Rapid and Efficient Screening of Helicobacter pylori in Gastric Samples Stained with Warthin-Starry Using Deep Learning.

Background/

Objectives: Helicobacter pylori is a major risk factor for gastric cancer. The incidence and prevalence of the pathogen are increasing worldwide, urging novel approaches to reduce detection turnaround times. H. pylori diagnosis relies on histological examination of gastric biopsies, but interobserver variability considerably impacts its identification. We present an algorithm combining a feature pyramid network and a ResNet architecture for automatic and rapid H. pylori detection in digitized Warthin-Starry-stained gastric biopsies.

Methods: Whole-slide images were segmented into manually annotated smaller patches and segments containing stomach tissue were analyzed for the presence of Gram-negative bacteria. Patches classified as positive were examined to confirm the presence/absence of bacteria in contact with the gastric epithelial surface (H. pylori).

Results: The algorithm exhibited 0.923 average precision and 0.982 average recall. The conducted efficiency study demonstrated that algorithm utilization significantly decreased (p < 0.001) diagnostic turnaround times for all participants (two pathologists, a pathology resident, a pathology technician, and a biotechnologist), observing an 88.13-91.76% time reduction. Implementation of the algorithm also improved diagnostic accuracy for the resident, technician, and biotechnologist, indicating that the tool remarkably supports less experienced personnel.

Conclusions: We believe that the incorporation of our algorithm into pathology workflows will help standardize diagnostic protocols and drastically reduce H. pylori diagnostic turnaround times.