A multi-center study published in Gut has evaluated the use of an artificial intelligence–based classifier to detect pancreatic cancer from cyst fluid samples. The approach combines genomic, proteomic, and clinical data to improve risk stratification of pancreatic cystic lesions, which remain diagnostically challenging.
Pancreatic cysts are frequently detected incidentally through imaging, but distinguishing benign from premalignant or malignant cysts remains difficult using current diagnostic tools. The new study aimed to assess whether a multiomic classifier could offer a more accurate method to support clinical decision-making, particularly in identifying cancer at an earlier, more treatable stage.
Researchers developed and validated a classifier – termed CompCyst-AI – using samples from 862 patients across four institutions. All patients underwent surgical resection of a pancreatic cyst, providing histopathologic diagnosis as ground truth. The cyst fluid samples were analyzed for mutations in common pancreatic cancer–associated genes (including KRAS, GNAS, and TP53), protein expression profiles, and standard clinical features such as imaging characteristics and cytology.
The classifier was trained on a subset of 380 cases and tested on 392 independent samples. Patients were assigned to one of three categories: surgery recommended, surveillance appropriate, or no surveillance needed. Results were compared against real-world clinical management decisions and final surgical pathology.
In the validation cohort, CompCyst-AI matched the pathology-determined management recommendation in 72 percent of cases, compared with 52 percent agreement using current clinical approaches. The classifier identified 90 percent of patients with high-grade dysplasia or invasive carcinoma who required surgery and reduced the number of unnecessary surgeries for patients with low-risk lesions.
Importantly, the model demonstrated strong negative predictive value, suggesting it may help reduce overtreatment of benign cysts. However, the authors noted that while the classifier improved diagnostic accuracy, certain limitations – such as the need for cyst fluid sampling – remain.
The study did not include patients managed non-operatively, meaning further work is needed to validate performance in routine surveillance populations. The researchers suggest that prospective studies will be essential to determine clinical utility, particularly for patients who are not surgical candidates or those under long-term monitoring.
Overall, these findings contribute to ongoing efforts to improve risk stratification of pancreatic cysts and reduce unnecessary surgeries while ensuring appropriate identification of cancer. The study also highlights the potential of combining multiomic data with AI algorithms in a structured, pathology-validated framework.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
