Clinical and methodological considerations in artificial intelligence-assisted endoscopic ultrasonography for pancreatic mass detection

10.1055/a-2701-6530

I read with interest the recent article on the prospective clinical validation of the PANCRAIEUS artificial intelligence (AI) system for real-time detection of solid pancreatic masses during endoscopic ultrasonography (EUS) [1]. While the study marks an important advance toward integrating AI into pancreatic imaging, several methodological and clinical limitations merit consideration.

The study does not address the critical distinction between exophytic pancreatic tumors and lesions arising from adjacent organs, such as gallbladder cancer infiltrating the pancreas. This differentiation is essential, as management strategies, surgical eligibility, and prognosis differ markedly between primary pancreatic neoplasms and secondary involvement. Without explicit evaluation of this boundary, detection rates may overestimate true pancreatic lesion identification [2].

Contrast-enhanced EUS is pivotal in characterizing pancreatic masses, enabling assessment of vascularity, necrosis, and enhancement patterns that aid in distinguishing adenocarcinoma from neuroendocrine tumors or inflammatory lesions [3]. The omission of contrast-enhanced EUS data limits applicability to real-world workflows, where multimodal EUS – including contrast enhancement – is increasingly standard practice.

Beyond detection, EUS plays a central role in staging and surgical planning by evaluating vascular involvement (portal vein, superior mesenteric vessels, celiac axis) and regional lymphadenopathy [4]. The study does not report these parameters, narrowing the utility of the AI system to detection alone, without addressing its potential contribution to comprehensive staging.

The discriminative architecture of the AI system is inherently prone to false positives when encountering atypical anatomy or artifacts [5]. The manuscript does not clarify whether safeguards such as uncertainty quantification, confidence thresholds, or secondary verification loops are incorporated to mitigate misclassification risk, particularly for less experienced operators.

While the study demonstrates promising detection rates, the absence of lesion origin differentiation, contrast-enhanced EUS integration, resectability assessment, and defined error-handling mechanisms limits immediate clinical translation. Future work should incorporate these elements to ensure AI-assisted EUS evolves into a comprehensive, reliable tool for pancreatic disease evaluation.

Publication History

Article published online:
22 January 2026

• References

• 1 Bang JY, Saftoiu A, Udriștoiu A. et al. Prospective clinical validation of a novel artificial intelligence system for real-time detection of solid pancreatic masses during endoscopic ultrasonography. Endoscopy 2025;
• 2 Schawkat K, Manning MA, Glickman JN. et al. Pancreatic ductal adenocarcinoma and its variants: pearls and perils. Radiographics 2020; 40: 1219-1239
• 3 Kitano M, Kamata K, Imai H. et al. Contrast‐enhanced harmonic endoscopic ultrasonography for pancreatobiliary diseases. Dig Endosc 2015; 27: 60-67
• 4 Tiwari A, Khanikar D, Sharma M. Endoscopy’s evolving role in locally advanced and borderline resectable pancreatic cancer. Ann Surg Oncol 2024; 31: 7086-7087
• 5 Kelly CJ, Karthikesalingam A, Suleyman M. et al. Key challenges for delivering clinical impact with artificial intelligence. BMC Med 2019; 17: 195