Effect of Helicobacter pylori infection on deep learning-assisted detection of gastric neoplastic lesions under white...

Artificial intelligence (AI)-assisted endoscopy facilitates upper gastrointestinal lesion detection. Whether Helicobacter pylori (H. pylori) infection influences its diagnostic performance remains unclear. This study evaluated the effect of H. pylori infection on an AI model's accuracy for diagnosing gastric neoplasms.

A deep convolutional neural network-based AI system was evaluated for gastric neoplasm detection (low-grade intraepithelial neoplasia (LGIN), high-grade intraepithelial neoplasia (HGIN), and early gastric cancer (EGC)) in a retrospective cohort study. White light endoscopy (WLE) images (n = 2347) were collected from 563 patients who underwent imaging from November 2019 to August 2024 at Taizhou Hospital of Zhejiang Province to assess H. pylori infection's impact on diagnostic performance. Additional WLE images (n = 447) from 117 patients (September 2024-June 2025) were used to compare the AI system's performance with that of expert and non-expert endoscopists.

The AI system achieved 85.0% accuracy, 82.0% sensitivity, 87.6% specificity, 85.2% positive predictive value (PPV), and 84.8% negative predictive value (NPV). The accuracy (87.1% vs. 80.2%), specificity (89.9% vs. 76.6%), and NPV (89.7% vs. 65.4%) were significantly higher in the H. pylori-negative group than in the H. pylori-positive group (all P < 0.001), whereas the PPV was lower (82.7% vs. 88.5%, P = 0.008), with a comparable sensitivity (82.3% vs. 81.6%, P = 0.790). Within the H. pylori-negative cohort, further stratification into never-infected and eradicated subgroups showed that the eradicated group had significantly higher accuracy, sensitivity, and NPV than the H. pylori-positive group (all P < 0.05). It exhibited significantly higher accuracy for detecting non-neoplastic lesions and LGIN in the H. pylori-negative group (P < 0.05), but not for HGIN or EGC (P > 0.05). Its diagnostic accuracy was comparable to the expert endoscopists' (84.8% vs. 81.9%, P = 0.247) and significantly higher than the non-expert endoscopists' (84.8% vs. 72.1%, P < 0.001).

This AI system exhibited excellent performance for gastric neoplasm detection, which was significantly affected by H. pylori infection (particularly for non-neoplastic lesions and LGIN). Eradication of H. pylori appeared to restore the diagnostic performance of the AI system. Its diagnostic accuracy in still image classification was comparable to expert endoscopists and superior to non-experts, supporting its potential as an adjunctive clinical endoscopy tool.