Evaluation of quantitative image analysis criteria for the high-resolution microendoscopic detection of neoplasia in Barrett's esophagus

Abstract

Early detection of neoplasia in patients with Barrett's esophagus is essential to improve outcomes. The aim of this ex vivo study was to evaluate the ability of high-resolution microendoscopic imaging and quantitative image analysis to identify neoplastic lesions in patients with Barrett's esophagus. Nine patients with pathologically confirmed Barrett's esophagus underwent endoscopic examination with biopsies or endoscopic mucosal resection. Resected fresh tissue was imaged with fiber bundle microendoscopy; images were analyzed by visual interpretation or by quantitative image analysis to predict whether the imaged sites were non-neoplastic or neoplastic. The best performing pair of quantitative features were chosen based on their ability to correctly classify the data into the two groups. Predictions were compared to the gold standard of histopathology. Subjective analysis of the images by expert clinicians achieved average sensitivity and specificity of 87% and 61%, respectively. The best performing quantitative classification algorithm relied on two image textural features and achieved a sensitivity and specificity of 87% and 85%, respectively. This ex vivo pilot trial demonstrates that quantitative analysis of images obtained with a simple microendoscope system can distinguish neoplasia in Barrett's esophagus with good sensitivity and specificity when compared to histopathology and to subjective image interpretation.

Figure 1

Illustration of HRME images (top row) compared to H&E images (bottom row). From left to right, diagnostic categories are: Barrett’s metaplasia and low-grade dysplasia (LGD), high-grade dysplasia (HGD), and esophageal adenocarcinoma (EAC). All images are sized to the same scale; scale bar represents 100 μm.

Figure 2

(a) Box and whiskers plot showing the GLCM correlation by diagnostic category. The central line in each box represents the mean, while the top and bottom edges represent the 25th and 75th percentile. The notches in the boxes represent a 95% confidence interval for the mean. (b) Features values for top two performing features by measurement site. Non-neoplastic sites are plotted as squares; neoplastic sites are plotted as crosses. The decision line is shown.

Figure 3

Classification performance. (a) Calculated classification performance versus number of features used; note the plateau at two features. (b) Scatter plot of calculated posterior probability generated by a two-feature linear discriminant analysis (LDA) algorithm. The decision line is shown at 0.24. (c) ROC curve of calculated classification performance using the two-class, two-feature LDA. The Q-point is shown, corresponding to a sensitivity of 87% and a specificity of 85%.