A new artificial-intelligence system is giving doctors a possible head start against pancreatic cancer, detecting subtle changes in routine CT scans long before a tumor becomes obvious. Called REDMOD, the experimental model identified warning patterns a median of 475 days before diagnosis and found nearly twice as many early cancers as expert radiologists reviewing the same images. The early-detection findings could be especially important for a disease that is often discovered only after it has spread.
Developed by a Mayo Clinic-led team of physicians, radiologists, and data scientists, REDMOD stands for Radiomics-Based Early Detection Model. Rather than searching for a visible mass, it measures tiny variations in tissue texture and other quantitative features that may reflect the pancreas changing around an emerging cancer. The model was designed to identify pancreatic ductal adenocarcinoma, the most common form of pancreatic cancer, on contrast-enhanced abdominal CT scans that had previously appeared normal to human readers. Mayo Clinic’s overview describes REDMOD as an investigational tool rather than an approved screening product.
In an independent test group, REDMOD detected 73 percent of prediagnostic cancers while correctly excluding most people who did not later develop the disease. Its area under the curve was 0.82, with specificity of 81.1 percent. By comparison, two board-certified abdominal radiologists detected 38.9 percent of the same early cases. The gap widened in scans taken more than two years before diagnosis, when the AI still identified a substantial share of cancers that were effectively invisible to conventional review. The American College of Gastroenterology’s coverage details the performance comparison and the study’s 475-day median lead time.
The stakes are high because pancreatic cancer frequently produces few recognizable symptoms until it is advanced. Standard imaging can miss the earliest biological changes, while the disease may progress rapidly between routine appointments. Earlier identification could create an opportunity for surgery or other treatment while the cancer remains localized, but a positive AI signal would still require follow-up testing rather than functioning as a diagnosis by itself.
That distinction is central to what comes next. REDMOD was evaluated retrospectively across nearly 2,000 CT scans from multiple institutions and imaging systems, but it is not currently cleared by the FDA for clinical use or general-population screening. Researchers are moving toward prospective testing through AI-PACED, a clinical study focused on people at elevated risk, including some with new-onset diabetes. The trial is expected to examine false-positive rates, repeat-scan performance, and whether acting on the model’s warnings actually improves patient outcomes.
