Abstract
BACKGROUND AND AIMS: Endoscopic Ultrasound (EUS)-guided needle-based confocal laser endomicroscopy (nCLE) enables real-time microscopic visualization of pancreatic cyst epithelium and can identify high-grade dysplasia/invasive adenocarcinoma (HGD/IC) in branch-duct (BD) intraductal papillary mucinous neoplasms (IPMNs). We aimed to compare the performance of experts (humans) with artificial intelligence (AI) in stratifying dysplasia in BD-IPMNs. METHODS: This post-hoc analysis involved BD-IPMNs with definitive diagnoses from prospective EUS-nCLE studies (2015-2023) enrolled at a single center. Dysplasia grade was reviewed by two pathologists. Blinded EUS-nCLE experts reviewed unedited nCLE videos to classify dysplasia without and with revised Fukuoka criteria (revised-FC). The AI model, nCLE-AI, was similarly analyzed. Diagnostic parameters and AUC were compared to evaluate human and nCLE-AI performance. RESULTS: Among 60 BD-IPMNs (mean size = 3.43 ± 1.00 cm), 23 (38.3 %) had HGD/IC. To detect HGD/IC using nCLE, interobserver agreement (IOA) among 16 nCLE experts was 'fair' (κ = 0.29, 95 % CI: 0.27-0.32), with a sensitivity of 58 %, specificity of 59 %, and AUC of 0.59 (95 % CI 0.55-0.62). Incorporating revised-FC improved the sensitivity to 72 % and AUC to 0.64 (95 % CI 0.61-0.68; p < 0.001), with similar IOA (κ = 0.36 'fair', 95 % CI: 0.33-0.38) and specificity (57 %). Comparatively, nCLE-AI achieved 87 % sensitivity, 54 % specificity, and an AUC of 0.70 (0.57-0.84). When combined with revised-FC, nCLE-AI reached 78 % sensitivity, 78 % specificity, and an AUC of 0.85 (95 % CI: 0.74-0.96, p = 0.02), significantly higher than humans with revised-FC (p < 0.01). CONCLUSIONS: Human dysplasia classification of BD-IPMNs using nCLE showed modest IOA and accuracy. In contrast, nCLE-AI classifications combined with clinical criteria offer superior accuracy for detecting HGD/IC while eliminating interobserver variability.