Pancreatic ductal adenocarcinoma is the third leading cause of cancer death in the US, and it is nearly always diagnosed too late. By the time patients have symptoms, over 80% have tumors that cannot be removed surgically — the only potentially curative option.
There is a crucial window: patients with stage I disease have a 26-month median survival, versus only 4.8 months for stage IV. Finding cancer earlier would not only allow surgery but would catch patients before cancer-related cachexia reduces their ability to tolerate an operation.
Researchers at Mayo Clinic hypothesized that even before a tumor is visibly detectable on CT scans, the surrounding pancreatic tissue undergoes subtle changes that AI — using radiomics — might be able to detect.
The study used a case-control design. Researchers identified patients who were later diagnosed with biopsy-proven PDAC, then found CT scans taken incidentally — for other reasons — before the cancer was ever suspected. These prediagnostic scans became the test dataset.
Radiomics features were extracted from the region of the pancreas on these prediagnostic scans, and machine learning models were trained to distinguish patients who would go on to develop PDAC from matched controls who did not.
The models were compared head-to-head against radiologists who reviewed the same scans, with key measurements including how far in advance the AI could detect signals (lead time) and whether its performance exceeded that of experienced radiologists.
The machine learning models successfully detected radiomics signatures of PDAC on prediagnostic CT scans taken before the cancer was clinically apparent. The models achieved performance significantly better than radiologists reviewing the same images.
Critically, the AI models were able to detect these signals at a substantial lead time — months to over a year before the cancer was clinically diagnosed — representing a genuine opportunity to catch pancreatic cancer at an earlier, more treatable stage.
This finding establishes proof of concept that the pancreatic microenvironment undergoes changes detectable by AI before a visible tumor forms — a conceptual breakthrough for pancreatic cancer screening research.
The findings are particularly relevant for high-risk populations — including people with new-onset diabetes, family history of pancreatic cancer, or known genetic mutations like BRCA2 or PALB2 — who already undergo some degree of surveillance.
One promising strategy involves patients who develop new-onset diabetes above age 50. These individuals have higher-than-average risk of harboring pancreatic cancer, and many already get abdominal CT scans for diabetes workup. AI could be applied to these routinely acquired scans.
The END-PAC model identifies high-risk new-onset diabetes patients who might benefit from CT-based screening. Integrating AI radiomics into this pathway could dramatically improve the proportion of patients diagnosed at a resectable stage.
The false positive rate — flagging people as high-risk when they do not have cancer — must be low enough to avoid unnecessary follow-up procedures and patient anxiety before this can be deployed clinically.
The study was retrospective, meaning it analyzed existing scans from patients already known to have developed cancer. Prospective validation — actually screening people and following them over time — is needed to confirm real-world utility.
Scanner variability, image quality differences, and the challenge of automating pancreas segmentation across diverse imaging datasets all represent technical hurdles that must be overcome for widespread clinical deployment.
This study provides compelling evidence that AI can detect subtle imaging signatures of pancreatic cancer before it becomes clinically visible — a finding with profound implications for a disease where late detection is the primary driver of poor outcomes.
If validated prospectively, this approach could shift pancreatic cancer from a disease almost always diagnosed too late into one where early, resectable detection becomes feasible for high-risk populations.
The work represents a significant advance in the quest for a viable pancreatic cancer screening strategy, motivating ongoing investment in AI-based early detection research.