An AI spotted pancreatic cancer on CT scans more than a year before diagnosis
The full story behind my Short: Mayo Clinic's REDMOD AI flagged 73% of hidden pancreatic cancers on CT scans that had looked normal, a median of 475 days before diagnosis, versus 39% for radiologists (Gut, 2026).
Watch the Short
This is a research tool, not an approved test. This page is not medical advice. If you're worried about symptoms, see your doctor.
The short version
Pancreatic cancer is one of the deadliest cancers, partly because it is usually found late. A Mayo Clinic team built an AI model called REDMOD that looks for subtle texture (radiomic) changes in the pancreas on CT scans, changes the human eye can't see. They tested it on scans that had looked normal at the time, from people who were later diagnosed with pancreatic cancer [1][2].
What they found
In an independent test set, REDMOD correctly flagged 73.0% of the scans from people who went on to be diagnosed, a median of 475 days before diagnosis. Radiologists reading the same scans identified 38.9% (p<0.001), so the AI was nearly twice as sensitive [1][2]. (ScienceDaily describes the 475 days as an average; the paper reports a median.)
On scans taken more than 24 months before diagnosis, the AI picked up 68% versus 23% for radiologists [1]. Its overall accuracy (area under the curve) was 0.82 [1].
Limits and caveats
- False alarms. Specificity was 81.3%, so about 1 in 5 scans from people without cancer were wrongly flagged (87.5% in a separate NIH set) [1].
- Small test set. The independent test set had only 63 cancer cases [1].
- Looked back. It was a retrospective study of old scans. It still needs testing in real time, in people at high risk [1][2].
- Limited diversity in the patients studied [1].
- Not approved. REDMOD is investigational and isn't approved for clinical use.
Why it matters
Most pancreatic cancers are found once they have already spread, when treatment options are limited. A tool that could spot early warning signs on CT scans people are already having for other reasons could, if confirmed in real-world testing, give doctors a head start. False alarms also matter, because they can lead to extra tests and worry, so the balance needs careful study.
Full narration (transcript)
An AI spotted pancreatic cancer on CT scans more than a year before patients were diagnosed. Early detection matters, because this deadly cancer is usually found late. Mayo Clinic's AI, called REDMOD, reads texture changes in the pancreas that the eye can't see. It was tested on scans that looked normal, from people later diagnosed. It flagged 73 percent of these hidden cancers, a median of 475 days before diagnosis. Radiologists caught 39 percent. Nearly twice as sensitive. More than two years out: 68 percent versus 23. The catch: it wrongly flagged about 1 in 5 cancer-free scans, and the test had just 63 cancers. It's not approved, it only looked back at old scans, and it still needs real-time testing in high-risk people. This isn't medical advice. Worried? See your doctor. Follow the link to the main website for much deeper information.
Sources
- [1] Mukherjee S, Goenka AH et al. (2026). Gut. doi:10.1136/gutjnl-2025-337266. https://doi.org/10.1136/gutjnl-2025-337266
- [2] ScienceDaily, October 2026 (Mayo Clinic). https://www.sciencedaily.com/releases/2026/10/261024222031.htm
Credits
Images via Wikimedia Commons (illustrative unless stated): Scanner Hôpital Laquintinie Douala 04.jpg (Happi Raphael, CC BY-SA 4.0); Human Pancreas (NIH BioArt 239).png (Courtesy of NIAID Ryan Kissinger, Public domain). On-screen graphics: own.
AI voice note: the narration in the Short is an AI clone of my own voice (Chatterbox TTS by Resemble AI, MIT licence), used with my consent. I wrote the script and checked the figures against the sources above.
This page is general information, not medical or dental advice. Please talk to a doctor, dentist or pharmacist about your own health.