Despite substantial progress in endoscopic ultrasound and fine needle techniques, critical diagnostic uncertainties persist in the evaluation of pancreatic cysts, solid lesions, and lymph nodes. Misclassification continues to drive overtreatment or delayed therapy, particularly in indeterminate cysts, low-grade IPMNs, and lymph node staging. Real-world evidence on how artificial intelligence could reshape decision making across diverse EUS scenarios remains limited. Addressing this gap is essential to advance precision, reduce unnecessary interventions, and strengthen the accuracy of pancreatic and nodal assessment.
This multicenter retrospective study applied a validated AI-CNN ensemble to real EUS images from three distinct patients, in a post hoc, counterfactual fashion. The models — previously trained and tested on separate multicenter datasets — included pancreatic cystic model (mucinous vs. non-mucinous cyst), pancreatic solid lesions model (PDAC vs. PNET), and lymph nodes model (benign vs. malignant LN). All predictions were made blinded to histopathology and outcomes. The aim was to evaluate whether AI predictions could have informed different more accurate clinical actions.
Case 1: A 71-year-old woman with a stable 3.5 cm pancreatic cyst underwent multiple EUS-guided FNAs over two years, all of which were non-diagnostic. The AI system classified the lesion as non-mucinous with over 90% confidence. In retrospect, this high-certainty prediction could have supported a decision to reduce the intensity of follow-up, avoiding repeat procedures and minimizing long-term patient anxiety associated with presumed risk of malignancy.
Case 2: A 48-year-old man with a 7.1 cm mixed-type IPMN and several worrisome features underwent duodenopancreatectomy. Final histopathology revealed only low-grade dysplasia. Retrospectively, the AI model classified the lesion as low-grade with consistent confidence levels above 80%. Had the tool been available in real time, this result might have supported a more conservative management approach, potentially sparing the patient from an extensive surgery with significant morbidity.
Case 3: A 59-year-old man with obstructive jaundice underwent two EUS-guided FNAs targeting a pancreatic mass and hilar lymph nodes. Pancreatic adenocarcinoma was confirmed only after the second attempt, while lymph node cytology remained negative. Retrospectively, the AI model identified the mass as PDAC with 98% probability. Furthermore, the AI model highlighted one accessible lymph node—previously unbiopsied—with fluctuating malignancy probabilities between 57% and 71%. Targeting this node might have altered staging, supporting the early initiation of neoadjuvant chemotherapy. The patient ultimately underwent surgery, which confirmed pancreatic ductal adenocarcinoma with N2 nodal status, underscoring the potential of AI to inform more accurate staging and improve the likelihood of achieving an R0 resection—an important prognostic factor in pancreatic cancer.
In this post hoc counterfactual analysis, the AI models demonstrated strong concordance with final histology and exposed pivotal decision moments where real-time guidance could have meaningfully altered management. The findings point to clear opportunities to reduce unnecessary procedures, prevent overtreatment, and accelerate timely systemic therapy. Despite its retrospective design, this innovative framework positions AI as a compelling second-opinion tool and calls for prospective studies to integrate AI-guided EUS into routine clinical workflows, aiming to elevate the precision and impact of pancreatic care.