Time-Lapse Deep Learning for Single-Cell Subcellular Structural Phenotypic Antimicrobial Susceptibility Testing

Antimicrobial resistance (AMR) is a global health concern that complicates the effective treatment of infections, resulting in an increased severity of illness and elevated healthcare costs. Traditional phenotypic antimicrobial susceptibility testing (AST) relies on laborious culturing and interpretation of visible growth, resulting in delays ranging from 24 h to several days. Genotypic assays detect only known resistance genes and cannot anticipate novel or emerging variants. Consequently, there is an urgent need for rapid, accurate AST methods that minimize culture time and increase detection resolution. We developed a rapid phenotypic AST platform that integrates structured illumination microscopy (SIM) imaging and deep learning to assess subcellular phenotypes in bacteria treated with antibiotics. Seven deep learning architectures, including C3D, DenseNet-121, MobileNet-V2, MobileNet-V3 Large, ResNet-50, ResNet-101, and MobileNet-V3 Small were trained on phenotypic image data sets. ResNet-50 achieved optimal performance, delivering AST results with 87% accuracy in under 20 min for <i>E. coli</i>, 4 h for <i>M. smegmatis</i>, and 15 h for BCG, all in strong agreement with conventional assays. We applied single-cell analysis at antibiotic concentrations near the minimum inhibitory concentration (MIC) and found that some cells were inhibited below the population MIC while others remained viable above it, revealing heterogeneity masked by conventional AST. Our method enables subcellular-level rapid phenotypic AST, with no need for culture requirements, and is suitable for assessing the fast effectiveness of antibiotics. Observing single-cell heterogeneity provides a tool for elucidating resistance mechanisms and informing timely clinical decision-making with the potential to curb the spread of AMR.