Could AI Help Doctors Predict Pancreatic Cancer? 

Written by: Beth Dougherty

As a cancer imaging fellow at Dana-Farber, Michael Rosenthal, MD, PhD, spent about two years working on a radiologist’s version of paint-by-number. Together with his colleagues, he annotated 687 computed tomography (CT) scans, manually differentiating skeletal muscle from fat tissue by labelling them with different colors.  

The work was part of a 2018 project to assess the body composition of patients with pancreatic cancer. The team, led by Brian Wolpin, MD, MPH, director of the gastrointestinal cancer center at Dana-Farber, studied the annotated images and discovered that patients with pancreatic cancer often show signs of skeletal muscle wasting at the time of diagnosis.  

The finding raised new questions. To answer them, Rosenthal was going to need to annotate a lot more CT scans. 

“I realized this was a great opportunity for automation with AI,” says Rosenthal, who has a PhD in computer science and is currently assistant director of radiology at Dana-Farber Brigham and Women’s Hospital Pancreas and Biliary Tumor Center.  

Since then, Rosenthal has used the AI tool he developed to analyze over 100,000 CT scans. Recently, he and Wolpin made another important discovery: Skeletal muscle wasting could be an early sign of pancreatic cancer. The finding, published in Nature Communications, is part of an ongoing line of work led by Wolpin at the Hale Family Center for Pancreatic Cancer Research at Dana-Farber. The goal is to find signals — in blood, in scans, or in medical records — that would aid in the early detection of pancreatic cancer. 

Approximately 80% of patients with pancreatic cancer discover the disease late, when it has already spread and cannot be surgically removed. Of these patients, almost none are still living five years after diagnosis. But when the disease is discovered early, it has the potential to be cured, and about 44% survive for 5 years. 

“If we’re going to really improve the lives of patients with this cancer, earlier detection has to be part of that program,” says Wolpin. “Treatments are really important, but earlier detection is essential.” 

An analysis of AI assessments by Dana-Farber researchers found that skeletal muscle wasting can be detected up to 18 months prior to a pancreatic cancer diagnosis.
An analysis of AI assessments by Dana-Farber researchers found that skeletal muscle wasting can be detected up to 18 months prior to a pancreatic cancer diagnosis.

Metabolism and muscle 

About ten years ago, Wolpin and colleagues discovered some cases of pancreatic cancer cause diabetes. The pancreas is responsible for secreting enzymes that aid in digestion, so they set out to learn more about how pancreatic cancer affects metabolism. 

In mouse studies performed in collaboration with Matthew Vander Heiden’s laboratory at MIT, they noticed a rise in levels of certain amino acids in the blood in the early stages of cancer development. They were coming from the deterioration of muscle and fat.  

“The cancer was causing the pancreas to improperly secrete the enzymes that normally break down food,” says Wolpin. “The mice end up malnourished, leading to muscle and fat loss.” 

This finding, published in 2018, also included the work Rosenthal had done to annotate CT scans of patients with pancreatic cancer. That investigation showed that about 65% of patients with pancreatic cancer had signs of muscle wasting at the time of diagnosis.   

“We needed to look in people before cancer diagnosis and find out if we can see the beginning of this breakdown of tissues that leads to muscle wasting,” Wolpin says. 

Deep learning 

To support this investigation, the team searched two health systems for records of patients who had CT scans between 2 months and 5 years before a pancreatic cancer diagnosis. They also searched for patients who had CT scans but did not have pancreatic cancer.  

To analyze the nearly 3,000 scans, Rosenthal built an AI model. It uses deep learning to perform the same paint-by-number assessment of CT scans that he’d done manually for the 2018 study. The tool labels the images pixel by pixel as muscle or fat, but does so in seconds, not hours. 

“The neural network works the same way the brain’s optical system works to process visual information, breaking it down into recognizable lines and shapes,” says Rosenthal. “It’s really pretty amazing.” 

An analysis of the AI assessments — all normalized to account for body composition differences across gender, race, and age — found that skeletal muscle wasting can be detected up to 18 months prior to a pancreatic cancer diagnosis. Adipose tissue (fat) wasting is detectable about 6 months before. 

The finding is intriguing, says Wolpin, “but I don’t think it can be used clinically on its own. People lose fat and muscle tissue for many reasons other than pancreatic cancer.” 

Exploring many avenues 

For next steps, the team wants to understand what other factors might be combined with tissue wasting measured on CT scans. Recently, they began a collaboration with Vander Heiden, MD, PhD, and Sangeeta Bhatia, MD, PhD, at MIT to develop stool tests that allow sensitive and specific detection of early pancreatic cancers based on the changes in enzyme secretion from the pancreas.  

They are also looking for ways to use AI to look for signals in medical records — such as medical codes, prescriptions, lab work, and scans — that might, taken all together, point to the presence of an early pancreatic cancer. 

2 thoughts on “Could AI Help Doctors Predict Pancreatic Cancer? ”

  1. For people who have pancreatic cancer in their family or who have a gene that predisposes them, this is an unbelievably helpful and very exciting tool in diagnosing the disease early enough to have success in treating it.

Comments are closed.