Abstract
BACKGROUND AND AIMS: Distinguishing pancreatic cancer from nonneoplastic masses is critical and remains a clinical challenge. The study aims to construct a deep learning-based artificial intelligence system to facilitate pancreatic mass diagnosis, and to guide EUS-guided fine-needle aspiration (EUS-FNA) in real time. METHODS: This is a prospective study. The CH-EUS MASTER system is composed of Model 1 (real-time capture and segmentation) and Model 2 (benign and malignant identification). It was developed using deep convolutional neural networks and Random Forest algorithm. Patients with pancreatic masses undergoing CH-EUS examinations followed by EUS-FNA were recruited. All patients underwent CH-EUS and were diagnosed both by endoscopists and CH-EUS MASTER. After diagnosis, they were randomly assigned to undergo EUS-FNA with or without CH-EUS MASTER guidance. RESULTS: Compared with manual labeling by experts, the average overlap rate of Model 1 was 0.708. In the independent CH-EUS video testing set, Model 2 generated an accuracy of 88.9% in identifying malignant tumors. In clinical trial, the accuracy, sensitivity, and specificity for diagnosing pancreatic masses by CH-EUS MASTER were significantly better than that of endoscopists. The accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were respectively 93.8%, 90.9%, 100%, 100%, and 83.3% by CH-EUS MASTER guided EUS-FNA, and were not significantly different compared to the control group. CH-EUS MASTER-guided EUS-FNA significantly improved the first-pass diagnostic yield. CONCLUSION: CH-EUS MASTER is a promising artificial intelligence system diagnosing malignant and benign pancreatic masses and may guide FNA in real time. TRIAL REGISTRATION NUMBER: NCT04607720.