This work addresses the detection of Helicobacter pylori, a bacterium classified since 1994 as class 1 carcinogen to humans. Due to its high specificity and sensitivity, the preferred diagnosis technique is the analysis of histological images with immunohistochemical staining, a process in which certain stained antibodies bind to antigens of the biological element of interest. This analysis is a time consuming task, which is currently done by an expert pathologist that visually inspects the digitized images.

We propose the use of a Convolutional Neural Network (CNN) at sample level, in conjunction with a simple classifier to determine the final diagnosis at patient level. The designed CNN architecture is able to discern intricate patterns at different levels, particularly focusing on the distinctive color and structure that H. pylori displays in the samples. We have tested our model using a cross-validation on a set of patients with annotated patches, at sample and diagnosis level, and assessed reproducibility of diagnosis prediction on an independent set of patients. In the 5-fold cross-validation, our CNN model has an overall 92% +/- 3% of accuracy and 91% +/- 6% sensitivity in classification of Helicobacter pylori at sample level, and an accuracy of 92% +/- 2%, with sensitivity of 92% +/- 5% at patient diagnosis level. In the independent set an accuracy of 84%, with a sensitivity of 84% at patient diagnosis level validates the reproducibility of results.