AI and Early Pancreatic Cancer
Every scan in the study had already been read and cleared as normal. Yet hidden in them were signs of a deadly cancer. Mayo Clinic researchers reported in the journal Gut that their AI model, REDMOD, found pancreatic cancer on routine CT scans an average of 475 days before patients were diagnosed.
Key Takeaways
- REDMOD analyzed earlier CT scans from 219 patients later diagnosed with pancreatic ductal adenocarcinoma and 1,243 cancer-free controls, all followed for at least three years.
- The model detected 73% of cancers, nearly double the 39% that experienced radiologists caught on the same scans, while correctly clearing about 81% to 88% of healthy scans.
- The results are retrospective, and a prospective trial called AI-PACED will test the model in high-risk patients.
Reading Texture, Not Tumors
REDMOD stands for Radiomics-based Early Detection Model. It does not look for a mass. It measures tissue texture.
The system first segments the pancreas automatically, building a 3D model from the 2D CT slices. That removes manual variation between readers. It then extracts quantitative radiomic features, subtle texture patterns linked to early biological changes as cancer starts to form.
“The key takeaway is that pancreatic cancer produces a detectable signal on CT scans well before a mass becomes visible, and that AI can read that signal with nearly twice the sensitivity of expert radiologists,” said senior author Ajit Harishkumar Goenka of Mayo Clinic's Department of Radiology.
How the Test Worked
The team trained the model on 969 pancreas CT scans. It then tested the model on separate sets of scans gathered from several hospitals, imaging systems and protocols.
The prediagnostic scans were taken three to 36 months before diagnosis. Radiologists had read all of them as negative for a mass or other suspicious features. In some scans, REDMOD flagged suspicious patterns more than two years ahead. For scans taken over two years before diagnosis, it found nearly three times as many early cancers as would otherwise go unnoticed.
Specificity held above 80% in an external multi-institution group of 539 patients. On a public National Institutes of Health pancreas dataset, the model reached 87.5%. Results also stayed consistent across repeat scans of the same patients over time.
“The greatest barrier to saving lives from pancreatic cancer has been our inability to see the disease when it is still curable,” Goenka said.
Why Timing Decides Survival
More than 85% of pancreatic cancers are diagnosed after they have spread. In the US, five-year survival hovers around 12% to 13%, and the disease is projected to become the country's second-leading cause of cancer death by 2030.
REDMOD uses scans that patients already get for other reasons. That makes it a candidate for flagging hidden disease without new imaging.
The authors list limits. The study is retrospective, and the population lacked ethnic diversity. They call prospective validation essential before clinical use.
From Scans to Screening
The next step targets people at higher risk, such as those with new-onset diabetes. The AI-PACED study will pair repeated AI-assisted CT imaging with biobanking to measure detection, false positives and outcomes. False alarms matter here. Even at 81% specificity, a screening program in the general population would flag many healthy people, so researchers aim it at enriched groups first. Gastroenterology experts reviewing the study describe it as grounds for more research rather than an immediate change in practice.
The work joins a fast-growing set of AI tools that find disease in data doctors already collect. An AI study of 100,000 patients recently linked brown fat on PET-CT scans to lower diabetes risk, and Microsoft's MAI-DxO diagnosed 83% of complex cases in a Science study. The challenge for all of them is the same: proving in real clinics that earlier flags lead to longer lives.
Written by Alius Noreika