Artificial intelligence has been trained to detect pancreatic cancer long before it becomes visible on scans, according to a new study.

This development opens up the possibility of identifying one of the deadliest tumors at an early enough stage for successful treatment.

The AI model, Redmod, was created by researchers at the Mayo Clinic in collaboration with colleagues. It identified subtle changes on standard CT scans an average of 475 days before a diagnosis was made.

Pancreatic cancer is rarely detected at an early stage because tumors do not cause symptoms and are often not visible on scans until the disease has progressed. More than 85% of cases are diagnosed when treatment is limited to symptom relief.

The findings suggest a potential shift in the approach to cancer diagnosis.

“This time window is crucial, as such early detection can significantly increase the chances of cure and improve survival rates,” the researchers wrote.

If the effectiveness of this new tool is confirmed in real-world screening studies, early detection of cancer could allow for surgical intervention or other treatment methods.

“Modeling shows that increasing the proportion of localized cases from 10% to 50% [of pancreatic ductal adenocarcinomas] could more than double survival rates. This once again emphasizes that the timeliness of diagnosis is the single most important factor determining treatment outcomes,” the experts stated.

Effectiveness

Redmod analyzes patterns on CT images that are not visible to the human eye. It has been trained and tested on scans from over 1,400 individuals, including 219 patients whose early scans were assessed as normal but later developed pancreatic cancer.

In direct comparisons, the AI significantly outperformed radiologists, correctly identifying 73% of cases compared to 39% by specialists.

For scans taken more than two years before diagnosis, the AI's advantage was even greater—68% versus 23%.

The model demonstrated consistency across different hospitals and scanners, correctly classifying over 80% of images from individuals who did not develop cancer.

Researchers emphasized that the tool could be used to identify high-risk patients. However, prospective trials are needed before it can be implemented in routine practice.

In October 2025, Google, in collaboration with Yale University, introduced a new foundational model with 27 billion parameters designed to understand the "language" of individual cells.