Rapid Bacterial Detection in Urine Using Laser Scattering and Deep Learning Analysis

Abstract

Images of laser scattering patterns generated by bacteria in urine are promising resources for deep learning. However, floating bacteria in urine produce dynamic scattering patterns and require deep learning of spatial and temporal features. We hypothesized that bacteria with variable bacterial densities and different Gram staining reactions would generate different speckle images. After deep learning of speckle patterns generated by various densities of bacteria in artificial urine, we validated the model in an independent set of clinical urine samples in a tertiary hospital. Even at a low bacterial density cutoff (1,000 CFU/mL), the model achieved a predictive accuracy of 90.9% for positive urine culture. At a cutoff of 50,000 CFU/mL, it showed a better accuracy of 98.5%. The model achieved satisfactory accuracy at both cutoff levels for predicting the Gram staining reaction. Considering only 30 min of analysis, our method appears as a new screening tool for predicting the presence of bacteria before urine culture. IMPORTANCE This study performed deep learning of multiple laser scattering patterns by the bacteria in urine to predict positive urine culture. Conventional urine analyzers have limited performance in identifying bacteria in urine. This novel method showed a satisfactory accuracy taking only 30 min of analysis without conventional urine culture. It was also developed to predict the Gram staining reaction of the bacteria. It can be used as a standalone screening tool for urinary tract infection.