Related Experiment Video For AI
Updated: Mar 24, 2026
Author Spotlight: Advancing Early Detection and Treatment of Gastrointestinal Tumors
Published on: February 16, 2024
Preoperative Prediction of Subtype and Artificial Intelligence-Driven Endomicroscopy Detection of Advanced Neoplasia
Bryn D Koehler1, Wei Chen2, Ashwini K Esnakula2
1Department of Internal Medicine, The Ohio State University Wexner Medical Center, Columbus, OH 43210, USA.
Background And Aims:
Non-gastric subtype of branch duct (BD)-intraductal papillary mucinous neoplasms (IPMNs) are associated with high-grade dysplasia/invasive adenocarcinoma (HGD/IC) and disease progression. We evaluated preoperative prediction of gastric vs. non-gastric BD-IPMN subtypes and assessed a needle-based confocal laser endomicroscopy-guided artificial intelligence algorithm (nCLE-AI) for detecting HGD/IC in pathologist-reclassified BD-IPMNs.
Methods:
Participants with resected BD-IPMNs were enrolled from prospective studies (2015-2024). Phase 1: lesions were reclassified by subtype and dysplasia grade through blinded pathologist review, with discordant cases receiving MUC immunostaining and consensus review. Phase 2: using this reclassified pathology data, preoperative clinical and morphological features were analyzed to predict BD-IPMN subtypes. Phase 3: nCLE-AI performance in detecting HGD/IC within reclassified gastric and non-gastric BD-IPMNs was evaluated using preoperative endomicroscopy videos.
Results:
Among 63 resected BD-IPMNs (mean diameter=35.0±10.1 mm), 38% were classified as HGD/IC. Phase 1: The interobserver agreement among pathologists for subtype classification was moderate (k=0.52; 95%CI: 0.27-0.77). Phase 2: Multivariable analysis or preoperative variables revealed Kyoto high-risk stigmata (aOR=11.568, p=0.007), unifocal lesions (aOR=8.354, p=0.041), and lower BMI (aOR=1.37, p=0.04) predicted non-gastric subtype. Phase 3: The nCLE-AI algorithm using presurgical endomicroscopy imaging showed comparable sensitivity for detecting HGD/IC in non-gastric and gastric IPMN subtypes (83% vs. 82%, p=0.92), but significantly higher specificity (100% vs. 44%, p=0.06) and accuracy (87% vs. 53%, p<0.02) in the non-gastric subtype.
Conclusion:
Moderate interobserver variability in BD-IPMN subtype classification among pathologists highlights the need for immunohistochemistry and consensus review in challenging cases. Preoperative clinical variables can predict non-gastric subtype which is associated with less favorable prognosis. nCLE-AI shows improved performance in detecting HGD/IC in non-gastric BD-IPMNs, where accurate risk stratification is particularly important due to higher risk of progression.
