Article by Barbara Toman
Only 9% of people with pancreatic cancer live for five years after diagnosis. "That is an abysmally low number, probably the worst in human cancers," says Michael Wallace, M.D., a digestive disease specialist at the Mayo Clinic campus in Florida. "We want to get that rate substantially higher."
Artificial intelligence is providing a way to do just that. In collaboration with computer scientists from the University of Central Florida, Mayo Clinic radiology and gastroenterology experts have developed an algorithm that can identify pancreatic cysts that are at higher risk of developing into pancreatic cancer. (Read about the research in a recent publication). Typically, pancreatic cancer is found when it's too advanced for curable surgery. But people who are identified as high risk can be monitored to catch cancer early.
"Outcomes from other cancers — colon, breast, prostate and lung — have improved dramatically in the past decades, largely through early detection programs such as colonoscopy and mammograms," Dr. Wallace says. "We are applying that model to pancreatic cancer."
Mayo Clinic is committed to a personalized medicine approach to assessing disease risk. These efforts are led by the Center for Individualized Medicine, in collaboration with other groups within and beyond Mayo Clinic. Artificial intelligence is key to evaluating the risk of pancreatic cancer because screening for that disease is challenging.
"The only effective screening modalities for pancreatic cancer are very expensive and somewhat invasive. We wouldn't want to screen the general population," Dr. Wallace says. "But identifying individuals who are at above-average risk for pancreatic cancer allows us to apply that screening only to them."
A first in artificial intelligence
Artificial intelligence is increasingly used to assist in image analysis. But the Mayo Clinic-University of Central Florida work is the first to address pancreatic cancer.
Recent studies have found that pancreatic cancer often starts with a precancerous cyst known as an intraductal papillary mucinous neoplasm (IPMN). Like a skin mole, an IPMN is capable of remaining harmless or developing into cancer. Pancreatic cysts are commonly seen on abdominal MRIs and CTs that people have performed for another purpose.
"About 40% of people have some sort of pancreatic cyst. The vast majority are benign," Dr. Wallace says.
Radiologists who analyze scans of pancreatic cysts look for certain imaging features such as a cyst's size and enhancement. But those factors aren't very accurate at predicting cancer risk. "If you sent people to surgery based on the existing criteria, only about half would turn out to have pancreatic cancer or an advanced precancerous cyst," Dr. Wallace says.
Like the human brain, Mayo Clinic's artificial intelligence tool learns from experience. The researchers fed into the algorithm MRIs of individuals who’s IPMNs progressed to cancer, and MRIs from a control group who’s IPMNs remained benign for many years. Once the algorithm was "trained," its classifications of high-risk and low-risk cysts were compared to classifications made by Mayo Clinic radiologists.
"We found that the algorithm reads a scan — which is about 1,200 images — in roughly half a second, versus the 20 to 30 minutes an average radiologist would need," Dr. Wallace says.
“The algorithm is able to call attention to cysts that may be of higher risk, bringing it to the attention of the radiologist for detailed review” says Candice Bolan, M.D., chair of the Division of Body MRI in Florida, who helped in the development of the software.
"This artificial intelligence has the potential to provide high-quality image interpretation to people anywhere in the world,” Dr. Wallace says.
The next step is further enhancing the algorithm's accuracy. Dr. Wallace, Dr. Bolan and their UCF colleagues recently received National Institutes of Health funding that will allow them to feed more pancreatic cyst scans into the algorithm.
"The more cases we have, the better we can train and refine the algorithm," Dr. Wallace says. "It's like an online photo collection — the more times you tag someone's face on your photo app, the better the app is at detecting that person on unknown photographs. An algorithm that is as good as our best radiologists isn't good enough. We want the algorithm to be better than that."
More tools for earlier cancer detection
In addition to applying artificial intelligence to imaging, Mayo Clinic is using patient questionnaires and genetic DNA testing to better characterize pancreatic cancer risk. Both approaches can help patients through earlier detection and treatment of cancer.
The patient questionnaire — designed in a multidisciplinary collaboration of the Center for Individualized Medicine and the Mayo Clinic Robert D. and Patricia E. Kern Center for the Science of Health Care Delivery, and supported by the Florida Pancreas Cancer Coalition and Champions for Hope — seeks to identify individuals with genetic syndromes that can increase the risk of pancreatic cancer. People who see Dr. Wallace for any gastrointestinal issue complete the questionnaire before their appointments.
"Although this tool is low-tech, it has already helped us direct people to genetic counseling and to identify individuals with pathogenic variants associated with pancreatic cancer," says Kristin Clift, who coordinates research for the Center for Individualized Medicine.
The questionnaire goes beyond pancreatic cancer to ask about a family history of other diseases, including breast cancer. "Many people understand that the BRCA1 and BRCA2 mutation can increase your risk for breast and ovarian cancer. But those mutations also increase risk for pancreatic cancer," Clift says.
Among 430 people who completed the questionnaire, 25% met National Comprehensive Cancer Network guidelines for referral to genetic counseling and testing. Three individuals were found to have pathogenic variants associated with pancreatic cancer, including one who was found to have the disease.
"The genetic testing helped determine the best treatment option for that individual. Her sister also came in for genetic testing and was found to have the variant," Clift says. "We were able to put the sister on a screening regimen so that we can catch the cancer earlier if it develops."
For Dr. Wallace, genetic testing and artificial intelligence are critical to improving pancreatic cancer outcomes. "They both allow for early detection. There is a strong need for better classification of individual risk, and we are committed to it."
In addition to providing expertise in survey design and implementation, the Mayo Clinic Kern Center for the Science of Health Care Delivery leads Mayo Clinic’s artificial intelligence strategy. And as Dr. Wallace mentions, radiology plays a critical role in many of the projects underway in AI. This work is an example of the team collaborations that exist across Mayo Clinic, and which are essential to Mayo’s efforts to provide the best possible care to patients.
This article was originally published on the Mayo Clinic Center for Individualized Medicine blog.
Tags: artificial intelligence, cancer, cancer prevention, cancer screening, Candice Bolan, Center for Individualized Medicine, Center for the Science of Health Care Delivery, gastroenterology, genetic counseling, genetic testing, imaging, Mayo Clinic Cancer Center, Michael Wallace, National Comprehensive Cancer Network, News, pancreatic cancer, radiology, republished