Pancreatic cancer is one of the deadliest common cancers largely because it hides in plain sight until it’s too late to cure. A new AI model developed at Mayo Clinic, described in a study published in the journal Gut, found signs of the disease on routine abdominal CT scans up to three years before patients were formally diagnosed — on the very same images that radiologists had already reviewed and read as normal.
The system, called REDMOD (Radiomics-based Early Detection Model), doesn’t look for tumors the way a human radiologist does. Instead, it measures hundreds of quantitative imaging features describing tissue texture, density and structure across the pancreas, picking up faint biological changes that occur as cancer begins developing well before a mass becomes visible to the eye.
Nearly 2,000 Scans, Read Twice
Researchers tested REDMOD on nearly 2,000 CT scans pulled from multiple institutions, using different imaging systems and protocols to mimic the messy variability of real-world clinical practice rather than a single pristine dataset. Critically, many of the scans came from patients who were later diagnosed with pancreatic cancer but whose earlier CT scans — taken for unrelated reasons — had been read by radiologists at the time as unremarkable.
When REDMOD analyzed those same “normal” scans retrospectively, it identified 73% of the prediagnostic cancers, at a median of roughly 16 months before the eventual clinical diagnosis. That detection rate nearly doubled what specialist radiologists had achieved on first read, according to the study’s authors, without the benefit of hindsight the AI system also didn’t have — it was scoring the scans blind to outcome.
Why Pancreatic Cancer Is Uniquely Suited to This Approach
Pancreatic cancer’s five-year survival rate remains under 13% nationally, largely because roughly 80% of patients are diagnosed after the cancer has already spread beyond the point where surgery can offer a cure. Unlike breast or cervical cancer, there is no recommended population-wide screening test for pancreatic cancer, because the disease is rare enough that broad screening of asymptomatic people would produce too many false positives to be worthwhile. That’s what makes an “opportunistic” detection approach like REDMOD appealing: rather than screening everyone, it re-analyzes CT scans that patients are already getting for other reasons — a kidney stone workup, a back pain evaluation — potentially catching pancreatic cancer as a byproduct of imaging nobody ordered to look for it.
Caution From the Study’s Own Authors
Mayo Clinic researchers involved in the work are notably careful about over-claiming. The current results are retrospective — the AI was tested against scans and outcomes that had already happened, not deployed in real time to change what doctors did next. That distinction matters enormously in medicine: a model can look impressive analyzing historical data and still fail to improve outcomes when integrated into live clinical workflows, where false alarms carry a real cost in unnecessary follow-up scans, biopsies and patient anxiety.
To address that gap, the team is preparing a prospective clinical trial, named AI-PACED, that will test REDMOD in real-world, high-risk patient populations going forward, rather than looking backward at scans that already have known outcomes. Only that kind of forward-looking trial can establish whether flagging these early signs actually leads to earlier treatment and better survival — or whether it mainly generates additional testing without changing outcomes.
Part of a Broader Push Into “Radiomics”
REDMOD sits within a growing field known as radiomics, which treats medical images as dense quantitative datasets rather than pictures to be visually interpreted — extracting patterns invisible to the human eye and feeding them into machine learning models trained on large outcome datasets. Similar approaches have been applied to lung nodules, cardiac imaging and now pancreatic cancer, part of a broader 2026 wave of AI tools aimed at catching cancer earlier from imaging that’s already being collected for other reasons.
What’s Next
If the AI-PACED trial confirms REDMOD’s retrospective performance in a prospective, real-world setting, the next hurdle would be regulatory: an FDA submission, likely as a Class II decision-support device requiring radiologist oversight, following the same path other AI imaging tools have taken in 2026. Mayo Clinic researchers describe REDMOD as part of a “multiyear research effort” rather than an imminent clinical product, but for a cancer where earlier detection is one of the only levers proven to improve survival, even a modest improvement in lead time could translate into meaningfully more patients eligible for curative surgery.
