This article is more than 1 year old

Algorithm can predict pancreatic cancer from CT scans well before diagnosis

Software picks up subtle clues human doctors miss

AI algorithms can predict whether a patient will develop pancreatic cancer years before an official diagnosis, or so this research suggests.

Tens of thousands of people in the US are diagnosed with pancreatic ductal adenocarcinoma – the most common type of pancreatic cancer – every year. Less than 10 percent of patients live more than five years after diagnosis.

Detecting the disease earlier could boost survival rates by up to 50 percent, it is believed. But doctors don't right now have any methods that screen patients for early signs of pancreatic cancer. Now, a team of researchers led by Cedars-Sinai Medical Center, a top non-profit hospital based in Los Angeles, California, believe AI could be up to the task.

They trained a classifier algorithm to predict whether a patient will go on to develop pancreatic cancer or not by analyzing CT scans. "There are no unique symptoms that can provide an early diagnosis for pancreatic ductal adenocarcinoma," Stephen Pandol, co-author of the study published in Cancer Biomarkers journal, and director of Basic and Translational Pancreas Research at Cedars-Sinai, said in a statement on Tuesday.

"This AI tool may eventually be used to detect early disease in people undergoing CT scans for abdominal pain or other issues." 

The team analyzed medical records to find patients who were diagnosed with pancreatic cancer over the past 15 years and had CT scans performed six months to three years before their diagnosis. They also selected medical data from patients who had undergone CT scans and did not develop the disease to form a control group. A total of 108 CT scans from 72 subjects were obtained at two health centers, according to the paper.

The dataset was split for training and testing purposes; 66 CT scans from 44 patients were evenly split in three groups, labeled as healthy, pre-diagnostic, and diagnostic, to teach the algorithm how to identify features from each group. The final 42 scans for the remaining 28 patients also evenly split across the same labels were used to test the system. It was able to predict whether a patient would go on to develop pancreatic cancer with 86 percent accuracy, the researchers claimed. 

Although the AI algorithm seems promising, the team doesn't really know what it is analyzing when it makes its predictions.

"To develop or train the model, we know in advance some image features are predictive of pancreatic cancer," Debiao Li, co-author of the study and professor of Biomedical Sciences and Imaging at Cedars-Sinai, told The Register.

We know some image features are predictive of pancreatic cancer ... they are very subtle and human eyes couldn't readily discern them so we use computers

"However, they are very subtle and human eyes couldn't readily discern them so we use computers to find them. Most other image features predictive of cancer are detected by computers without our advance knowledge.

"We select some of these features using data analysis tools to form the prediction model. Once the model is developed, we know what features are included in the prediction model and computers will find these features in new images to make a prediction of cancer or assess cancer risk. In fact, the system looks at 4,000 different features when it processes each scan."

Doctors, on the other hand, can't keep track of that many cues. The researchers reckon the model is able to detect textural differences in CT scans that change during the development of pancreatic cancer that are too subtle for the human eye to see.

They are collecting more data to continue improving and testing their algorithm on more patients beyond this initial, prototype experiment. America's National Institutes of Health has given them a grant to continue their research, we're told.

Touseef Ahmad Qureshi, first author of the study and a scientist at the non-profit hospital, said he hopes it will be used in real clinical settings one day. "Our hope is this tool could catch the cancer early enough to make it possible for more people to have their tumor completely removed through surgery," he concluded. ®

More about

TIP US OFF

Send us news


Other stories you might like