Four hundred and seventy-five days. That is how far ahead an AI system built at Mayo Clinic runs when reading the same CT scans that radiologists already review. Called REDMOD, for radiomics-based early detection model, the system identifies pancreatic cancer with 73% sensitivity in scans where no tumor is visible to the human eye, according to a multi-institutional validation study published April 28 in Gut, a BMJ journal. Experienced radiologists hit 39%. Not close.
Pancreatic cancer kills 52,740 Americans per year, according to the American Cancer Society's 2026 projections, and carries a five-year survival rate of 13%. Only 10% of cases are caught while the tumor is still localized, where five-year survival jumps to 44%. Most patients feel nothing until the cancer has already spread, and projections show it becoming the second leading cause of cancer death in the United States by 2030. Early detection changes everything, and nobody screens for it.
REDMOD does not require a new scan, does not require new hardware, and runs as a software overlay on abdominal CT images that hospitals are already acquiring for other clinical reasons, extracting radiomic features from the pancreatic region that fall below the threshold of human visual perception. Mayo's AI-PACED prospective clinical trial is now underway to validate these retrospective results in a real-world screening workflow, with results expected to determine whether the sensitivity and specificity numbers hold outside research conditions.
What REDMOD Actually Sees
The study, led by Surya Mukherjee and colleagues, analyzed 219 patients who were subsequently diagnosed with pancreatic cancer and 1,243 matched controls across multiple institutions. Forty percent of the cancer patients were diagnosed three to twelve months after the CT scan REDMOD flagged, thirty-five percent were diagnosed twelve to twenty-four months later, and 25% were diagnosed more than twenty-four months out, meaning the AI identified cancer signals up to three years before clinical detection.
At the 24-month-plus window, where the cancer is at its most invisible, REDMOD maintained 68% sensitivity. Radiologists? Twenty-three percent. That gap widens further with time. Reproducibility ranged from 90 to 92% on repeat scans of the same patients, meaning the system was not capturing noise or scanner artifacts but consistent biological signals embedded in the tissue's texture, patterns so subtle that even an expert radiologist who knows where to look cannot reliably distinguish them from normal pancreatic parenchyma.
Specificity was 81% in an independent 539-patient validation cohort and 87.5% in a separate 80-patient cohort drawn from the NIH Pancreatic Cancer Trial dataset, numbers that sit at the high end of diagnostic AI tools currently in development but below the specificity levels that screening programs typically require before the US Preventive Services Task Force issues a recommendation. Those numbers matter enormously for the cost calculation that follows, because specificity determines how many false alarms a population-wide deployment would generate.
The Calculation Nobody Has Published
Here is the math. Hospitals across America perform approximately 93 million CT scans per year, according to NPR's reporting on radiation exposure data. Roughly 30 to 35% of those are abdominal or pelvic scans, yielding approximately 30 million abdominal CTs annually. With a US adult population of about 260 million, that means roughly 11.5% of American adults receive an abdominal CT in any given year, and over REDMOD's three-year detection window, somewhere between 30 and 35% of adults will have had at least one scan the system could analyze.
Now apply that to pancreatic cancer incidence: of the 67,530 new cases diagnosed annually, if 35% of those future patients had an abdominal CT within the preceding three years, that gives REDMOD access to approximately 23,636 cases. At 73% sensitivity, the system would flag 17,254 of them before clinical symptoms appeared, which is 25.5% of all pancreatic cancer cases caught early by AI overlay alone, on top of the 10% currently found while localized.
Run the survival arithmetic. Currently, 67,530 annual diagnoses at 13% five-year survival produces 8,779 survivors. With REDMOD catching 23,636 cases at the localized-equivalent stage where 44% survive five years, and the remaining 43,894 cases continuing at late-stage rates of roughly 5%, the projected total climbs to 12,595 five-year survivors. Net gain: 3,816 people. Every year.
Incremental cost is almost absurdly low by medical screening standards, the kind of number that makes health economists do a double take. REDMOD is a software pipeline running radiomics analysis on existing DICOM images, requiring no new scanners, no additional patient visits, no contrast agents. Computational cost for running radiomic feature extraction on a standard abdominal CT is estimated at $5 to $15 per scan, depending on cloud versus on-premise infrastructure. At the midpoint of $10 per scan across 30 million abdominal CTs, annual deployment cost reaches $300 million, less than what the United States currently spends treating roughly 2,000 late-stage pancreatic cancer patients.
| Screening Program | Cost per QALY |
|---|---|
| Colonoscopy (colon cancer) | $20,000 - $50,000 |
| Mammography (breast cancer) | $30,000 - $60,000 |
| REDMOD on existing CTs (estimated) | $50,000 - $80,000 |
| Low-dose CT (lung cancer) | ~$81,000 |
Divide $300 million by 3,816 additional survivors: $78,616 per life saved. In cost-per-QALY terms, that falls between $50,000 and $80,000, well within the range that health economists consider cost-effective (the commonly cited US threshold is $50,000 to $150,000 per QALY) and competitive with established cancer screening programs that nobody questions. Mammography costs $30,000 to $60,000 per QALY, colonoscopy runs $20,000 to $50,000, and low-dose CT for lung cancer screening comes in at approximately $81,000. REDMOD sits squarely in that bracket, with a structural advantage that none of those programs share: it requires zero patient behavior change and zero additional medical procedures to generate the initial flag.
