Pancreatic cancer remains one of the most life-threatening malignancies because it is so difficult to detect in its earliest, most manageable stage. Recent developments in AI, however, suggest it may be possible to improve the diagnostic process.
By John Halamka, M.D., M.S., Diercks President, Mayo Clinic Platform and Paul Cerrato, MA, senior research analyst and communications specialist, Mayo Clinic Platform and a Professor at Northeastern University.
Although pancreatic ductal adenocarcinoma (PDA) doesn’t get a lot of attention in the popular press, it’s actually the leading cause of cancer-related deaths worldwide, accounting for over 467,000 deaths each year. By one estimate, about 80% of patients with PDA have already experienced localized advanced disease or metastasis by the time they are diagnosed; only 2-9% of patients survive for 5 years. Despite the urgency of the situation, two perplexing issues remain: Who should undergo diagnostic testing and which procedures are most likely to detect the cancer? The problem is that the most common signs and symptoms of PDA are nonspecific, including abdominal pain, weight loss, loss of appetite and fatigue, all of which can be caused by many other conditions. Other possible signposts include painless jaundice, new-onset or worsening diabetes, and pancreatitis. Several expert organizations recommend contrast-enhanced CT scans for at-risk patients but as Alves et al have pointed out: “… detection on CT is not standardised and is susceptible to high inter-reader variability.”
With these concerns in mind, several investigators have explored the potential value of AI-driven algorithms to increase the accuracy of CT scanning for the disease. For instance, Mayo Clinic investigators conducted a case/control study that used a convolutional neural network (CNN) to detect pancreatic cancer in diagnostic CT scans and to detect visually occult pre-invasive cancer on pre-diagnostic CTs. The CNN accurately classified 360 scans (88%) as cancer and 783 of the control patients’ scans as non-cancerous (94%). Similarly, the algorithm was able to identify occult PDA on the pre-diagnostic scans (accuracy 84%) more than a year prior to clinical diagnosis (475 days).
Chinese researchers have found deep learning models a valuable adjunct in diagnosing PDA. Cao et al used an algorithm they called PANDA to accurately detect and classify pancreatic lesions using non-contrast CT scans. PANDA was trained on a dataset of over 3,000 patients from a single center but was nonetheless able to detect the cancer during multicenter validation involving over 6,000 patients across 10 centers with an area under the curve (AUC) of 0.986-0.996. They reported that the model “outperforms the mean radiologist performance by 34.1% in sensitivity and 6.3% in specificity for PDAC identification, and achieves a sensitivity of 92.9% and specificity of 99.9% for lesion detection in a real-world multi-scenario validation consisting of 20,530 consecutive patients.”
European researchers have also successfully developed an AI model based on a dataset called PANORAMA to help detect PDA. Their noninferiority observational investigation used data from the Netherlands, USA, Sweden, and Norway concluding that “AI demonstrated substantially improved PDAC detection on routine CT scans compared to radiologists on average, showing potential to detect cancer earlier and improve patient outcomes.” While the study has limitations that go beyond the scope of our short column and it did not directly demonstrate that its model can improve early detection, their model improved sensitivity and reduced the number of unnecessary referrals, by generating up to 38% fewer false positives when compared to the performance of 68 radiologists.
Of course, in order for AI models to enhance the accuracy of CT scans, at-risk patients need to undergo a scan in the first place, and that poses difficult questions: How does one define high risk and what testing facilities are most capable of performing an accurate procedure? Symptoms are often vague and not all radiology departments have the same expertise or available equipment. However, there is evidence to suggest that certain biomarkers may help identify persons most at risk. Mayo Clinic and others, for instance, have used glycemically identified new onset diabetes and an END-PAC score of 3 or more to choose candidates for CT scanning. END-PAC refers to Enriching New Onset Diabetes for Pancreatic Cancer. The score is derived from changes in body weight and blood glucose and the age when a patient first develops diabetes. In one study, END-PAC was able to identify pancreatic cancer within 3 years of the onset of diabetes (AUC, 0.87).
Although we have a long road ahead in our battle to defeat pancreatic cancer, the research to date shows that with the assistance of the best AI models, we are clearly making progress.
Share this:
