352 Volumetric Laser Endomicroscopy Feature Discovery and Diagnostic Performance Optimization for Barrett’s Esophagus Dysplasia Using Laser-Targeted Biopsies With Histopathologic Correlation

Abstract

INTRODUCTION: Volumetric laser endomicroscopy (VLE) is a high-resolution imaging modality used in Barrett’s esophagus (BE) surveillance. Established VLE scoring systems use a combination of features to guide BE dysplasia diagnosis but it remains unclear whether these features are sufficient to reach optimal VLE diagnostic performance. This study explores an exhaustive list of VLE features using one-to-one imaging to pathology correlation to develop optimal VLE diagnostic models for BE dysplasia. METHODS: In this prospective, multi-center (5 sites) study, patients with dysplastic BE underwent VLE with laser targeted regions of interest (ROIs), followed by standard surveillance endoscopy and biopsy directed to laser-marked sites. Histology was reviewed by two GI pathologists (MV, RO) who achieved a consensus diagnosis (gold-standard). ROIs were dysplastic if low-grade dysplasia (LGD), high-grade dysplasia (HGD) or intramucosal adenocarcinoma (IMC) was present. Two VLE experts (CLL, GJT) reviewed ROIs and corresponding endoscopic images to identify and define a set of VLE features associated with BE dysplasia. A multi-variable logistic regression model based on least absolute shrinkage and selection operator (LASSO) method was developed to determine the optimal diagnostic performance of these features. A decision tree model with high clinical usability also was developed to evaluate VLE diagnostic performance for real-time image interpretation. RESULTS: A total of 461 VLE ROIs with matching histology were obtained from 150 patients (Table 1). A final dataset of 55 neoplastic (LGD, 12; HGD, 24; IMC, 19) and 56 randomly selected non-dysplastic ROIs were reviewed. Twenty-eight VLE features were identified describing layering, signal intensity, surface topography, and glandular characteristics. LASSO logistic regression modeling identified 5 VLE features which when utilized optimized diagnostic performance with an area under the curve (AUC) for BE dysplasia of 0.95 (Figure 1). To improve clinical usability, a decision tree model was optimized to three levels with an AUC of 0.90 (Figure 2). CONCLUSION: This study represents the most comprehensive review of VLE features associated with BE dysplasia to date. The proposed high-performing diagnostic models suggest that VLE can be used efficiently and effectively to enhance BE surveillance by identifying ROIs for targeted biopsy. Further validation of the proposed VLE models will determine their clinical utility.