Deep learning model for automated detection of Helicobacter pylori and intestinal metaplasia on gastric biopsy digital whole slide images.

Objective: To develop an automated detection tool for Helicobacter pylori (HP) microorganisms (HPOrg) and intestinal metaplasia (IM) identification on gastric biopsy specimens on hematoxylin and eosin (H&E) whole-slide images (WSIs), incorporating background histopathologic features.

Methods: A total of 180 H&E gastric biopsy WSIs, archived at the Department of Anatomical Pathology, Singapore General Hospital, were used to train, validate, and test (60:20:20) a decision support tool. Eighty WSIs displayed non-HP inflammation; 100 were annotated for HP-associated gastritis, HPOrg, and IM. A 2-stage model was employed-a Vision Transformer-based model filtered artifacts after stain normalization, and then a Graph Attention Network component aggregated patch-level features, giving a prediction for each of 6 tissue sections within each WSI, with a majority vote determining the final WSI prediction.

Results: A total of 776 636 patches were used for training/validation and testing. The optimized model showed HPOrg classification (precision: 0.604, F1-score: 0.617, and top 10 micro F1-score: 0.714) and IM classification (precision: 0.905, F1-score: 0.861, and top 10 micro F1-score: 1.0). The macro average F1-score was 0.739, section-level precision was 0.981, and the F1-score was 0.945. The WSI-level precision achieved was 1.0, with a F1-score of 0.96.

Conclusions: We demonstrate a 2-stage model to detect HP and IM in gastric biopsy specimens, considering background inflammation, which more closely reflects real-world clinical diagnosis.