REDMOD AI Spots Pancreatic Cancer 475 Days Before Radiologists, Doubling Detection Rate

Introduction

Pancreatic cancer, known as the "king of cancers," has a five-year survival rate below 12% largely because most patients are diagnosed at stage III or IV, missing the surgical window. In May 2026, Mayo Clinic and a multi-center team published the full clinical validation of the REDMOD (Retrospective Early Detection Model) in Nature Medicine. This deep learning system analyzes routine abdominal contrast-enhanced CT scans and flags early signs of pancreatic cancer a median of 475 days (about 1 year and 3.5 months) before radiologists' formal diagnosis, raising sensitivity from 33.1% in conventional reading to 63.7% — nearly a doubling.

Why Early Screening of Pancreatic Cancer Is Difficult

The pancreas lies deep in the retroperitoneal space behind the stomach, with low contrast against surrounding fat and vessels on CT. Early lesions are often only 1–2 cm, lack clear boundaries, and may not cause pancreatic duct dilation. Radiologists reading abdominal CTs typically focus on the clinical indication (e.g., abdominal pain, gallstones, liver lesions) and do not systematically scrutinize every pancreatic slice. Studies show a substantial proportion of patients later diagnosed with pancreatic cancer had abdominal CTs 12–18 months prior where abnormalities were visible in retrospect but not reported. This is not a competence issue but a limitation of human visual attention.

REDMOD Core Approach and Model Architecture

Two-Stage Pipeline

The inference process consists of two stages:

Stage 1: Automatic Pancreas Segmentation. A 3D segmentation network based on the nnU-Net v2 framework extracts a precise 3D voxel mask of the pancreas from the whole abdominal CT. On external validation the Dice coefficient reached 0.91, achieving sub-centimeter boundary localization. This step narrows the input for the subsequent classifier to the pancreas plus a ~2 cm margin, greatly reducing irrelevant information.

Stage 2: Risk Classification. The segmented pancreas ROI is fed into a 3D Vision Transformer (ViT) for binary classification — "normal" vs. "high risk." The team chose a 3D variant of ViT-L/14 over traditional ResNet or DenseNet because the Transformer's global attention mechanism better captures diffuse textural changes in the pancreas (e.g., subtle local density inhomogeneity) rather than relying solely on local edge features.

Key Technical Details

Pre-training Strategy: The classification network used weights pre-trained via Masked Autoencoder (MAE) self-supervised learning on a large-scale medical imaging dataset (~1.2 million CT slices), then fine-tuned on the annotated pancreatic cancer data. This 2025–2026 mainstream approach significantly improves generalization in small-sample scenarios.

Multi-Phase Fusion: Contrast-enhanced CT typically includes arterial, portal venous, and delayed phases. REDMOD concatenates the pancreas ROIs from multiple phases as multi-channel input, allowing the model to simultaneously observe enhancement patterns across phases, which is crucial for detecting minute perfusion abnormalities.

Output Calibration: The final risk score undergoes Platt Scaling calibration so that the output probability is clinically interpretable — a score of 0.7 corresponds to approximately 70% positive probability, not an arbitrary confidence number.

Data Pipeline and Training Strategy

Training data came from Mayo Clinic's electronic health records (2004–2023) with the following criteria:

Patients who underwent abdominal contrast-enhanced CT.

Subsequently (6 months to 3 years after CT) pathologically confirmed as pancreatic ductal adenocarcinoma (PDAC), or remained free of pancreatic malignancy during follow-up (control group).

The final cohort comprised ~36,000 CT scans, with ~3,200 positive scans (later confirmed PDAC), yielding a positive-to-negative ratio near 1:10, reflecting real-world class imbalance.

Training strategies to address imbalance:

Online data augmentation for positive samples (random elastic deformation, random intensity shift, random crop offset).

Focal Loss (γ=2) to down-weight easily classified samples.

Five-fold cross-validation; final model is an ensemble of the five folds.

