AI Tool Created That Can Triple Sensitivity of Early Pancreatic Cancer Detection
Mayo Clinic researchers developed the REDMOD model, which detects pancreatic cancer on CT scans at a stage when the tumor is still invisible to doctors. In tests, the model tripled detection sensitivity compared to radiologists and found cancer an average of 475 days before standard diagnosis.
Not just a detector, but a time machine: how REDMOD from Mayo Clinic rewrites the rules of the game against pancreatic cancer
[The Gist]: What's Really Happening
In late April 2026, Mayo Clinic researchers published a study in the journal Gut that, by all standards, should be considered not a scientific paper but a declaration of war. The enemy is pancreatic ductal adenocarcinoma (PDA), a disease with a five-year survival rate of 13% and 8% for the most common form. The weapon is REDMOD (Radiomics-based Early Detection Model), an AI model that detects cancer on CT scans when the pancreas looks completely normal to the human eye.
The numbers that make oncologists read the paper twice: the model's sensitivity at the stage when the tumor is not yet visible is 73% versus 39% for experienced radiologists. The median lead time is 475 days. And for cases detected more than two years before clinical diagnosis, REDMOD's sensitivity is three times higher than human (68% vs 23%). This is not an improvement of existing technology. It is a paradigm shift: from searching for visible lesions to detecting textural fingerprints, from reacting to disease to predictive interception.
Timeline and Context
The story of REDMOD didn't start yesterday. Back in 2022, a proof-of-concept study by the same group was published in Gastroenterology, showing that radiomic features could distinguish pre-diagnostic CT scans from controls. But that first-generation model had critical limitations: it required manual pancreas segmentation, was tested on an artificially balanced 1:1 (cancer/normal) sample, and was not ready for the real world.
The group leader, radiologist and nuclear medicine specialist Dr. Ajit Goenka, described four technological breakthroughs that turned the concept into REDMOD:
- Automatic 3D segmentation based on the nnU-Net architecture — eliminated the human factor and made the system scalable.
- Multiscale wavelet transform — 968 radiomic features, from which the mRMR algorithm selected the 40 most predictive. Key detail: 90% of the selected features were derived from filtered textures, not from visible grayscale characteristics.
- Ensemble classifier — logistic regression, random forest, and XGBoost combined via a soft-voting mechanism.
- Adjustable threshold — the model outputs a continuous probability (0–1), and the threshold can be moved: lower for maximum sensitivity in screening, higher for specificity before invasive procedures.
The results were published in Gut on April 28, 2026. And on May 7, Mayo Clinic announced the launch of the first clinical trial in the US of an AI radiomic algorithm for early detection of pancreatic cancer — enrollment of 100 patients aged 50–85 has already begun.
Who Wins and Who Loses
Winners:
- High-risk patients — primarily people with newly diagnosed diabetes and elevated ENDPAC score. This is the population REDMOD targets, and here the prevalence of cancer is 3–4%, justifying screening. Modeling shows that if the proportion of localized PDA increases from the current 10% to 50%, survival would double.
- Mayo Clinic as an ecosystem. The clinic is already building tools to automatically extract variables from electronic medical records to identify eligible patients without manual labor. The system is designed for multi-institutional deployment: the trial will be expanded to campuses in Arizona and Florida, and then to partner centers worldwide. This is a strategic bet on creating a new standard of care, where Mayo becomes not just a provider but a platform.
- The AI-assisted diagnostics market in general. 2026 became the year when CPT codes first recognized AI-assisted image analysis as a separate reimbursable service. REDMOD fits perfectly into this regulatory framework: there is a specific algorithm, documentation, and a billing code.
Losers:
- Manufacturers of "late" diagnostic tests. REDMOD targets stage 0 — before mass formation, before clinical symptoms. If the model is prospectively validated, the market for tests designed to confirm already suspected cancer (biopsies under EUS, some molecular panels) will shrink: patients will enter the system earlier, and the share of late diagnostics will decrease.
