Validation and optimization of enhanced volumetric laser endomicroscopy scoring systems for Barrett's esophagus dysplasia

Abstract

BackgroundVolumetric laser endomicroscopy (VLE) is an advanced imaging modality used in Barrett's esophagus (BE) surveillance. VLE interpretation relies on identification of specific features associated with BE dysplasia and can be enhanced by intelligent real-time image segmentation (IRIS). The diagnostic performance of established VLE interpretation systems has not been validated. This study uses a dataset of histology-matched regions of interest (ROI) to determine the diagnostic performance of VLE with and without IRIS enhancement. MethodsVLE scans with ROIs laser-marked for targeted biopsy and matching histopathology were performed in patients undergoing BE surveillance. ROIs were rated independently for BE dysplasia by 3 blinded VLE experts using established VLE interpretation systems. Following a washout period, ROIs were rated using IRIS adapted to the VLE interpretation systems. Inter-rater agreement, diagnostic performance, and optimal diagnostic thresholds were calculated for each scoring system. ResultsA total of 71 laser-marked ROIs from 44 patients (mean age: 68 years) and BE length of 5 cm were included. About 45.1% of ROIs contained dysplasia. Inter-rater agreement was uniformly very good or excellent and only the surface intensity scoring agreement was enhanced by IRIS. The scoring systems’ AUC ranged from 0.59 to 0.74 and 0.60 to 0.75 for VLE with and without IRIS, respectively. When diagnostic thresholds of the existing scoring systems were optimized, the adjusted AUC ranged from 0.70 to 0.78. ConclusionsThis is the first study to validate and optimize the diagnostic performance of established VLE interpretation systems for BE dysplasia. IRIS image enhancement adapted to these interpretation systems demonstrated similar diagnostic performance.