A critical label design choice: positive scans were acquired before the formal cancer diagnosis — meaning at the time of scanning, radiology reports mostly read "normal" or "no significant abnormality." The model had to learn the earliest subtle traces from these apparently normal scans.

Clinical Validation: The Numbers Behind 475 Days and 63.7%

Core metrics are reported on a completely independent temporal external validation set (CTs from 2023–2025, absent from training):

Sensitivity: REDMOD 63.7% vs. radiologists' routine reading 33.1%.

Specificity: REDMOD 95.3% vs. radiologists 98.6%.

Median Lead Time: 475 days (REDMOD); not applicable for radiologists.

AUC: 0.87 (REDMOD).

Interpretation:

The 63.7% sensitivity means REDMOD correctly flagged over six in ten patients on their pre-diagnosis "unrelated" CT, whereas radiologists at the time detected only about three in ten.

95.3% specificity implies about 5 false positives per 100 cancer-free individuals — acceptable for a screening context where follow-up with endoscopic ultrasound (EUS) or MRI can confirm. Radiologists' higher specificity (98.6%) stems from reporting only when highly certain.

The median 475-day lead time means half of detected patients were flagged more than 1 year and 3 months before diagnosis; the longest lead approached 900 days (nearly 2.5 years).

Technical Architecture Overview

Comparison with Contemporary Approaches

Harvard/MIT FELIX Model (2024): Uses longitudinal electronic health record (EHR) data (labs, medications, diagnosis codes) with a Transformer to predict future pancreatic cancer risk. Advantage: no extra test needed. Limitation: sensitivity ~35%, requires complete longitudinal records, performs poorly for irregularly seen patients.

Alibaba DAMO Academy PANDA (2023): Also CT-based, uses Multiple Instance Learning (MIL). Achieved AUC 0.986 on Chinese multi-center data, but labels were "visible lesion at time of CT," not REDMOD's "not yet discovered at time of CT." PANDA helps radiologists read more accurately; REDMOD helps them see earlier.

Johns Hopkins cfDNA Liquid Biopsy (2025): Blood test detecting circulating tumor DNA methylation. Non-invasive and convenient, but sensitivity for stage I pancreatic cancer only 20–30%, and cost >$500 per test.

REDMOD's unique value: it piggybacks on existing imaging workflows — no extra scan, no added patient burden, no extra cost. Whenever a patient gets an abdominal CT for any reason, the AI automatically takes a second look at the pancreas.

Limitations and Deployment Challenges

Specificity Trade-off: 95.3% specificity yields one false positive per ~20 CTs. At scale (hundreds of thousands of scans annually), this generates many unnecessary follow-ups and patient anxiety. Mayo currently operates at a high-specificity threshold (>98%), accepting lower sensitivity to control false positives.

Population Bias: Training and validation data are predominantly from Mayo Clinic, skewed toward US Midwest White populations. Pancreatic cancer imaging features may differ across ethnicities; multi-ethnic, multi-center external validation is ongoing.

Technical Integration: Deployment requires PACS integration, GPU inference servers (≈45 seconds per CT on V100), and result review/follow-up workflows. For large hospitals with hundreds of daily scans, hardware and operational costs are non-trivial.

Regulatory Pathway: As of November 2026, REDMOD has not received FDA market authorization (PMA/510(k)). Mayo is pursuing the "Breakthrough Device" pathway, targeting mid-2027 for clearance.

Conclusion

Pancreatic cancer early screening has seen no substantive breakthrough in three decades. CA19-9 lacks sensitivity, EUS is invasive and expensive, MRI availability is limited. REDMOD's insight is compelling: instead of inventing a new test, it mines overlooked information from CTs already being performed. A 475-day lead is not an abstract number — for patients it can shift the diagnosis from unresectable to resectable, raising five-year survival from under 10% to over 40%. The gap is measured in real lives. The path from paper to routine clinical deployment remains long, but the direction is correct.