AI detects pancreatic cancer signals on CT scans taken before diagnosis

An AI model detected subtle pancreatic tissue changes on CT scans taken a median 475 days before patients were diagnosed with cancer. In an independent retrospective test, REDMOD identified 46 of 63 people who later developed pancreatic ductal adenocarcinoma, the study’s authors report.[1][2]

The work was first published in Gut on 28 April 2026. It examines an important diagnostic gap: scans can contain quantitative changes before a tumour is visually apparent. The result suggests that existing imaging may hold useful earlier signals, while leaving open how those signals should guide care.[1][2]

What the independent test found

The researchers developed the model using 969 scans: 156 from people imaged before their cancer diagnosis and 813 controls. Their independent test included 493 people, comprising 63 prediagnostic cases and 430 controls.[1][2]

REDMOD’s sensitivity was 73.0%, meaning it flagged 46 of the 63 people who subsequently developed cancer. Specificity was about 81%: it correctly classified 349 of 430 controls. Its area under the receiver-operating-characteristic curve was 0.82, a measure of discrimination across thresholds.[1][3]

The comparison with radiologists is striking: the study reports 38.9% sensitivity for the human readers, compared with 73.0% for the model. For scans taken more than 24 months before diagnosis, the reported figures were 23% and 68% respectively. These comparisons concern the study’s reading task and cases, not the accuracy of all pancreatic cancer care.[1][2]

Why a positive signal is only the start

The 81 false positives are as important to a possible service as the 46 detected cancers. Among the positive results in this test cohort, 46 of 127 were future cancer cases—about 36%, calculated from the published counts. That proportion would change with the prevalence and case mix of the population being tested.[1][3]

In a population where pancreatic cancer is much rarer, even the same sensitivity and specificity would produce a larger share of false alarms. A positive result would therefore need a defined follow-up pathway rather than being treated as a diagnosis. The choice of population—such as people already having abdominal imaging versus an organised high-risk group—would affect both the value and burden of using the model.

The authors also report external specificity checks of 81.3% in a 539-person cohort and 87.5% in an 80-person cohort. Those are separate datasets, not interchangeable denominators for the main test. Their differences reinforce the need to evaluate performance in the setting where the system would actually be used.[1]

The clinical question is what earlier detection changes

This was a retrospective study of scans from people whose later outcomes were already known. It does not show that using REDMOD prospectively leads to earlier treatment, fewer deaths or an acceptable balance of follow-up procedures and missed disease. Those outcomes require a study in which the system is introduced into care.[1][2]

Mayo Clinic says it is taking the work into the prospective AI-PACED programme, which evaluates REDMOD in people at elevated risk, including those with new-onset diabetes and validated risk scores. The programme combines serial AI-augmented CT imaging with follow-up to examine detection, false positives and clinical outcomes. That connects the retrospective signal to a defined clinical evaluation, while the tool remains investigational.[4]

Report an error or suggest a correction