The Strongest Case Against Universal Deployment
False positives are real, and the numbers are large. At 81% specificity applied to 30 million abdominal CTs, REDMOD would generate approximately 5.7 million false positive flags per year. Five point seven million. Each flag triggers clinical anxiety, follow-up imaging, potential endoscopic ultrasound, and in some cases pancreatic biopsy, a procedure that carries risks including pancreatitis, bleeding, infection, and in rare cases perforation that requires emergency surgery. Base rates make the math brutal: pancreatic cancer's lifetime risk is roughly 1 in 56, meaning that in any screening population, the vast majority of flagged patients will not have cancer.
This is the standard objection to any population-level screening tool for a rare disease, and it deserves serious engagement rather than dismissal. PSA testing for prostate cancer was enthusiastically adopted in the 1990s and then partially walked back by the US Preventive Services Task Force in 2012 precisely because false positives led to unnecessary biopsies and surgeries that caused more aggregate harm than the cancers they caught. Same trap. Any deployment of REDMOD would need a carefully designed triage pathway, likely involving a secondary AI or biomarker confirmation step, to filter the 5.7 million false positives down to a manageable clinical workload before invasive procedures are considered.
What This Analysis Does Not Show
Several limitations constrain these projections. Our cost-effectiveness calculation uses estimated computational costs, not validated implementation data from actual hospital deployments, because no such deployment exists yet, and the per-scan compute estimate of $5 to $15 could shift substantially depending on whether hospitals process locally or route to cloud infrastructure with associated data privacy and HIPAA compliance overhead. Our 30-million abdominal CT figure is derived from aggregate national scan volume data, not a precise registry count of qualifying abdominal acquisitions. Our assumption that 35% of future pancreatic cancer patients had a prior abdominal CT within three years is an estimate based on population scan rates, not a matched patient dataset. REDMOD's study, while multi-institutional, was retrospective and used a cohort that was not ethnically diverse, raising questions about generalizability across populations. And the false positive management costs, including follow-up imaging, specialist consultations, and unnecessary procedures triggered by those 5.7 million annual flags, are not modeled in the $300 million figure, meaning real-world deployment would cost meaningfully more.
The Bottom Line
A piece of software that reads scans hospitals are already performing can detect the deadliest common cancer 475 days before doctors can see it, at an incremental cost of approximately $10 per scan, with cost-effectiveness competitive with mammography and colonoscopy. Mayo Clinic's AI-PACED prospective trial will determine whether those retrospective numbers hold in clinical reality.
If you have a family history of pancreatic cancer or carry known risk factors (BRCA2 mutation, chronic pancreatitis, new-onset diabetes after age 50, Lynch syndrome), ask your gastroenterologist about enrollment in the AI-PACED trial or request that any abdominal CT you receive be flagged for radiomics analysis if your institution offers it. If you are a hospital CIO evaluating radiology AI platforms, REDMOD's published specificity and reproducibility data put it in the top tier of diagnostic AI tools that have reached multi-institutional validation, and integration requires a DICOM routing rule, not a capital equipment purchase. If you are a health insurer running the numbers: $300 million spread across 30 million scans is $10 per scan, while late-stage pancreatic cancer treatment costs $150,000 to $250,000 per patient, so catching 3,800 of those cases early is not philanthropy but arithmetic.
Sources
- Mukherjee et al. (April 28, 2026). Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability. Gut. gut.bmj.com
- Mayo Clinic News Network (April 28, 2026). Mayo Clinic AI detects pancreatic cancer up to 3 years before diagnosis in landmark validation study. newsnetwork.mayoclinic.org
- American Cancer Society (2026). Cancer Facts & Figures 2026. cancer.org
- NPR / WCMU (April 16, 2025). Study highlights cancer risk from millions of CT scans performed annually. radio.wcmu.org
- News-Medical.net (April 28, 2026). AI model detects pancreatic cancer years before clinical diagnosis. news-medical.net
- Pancreatic Cancer Action Network. Pancreatic cancer statistics. pancan.org