Could AI Help Detect the Pancreatic Cancers We Usually Miss?

For decades, one of the greatest frustrations in pancreatic cancer has been that most patients never knew they were at risk.

Unlike breast or colon cancer, there isn’t a screening test for the general population. People with inherited genetic mutations or a strong family history may qualify for surveillance, but they account for only a small percentage of pancreatic cancers.

The vast majority—about 85 percent—are considered sporadic; the cancer develops in people with no known inherited risk, no family history, and often no warning signs until the disease is already advanced.

Now, researchers at Mayo Clinic believe artificial intelligence (AI) could help change that, but not by replacing radiologists or by spotting tumors that doctors have somehow overlooked. Instead, AI may provide something pancreatic cancer has long been missing: a way to identify which seemingly healthy people deserve a closer look before a tumor can even be seen.

The technology is still experimental and must prove itself in prospective clinical trials before it could become part of routine care. But if successful, it could offer a practical new pathway toward earlier detection for a broader group of patients with pancreatic cancer, particularly those who randomly develop the cancer outside currently recognized hereditary risk pathways.

A Clue Known About for Years

One of the biggest challenges with sporadic pancreatic cancer is that doctors have had very little to work with.

There is, however, one important clue. As covered previously in Let’s Win, Suresh Chari, M.D., has led research into new-onset diabetes as an early sign of pancreatic cancer. Rather than a risk factor, this type of diabetes is now believed to be a hidden syndrome that occurs in people whose pancreas is already undergoing cancerous changes.

Chari, the former director of Mayo’s Pancreas Clinic and currently professor at The University of Texas MD Anderson Cancer Center (Houston, Texas), suspected that pancreatic cancer was behind progressive hyperglycemia (high blood sugar) and metabolic dysregulation, leading to new-onset diabetes.

He found this to be the case in approximately 25 percent of patients in the three years preceding a cancer diagnosis. His study, published in January 2026 in Gastroenterology, delved into this further. 

In the U.K., guidelines from the National Institute for Health and Care Excellence (NICE) now recommend urgent abdominal imaging for people with new-onset diabetes accompanied by unexplained weight loss. The problem, however, is that those CT scans frequently look completely normal.

So Ajit H. Goenka, M.D., co-leader of the Risk Assessment, Early Detection and Interception (REDI) program at Mayo Clinic Comprehensive Cancer Center, set out to see if AI could detect subtle changes beyond human capabilities. “The problem we are trying to address is what we call the ‘last mile’ logistics,”  he says. “You have the right individual, you have the right test, but you’re not getting the answer. That’s where AI comes into the picture.”

Looking Beyond What the Eye Can See

Goenka and his team developed the Radiomics-based Early Detection MODel (REDMOD). It is not a diagnostic test for pancreatic cancer; rather, it is being studied as a risk-enrichment tool that may help identify which patients with otherwise normal-appearing CT scans should undergo closer follow-up.

Unlike a radiologist, it isn’t looking for an obvious mass or suspicious shadow. Instead, it analyzes thousands of tiny mathematical features hidden within every CT image. “Think of the image as a mathematical puzzle,” Goenka explains. “Radiologists are trained to recognize visual patterns. AI converts that image into a mathematical encoding and looks at relationships between those features that we simply cannot perceive with the human eye.” The system generates a risk score, a process fully outlined in a paper published in April 2026 in the BMJ journal Gut.

By overlaying the AI analysis on top of other factors, the baseline risk of a patient developing pancreatic cancer in the next three years might increase from 3 percent to a potential 12 to 15 percent. Such a score should then trigger further evaluation, such as endoscopic ultrasound, biomarker testing, MRI, or repeat imaging, depending on the clinical context.

The Mayo team has already completed a follow-up computer modeling study to explore exactly how this might work in practice. Rather than simply asking whether the AI is accurate, the researchers modeled what should happen after a positive result, including which follow-up tests would be most effective and how many patients would need to be evaluated to save one life. Those answers will ultimately help determine whether the technology provides enough benefit for healthcare systems, regulators, and insurers to adopt it.

While it may sound appealing to automatically run the AI tool on every CT scan performed in a hospital, Goenka notes that such a practice would not be aligned with the realities of the disease. Because pancreatic cancer is relatively uncommon, screening everyone would generate far too many false positives, leading to unnecessary anxiety, testing, and costs. “Unfortunately, there is no such thing as opportunistic detection for pancreatic cancer,” he adds.

Instead, the tool is most likely to be useful in carefully selected groups, used in sync with the full pancreatic cancer pathway. “I see AI as just another tool that will amplify the impact of individuals and the whole healthcare system,” he explains. “I think it’s an important tool that has become available to our generation, and I think we should leverage it to the extent possible in a safe and responsible way.”

Putting the AI Through a Tough Test

Many AI studies train algorithms using scans where a tumor is already present, even if it was initially overlooked by a radiologist. But Goenka’s team was very stringent in their definition of what could qualify as true prediagnostic CT scans, excluding any that showed tumors missed at first examination, in order to ensure such scans would not inflate the performance of the AI.

“There is no point in creating an AI that can do what a radiologist can already do,” Goenka says. “We went even further back. We used CT scans where even the radiologist couldn’t see anything. Only then can you actually demonstrate the value of AI.”

The team also deliberately built a diverse dataset using CT scans from multiple institutions and different scanner manufacturers to improve the model’s robustness and reduce the chance that it was simply learning quirks of one hospital or one imaging system. “The fact that it was able to survive all those confounding factors and be resilient to so many variants was very impressive to me.”

Goenka notes that the work started well before ChatGPT and other LLMs became all the rage. His team had been exploring the use of AI for years before first publishing a proof of concept study in 2022. “Our approach is always to find the right problem, rather than to take a technology and try to find where it can fit. We start with the problem, and if the problem is significant enough, then we deploy all the resources and the tools you have at your disposal to solve it.”

A Growing Reason for Optimism

The study represents a significant milestone, but the work is far from done. In addition to additional trials, the team is continuing to refine and augment their AI model. Goenka says, “Our goal is early detection. AI is just a tool. We already have a next generation AI model, and we are also comparing it to some of the other open-source frontier models. What we want is the best tool available to be deployed into our clinical care.”

The findings also hint at something larger. For years, pancreatic cancer has largely been viewed as a disease that begins in one small spot before growing into a visible tumor. This research supports a growing belief in the field that the earliest changes may involve the entire pancreas.

Because the AI analyzes signals from the whole gland, not just one location, and can still predict future cancer risk, it supports the idea that subtle biological changes may be occurring throughout the pancreas years before a tumor becomes visible. If future studies further support such findings, it could fundamentally change how researchers think about the earliest stages of pancreatic cancer.

As new treatments—from targeted drugs to personalized cancer vaccines—continue to emerge, tools like REDMOD could also help identify more patients who might benefit from those therapies before the disease has spread. This would offer new hope for the patients with sporadic pancreatic cancer, who represent the majority of cases but, until now, have largely remained beyond the reach of screening.

Goenka believes this is a reason for optimism for a cancer that is often viewed through a lens of therapeutic nihilism, a disease discovered too late to meaningfully change the outcome. “The fact that we’re talking about early detection in a cancer where the conclusion is almost always foregone is itself something I find very rewarding. Because unless we do that, we won’t be able to change the scenario.”