An artificial intelligence system can spot pancreatic cancer long before it shows up on scans, raising the prospect of catching one of the deadliest tumours early enough to successfully treat, a study found.
The model, developed by researchers at the Mayo Clinic and collaborators, identified subtle changes in routine CT scans an average of about 475 days before patients were diagnosed, according to the study, published Tuesday in the journal Gut.
Pancreatic cancer is rarely detected early because tumours typically don’t cause symptoms and often aren’t visible on imaging until the disease is advanced. More than 85% of cases are found at a stage where treatment is largely limited to easing symptoms, helping explain why five-year survival is about 10% globally.
The findings point to a potential shift in how pancreatic cancer is diagnosed – from reacting to symptoms late in the disease to identifying patients at risk years earlier.
"This temporal window holds profound significance, as attaining such early detection would substantially augment the probability of cure and improved survival,” the researchers wrote.
If confirmed in real-world screening, such tools could help move more cases into a window where surgery or other treatments are possible, which modelling studies suggest could significantly improve survival.
"Modelling studies indicate that increasing the proportion of localised [pancreatic ductal carcinomas] from 10% to 50% would more than double survival rates, thereby underscoring that the timing of diagnosis is the single most critical determinant of survival outcomes,” they said.
The system, called Redmod, analyses patterns in CT images that aren’t visible to the human eye. It was trained and tested on scans from more than 1,400 people, including 219 patients whose earlier scans had been read as normal but who later developed pancreatic cancer.
In a head-to-head comparison, the AI was markedly better than radiologists at picking up these early signs. It correctly identified 73% of cases, compared with about 39% for doctors reviewing the same images. The advantage widened for scans taken more than two years before diagnosis, where the system detected 68% of cases versus 23% for radiologists.
The model also performed consistently across different hospitals and scanners, and correctly classified more than 80% of scans from people who didn’t develop cancer.
The tool could eventually be used to flag high-risk patients – such as older adults with unexplained weight loss and new-onset diabetes – for closer follow-up, the researchers said, but it needs prospective testing to confirm it improves outcomes before routine use. – Bloomberg
