Study reveals AI detects pancreatic cancer years earlier

Pancreatic cancer is projected to become the second-leading cause of cancer-related deaths in the US by 2030, partly because 85% of cases aren't identified until the illness has advanced.

With a new AI model from Mayo Clinic and University of Texas MD Anderson Cancer Center researchers, this scenario might change.

Called REDMOD (radiomics-based early detection model), this system was evaluated using CT scans of individuals later diagnosed with pancreatic cancer.

In approximately 3 out of 4 cases, REDMOD effectively detected the most prevalent form of pancreatic cancer about 16 months before diagnosis. This nearly doubles the detection rate compared to experts examining the scans without AI help.

In some instances, REDMOD detected concerning tissue patterns over two years before diagnosis, and the team believes it may spot cancer as early as three years beforehand.

"The biggest obstacle in saving lives from pancreatic cancer has been our inability to detect the disease while it remains treatable," says radiologist and nuclear medicine expert Ajit Goenka, from the Mayo Clinic.

"This AI can now precisely identify cancer markers in a seemingly normal pancreas, consistently over time and across various clinical environments."

The team trained REDMOD using 969 CT scans of the pancreas, teaching it to recognise the subtle signs of cancer in its earliest phases.

Instead of searching for a visible tumor, the model seeks radiomic patterns, slight changes in tissue texture and structure often too subtle for human vision to detect.

Many cancers begin when normal cells acquire DNA mutations that impact their growth and division, but it can take years for these changes to form a tumor large enough to show symptoms or be visible on a scan.

Following the training, REDMOD was tested on another set of CT scans: 63 from individuals who developed cancer post-scan, and 430 healthy controls who were cancer-free.

All these scans had previously been deemed clear by human radiologists, and when two radiologists reviewed the scans alongside REDMOD, they identified early cancer signs in only 38.9 percent of instances.

Among the 430 healthy controls, REDMOD incorrectly identified 81 as suspicious – suggesting in a real-world scenario, these people might have been called in for further testing before being cleared.

A similar performance level was revealed in two other dataset evaluations, using various equipment across different hospitals.

Moreover, for patients with several scans available, the AI yielded largely consistent outcomes – even when the scans were conducted months apart.

"These characteristics set it up for prospective validation in high-risk groups, a crucial step towards transitioning from late-stage diagnosis to proactive early interception," write the study authors in their paper.

The idea is that the earlier REDMOD can review CT scans – possibly taken routinely for other health reasons – the more beneficial it could be. It's feasible that it could discover pancreatic cancer when curative treatment remains viable.

However, there's further work to be done before that is feasible. Next steps include testing the AI on larger, more varied populations and evaluating how seamlessly doctors can integrate it into existing protocols.

Researchers are optimistic about these early findings and hope that with continued refinement, we might possess an extremely valuable tool against one of the most life-threatening cancers.

"The proven capability of the framework to consistently detect these hidden signals on a large clinically-focused dataset, coupled with its outstanding longitudinal stability and confirmed specificity, establishes a solid foundation for AI-enhanced early detection," write the researchers.