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

Area of Science:

• Microbiology and Infectious Diseases
• Biotechnology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Antimicrobial resistance (AMR) poses a significant global health threat, increasing illness severity and healthcare costs.
• Traditional antimicrobial susceptibility testing (AST) methods are slow (24 hours to days) and rely on culturing.
• Genotypic assays are limited to known resistance genes and cannot detect novel variants.

Purpose of the Study:

• To develop a rapid, accurate phenotypic AST platform to overcome limitations of current methods.
• To utilize structured illumination microscopy (SIM) and deep learning for subcellular bacterial phenotype assessment.
• To enable faster and more resolved AST for timely clinical decision-making and AMR containment.

Main Methods:

• Development of a rapid phenotypic AST platform integrating SIM imaging and deep learning.
• Training of seven deep learning architectures (ResNet-50, C3D, DenseNet-121, etc.) on bacterial image datasets.
• Application of single-cell analysis near the minimum inhibitory concentration (MIC) to assess bacterial response.

Main Results:

• ResNet-50 achieved 87% accuracy in AST results in under 20 minutes for E. coli, 4 hours for M. smegmatis, and 15 hours for BCG.
• The deep learning platform demonstrated strong agreement with conventional AST assays.
• Single-cell analysis revealed heterogeneity in bacterial response to antibiotics, masking population-level MIC.

Conclusions:

• The developed method enables rapid, subcellular-level phenotypic AST without culture requirements.
• This platform accurately assesses antibiotic effectiveness and reveals single-cell heterogeneity, crucial for understanding resistance.
• The findings support timely clinical decision-making and offer a tool to combat the spread of AMR.