Precision bacterial quantification via dual-enhanced U-net: Addressing division artifacts and spatial bias in...

Bacterial quantification is fundamental to biomedical research and clinical diagnostics. Traditional methods, such as plate counting and optical density (OD) measurement, are limited by poor timeliness and lack of scalability for high-throughput applications. Although neural network (NN)-aided analysis of optical microscopy images enables rapid bacterial enumeration, its counting accuracy remains limited. Here, we developed an enhanced U-Net-based framework that included two key methodological innovations: (1) an area normalization method (ANM) to address the issue of incomplete bacterial separation during division, and (2) a multi-view averaging strategy (MVAS) to compensate for uneven spatial distribution-both designed to improve counting accuracy. The optimized U-Net model achieved a pixel-wise segmentation accuracy of above 98% and a Dice similarity coefficient (DSC) of above 83% for Escherichia coli. The implementation of our ANM reduced the relative error (RE) of bacterial counts from approximately 50% to below 7% compared to baseline neural network performance. Validation of the dual-strategy approach (incorporating both ANM and MVAS) against conventional OD measurements showed a strong linear correlation (R² = 0.973). When monitoring bacterial proliferation dynamics, the method demonstrated robust adherence to the exponential growth model (R² = 0.951), accurately capturing binary fission characteristics. In conclusion, this study establishes a method for high-accuracy, high-throughput bacterial quantification, with immediate applications in real-time microbial monitoring and scalable diagnostic workflows.