- Radiologists not ready for AI integration. REDMOD does not replace the radiologist, but radically changes their role: from "finding a needle in a haystack" to "checking algorithm hints." The sensitivity gap (73% vs 39%) is too large to ignore in missed diagnosis lawsuits. In 3–5 years, the standard of care will require AI assistance in interpreting abdominal CT scans in high-risk patients — and those who do not adapt will face professional liability risks.
What the Media Isn't Saying
Non-obvious insight: REDMOD is not a cancer detector. It is a detector of the stromal reaction that precedes cancer.
Most publications focus on "476 days before diagnosis" and "threefold superiority over radiologists." But from Goenka's interview, a fundamentally different view of the signal's nature emerges.
90% of REDMOD's significant radiomic features are filtered texture characteristics derived from wavelet transforms and Laplacian of Gaussian filters. They capture not the tumor itself (it's not there yet), but changes in tissue microarchitecture: extracellular matrix remodeling, fibrotic changes, shifts in cell density — that is, the stroma reacting to carcinogenesis. Ablation analysis showed that the model on filtered features alone has an AUC of 0.82, while on unfiltered features only 0.74. The 8-point difference is statistically significant.
This redefines the very concept of "early diagnosis." Previously, we looked for the tumor. Now we look for the tissue reaction to a process that has not yet formed a tumor. It's like detecting a fire by the smell of smoke, not by visible flames — and doing it 16 months before anyone notices the fire.
Second insight: predicting invisible cancer opens Pandora's box for screening in asymptomatic individuals.
Currently, REDMOD is being prospectively tested on a high-risk population (ENDPAC-positive patients with new-onset diabetes). But technologically, nothing prevents running all abdominal CT scans done for any reason — whether for a kidney stone or appendicitis — through the model. This is called "opportunistic screening," and Mayo Clinic is already discussing this possibility.
The problem is that the regulatory and ethical framework for this is lacking. What to do with a patient for whom REDMOD shows a cancer probability of 0.5, but the CT shows a completely normal pancreas, and EUS finds nothing? How often to repeat scanning? Who pays for the "false alarm"? These questions have no answers, but the emergence of REDMOD forces us to start formulating them.
Forecast: Next 30 Days and 90 Days
30 days (by mid-June 2026):
Several major academic centers (MD Anderson, Johns Hopkins, Memorial Sloan Kettering) are expected to announce their own validation programs for REDMOD-like models on retrospective data. Mayo Clinic will likely present the first interim results of patient enrollment in the AI-PACED trial — even with N=100, the fact of accrual itself will be interpreted by the market as a signal that the technology is moving from science to practice.
Simultaneously, reimbursement discussions will intensify. The 2026 CPT codes already recognized AI-assisted diagnostics as a reimbursable service, and REDMOD is an ideal candidate for a separate code pending FDA clearance. Investment banks will begin estimating the potential market: 600,000+ abdominal CT scans per year among the high-risk population in the US × $100–200 per AI analysis = $60–120 million in annual revenue just in the domestic market.
90 days (by mid-August 2026):
The key catalyst is the FDA's decision on REDMOD classification. If the regulator goes the De Novo clearance route (rather than PMA), the path to market will shorten to 12–18 months. Simultaneously, the first external validation study from an institution not affiliated with Mayo is expected. This is critical: the current validation on the NIH-PCT dataset showed specificity of 87.5%, but the sample is limited. External confirmation on a European or Asian population will either strengthen confidence or reveal hidden biases.
By the end of summer 2026, an announcement of the first commercial partnership between Mayo Clinic and a PACS system manufacturer or radiology workflow platform (Sectra, Visage Imaging, or similar) to integrate REDMOD into routine radiologist workflow is also likely. This integration, not model accuracy, will determine the speed of adoption.
Bottom line: REDMOD is not just an AI tool. It is a demonstration that the most deadly form of cancer can be intercepted at a stage when it is curable. If prospective trials confirm the results, the definition of "pancreatic cancer diagnosis" will change forever — from "when it's already too late" to "16 months before it would have been too late."
— Editorial Team
