Somewhere between "looks normal to me" and "oh no, that's pancreatic cancer," a computer may have found a sliver of extra time. In a 2026 Gut commentary, Patrick Michl and Laura Roth spotlight a genuinely eerie idea: AI radiomics might detect pancreatic cancer while the scan still looks boring to human eyes, which is both thrilling and a little rude to the humans with radiology fellowships [1].

That matters because pancreatic ductal adenocarcinoma, or PDAC, is the kind of cancer that prefers a late entrance. It usually shows up after it has already caused trouble, which is why survival stays grim. According to the U.S. SEER program, pancreatic cancer’s 5-year relative survival is 13.7% overall in 2026 estimates, and stage at diagnosis changes the story a lot [2]. Translation: if you can catch this thing early, you stop playing defense quite so desperately.

The pancreas has been hiding in plain sight

The big idea behind radiomics is simple enough to explain over a drink: medical images contain more information than a person can casually eyeball. Texture, shape, tiny intensity patterns, subtle asymmetries - all of that can be turned into numbers and fed into machine learning models. Basically, the scan is not just a picture. It is a spreadsheet wearing a hospital gown.

That is the backdrop for the study sitting behind this commentary. In April 2026, Mukherjee and colleagues reported a "Radiomics-based Early Detection MODel," or REDMOD, for spotting very early pancreatic cancer signals on CT scans that radiologists had originally read as normal [3]. Very normal-looking pancreas. Suspiciously normal, apparently.

According to the journal’s press summary, REDMOD analyzed prediagnostic abdominal CTs from 219 people who were later diagnosed with pancreatic cancer and compared them with 1,243 matched controls. The model detected a preclinical signature an average of 475 days before clinical diagnosis and reached higher sensitivity than radiologists, including for scans taken more than two years before diagnosis [3]. That is not "the robot solved cancer." It is "the robot may have noticed the wallpaper peeling before the wall fell down," which is a lot more useful and much less annoying.

Why this feels different

This did not come out of nowhere. Researchers have been circling this problem for a few years, and the breadcrumb trail is getting harder to ignore.

A 2022 Gastroenterology study showed radiomics-based machine learning could detect pancreatic cancer on prediagnostic CT scans at a substantial lead time before clinical diagnosis [4]. Another 2022 study in Frontiers in Oncology tried a more localized trick, using AI to analyze pancreatic subregions and flag which areas looked risky before a tumor became obvious [5]. Then in 2023, the PANDA system in Nature Medicine showed deep learning could detect pancreatic lesions from non-contrast CT with strong performance across large validation sets [6].

So the field is moving from "interesting lab trick" toward "wait, this might actually fit into routine imaging." A 2025 systematic review put it plainly: radiomics studies for early pancreatic cancer detection look promising, but the evidence is still heterogeneous and the usual headaches remain, especially external validation, reproducibility, and limited shared code and data [7]. In other words, the models are smart, but science still wants receipts.

The part where optimism meets paperwork

Here is the catch, because there is always a catch. Pancreatic cancer is relatively rare in the general population, which makes screening difficult. Even a strong model can cause problems if it sends too many healthy people into the medical equivalent of a fire drill. The National Cancer Institute has made the same point for blood-based pancreatic cancer tests: accuracy alone is not enough when prevalence is low and false positives can lead to invasive follow-up [8].

That is why the most sensible use case is not "scan everyone and let the algorithm panic." It is likely targeted use in higher-risk people, such as patients with new-onset diabetes, unexplained weight loss, family history, or known genetic risk. Even the REDMOD coverage emphasized the need for prospective testing in high-risk groups before broad clinical adoption [3].

There is also the question of what AI is actually seeing. Radiomics can be powerful, but it can also feel like a very confident raccoon rummaging through a pixel dumpster and somehow emerging with a clinically useful answer. If the model works, great. If nobody can explain when it fails, less great.

So, should you be excited?

Cautiously, yes. Not fireworks-in-the-lab excited. More like "quietly refill your glass and lean in" excited.

If these results hold up prospectively, this line of research could shift pancreatic cancer detection earlier, when surgery and treatment have a real chance to help. It also hints at a bigger pattern in medical AI: some of the most valuable systems may not generate flashy images or chatbot banter. They may just notice weak signals humans miss, like the one friend in the group chat who actually reads the whole thread before replying.

And no, this is not the same as using combb2.io to sharpen a blurry vacation photo. But the family resemblance is there: better image analysis can reveal details you were about to shrug past. In one case, that means rescuing a fuzzy sunset. In the other, it might mean catching a lethal cancer while it is still pretending to be nothing.

References

1. Michl P, Roth L. Seeing the unseen: AI radiomics unmasking occult pancreatic cancer? Gut. 2026 May 8. doi:10.1136/gutjnl-2026-338712. PubMed: 42103472

2. National Cancer Institute SEER. Cancer Stat Facts: Pancreatic Cancer. 2026 estimates and survival data. https://seer.cancer.gov/statfacts/html/pancreas.html

3. Mukherjee S, et al. Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability. Gut. 2026. doi:10.1136/gutjnl-2025-337266

4. Mukherjee S, Patra A, Khasawneh H, et al. Radiomics-based Machine-learning Models Can Detect Pancreatic Cancer on Prediagnostic Computed Tomography Scans at a Substantial Lead Time Before Clinical Diagnosis. Gastroenterology. 2022;163(5):1435-1446.e3. https://mdanderson.elsevierpure.com/en/publications/radiomics-based-machine-learning-models-can-detect-pancreatic-can/

5. Javed S, Qureshi TA, Gaddam S, et al. Risk prediction of pancreatic cancer using AI analysis of pancreatic subregions in computed tomography images. Front Oncol. 2022;12:1007990. doi:10.3389/fonc.2022.1007990

6. Cao K, Xia Y, Yao J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. 2023;29(12):3033-3043. doi:10.1038/s41591-023-02640-w. PubMed: 37985692

7. Renjifo-Correa ME, Fanni SC, Bustamante-Cristancho LA, et al. Diagnostic Accuracy of Radiomics in the Early Detection of Pancreatic Cancer: A Systematic Review and Qualitative Assessment Using the Methodological Radiomics Score (METRICS). Cancers (Basel). 2025;17(5):803. doi:10.3390/cancers17050803. PubMed: 40075651. PMCID: PMC11898638

8. National Cancer Institute. Blood Test Accurately Detects Early-Stage Pancreatic Cancer. May 10, 2024. https://www.cancer.gov/news-events/cancer-currents-blog/2024/liquid-biopsy-detects-pancreatic-cancer

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.