Fig. 1.

Clinical relevance of H. pylori and illustration of strategy to build a model detecting H. pylori. a Illustration of the diagnostic spectrum for type B-gastritis (bacterial gastritis, or H. pylori -gastritis) linked to H. pylori infection. While endoscopic evaluation of the stomach remains, and consequently harvesting gastric biopsies, other diagnostic tests, such as H2 breathing test and serological testing for H. pylori can be applied but may not differentiate for an active H. pylori infection. The tissue of gastric biopsies can histologically be reviewed, but also further tests can be applied, such as Immunohistochemistry (IHC) and Polymerase chain reaction (PCR), which are more sensitive. b Schematic representation of the approach to build a H. pylori classifier. Initially, areas within gastric biopsies of H. pylori presence were extracted (H. pylori hot spots, circled with green). Then, these hot spots were annotated according to their presence or absence of H. pylori, following a training step of an initial model. To further improve the detection sensitivity and specificity, this step was repeated for several times to generate a larger training dataset. The final model was trained containing several thousand H. pylori hot spots. Lastly, data was augmented by using color augmentation