An artificial intelligence model may help detect pancreatic cancer-related changes on routine CT scans before they are visible to radiologists, according to findings published in Gut.
The study evaluated the Radiomics-based Early Detection MODel, or REDMOD, an automated AI framework developed to identify subtle pancreatic tissue patterns associated with prediagnostic pancreatic ductal adenocarcinoma (PDAC). Although the findings require prospective validation, they suggest that routine CT scans may contain early cancer-associated information that conventional visual interpretation does not capture.
“The key takeaway is that pancreatic cancer produces a detectable signal on CT scans well before a mass becomes visible, and that AI can read that signal with nearly twice the sensitivity of expert radiologists,” Ajit Harishkumar Goenka, MD, of the Department of Radiology at Mayo Clinic and senior author of the study, told GI & Hepatology News.
“For gastroenterologists managing patients with new-onset diabetes or unexplained weight loss, this matters because these are precisely the patients who would benefit most from a structured AI-augmented screening protocol using CT.”
According to the authors, PDAC is often diagnosed after curative treatment is no longer possible, partly because early disease may not produce an obvious imaging abnormality. The authors distinguished visually occult cancers from missed cancers, noting that visually occult disease shows no discernible mass even on expert re-review.
“The broader implication is that conventional imaging fails in pancreatic cancer not because the scan lacks information, but because the relevant information exists below the threshold of human visual perception,” Dr. Goenka said. “REDMOD accesses that layer, enabling the human radiologist to work with more information.”
REDMOD was trained and tested on multi-institutional CT scans, including scans obtained three to 36 months before PDAC diagnosis that had originally been read as negative for a focal pancreatic mass or other suspicious features.
In the independent test set, the model detected prediagnostic disease with an area under the curve of 0.82, sensitivity of 73.0%, and specificity of 81.1%, with a median lead time of 475 days before diagnosis. In a blinded comparison, REDMOD was more sensitive than two board-certified abdominal radiologists, 73.0% vs 38.9%, although the authors cautioned that longer lead-time subgroup analyses were limited by small case numbers.
The radiologists had higher specificity in some comparisons, highlighting the potential role of AI as an adjunct rather than a replacement.
“REDMOD is not designed to replace the radiologist’s read or the gastroenterologist’s clinical assessment,” Dr. Goenka said. “Rather, it occupies a complementary layer. The radiologist interprets the scan and generates a standard clinical report, while the AI independently analyzes the segmented pancreas and produces a risk score. In a future deployment model, the AI flag would function as an additional data point that informs the next clinical step without dictating it.”
For gastroenterologists, Dr. Goenka said the most immediate pathway would likely involve risk-enriched populations already seen in clinical practice, including patients with glycemically defined new-onset diabetes and elevated Enriching New-Onset Diabetes for Pancreatic Cancer, or ENDPAC, scores.
“In a structured screening protocol, an AI layer could analyze CT scans from these patients and flag those whose tissue texture patterns suggest occult disease,” he said. “This could trigger a clinical cascade: closer surveillance, molecular imaging, or endoscopic ultrasound with tissue sampling.”
The model also showed evidence of longitudinal stability, an important feature for any tool being considered for screening or surveillance. REDMOD produced largely consistent results on serial CT scans and maintained specificity above 80% in both an external multi-institutional cohort and a public National Institutes of Health pancreas CT dataset.
The findings remain limited by the retrospective design and the higher prevalence of prediagnostic disease in the test set compared with what would be expected in real-world high-risk populations.
“REDMOD was validated retrospectively,” Dr. Goenka said. “It performed well across institutions, scanner vendors and temporal intervals, but the test-set prevalence is higher than what would be encountered in a real-world high-risk cohort. Performance at true deployment prevalence has not yet been tested.”
The model also had an approximately 19% false-positive rate at its default operating threshold, and the downstream pathway for AI-flagged cases remains unresolved.
“Established screening programs that accept similar tradeoffs, such as low-dose CT and mammography, have well-defined imaging-based adjudication steps for borderline findings,” Dr. Goenka said. “The equivalent adjudication pathway for AI-flagged pancreatic cases has not yet been defined.”
The authors also noted potential deployment risks, including automation bias, anchoring and the need for informatics infrastructure to identify eligible patients from electronic medical records. A prospective study, AI-PACED, is being developed to evaluate REDMOD in a clinical validation setting. For now, the findings support further study of AI-augmented CT analysis in risk-enriched patients, rather than an immediate change in practice.
The authors reported no conflicts of interest.