In June, Mayo Clinic radiologist Ajit Goenka gave the American Gastroenterological Association’s newsletter two sentences about his pancreatic cancer software, and they belong side by side. One was the pitch: “The key takeaway is that pancreatic cancer produces a detectable signal on CT scans well before a mass becomes visible.” The other was the caveat: “Performance at true deployment prevalence has not yet been tested.”
This week the pitch got picked up and the caveat didn’t. ScienceDaily ran a BMJ Group release headlined “AI detects ‘invisible’ signs of pancreatic cancer years before diagnosis.” It covered REDMOD, a model Goenka’s team described in Gut in April, and said the software found warning signs “an average of 475 days before diagnosis” and was nearly twice as sensitive as experienced radiologists. The phrase “false positive” never appears in the release.
REDMOD outlines the pancreas on an ordinary abdominal CT, measures the texture of the tissue and turns those measurements into a risk score, according to the ASCO Post’s summary of the paper. The team built and tested it on 1,462 scans. Of those, 219 were taken three to 36 months before a pancreatic cancer diagnosis, at the stage the paper’s title calls visually occult, and 1,243 came from people who stayed cancer-free for at least three years. In the independent test set the model caught 73 percent of the future cancers, while two board-certified abdominal radiologists reading blind caught 38.9 percent. On scans taken more than two years before diagnosis, the split was 68 percent to 23. The median lead time was 475 days, about 16 months, and the model’s calls showed 90 to 92 percent concordance on repeat imaging.
Those catches came at a price that appears in the same results table. Specificity was 81.1 percent, which the AGA write-up translates plainly as an approximately 19 percent false-positive rate at the default threshold, and the same write-up notes that the radiologists had higher specificity in some comparisons. The BMJ release did report the figure, phrased its own way: the model “correctly classified just over 81% of scans as cancer free.” Readers were left to work out the other 19 percent for themselves.
| Measure | Share |
|---|---|
| of cancer-free scans flagged at the default threshold | 19 percent |
None of this would get much attention if the disease were more forgiving. More than 85 percent of pancreatic cancer patients are diagnosed after the disease has spread, and five-year survival remains below 15 percent. A program that could read a CT ordered for something else entirely and flag the pancreas 16 months early would change the odds for the few patients who are still caught in time.
Three years, twice
Mayo has made this promise before. In July 2022 the same pair, Sovanlal Mukherjee and Goenka, reported in Gastroenterology that an earlier radiomics model separated prediagnostic scans from controls with a sensitivity of 95.5 percent, a specificity of 90.3 percent and an area under the curve of 0.98 on a test set of 128 studies, while two radiologists caught about a third of the cancers. The coverage at the time said the disease might be detectable as early as three years before diagnosis. The median interval in that study was 398 days, and three years was the top of its range.
Four years on, the new model posted an area under the curve of 0.82, a sensitivity of 73 percent and a specificity of 81.1 percent. These are different models on different datasets, so the two papers are not a controlled comparison, though the newer one was built to hold up across institutions, scanner vendors and time, which is a harder test. The numbers moved. The headline did not. Mayo’s April rollout, as FOX 9 carried it, promised detection “up to three years before a clinical diagnosis”, and this month’s release turned the 475 days, which the ASCO Post and the AGA both report as a median, into “an average.”
Thirty false alarms for every catch
In the study, prediagnostic scans made up 219 of 1,462, about 15 percent of the dataset. Goenka acknowledges the problem: “the test-set prevalence is higher than what would be encountered in a real-world high-risk cohort.” No one is proposing to screen the general public. The researchers point instead to people with unexplained weight loss or recently diagnosed diabetes.
That diabetic group has been measured before. In 2005 a team led by Suresh Chari followed 2,122 residents of Rochester, Minnesota, who first met the criteria for diabetes at age 50 or older, and 18 of them, 0.85 percent, were diagnosed with pancreatic cancer within three years, roughly eight times the expected rate. A later prospective study of 18,838 adults with new-onset diabetes put the three-year figure at one in 160.
Now take a hypothetical 10,000 such patients at the 2005 rate, give each one a single scan, and assume REDMOD’s published sensitivity and specificity carry over unchanged, which is exactly the assumption nobody has tested yet. About 85 of them would develop pancreatic cancer within three years, roughly the study’s own window, and the model would flag about 62. Of the 9,915 who stay cancer-free, about 1,870 would be flagged as well, which works out to about 30 false alarms for every cancer caught. Using the 87.5 percent specificity the model posted on a small external set of 80 patients brings that down to about 20. At the one-in-160 rate with 81.1 percent specificity, it climbs to about 40.
Each of those flags is a person, and Goenka has described what a flag could set off: “closer surveillance, molecular imaging, or endoscopic ultrasound with tissue sampling.” He told the same newsletter that mammography and low-dose lung CT have well-defined steps for borderline findings, and that “the equivalent adjudication pathway for AI-flagged pancreatic cases has not yet been defined.” For now, REDMOD gives a risk score and stops there.
The other limits are the ordinary ones, and the study’s materials state them. It was retrospective, and its population did not include a sufficiently diverse range of ethnic backgrounds. The work was funded by the National Institutes of Health and the Mayo Clinic Comprehensive Cancer Center, among others, according to the ASCO Post.
Mayo says the next step is a prospective study called AI-PACED, which pairs the model with follow-up in higher-risk patients and assesses “early detection rates, false positives and clinical outcomes”. Those are the right endpoints. The April announcement gave no enrollment figure or timeline, and two months later the AGA described the study as still being developed. Until it reports, every “years before diagnosis” headline describes software reading old scans whose outcomes were already known from later diagnoses. Mayo has now announced the three-year horizon twice, four years apart. In its own researcher’s words, what happens to the patient whose pancreas the software flags has not yet been defined.
Sources
- Gut: Mukherjee et al., “Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability” (2026)
- ASCO Post: AI model enables earlier detection of pancreatic cancer on routine CT scans (May 2026)
- AGA News: AI model flags pancreatic cancer on CT more than a year before diagnosis (June 2026)
- ScienceDaily: AI detects “invisible” signs of pancreatic cancer years before diagnosis (October 2026)
- AuntMinnie: CT radiomics, AI can yield early detection of pancreatic cancer, on Mukherjee et al., Gastroenterology (2022)
- Gastroenterology: Chari et al., probability of pancreatic cancer following diabetes, a population-based study (2005)
- MD Anderson: New-onset diabetes may serve as an early marker of pancreatic cancer (2025)
- FOX 9: AI detects pancreatic cancer up to 3 years earlier, Mayo study shows (April 2026)