Automated diagnosis of Helicobacter pylori infection from routine endoscopic images using deep learning: a development and validation study

BACKGROUND: Real-time endoscopic diagnosis of Helicobacter pylori infection remains challenging and often requires biopsy-based testing, delaying treatment decisions. Although deep learning (DL) approaches have shown promise, most prior studies have relied on retrospective datasets or image-enhanced modalities, limiting their applicability to routine white-light endoscopy. We aimed to develop and prospectively validate a research-stage DL model for H. pylori detection using only standard white-light endoscopic images. METHODS: In this single-center prospective study, consecutive adults undergoing diagnostic gastroscopy were enrolled. Six standardized gastric images per patient were targeted under standard white-light endoscopic imaging. Histopathology from antral and corpus biopsies served as the reference standard. An EfficientNet-B0-based DL model was developed to classify H. pylori infection at the patient level by aggregating image-level predictions. Model performance was assessed using five-fold cross-validation within the development cohort, followed by evaluation in an independent temporally separated validation cohort (70% development / 30% temporal validation). RESULTS: A total of 172 patients (1,000 images) were included; 94 patients (54.7%) were H. pylori-positive. In five-fold cross-validation, the model achieved a patient-level AUC of 0.901 (95% CI: 0.863-0.936), with 85.1% sensitivity and 81.4% specificity. In the independent temporal validation cohort (n = 52; prevalence 48.1%), the AUC was 0.889 (95% CI: 0.793-0.960), with 84.0% sensitivity and 85.2% specificity. Aggregating predictions across multiple gastric views improved discrimination compared with single-image inference. CONCLUSION: In this prospective study, a deep learning model using routine white-light endoscopic images demonstrated reasonable patient-level discrimination for H. pylori detection, including in an independent temporally separated validation cohort. At present, the model should be viewed as a research and decision-support tool rather than a standalone diagnostic system. Multicenter external validation and prospective video-based studies are warranted before routine clinical deployment.