Skip to main content
Premium Trial:

Request an Annual Quote

AI-Based CT Imaging Model Could Catch Pancreatic Cancer Early


NEW YORK – An artificial intelligence tool under development may detect early-stage pancreatic cancer in patients incidentally receiving an abdominal CT scan and could become a broader screening tool for patients with risk factors when paired with a blood and urine biomarker kit.

Pancreatic cancer is very aggressive with a five-year survival rate of 11 percent. However, if caught early, around 50 percent of pancreatic cancer patients live at least five years after diagnosis, according to the American Cancer Society. However, the early signs of cancer, such as unexplained weight loss, jaundice, and abdominal pain, tend to be vague and overlap with symptoms of other illnesses. Many people presenting with those symptoms will find themselves inside a CT scanner. However, conventional, manual evaluation by a radiologist still misses many patients in early stages, who go on to develop late-stage cancer. 

Researchers at Cedars-Sinai Medical Center have developed an AI model that picks up features from CT scans at that "pre-diagnostic" stage, which are not appreciated by the human eye. There would be ample opportunity to deploy the tool. According to the CDC, abdominal pain is the single most common reason that Americans visit the emergency room, adding up to about 7 million visits per year, and those patients often receive a CT scan. The CT scans, whether ordered for abdominal pain or for something completely unrelated, like a car accident, could then automatically be screened for pancreatic cancer.

The researchers developed the tool by training it on CT scans from people with pancreatic cancer and comparing those scans with healthy controls. The method, called radiomic analysis, involves extracting a large number of features from the images and then analyzing them in combination with clinical data. The resulting model was 86 percent accurate at identifying scans from patients who would later go on to develop pancreatic cancer based on textural differences on the surface of the pancreas.

Debiao Li, director of the Biomedical Imaging Research Institute at Cedars-Sinai, said his team's next step will be an independent, prospective validation of the model. One challenge will be getting enough data to do that study. "We are solving that problem by teaming up with other medical centers," said Li. The initial study was in less than 100 patients. Li said the research team hopes to increase that number tenfold, to about 1,000 patients, for the validation phase.

"Once developed, the tool would run in the background and alert the reading radiologist of a possible small cancer or predict an area that would develop into cancer," said study author Stephen Pandol, director of Basic and Translational Pancreas Research at Cedars-Sinai. He and his collaborators are working to add clinically available data such as age, smoking history, diabetes, and other factors to increase sensitivity and specificity of the model.

Additionally, Pandol said his group is working with a team led by Tatjana Crnogorac-Jurcevic at Barts Cancer Institute, Queen Mary University of London, to develop a panel of three urinary biomarkers and one blood biomarker to use in conjunction with the AI imaging tools. Pandol and Crnogorac-Jurcevic said their work on the panel is nearly complete and they will soon be preparing kits for clinical use and final regulatory approval. "We posit that the combination of our prediction model based on imaging and the urine and blood tests can further boost a combined performance for specific early detection for effective treatment and possibly prevention," said Pandol.

The urinary biomarker panel screens for three proteins: lymphatic vessel endothelial hyaluronan receptor 1 (LYVE1), regenerating family member 1 beta (REG1B), and trefoil factor 1 (TFF1). Crnogorac-Jurcevic is testing that panel with an affiliated PancRISK score alone and in combination with plasma carbohydrate antigen 19-9 (CA 19-9) for early detection of pancreatic cancer in the UroPanc clinical trial. CA 19-9 has been shown to predict outcomes for patients with advanced pancreatic cancer who are treated with gemcitabine-containing chemotherapy regimens.

The study will determine the urinary biomarker test's accuracy using samples from symptomatic and asymptomatic people at risk of pancreatic cancer through collaborations with biobanks and registries at University College London and University of Liverpool. The symptomatic group will comprise patients who are suspected of having pancreatic cancer based on their symptoms. Asymptomatic participants will be drawn from the EUROPAC registry of more than 400 people who are at high risk of developing pancreatic cancer due to family history.

In a preclinical study, sensitivity and specificity of the urinary biomarker panel were both over 85 percent. Crnogorac-Jurcevic said combining the urine biomarker panel with imaging or using imaging to follow up on urine biomarker testing would improve the accuracy of cancer detection compared to either test alone, and imaging would add important information about the location of the tumor.

Once fully developed, both the CT imaging screen and the urinary biomarker test would be used mainly in two groups of patients. First, in people with increased risk factors, such as a family history of pancreatic cancer or inherited mutations in BRCA1/2 linked to pancreatic cancer risk.

The second group targeted for screening would include those with symptoms suggestive of pancreatic cancer like lower back pain, indigestion, and changes in bowel movements. "We are constantly looking for combinations of different modalities to achieve as soon as possible correct detection of pancreatic cancer," said Crnogorac-Jurcevic.

These tests could be ordered by primary care physicians, internists, gastroenterologists, oncologists, or other types of practitioners. According to Pandol, the AI software would run in all CT scanner imaging data output. "In other words, in the future, algorithms will be applied to CT and other imaging modalities to assist in interpretation of the imaging data," said Pandol.

Clinical, genetic, and other data added to the algorithm would further increase performance. The method is theoretically applicable to other cancers and diseases, but the algorithms would have to be tailored to those diseases.

Patients with positive screens would have several options. "At this point, careful monitoring including repeating the [test] would be the best option," Pandol said. "This is because performing biopsies and surgery have risk. I have been thinking about how we can show that instituting a safe, preventive therapy can reduce the risk in such patients." That therapy would be an agent with a high safety profile and it would be used in patients whose urine and blood tests or AI-based imaging screen indicate a high risk but do not have enough evidence to justify surgery.

For patients testing positive on the PancRISK panel, Crnogorac-Jurcevic said they might be referred to a specialist for more invasive testing including endoscopy.

Pandol said the imaging screen is not intended for routine screening of healthy people. The researchers cautioned that in implementing such an AI tool, the screening modalities must be designed to minimize unnecessary procedures and associated harms in people without pancreatic cancer risk factors.

"As with any test, there is a false-positive rate, which will certainly increase when large numbers of patients with low risk are tested," said Pandol. "False positives have multiple negative consequences from emotional trauma to application of needless interventions that could add risk to the patient."