Christian Walter (Heidelberg / DE), Christoph Weissert (St. Gallen / CH), Eve Gizewski (Altlussheim / DE), Irene Burckhardt (Heidelberg / DE), Heiko Mannsperger (Altlussheim / DE), Siegfried Hänselmann (Altlussheim / DE), Winfried Busch (Altlussheim / DE), Oliver Nolte (St. Gallen / CH), Stefan Zimmermann (Heidelberg / DE)
Background: Bloodstream infections are an important cause of severe morbidity and mortality. Manual microscopy of Gram stains from positive blood cultures (PBC) is pivotal in diagnosing bloodstream infections but remains labor-intensive, time-consuming, and subjective. The objective of the present study was to evaluate the potential of a scan and analysis system that combines fully automated digital microscopy with deep convolutional neural networks (CNN) to assist Gram stain interpretation from PBCs for routine laboratory use.
Methods: The CNN was trained to categorize images of Gram stains based on staining and morphology into seven different classes: negative/false-positive, gram-positive cocci in pairs (GPCP), gram-positive cocci in clusters (GPCCL), gram-positive cocci in chains (GPCC), rod-shaped bacilli (RSB), yeasts, and polymicrobial specimens (FIG1 exemplary image of the graphical user interface). 1555 Gram-stained slides of PBCs were scanned, pre-classified, and reviewed by medical professionals. The assisted Gram stain results were compared to manual microscopy as well as cultural species identification by MALDI-TOF MS as the corresponding reference standard. Furthermore, site-to-site reproducibility between three sites, repeatability, and limit of detection (LOD) were determined.
Results: Comparison of assisted Gram stain interpretation to manual microscopy yielded sensitivity and specificity of 95.8% / 98.0% (GPCCL), 87.6% / 99.3% (GPCP/GPCC), 97.4% / 97.8% (RSB), 83.3% / 99.3% (yeasts) and 87.0% / 98.5% (negative/false-positive), respectively. The comparison of assisted Gram stain interpretation to MALDI-TOF MS yielded similar results as the previous comparison. In the analytical performance study, assisted interpretation showed excellent reproducibility and repeatability. With the determined LOD of 105 CFU/ml, any microorganism in PBCs should be detectable.
Conclusion: The application has demonstrated its ability to classify microorganisms based on their Gram reaction, arrangement, and morphology. It has shown potential for implementation in clinical laboratories in the future. Furthermore, digital review allows off-site image interpretation.