Pancreatic cystic lesions are increasingly detected as incidental findings, with intraductal papillary mucinous neoplasms representing the most prevalent and clinically significant subtype. Their management hinges on the ability to differentiate low grade dysplasia from high grade dysplasia or carcinoma, a distinction that carries direct implications for surveillance strategies and surgical selection. However, accurate preoperative stratification remains challenging, and current diagnostic tools offer only limited precision. In this context, this study aimed to develop and validate a deep learning model capable of distinguishing high-grade dysplasia or carcinoma from low grade dysplasia using endoscopic ultrasound images, with the objective of enhancing diagnostic accuracy and informing evidence based clinical decision making.

An artificial intelligence (AI) based on a convolutional neural network was developed, including both detection and classification components. Lesions identified by the AI algorithm were identified using a bounding box, and further classified into LGD or HGD/C (Figure 1A). The ground truth classification was based on cytological analysis of the pancreatic cyst fluid, EUS-guided through-the-needle biopsy, or surgical specimens. Model performance was assessed based on mean average precision with an intersection-over-unit threshold of 50% (mAP50), sensitivity, precision, accuracy, and area under the precision-recall curve (AUPRC).

A total of 22,515 EUS images were extracted from 40 exams performed at 5 centers in Portugal, Spain, Brazil and the United States. The model detected IPMNs with a mAP50 of 0.885. Regarding the characterization module, the AI system distinguished IPMNs with HGD/C from those with LGD with an overall sensitivity of 88.2%, a precision of 89.4%, and an accuracy of 89.1%. The AUPRC was 0.874 (Figure 1B).

This study constitutes one of the earliest investigations into an AI system capable of concurrently detecting IPMN and classifying the degree of dysplasia. Precise differentiation between HGD/C and LGD remains essential for guiding clinical decisions, ensuring timely identification of high-risk lesions while avoiding unnecessary surgical intervention in patients with lower malignant potential. These findings underscore the need for continued development and validation of such tools to strengthen risk stratification in routine practice.