AI Spots Hidden Signs of Pancreatic Cancer Years Before Diagnosis, Nearly Twice as Sensitive as Radiologists

LONDON, — An artificial intelligence system can detect subtle signs of pancreatic cancer on routine CT scans more than a year before patients are diagnosed, according to new research. The findings could one day help doctors catch one of the deadliest cancers while it is still treatable.

The model, called REDMOD, identified hidden warning signs an average of 475 days before diagnosis. It was nearly twice as sensitive as experienced radiologists reviewing the same scans. The study, led by researchers at Mayo Clinic, was published in the journal Gut and highlighted Monday by the BMJ Group, which publishes the journal.

Researchers cautioned that the technology still needs further testing before it can be used in routine care.

A deadly, hard-to-spot cancer

Pancreatic cancer is notoriously difficult to detect early. It often causes few symptoms in its early stages, and the tissue changes it produces can be too faint to see on standard imaging. By the time many patients are diagnosed, the disease has already spread.

An estimated 67,530 Americans will be diagnosed with pancreatic cancer in 2026 and about 52,740 will die from it, according to American Cancer Society figures cited by the Pancreatic Cancer Action Network. It is the third-leading cause of cancer death in the United States. Its five-year survival rate is 13%, the lowest among major cancers.

More than 85% of patients are diagnosed after the disease has spread, according to the National Cancer Institute, as cited by Mayo Clinic.

"Modelling studies indicate that increasing the proportion of localized [pancreatic ductal carcinomas] from 10% to 50% would more than double survival rates," the researchers wrote. They said this underscores "that the timing of diagnosis is the single most critical determinant of survival outcomes."

How REDMOD works

REDMOD, short for Radiomics-based Early Detection Model, analyzes tiny patterns in the texture of pancreatic tissue on computed tomography scans. This approach, known as radiomics, measures features that are not visible to the human eye.

The system also automatically identifies and outlines the pancreas, separating it from surrounding organs. That removes the need for specialists to trace the organ by hand, a step that can vary from one person to another.

The model targets pancreatic ductal adenocarcinoma, the most common form of the disease.

Study design

Researchers analyzed abdominal CT scans from 219 patients treated at several hospitals. When the scans were taken, radiologists had found no evidence of cancer, but every one of these patients was later diagnosed with pancreatic cancer.

The scans were taken at different points before diagnosis:

• 87 patients (40%) were scanned three to 12 months before diagnosis.
• 76 patients (35%) were scanned 12 to 24 months before diagnosis.
• 56 patients (25%) were scanned more than 24 months before diagnosis, up to about three years.

The researchers compared those images with scans from 1,243 people who remained free of pancreatic cancer for up to three years, matched by age, sex and scan date. Patients who went on to develop cancer had an average age of 69, compared with 64 in the comparison group.

Results

REDMOD outperformed radiologists by a wide margin:

• It detected cancer-related signs in 73% of cases, compared with 39% for radiologists.
• In scans taken more than two years before diagnosis, the AI detected 68% of cases. Radiologists caught 23%, giving the AI nearly three times their rate.
• On average, the AI flagged warning signs 475 days before clinical diagnosis.

The system also performed well at recognizing healthy scans. In an independent group of 539 patients from several hospitals, it correctly classified just over 81% as cancer-free. On a public U.S. National Institutes of Health dataset of 80 patients, it correctly identified 87.5%.

When researchers analyzed scans from the same patients taken months apart, REDMOD gave consistent results in 90% to 92% of cases. That consistency, they said, suggests the system is detecting real disease-related changes rather than random noise.

"This temporal window holds profound significance, as attaining such early detection would substantially augment the probability of cure and improved survival," the researchers wrote.

Seeing the unseen

Dr. Ajit Goenka, a Mayo Clinic radiologist and the study's senior author, has said that the central problem in pancreatic cancer has been "our inability to see the disease when it is still curable."

"This AI can now identify the signature of cancer from a normal-appearing pancreas," Goenka said in a Mayo Clinic statement.

In their paper, the researchers concluded that REDMOD is "a fully automated AI framework capable of identifying the imaging signatures of stage 0 [pancreatic ductal adenocarcinoma] in normal pancreas, achieving this with substantial lead times and performance superior to expert radiologists."

Limitations and next steps

The researchers acknowledged several limitations. The study population did not include a sufficiently diverse range of ethnic backgrounds, which could affect how well the model performs in other groups.

The findings are also retrospective, meaning the researchers looked back at existing scans. The model must be tested prospectively, especially in people at higher risk, such as those with unexplained weight loss or newly diagnosed diabetes. Both can be early signs of pancreatic cancer.

Mayo Clinic said it is launching a prospective study called AI-PACED, short for Artificial Intelligence for Pancreatic Cancer Early Detection. The study will evaluate how doctors can use AI-guided detection in caring for high-risk patients, tracking early detection, false positives and outcomes over time.

"While prospective validation is paramount to confirm clinical utility, the REDMOD framework represents a significant advance towards shifting the paradigm for sporadic [pancreatic ductal adenocarcinoma] from a late-stage symptomatic diagnosis to proactive pre-clinical interception," the researchers wrote.

Because the system works on CT scans that patients often already have for other reasons, experts say it could eventually be used to screen existing images without extra cost or radiation exposure.

The research was funded in part by the National Institutes of Health, according to Mayo Clinic.