Abstract
BACKGROUND AND AIMS: Endoscopic ultrasound-guided needle-based confocal laser endomicroscopy (nCLE) provides real-time optical biopsies enabling diagnosis and risk stratification of intraductal papillary mucinous neoplasms (IPMNs). However, the clinical implementation of Endoscopic ultrasound-nCLE is hindered by time-consuming image review and interobserver variability. We aimed to optimize artificial intelligence (AI) models to accurately detect diagnostic structures to address these gaps. METHODS: Participants with definitive IPMN diagnoses were selected from prospective studies (2015-2023). Two observers labeled endomicroscopy images used to develop AI models to detect informative segments. Performance was assessed using area under the curve, sensitivity, specificity, and accuracy. RESULTS: In 66 endomicroscopy videos of IPMNs, 291,045 frames were analyzed, with 15.5% showing papillary or vascular structures and 84.5% lacking structures. Four classification (pattern recognition) and segmentation (image division) models were tested to evaluate binary detection outcomes (papillae vs other) and ternary detection outcomes (papillae, vascularity, and nonstructure). The classification model DINOv2-ViT-G outperformed all others for the binary outcome, achieving an area under the curve of 0.942, sensitivity of 80.6%, specificity of 90.6%, and accuracy of 89.3%. For the ternary outcome, only classification models were used because labeling areas for segmentation models to detect vascularity was impractical. DINOv2-ViT-G similarly demonstrated the best performance, with sensitivities for detecting papillae, vascularity, and nonstructure of 81.7%, 82.0%, and 80.5%, respectively. The DINOv2-ViT-G model reduced nCLE video duration to 1.85 minutes of high-yield, structure-containing segments, saving 4.27 minutes (70%) per IPMN case (P < 0.001). CONCLUSION: Optimized AI models for structure identification enhance the clinical utility of nCLE by generating high-yield diagnostic segments, ensuring consistent and accurate interpretation, reducing manual effort, and enabling the development of fully autonomous systems in the future.