P1446 - Comparing the Performance of Revised International Consensus Guidelines, Manual and Artificial Intelligence Interpretation of Needle-Based Confocal Endomicroscopy in Predicting Advanced Neoplasia of IPMNs

Award: Presidential Poster Award

Matt Leupold, MD¹, Troy Cao, BS², Stacey Culp, PhD³, Wei-Lun Chao, PhD⁴, Ronald Turner, ¹, Jared Melynchuck, ¹, Nehaal Ahmed, ¹, Aayush Vishwanath, ³, Somashekar Krishna, MD^5
1The Ohio State University Wexner Medical Center, Columbus, OH; ²Ohio State University Wexner Medical Center, Columbus, OH; ³The Ohio State University, Columbus, OH; ⁴The Ohio State University, Colubmus, OH; ⁵The Ohio State Wexner Medical Center, Columbus, OH

Introduction: The risk stratification of branch duct intraductal papillary mucinous neoplasms (BD-IPMNs) is typically based on the Fukuoka 2017 International Consensus Guidelines (ICG) high-risk (HR) criteria. EUS-guided needle-based confocal laser endomicroscopy (EUS-nCLE) has emerged as an advanced diagnostic modality for distinguishing high-grade dysplasia/adenocarcinoma (HGD-Ca) in BD-IPMNs. A convolutional neural network-based artificial intelligence (AI) algorithm was recently developed to edit and interpret full-length nCLE videos from IPMNs. In this study, we applied blinded manual interpretation (manual-nCLE) and AI algorithms (AI-nCLE) to full-length unedited nCLE videos obtained from BD-IPMNs. Our objective was to compare the performance of the ICG-HR, manual-nCLE, and AI-nCLE for the risk-stratification of BD-IPMNs.

Methods: Subjects with BD-IPMNs were enrolled from a prospective database at a single center evaluating pancreatic cystic lesions (2015-2023). All nCLE videos were reviewed by a blinded expert documenting low-grade (LGD) vs. HGD-Ca (Fig A-B). The diagnostic interpretations of ICG high-risk criteria, manual-nCLE, and AI-nCLE were compared to evaluate their performance. Combined accuracies for the three interpretations were also analyzed.

Results: A total of 60 subjects were included in the analysis, with 25 cases used for developing the AI algorithm and 35 cases used for testing. In the test group (34.3% female, mean age 67±9 years, mean cyst diameter 36.3 ± 10.3 mm), 20% of the BD-IPMNs were histopathologically diagnosed as HGD-Ca.

Table 1 presents the performance of ICG-HR criteria, manual-nCLE, and AI-nCLE for differentiating BD-IPMNs. Sensitivity was lowest for ICG-HR criteria (43%) and highest for AI-nCLE (71%). Specificity was lowest for AI-nCLE (64%) and highest for manual-nCLE (93%). Overall accuracy was highest for manual-nCLE at 86%. The combination of ICG-HR and AI-nCLE demonstrated the highest sensitivity of 86%.

Discussion: The ICG-HR criteria alone demonstrated poor sensitivity, suggesting a higher likelihood of missing BD-IPMNs with HGD-Ca. While manual-nCLE had the highest overall specificity and accuracy, it is susceptible to interobserver variability. The combination of AI-nCLE and ICG-HR criteria resulted in the greatest sensitivity. The performance of AI-nCLE can be further enhanced through the accumulation of more data and images. This highlights the need for ongoing research to optimize the AI-based approach and improve its accuracy.



Disclosures:


Matt Leupold, MD¹, Troy Cao, BS², Stacey Culp, PhD³, Wei-Lun Chao, PhD⁴, Ronald Turner, ¹, Jared Melynchuck, ¹, Nehaal Ahmed, ¹, Aayush Vishwanath, ³, Somashekar Krishna, MD⁵. P1446 - Comparing the Performance of Revised International Consensus Guidelines, Manual and Artificial Intelligence Interpretation of Needle-Based Confocal Endomicroscopy in Predicting Advanced Neoplasia of IPMNs, ACG 2023 Annual Scientific Meeting Abstracts. Vancouver, BC, Canada: American College of Gastroenterology.