Quantitative analysis of high-resolution microendoscopic images for diagnosis of neoplasia in patients with Barrett's esophagus

Abstract

Background and aims: Previous studies show that microendoscopic images can be interpreted visually to identify the presence of neoplasia in patients with Barrett's esophagus (BE), but this approach is subjective and requires clinical expertise. This study describes an approach for quantitative image analysis of microendoscopic images to identify neoplastic lesions in patients with BE.

Methods: Images were acquired from 230 sites from 58 patients by using a fiberoptic high-resolution microendoscope during standard endoscopic procedures. Images were analyzed by a fully automated image processing algorithm, which automatically selected a region of interest and calculated quantitative image features. Image features were used to develop an algorithm to identify the presence of neoplasia; results were compared with a histopathology diagnosis.

Results: A sequential classification algorithm that used image features related to glandular and cellular morphology resulted in a sensitivity of 84% and a specificity of 85%. Applying the algorithm to an independent validation set resulted in a sensitivity of 88% and a specificity of 85%.

Conclusions: This pilot study demonstrates that automated analysis of microendoscopic images can provide an objective, quantitative framework to assist clinicians in evaluating esophageal lesions from patients with BE. (

Clinical trial registration number: NCT01384227 and NCT02018367.).

Conflict of interest statement

Competing interests

None reported.

Copyright © 2016 The Authors. Published by Elsevier Inc. All rights reserved.

Figures

Figure 1

Image analysis procedure: (A) A circular ROI is selected. (B) Fiber pattern is removed using Gaussian filtering. (C) The contrast of an image is enhanced using adaptive histogram equalization. (D) Nuclei are segmented. The size distribution of glandular features is determined by using granulometry. (E) Quantitative image features are calculated.

Figure 2

Flowchart for visual classification. Scale bars represent 100 μm.

Figure 3

Resulting classification trees of (A) the training set and (B) the validation set. A bar graph on top indicates the total number of images in the data set. Bar graphs on bottom indicate the number of images classified as one of the categories.