Gut
AI may spot hidden pancreatic cancer changes early
Clinical takeaway: AI-based computed tomography analysis could eventually help shift pancreatic cancer detection earlier, but it still needs prospective testing in high-risk patients before clinical use.
Pancreatic ductal adenocarcinoma (PDA) is often diagnosed late because early disease may cause few symptoms and little visible change on standard imaging. Investigators developed REDMOD, a radiomics-based artificial intelligence model, to detect subtle pancreatic texture changes that are difficult for radiologists to see.
Researchers tested REDMOD on abdominal computed tomography scans from 219 patients whose scans were initially read as showing no disease, but who were diagnosed with pancreatic cancer months to years later. Those scans were compared with scans from 1,243 matched patients who didn't develop pancreatic cancer within three years.
REDMOD detected a preclinical pancreatic cancer signature an average of 475 days before clinical diagnosis. It was more sensitive than experienced radiologists, detecting 73% of cases vs. 39% overall. For cancers diagnosed more than two years later, sensitivity was 68% with REDMOD vs. 23% with radiologists.
In external testing, REDMOD correctly identified cancer-free scans in 81% of patients from an independent multi-hospital group and 87.5% of patients in a public National Institutes of Health dataset. Repeated scans from the same patients produced consistent results in about 90% to 92% of cases.
Next, REDMOD needs prospective validation in high-risk patients, such as those with unexpected weight loss and newly diagnosed diabetes before it can be used broadly in practice.
“This study validates REDMOD as a fully automated AI framework capable of identifying the imaging signatures of stage 0 PDA in normal pancreas,” the researchers concluded, calling it “a scalable objective tool that addresses a critical diagnostic gap.”
Source: Paiella S. Gut. 2026 Apr 28. Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability