Deep Learning for Anatomical Site Classification in Upper GI Endoscopy

Overview

This study presents a novel workflow using deep learning to classify anatomical sites in upper gastrointestinal endoscopy images. A new annotated dataset of 3704 images from routine clinical care in China was developed based on a modified guideline balancing British and Japanese standards, demonstrating feasibility for improving photo-documentation quality.

Background

Oesophagogastroduodenoscopy (EGD) is the gold standard for diagnosing upper GI diseases, including early gastric cancer, which is often missed due to blind spots. Quality photo-documentation of anatomical landmarks is a key quality indicator in EGD. Artificial intelligence, particularly deep learning, offers potential to enhance endoscopic examination completeness and lesion detection by automating anatomical site classification and improving procedural analysis.

Data Highlights

The dataset comprises 3704 EGD images collected from Tianjin Medical University General Hospital, China, annotated by clinical experts following a modified guideline merging British and Japanese standards. Images vary in resolution and region of interest, with an automatic multi-resolution ROI extraction method applied prior to classification by a convolutional neural network (CNN).

Key Findings

• A modified guideline for upper GI endoscopic photo-documentation was proposed, balancing detailed Japanese SSS and pragmatic British standards.
• A new annotated dataset of 3704 upper GI endoscopic images was created, addressing a gap in publicly available data for AI development.
• An automatic multi-resolution ROI extraction method was developed to standardize image inputs from various clinical systems.
• A complete workflow for EGD image classification using deep learning was introduced, encompassing data collection, ROI extraction, annotation, and CNN-based classification.
• The approach demonstrated feasibility for anatomical site classification, supporting improved quality assurance in endoscopic photo-documentation.

Clinical Implications

Implementing AI-based anatomical site classification can enhance the completeness and quality of upper GI endoscopic examinations by ensuring standardized photo-documentation. This may reduce missed lesions and improve early detection of gastric diseases, particularly in regions with high gastric cancer incidence. The proposed workflow and dataset provide a foundation for integrating AI tools into routine clinical practice.

Conclusion

This study establishes a clinically relevant AI workflow for anatomical site classification in upper GI endoscopy, supported by a novel annotated dataset and a balanced photo-documentation guideline. These advances have the potential to improve endoscopic quality assurance and diagnostic accuracy.

References

1. Yao et al. 2013 -- Systematic Screening Protocol for the Stomach (SSS)
2. European Society of Gastrointestinal Endoscopy (ESGE) 2001 -- Standardised Image Documentation in Upper GI Endoscopy
3. GLOBOCAN 2018 -- Global Cancer Incidence Data