Artificial intelligence (AI) is rapidly being applied to medical imaging; however, the evidence base for endoscopic ultrasonography-based AI (EUS-AI) remains limited. We conducted a structured literature search (PubMed, Embase, and the Cochrane Library) for studies on AI for the diagnosis of pancreatic diseases using EUS images. Overall, 1 detection and 17 classifications of pancreatic tumors, 4 classifications of cystic lesions, and 4 focused on parenchymal or station recognition were reported in recent peer-reviewed publications. Deep learning architectures, such as ResNet, EfficientNet, VGG, UNet++, YOLO, and custom convolutional networks, were utilized. Multimodal models combine imaging with clinical or cytological data. Reported accuracy ranged from 0.84 to 0.94; however, only a small number of studies performed external validation with enough data to confidently support their findings. Only one model has been approved in Japan with unpublished technical details. Significant challenges include limited data from single institutions, inconsistent case labeling, varying diagnostic criteria, and internal validation, which may overestimate the actual performance of the AI models. Future directions include establishing nationwide systems for collecting and sharing EUS images, applying large language models to assist in reporting and patient explanation, and exploring the use of AI to support or partially automate EUS procedures through integration with robotics. With the first commercial systems now appearing, continued innovation and rigorous validation are expected to accelerate the clinical usage of EUS-AI.
Kuwahara et al. (2026) studied this question.