August 23, 2022 | Radiomics-based machine learning models have been shown to do an impressive job of detecting pancreatic cancer on pre-diagnostic CT scans several months before symptoms appear, suggesting that one day the disease will more often be diagnosed at a stage where a surgical cure may be possible, according to Ajit H. Goenka, M.D., diagnostic radiologist at the Mayo Clinic. Currently, no more than one-quarter of the overall patient population presents with disease that is localized to the pancreas and therefore has hope of potential long-term survival.
Early pancreatic tumors are typically small and oddly situated, making it easy for even top-notch radiologists to miss them on a CT scan, Goenka says. Up to 40% of small pancreas cancers may not even show up on standard imaging. Moreover, some findings suggestive of early cancer are also seen in a normal pancreas.
Artificial intelligence (AI) may be able to help overcome the diagnostic conundrum, as demonstrated in a study that published recently in Gastroenterology (DOI: 10.1053/j.gastro.2022.06.066). Machine learning models could reliably differentiate between patients who went on to develop pancreatic cancer versus those who had a normal pancreas.
The image analysis exercise was done on 155 patients with pancreatic cancer who happened to have an incidental CT scan done no more than three years prior to diagnosis, together with an age-matched cohort of 265 subjects with a normal pancreas. Support vector machine model performed particularly well, with 95.5% sensitivity and 90.3% specificity, and the high specificity was validated (at 96.2%) on an open-source dataset of 80 normal CTs maintained by the National Institutes of Health (NIH).
Machine learning models were able to predict future risk of pancreatic cancer at a median time of 386 days prior to clinical diagnosis, with accuracies ranging from 94% to 98%. Notably, those results were unaffected by variations in image noise, scanner models, image acquisition protocols, and postprocessing parameters, says Goenka.
All four machine learning models—also including k-nearest neighbor, random forest, and XGBoost—outperformed a pair of fellowship-trained radiologists, who also recorded false positive indirect findings of pancreatic cancer in control subjects. This is certainly not due to incompetence of the experts reading the scans, Goenka stresses. “It is just the nature of this particular pathology.”
The radiologists were blinded to information about whether the CTs belonged to the prediagnostic group or the control cohort and were also instructed to look only at the pancreas, he points out. If anything, performance of the radiologists was overestimated since in the real world the doctors would also have been paying attention to multiple other organs.
It has become commonplace for physicians like Goenka to drive the kind of AI that gets deployed in clinical practice. This project was an interdisciplinary collaboration with AI scientists and gastroenterologists from the Mayo Clinic.
Goenka was introduced to AI a few years ago after an informal conversation with a few of his gastroenterologists colleagues about the early-detection crisis in pancreatic cancer and the ability of machine learning to pick up on imaging features undetectable by the human eye. “It was a learning curve for all of us,” he says, noting that about 80% of the study time was spent finding the right kind of datasets and making sure they were free of biases.
“We started off with 3,000 patients with pancreatic cancer and then we went back in time for every one of them, to see who had an incidental CT scan done,” explains Goenka. These prediagnostic scans would be on patients who, for example, had gone to the emergency room with food poisoning where a CT scan was done to rule out other conditions.
The problem being addressed is twofold, Goenka says. Cancer is starting much earlier than can be detected by the most well-trained CT reader, and false-positive findings exact a heavy penalty both on limited healthcare resources and patients who unnecessarily undergo complex and costly examinations with all its associated complications.
One of the challenges in advancing AI is the scarcity of deidentified data sets in the public domain, says Goenka, which was important to use in the case-control study as a benchmark against which future machine learning models can likewise be compared or validated.
The NIH dataset contains the CTs of subjects without any pathology who were scanned prior to making a kidney donation. The incidence of pancreas cancer relative to lung cancer or breast cancer is not very high, he says, so being able to weed out people without the disease was important.
“We are not close to clinical translation right now,” Goenka says. “You have to get to base camp before you can climb to Mount Everest, and we certainly are on the right trajectory.”
Early Detection Initiative
Next steps for the research team include exploring the option of further validating the AI models on CTs being done as part of the Early Detection Initiative, a large prospective clinical trial sponsored by the Pancreatic Cancer Action Network. The study, which launched last fall, will enroll 12,500 participants in the U.S. at high risk for pancreatic cancer and evaluate the impact of a screening strategy using CT.
The study chair is Suresh Chari, M.D., professor of medicine at the University of Texas MD Anderson Cancer Center and formerly director of the pancreas clinic at the Mayo Clinic. He holds an emeritus position at Mayo.
Goenka names senior data science scientists Sovanlal Mukherjee, Ph.D., Panagiotis Korfiatis, Ph.D., and gastroenterologist Shounak Majumder, M.D., as his key partners in the early-detection process at the Mayo Clinic. They are anticipating institutional review board approval to start enrolling patients in the Early Detection Initiative.
Study participants will be followed to see how many end up having pancreatic cancer based upon disease risk factors. Inclusion of the machine learning models would presumably help ensure more true positives are detected at an earlier, more treatable stage, says Goenka.
It will be easy enough to validate the machine learning models on participants enrolled at Mayo, and it is expected that the patient-friendly Pancreatic Cancer Action Network and the other trial sites will be in favor of having all images run through the models, he continues. It wouldn’t require any modifications to the protocol since the study involves standard-of-care CT scans.
It is in any case “just a matter of time” before AI becomes a part of the early-detection toolkit for pancreatic cancer, says Goenka. All the major institutions are experimenting with the technology, and the intention at Mayo as elsewhere is to see it translated widely. “That is the only way we can make a big difference.”
Once the models have been prospectively validated, they could be deployed almost immediately into clinical care practices at the Mayo Clinic, he adds. “The idea is not for AI to function independently, but with the oversight of radiologists, because at the end of the day machines are machines and they need human supervision.”
That approach also lowers the bar for the software to be approved by the U.S. Food and Drug Administration in terms of risk-benefit profile, says Goenka. Mayo has experience taking machine learning models through that process, and currently has more than 200 AI projects in various stages of maturity in addition to a research department for AI and informatics, a Center for Digital Health, and its own platform for developing and deploying healthcare solutions.