Pancreatic cancer is one of medicine's hardest diagnostic problems. By the time symptoms drive someone to a scan, the disease is usually advanced and survival odds are bleak. A team at Mayo Clinic now claims its AI tool can catch the cancer on ordinary abdominal CT scans up to three years before a clinical diagnosis arrives.
The model, called REDMOD, was published on April 28 in the journal Gut. In a validation set of around two thousand scans drawn from several institutions, it flagged 73 percent of prediagnostic cases at a median of sixteen months before diagnosis. Radiologists working without AI assistance flagged 39 percent of the same set.
Reading scans no one ordered for cancer
The trick is that the scans were not ordered to look for pancreatic cancer at all. They were routine abdominal CTs taken for unrelated reasons, the kind a hospital generates by the thousand every week. REDMOD is a radiomics model, which means it reads patterns of tissue density and texture that fall below the threshold a human radiologist can reliably register on a busy reporting shift.
The implication is operational rather than headline-grabbing. If a model can pull early warnings from imaging that already exists in the system, screening for pancreatic cancer no longer requires a dedicated programme to find candidates. The infrastructure is the existing CT archive.
Part of a wider pattern
That is the same thinking driving the wider wave of AI radiology validation this year. The University of Michigan's Prima system, published in Nature Biomedical Engineering in February, showed similar gains for brain MRI in emergency settings, with up to 97.5 percent accuracy and second-scale read times. Different anatomy, different workflow, same underlying argument: the diagnostic value sitting unread inside hospital archives is enormous, and AI is the first technology positioned to extract it cheaply.
Caveats remain. The Mayo paper is a retrospective validation, not a deployed screening trial, and false-positive behavior at population scale is the next thing to watch. But the central claim is now in the literature: REDMOD meaningfully outperforms unaided radiologists on a disease almost no one currently catches early.
Commentarii · 0