosis. For scans taken more than two years before diagnosis, the model’s advantage over unaided specialists grew to nearly three times. Pancreatic cancer carries a five-year survival rate of around 9%, largely because it is almost always caught after it has spread, which is exactly the problem REDMOD targets.
Mammography Signals, Years Out
A separate 2026 study in Radiology tested three commercially available AI systems against mammograms from a large Swedish screening population. AI-generated cancer scores were elevated years before conventional diagnosis: the systems flagged early signals in 19.7% of individuals six years before their recorded diagnosis, rising to 39.3% two years out. AI tools cleared for breast cancer screening have been moving in this direction for several years, but these figures from a population-scale dataset clarify how far in advance the signals can appear. Earlier flags mean earlier intervention and, potentially, less aggressive treatment by the time a patient enters the clinical pathway.
Microsoft’s Diagnostic AI
Microsoft’s MAI-DxO took a different approach to the diagnostic problem. Rather than analysing images, the sequential orchestrator model works through a case the way a physician does: asking questions, ordering tests and updating its working diagnosis step by step. Research published in Science in November 2025 showed MAI-DxO achieved 80% accuracy on typical emergency room cases, compared to 47% for general physicians using textbooks and internet resources, according to Microsoft. On highly complex diagnostic dilemmas, it scored more than twice the accuracy of its human counterparts, the company says. The model also reduced unnecessary lab test orders, which Microsoft says improves cost efficiency. AI systems deployed in high-stakes settings carry their own failure modes, and MAI-DxO faces a real test when it enters the emergency department waiting room at Beth Israel Deaconess Medical Center in Boston to triage actual patients.
Where AI Fits in the Clinic
The FDA had cleared more than 1,000 AI imaging algorithms by late 2025, and that volume reflects where the clinical impact is currently most concentrated: earlier and more precise detection. REDMOD’s performance on pancreatic cancer and the Swedish mammography study’s population-scale findings both point in the same direction, signals that were always present in the imaging data, but previously undetectable at the speed and consistency required for routine screening. Beyond detection, the operational case is also building. AI chest X-ray systems have cut radiologist turnaround times in emergency settings while maintaining accuracy above 90%, according to published research, and MAI-DxO’s reduction in unnecessary lab orders points toward a secondary benefit that matters to procurement teams: lower per-case cost. None of these tools operate without a clinician in the loop, and regulators in multiple jurisdictions are still working through how to evaluate AI medical systems reliably. What the 2026 studies establish is that the detection window for at least two major cancer types is wider than previously documented, and that AI is what makes it visible.
