REDMOD: Radiomics AI Flags Pancreatic Cancer on CT a Median 16 Months Early

A radiomics model developed at Mayo Clinic identified 73% of pancreatic cancers on abdominal CT scans that radiologists had read as normal. The median gap between the AI flag and histopathological diagnosis was 475 days, about 16 months. According to the study in Gut by Mukherjee and colleagues, the model, called REDMOD (Radiomics-based Early Detection MODel), caught nearly twice as many cases as expert readers. The result is retrospective, and the model is not a screening tool yet. Its false-positive rate is the main reason, and the study points to a narrower route to clinical use.

01What the model measures

REDMOD does not look for a mass. It analyses routine abdominal CT for subtle changes in pancreatic tissue texture that may reflect the earliest biological effects of the disease. Radiomics converts image voxels into quantitative descriptors of intensity, shape and texture. Some of those descriptors vary in ways a human reader cannot perceive.

The pipeline has three stages, according to the ASCO Post's summary of the paper:

• Segmentation. Deep-learning segmentation of the pancreas feeds a radiomic extraction step that produces 968 quantitative features per scan. Automating the outline matters because, as the BMJ Group release notes, manual tracing by specialists introduces inconsistencies between readers. Removing it also makes the tool easier to run on routine scans.
• Feature selection. Minimum-redundancy maximum-relevance selection cuts the 968 features to 40.
• Classification. The 40 features go into a heterogeneous ensemble combining logistic regression and gradient boosting.

The selected features point to a mechanism. About 90% came from multiscale, wavelet-filtered images, which outperformed unfiltered data (AUC 0.82 vs 0.74; P = .007). Wavelet filtering separates image structure by spatial scale. That suggests the signal sits in fine-grained heterogeneity of the parenchyma rather than in anything a reader could see directly. This is an inference from the feature distribution. It does not show what is happening in the tissue. The authors describe the model as mechanistically grounded, but the biology behind the texture change is not established by this work.

02Training and validation on scans taken before diagnosis

The dataset is the unusual part. It comprises 1,462 CT scans: 219 pre-diagnostic scans from patients later diagnosed with pancreatic ductal adenocarcinoma (PDAC), and 1,243 controls with at least three years of follow-up without cancer. Each case scan had been read as normal when it was taken. The BMJ Group release reports that controls were matched on age, sex and scan date.

Lead times varied. According to the release, 87 patients (40%) were scanned 3–12 months before diagnosis, 76 (35%) 12–24 months before, and 56 (25%) more than 24 months before. About 64% of tumours were in the pancreatic head.

The split was as follows:

These figures come from Diagnostic Imaging's report on the paper. In the test set, cases were about 13% of scans. That is a far higher prevalence than a real screening population would have.

Two further checks were run beyond the test split:

• External specificity. Specificity was 81.3% in a multi-institutional set of 539 patients and 87.5% in a public set of 80.
• Longitudinal stability. Concordance was 90–92% on repeat scans from the same patients. The authors say this supports a real biological signal rather than noise.

One caveat from independent commentary: the held-out test set was independent but came from the same multi-institutional cohort that shaped the model's design. The external sets validated specificity only. Sensitivity was never checked against an outside cohort of pre-diagnostic cancers.

03Performance against radiologists

At the median lead time, independent testing gave an AUC of 82%, sensitivity of 73% and specificity of 81.1%. The comparators were two abdominal radiologists with three years of post-residency experience. They reached 38.9% sensitivity on the same scans, so REDMOD's was 34.1 percentage points higher. For scans taken more than 24 months before diagnosis, REDMOD's sensitivity was 68% against 23% for the readers.

That gap is plausible. A reader looking for a mass, duct cut-off or contour change has little to find two years before diagnosis, while a texture model may still pick up a signal.

The trade-off is in specificity. Radiologists correctly cleared 92.2% of disease-free patients in the main test set, against 81.1% for REDMOD. That makes the model's false-positive rate about 18.9%, roughly 2.4 times the readers' 7.8%.

A hypothetical shows why this matters. Suppose 1 in 100 scanned people go on to develop PDAC. That figure is illustrative and does not come from the study. At 73% sensitivity and 81.1% specificity, 1,000 scans would yield about 7 true positives and about 187 false positives. Fewer than 4% of flags would be real cancers. Positive predictive value depends heavily on prevalence, and the study's enriched test set cannot show how the model will behave in an unselected population.

04What stands between REDMOD and clinical screening

The authors and outside commentators raise several limits.

1. Retrospective design. Retrospective data can show that a signal existed in old scans. They cannot show how the model behaves when used on real patients in real time. Lead-time bias is also unaddressed. Detecting a cancer 16 months earlier only matters if the extra time changes the outcome.
2. Population diversity. The study population lacked sufficient ethnic diversity, which could affect performance in other groups.
3. Target population. Mayo says it is launching a prospective study called AI-PACED. It will track early detection, false positives and outcomes over time in high-risk patients. Candidates include people with new-onset diabetes. New-onset diabetes in people aged 50 or older has been identified as a possible early indicator of PDAC. Population screening is a different matter. Widespread screening is unlikely to be effective because the cancer is rare, and the US Preventive Services Task Force does not recommend it.
4. Regulatory status. The model is investigational and not approved for clinical use or population screening. Goenka has suggested it could take three to five years to reach the public. That is an estimate, not a commitment.

A positive flag is also a risk signal rather than a diagnosis. What follows depends on clinical history. It might be repeat CT, endoscopic ultrasound, tissue sampling or continued surveillance. The workflow for handling a flagged scan is as important as the model, and nobody has tested it yet.

The stakes explain the interest. The authors cite modelling suggesting that raising the share of localised PDAC from 10% to 50% would more than double survival. REDMOD shows that a computable signal exists in scans that looked normal. Whether acting on it saves lives will be answered by prospective data, not by retrospective sensitivity.