An explainable artificial intelligence system for diagnosing Helicobacter Pylori infection under endoscopy: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Gastric mucosal changes from Helicobacter pylori (H. pylori) infection can obscure early gastric cancer detection during endoscopy.
• Computer-aided diagnosis (CAD) systems show promise for H. pylori detection, but their explainability remains a challenge.

Purpose of the Study:

• To develop an explainable artificial intelligence system for diagnosing H. pylori infection (EADHI) directly from endoscopic images.
• To provide a diagnostic basis for H. pylori infection under endoscopy.

Main Methods:

• A case-control study involving 47,239 endoscopic images from 1826 patients.
• Development of EADHI using ResNet-50 for feature extraction and long short-term memory networks for classification.
• Inclusion of nine endoscopic features and analysis of feature contributions using a gradient-boosting decision tree model.

Main Results:

• EADHI achieved an overall accuracy of 78.3% in feature extraction.
• Internal testing showed EADHI's diagnostic accuracy for H. pylori infection at 91.1%, significantly higher than endoscopists (15.5% increase).
• External testing demonstrated EADHI's robustness with 91.9% accuracy; mucosal edema and regular collecting venules were key diagnostic features.

Conclusions:

• EADHI accurately diagnoses H. pylori gastritis with high explainability, potentially increasing endoscopist trust and acceptance of CAD systems.
• The system's ability to identify diagnostic features enhances its clinical utility.
• Future multicenter, prospective studies are recommended to validate clinical applicability.