Real-Time AI During Pancreatoscopy for Detection of Pancreatic Neoplasia

This prospective pilot diagnostic-accuracy study will evaluate the performance of a previously validated artificial intelligence (AI) model when applied in real time during digital per-oral pancreatoscopy (POPS) for the identification of pancreatic neoplastic lesions and intraductal papillary mucinous neoplasm (IPMN). Adults undergoing clinically indicated digital POPS at the Instituto Ecuatoriano de Enfermedades Digestivas (IECED) will be prospectively enrolled. Eligible patients will include those undergoing pancreatoscopy for suspected or known main-duct or mixed-type IPMN, branch-duct IPMN with suspected main-duct communication and concerning features, indeterminate pancreatic-duct strictures or filling defects, or preoperative assessment and mapping of IPMN extent. During each POPS examination, the endoscopist will first perform and record a conventional visual assessment while the AI system remains hidden. The AI system, AIWorks-Cholangioscopy, will subsequently be activated during the same examination. The previously validated model was developed and validated using digital cholangioscopy data and will be applied to digital pancreatoscopy video without modification of its model weights. AI-generated findings will be recorded independently and compared with the endoscopist's initial assessment. The AI system will provide real-time visual information, including detection and localization of suspected abnormal areas. AI findings will be documented as an index diagnostic test and will not independently determine patient management. Tissue sampling and subsequent clinical management will remain at the discretion of the treating endoscopist and multidisciplinary team according to standard clinical practice. When feasible, findings identified by either the endoscopist or AI may be documented for correlation with subsequent tissue sampling. The study will also record whether AI findings were concordant or discordant with the initial endoscopist assessment and whether the information was considered during the procedure. The reference standard will consist of surgical histopathology when pancreatic resection is performed. In patients who do not undergo surgery, the reference assessment will be based on available intraductal tissue sampling and/or cytology together with clinical, imaging, and endoscopic follow-up for up to 6 months. Histopathologic assessment will be performed independently of the AI findings and the endoscopist's locked pre-AI assessment whenever feasible. The primary diagnostic endpoint will be patient-level identification of high-grade dysplasia or invasive carcinoma, classified as a binary outcome of high-grade dysplasia/invasive carcinoma versus all other diagnostic categories. Diagnostic performance of real-time AI will be estimated using sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy, with corresponding 95% confidence intervals. Secondary analyses will evaluate the diagnostic performance of the endoscopist's initial visual assessment, agreement and discordance between AI and endoscopist assessments, identification of IPMN epithelium, segment-level findings, technical feasibility of real-time AI application, and the relationship between AI findings and subsequent tissue sampling or clinical decision-making. Procedural safety will also be assessed through recording of adverse events occurring within 30 days of pancreatoscopy. The study is designed as a pilot investigation. The planned evaluable sample is 60 participants, with up to approximately 70 participants potentially screened or enrolled to account for exclusions and non-evaluable examinations. The pilot is intended to generate preliminary patient-level diagnostic-accuracy estimates, evaluate the feasibility of real-time AI application during digital POPS, characterize AI-endoscopist concordance, and provide parameters for the design and sample-size planning of a future confirmatory diagnostic-accuracy study.