Analytics in Action News

Mayo Clinic ML Can Predict Pancreatic Cancer Earlier Than Usual Methods

Mayo Clinic researchers have developed machine-learning models that can detect pancreatic cancer on pre-diagnostic CT scans substantially sooner than traditional methods for clinical diagnosis.

A digital representation of a human full-body scan, representing AI in medical imaging

Source: Getty Images

By Shania Kennedy

- A Mayo Clinic-led study published in Gastroenterology shows that radiomics-based machine-learning (ML) models may help diagnose pancreatic cancer at an earlier, more treatable stage than standard diagnosis methods.

According to the American Cancer Society (ACS), pancreatic cancer accounts for approximately 3 percent of all cancers in the US and about 7 percent of all cancer deaths. ACS estimates that 62,210 people will be diagnosed with pancreatic cancer, and 49,830 people will die of the disease this year. Pancreatic cancer is slightly more common in men than women, but the average lifetime risk of getting this cancer is about one in 64.

Mayo Clinic’s overview of pancreatic cancer states that it is rarely detected in its earliest stages when it’s most curable because it often doesn’t cause symptoms until the cancer has already spread to other organs. In a recent Mayo Clinic press release, the research team who led the study noted that early detection of pancreatic cancer improves the chances of successful treatment, but early detection is almost impossible using standard medical imaging. Up to 40 percent of small pancreas cancers are unlikely to show up on standard imaging, meaning that most patients present with advanced and non-curable pancreatic cancer, according to the researchers.

Because standard imaging is so limited in this area, the researchers sought to utilize a combination of artificial intelligence (AI) and radiological screening to detect pancreatic cancer in its early stages. The team computationally extracted the imaging signature of early cancer from pre-diagnostic computed tomography (CT) scans for 155 patients. These scans were done for reasons unrelated to pancreatic cancer between three months and three years prior to cancer occurrence.

The researchers also gathered CTs from an age-matched cohort group that did not develop pancreatic cancer during the three years of follow-up. Two expert radiologists were brought in to segment the pancreas on CTs from both groups and to computationally extract and quantify pancreas tissue heterogeneity.

The research team then built ML models to predict the future risk of pancreatic cancer at 97 to 1,092 days between pre-diagnostic CT and cancer diagnosis. The ML models achieved accuracies between 94 and 98 percent, with the average prediction at 386 days before clinical diagnosis. In comparison, the two radiologists could not reliably differentiate between patients who went on to develop pancreatic cancer and those who had normal pancreas results.

"Our study demonstrates that artificial intelligence can identify those asymptomatic people who may harbor an occult cancer at a stage when surgical cure may be possible," said Ajit Goenka, MD, a Mayo Clinic diagnostic radiologist, and the study's senior author, in the press release. "These findings may help overcome one of the key barriers to improving survival for patients with pancreatic cancer."

The research team is now exploring the possibility of validating their models from this study in the large prospective clinical trial known as the Early Detection Initiative (EDI). The EDI, led by a Mayo Clinic gastroenterologist, will evaluate the impact of a pancreatic cancer screening strategy using CTs in 12,500 participants.