Deep Neural Network-Based Algorithm Successful at Detecting Helicobacter pylori Infection

A deep learning-based decision support algorithm was found to successfully stratify biopsy images for Helicobacter pylori infection, according to results of a study published in BMC Gastroenterology.

Researchers at University Hospital Cologne in Germany digitized stained whole slide images of gastric biopsy specimens. H pylori “hotspots” (containing hundreds to thousands of H pylori patch candidates) or regions devoid of H pylori were identified by image processing, and images were cropped to a standardized size. These images (Giemsa: n=286; H&E: n=191), primarily of tissue from the antrum, were used to train a deep neural network and similar images (Giemsa: n=87; hematoxylin-eosin (H&E): n=71) were used as a validation cohort to assess the model.

Compared with microscopic assessment of the same images, the machine learning algorithm correctly identified H pylori infection with an area under the curve (AUC) of 0.92 for biopsies stained with Giemsa and 0.81 for H&E-stained biopsies.

Among the Giemsa-stained biopsies, the machine learning algorithm in general identified more positives (true: 19 vs 13; false: 36 vs 5) and fewer negatives (true: 32 vs 63; false: 0 vs 6) compared with the microscope-based assessment. When comparing between methods, these values corresponded with a higher sensitivity for the machine learning algorithm (100% vs 68.4%) but lower specificity (47.1% vs 92.6%).

The algorithm performance was compared with the ground-truth by testing biopsy specimens that were ambiguously classified by the 2 methods by a polymerase chain reaction analysis for evidence of infection. These results tended to coincide with the machine-learning algorithm findings, in which 30% of the cases identified as negative by the clinician and positive by the algorithm had genetic evidence of H pylori, increasing the AUC for the Giemsa-stained biopsies to 0.95.

This study may have been limited by the choice to pre-process images and select H pylori hotspots. Although this was done to reduce the computational burden, it may have hindered an easy translation of the algorithm into the clinical setting.

These data indicated that microscope-based assessment of gastric biopsy specimens for H pylori may be underpowered compared with a deep learning-based algorithm, and that more sensitive H pylori screening methods are likely needed.

Disclosure: An author declared affiliations with industry. Please refer to the original article for a full list of disclosures.

Reference

Klein S, Gildenblat J, Ihle MA, et al. Deep learning for sensitive detection of Helicobacter Pylori in gastric biopsies. BMC Gastroenterol. 2020;20(1):417. doi:10.1186/s12876-020-01494-7