Integrating deep learning and handcrafted radiomics features for Helicobacter pylori infection identification on multi‑region gastroscopic images: a comparative study

Helicobacter pylori ( H. pylori ) infection is a major pathogenic factor of chronic gastritis, peptic ulcer, and gastric cancer. Early non-invasive screening is clinically important. Conventional gastroscopic diagnosis relies heavily on endoscopists’ experience with high subjectivity. This study aimed to develop a fusion model of deep learning and radiomics based on gastroscopy images for intelligent, objective, and accurate identification of H. pylori infection. A total of 6235 gastroscopic images from 151 patients at Longgang District People’s Hospital of Shenzhen were retrospectively collected. Datasets were constructed according to the gastric body, fundus, antrum, angle, and a combined four-region set. Five CNN models, including AlexNet and DenseNet121, were used with transfer learning and fivefold cross-validation. A dual-branch deep learning–radiomics fusion framework was established, and performance was evaluated by accuracy, AUC, and other metrics. ResNet50 and DenseNet121 were the best-performing single models, with ResNet50 reaching an AUC of 0.8563 in the gastric body. Integrating radiomics features improved classification performance in some regions, although the benefit was region-dependent; the DenseNet121 fusion model achieved the highest fusion AUC of 0.8184 in the gastric fundus. Despite imperfectly balanced single-region sample sizes, the combined four-region dataset achieved performance comparable to that of the best single-region subset, supporting the model’s robustness to moderate inter-region heterogeneity. In an independent external cohort of 50 patients, the gastric body and fundus remained the best-performing regions (highest AUCs: 0.7836 for single models and 0.7085 for fusion models), although overall external performance was modest and requires confirmation in larger multicentre studies. The combination of deep learning and radiomics effectively improves the diagnostic performance of H. pylori infection using gastroscopy images. The gastric body and fundus are the most valuable regions. This non-invasive and interpretable approach can provide real-time intelligent assistance for routine gastroscopy and shows promising clinical application prospects.