Label-free Pathogen Identification with Microscopy Imaging and Deep Learning

Rapid and accurate pathogen identification is crucial for the clinical management of infectious diseases, particularly sepsis and severe respiratory infections, yet standard clinical workflows remain slow and resource-intensive. Here, we developed an automated, high-throughput imaging platform built on standard, clinically accessible bright-field microscopy, and generated a large dataset comprising 24.9 million label-free bacterial cells across six focal pathogens. Leveraging this resource, we trained a neural network (ESKAPe-ResNet) to identify ESKAPe species at the single-bacterium level. The model achieved >92% accuracy in species-level classification and >82% accuracy in quantifying ESKAPe abundance in mock mixtures, with high specificity against non-ESKAPe bacteria. In clinical validation using sputum, bronchoalveolar lavage fluid and blood samples from patients with respiratory infections and sepsis, the approach correctly identified the dominant ESKAPe pathogen in >78% of samples after minimum broth culture enrichment. The imaging-to-identification pipeline was completed in under 10 minutes, and coupled with brief cultivation, the median time to accurate identification was reduced to 5-6 hours, compared with days for conventional blood culture-based workflows. This work establishes the proof-of-principle for label-free, hardware-minimal rapid pathogen identification, providing a clinically deployable workflow to expedite diagnosis and reduce mortality in severe bacterial infections.