Identification of Barrett's Esophagus in Endoscopic Images Using Deep Learning

Wen Pan, Xujia Li, Weijia Wang, Linjing Zhou, Jiali Wu, Tao Ren, Chao Liu, Muhan Lv, Song Su, Yong TangResearch 2021 · DOI

Abstract

Development of a deep learning method to identify Barrett's esophagus (BE) scopes in endoscopic images. 443 endoscopic images from 187 patients of BE were included in this study. The gastroesophageal junction (GEJ) and squamous-columnar junction (SCJ) of BE were manually annotated in endoscopic images by experts. Fully convolutional neural networks (FCN) were developed to automatically identify the BE scopes in endoscopic images. The networks were trained and evaluated in two separate image sets. The performance of segmentation was evaluated by intersection over union (IOU). The deep learning method was proved to be satisfying in the automated identification of BE in endoscopic images. The values of the IOU were 0.56 (GEJ) and 0.82 (SCJ), respectively. Deep learning algorithm is promising with accuracies of concordance with manual human assessment in segmentation of the BE scope in endoscopic images. This automated recognition method helps clinicians to locate and recognize the scopes of BE in endoscopic examinations.

Introduction

Barrett’s Esophagus (BE) is a precancerous condition where the normal lining of the esophagus is replaced by tissue similar to the intestinal lining. It is the primary precursor to esophageal adenocarcinoma (EAC), a cancer with a low five-year survival rate.

The clinical challenge lies in the "Prague classification"—the standard method for measuring the extent of BE. To do this accurately, endoscopists must identify two specific anatomical landmarks: the gastroesophageal junction (GEJ) and the squamous-columnar junction (SCJ). However, these junctions vary wildly in shape and appearance between patients. Identifying them is currently a subjective process, leading to potential misjudgments and inconsistent biopsies. This paper seeks to provide an objective, automated deep learning approach to standardize these measurements.

Methods

The authors employed Fully Convolutional Networks (FCN), an architecture specifically designed for semantic segmentation. Key technical aspects include:

  • Architecture: The model uses multiple convolutional layers to extract abstract features, followed by deconvolutional layers to upsample the data back to the original image size.
  • Skip Architectures: Fusing deep and shallow layers allowed the model to capture both high-level semantic information and fine-grained spatial details.
  • Independent Network Strategy: The researchers found that training two independent networks (one for GEJ, one for SCJ) significantly outperformed a single-network model.

Datasets

The study utilized a retrospective dataset of 443 endoscopic images collected from 187 patients. To ensure high-quality data, two senior endoscopists manually annotated the GEJ and SCJ boundaries. The data was split by patient to prevent the model from simply memorizing specific cases.

Distribution of patients and images across the training and test datasets.

Key Results

The performance was measured using Intersection over Union (IoU) and Dice Similarity Coefficient (DSC). The model demonstrated high reliability, particularly for the SCJ boundary.

Performance metrics (IoU and DSC) for the segmentation of GEJ and SCJ landmarks.

The high DSC of 0.90 for the SCJ indicates a very strong overlap with expert annotations. While the GEJ proved more difficult to identify (IoU 0.56), expert review concluded that the predicted shapes were still clinically useful for locating the junctions.

Discussion and Conclusion

The study reveals an "anatomical difficulty gap," where the visual transition at the SCJ is much more distinct for AI than the GEJ, which can be obscured by tissue folds. A major takeaway is that specialized, independent networks are more effective than general multi-class approaches in complex anatomical environments.

While promising, the study is currently limited by its single-center data and the lack of differential diagnosis between BE and other lesions. Future work will focus on validating the model against external cohorts and integrating the tool into real-time endoscopic surveillance to assist doctors in performing more accurate, targeted biopsies.

Summary updated

Abstract

Abstract Background: Development of a deep learning method to identify Barrett's esophagus (BE) scopes in endoscopic images. Methods: 443 endoscopic images from 187 patients of BE were included in this study. The gastroesophageal junction (GEJ) and squamous-columnar junction (SCJ) of BE were manually annotated in endoscopic images by experts. Fully convolutional neural networks (FCN) were developed to automatically identify the BE scopes in endoscopic images. The networks were trained and evaluated in two separate image sets. The performance of segmentation was evaluated by intersection over union (IOU). Results: The deep learning method was proved to be satisfying in the automated identification of BE in endoscopic images. The values of the IOU were 0.56 (GEJ) and 0.82 (SCJ), respectively. Conclusions: Deep learning algorithm is promising with accuracies of concordance with manual human assessment in segmentation of the BE scope in endoscopic images. This automated recognition method helps clinicians to locate and recognize the scopes of BE in endoscopic examinations.
BibTeX
@inproceedings{pan2021identification,
  author = {Wen Pan and Xujia Li and Weijia Wang and Linjing Zhou and Jiali Wu and Tao Ren and Chao Liu and Muhan Lv and Song Su and Yong Tang},
  title = {{Identification of Barrett's Esophagus in Endoscopic Images Using Deep Learning}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-698574/v1},
}
Figure