P1759 - EUS-Guided Needle-Based Confocal Laser Endomicroscopy (nCLE) for Dysplasia Classification in IPMNs: Assessment of Artificial Intelligence (AI) Algorithms to Reduce Interobserver Variability

Ahmed RM. Abdelbaki, MBBCh¹, Li Ziwei, ², Wei-Lun chao, PhD², Stacey Culp, PhD³, Phil Hart, MD¹, Jordan Burlen, MD¹, Samer Eldika, MD⁴, Helga Bertani, ⁵, Yunlu Feng, ⁶, Geri keane, MBBS⁷, Tara Keihanian, MD⁸, Antonio Ladd, MD⁹, Pradermchai Kongkam, ¹⁰, Sarah Leblanc, ¹¹, Anjuli Luthra, MD¹², Jorge D. Machicado, MD¹³, Shaffer Mok, MD¹², Thiruvengadam Muniraj, MD¹⁴, Bertrand Napoléon, ¹⁵, Damien Tan, ¹⁶, Kavel H. Visrodia, MD¹⁷, Somashekar G Krishna, MD^1
1The Ohio State University Wexner Medical Center, Columbus, OH; ²The Ohio State University, Columbus, OH; ³Ohio State University, Columbus, OH; ⁴Stanford Health Care, Palo Alto, CA; ⁵Baggiovara Hospital, Modena, Emilia-Romagna, Italy; ⁶Peking Union Medical College Hospital, Beijing, Beijing, China; ⁷Johns Hopkins University School of Medicine, Washington, WA; ⁸Baylor College of Medicine, Houston, TX; ⁹University of California Davis, Albuquerque, NM; ¹⁰King Chulalongkorn Memorial Hospital, Bangkok, Buriram, Thailand; ¹¹Cochin Hospital, Paris, Bourgogne, France; ¹²Moffitt Cancer Center, Tampa, FL; ¹³University of Michigan Health, Ann Arbor, MI; ¹⁴Yale New Haven Health, New Haven, CT; ¹⁵Hôpital Privé Jean Mermoz, Lyon, Bretagne, France; ¹⁶Singapore General Hospital, Central Region, Singapore; ¹⁷New York-Presbyterian/Columbia University Irving Medical Center, New York, NY

Introduction: EUS-nCLE enables real-time in vivo visualization of papillae for risk stratification of IPMNs, but interobserver agreement (IOA) is a limitation. We have developed a fully functional nCLE-based AI algorithm to detect high-grade dysplasia/invasive adenocarcinoma (HGD-IC) in branch duct (BD)-IPMNs. We compared the performance of nCLE-based dysplasia classification of BD-IPMNs between an expert panel of EUS-nCLE users and an improved nCLE-AI algorithm.

Methods: We selected consecutive IPMNs with definitive diagnoses from prospective studies (2015-2023) evaluating EUS-guided nCLE for the diagnosis of pancreatic cysts.
IOA study: An international panel of 16 blinded interventional endoscopists reviewed full-length, unedited nCLE videos to classify dysplasia.
AI assessment: A fully functional two-stage AI model was used: Stage 1 selected high-yield frames, and Stage 2 extracted nCLE features for risk stratification.
Comparisons: The diagnostic parameters of the observers and the nCLE-AI algorithm were compared to the Kyoto guidelines’ high-risk criteria (HRC).

Results: Among 60 IPMNs (HGD-IC: 23), the mean size was 3.43±1.00 cm and mean age of subjects was 68.2±9.0 years. In the observer group (n=16; 960 classifications), the mean sensitivity, specificity, and accuracy for the detection of HGD-IC were 58%, 59%, and 59%, respectively (Table 1). The interobserver agreement (IOA) was ‘fair’; κ=0.29 (95%CI: 0.27–0.32).

For the fully functional nCLE-AI algorithm, the sensitivity, specificity, and accuracy for detecting HGD-IC were 78%, 70%, and 73%, respectively. To detect HGD-IC, the sensitivity, specificity, and accuracy of Kyoto-HRC was 57%, 95%, and 80%, respectively (Table 1).

Discussion: The issue of suboptimal interobserver agreement among experts in classifying HGD-IC using nCLE imaging is resolved by introducing an nCLE-AI algorithm that does not rely on human interpretation. In this cohort, the nCLE-AI demonstrated superior sensitivity compared to the Kyoto-HRC for IPMN dysplasia classification while maintaining similar accuracy. Given the inherent ability of AI to improve with larger datasets, the algorithm's (nCLE-AI) accuracy and the validation of these findings would benefit from further refinement using extensive, real-time imaging data.



Disclosures:


Ahmed RM. Abdelbaki, MBBCh¹, Li Ziwei, ², Wei-Lun chao, PhD², Stacey Culp, PhD³, Phil Hart, MD¹, Jordan Burlen, MD¹, Samer Eldika, MD⁴, Helga Bertani, ⁵, Yunlu Feng, ⁶, Geri keane, MBBS⁷, Tara Keihanian, MD⁸, Antonio Ladd, MD⁹, Pradermchai Kongkam, ¹⁰, Sarah Leblanc, ¹¹, Anjuli Luthra, MD¹², Jorge D. Machicado, MD¹³, Shaffer Mok, MD¹², Thiruvengadam Muniraj, MD¹⁴, Bertrand Napoléon, ¹⁵, Damien Tan, ¹⁶, Kavel H. Visrodia, MD¹⁷, Somashekar G Krishna, MD¹. P1759 - EUS-Guided Needle-Based Confocal Laser Endomicroscopy (nCLE) for Dysplasia Classification in IPMNs: Assessment of Artificial Intelligence (AI) Algorithms to Reduce Interobserver Variability, ACG 2024 Annual Scientific Meeting Abstracts. Philadelphia, PA: American College of Gastroenterology.