Rapid species identification of pathogenic bacteria from a minute quantity exploiting three-dimensional quantitative phase imaging and artificial neural network

Infections by microorganisms are a global healthcare issue associated with a large number of deaths and substantial healthcare costs. Bacteria account for approximately half of reported infections. Multiple studies indicate that administering an antibiotic treatment appropriate to the pathogen during the early hours of infection significantly reduces mortality. However, in clinical settings early antibiotic treatments are often empirical and imperfect due to the long turnaround time (>24 h) of routine microbial identification, leading to increased mortality risk. Conventional culture-based approaches are nonspecific and time-consuming. Molecular diagnostics can be faster but are not scalable to arbitrary pathogens. MALDI-TOF MS is the gold standard but typically requires sufficient sample quality after about 24 h of culture. Image-based methods like fluorescence microscopy and FISH can detect or identify bacteria but require labels and can be destructive; label-free alternatives such as autofluorescence are limited by intrinsic fluorophore specificity. This study tackles rapid microbial identification by combining three-dimensional quantitative phase imaging (QPI) to obtain refractive index (RI) tomograms of live bacteria with an artificial neural network (ANN) classifier. The approach aims to identify pathogens from single to few bacterial cells, potentially providing preliminary results early in infection and guiding antibiotic treatment before culture-based diagnostics are available.

Prior work includes conventional culture tests (simple but slow and nonspecific), molecular diagnostic methods (faster but limited in scalability to arbitrary pathogens), and MALDI-TOF MS (current gold standard that detects molecular markers but typically requires ~24 h of culture to reach sufficient sample quality). Image-based approaches have used fluorescence microscopy for detection/counting and fluorescence in situ hybridization (FISH) for screening targeted bacteria via labeled genomic patterns, but these require destructive labeling and optimized probes. Label-free imaging such as autofluorescence has been explored but offers limited specificity due to reliance on intrinsic fluorophores. Quantitative phase imaging (QPI) has been applied for quantitative cell profiling, offering non-destructive, label-free measurements.

- Imaging system: A three-dimensional quantitative phase imaging (3D QPI) system (holotomography, HT-2H, Tomocube Inc.) based on a simplified Mach-Zehnder interferometer equipped with a digital micromirror device (DMD) to scan illumination angles. Two-dimensional QPI measurements across angles are acquired to form a sinogram. The 3D refractive index (RI) tomogram is reconstructed from the sinogram using optical diffraction tomography followed by iterative regularization. - Specimens and dataset: Isolates from 19 bacterial species associated with bloodstream infections (BSIs) were measured. The species include Acinetobacter baumannii, Bacillus subtilis, Enterobacter cloacae, Enterococcus faecalis, Escherichia coli, Haemophilus influenzae, Klebsiella pneumoniae, Listeria monocytogenes, Micrococcus luteus, Proteus mirabilis, Pseudomonas aeruginosa, Serratia marcescens, Staphylococcus aureus, Staphylococcus epidermidis, Stenotrophomonas maltophilia, Streptococcus agalactiae, Streptococcus anginosus, Streptococcus pneumoniae, and Streptococcus pyogenes. A database of 10,556 3D RI tomograms was built, each containing a single bacterium or small clusters. Bacilli were typically single cells, while cocci and coccobacilli often appeared as small clusters (e.g., Streptococcus chains). - ANN architecture and training: A 3D convolutional neural network with dense connectivity was trained to classify 3D RI tomograms into one of the 19 species. The network includes: an initial 3D convolution (3×3×3 kernels, stride 2×2×2), followed by four dense blocks with transition units between them. Each dense block repeats pairs of convolutions (1×1×1 then 3×3×3, both stride 1×1×1) with concatenation of feature maps. Transition units contain 1×1×1 convolutions (stride 1×1×1) to adjust feature scales. Additional components include batch normalization, leaky ReLU activations, global average pooling, and a final fully connected layer for classification. Network parameters were optimized via gradient-based training on labeled tomograms. - Inference and evaluation: Single 3D RI tomograms were fed to the trained ANN to predict species. Blind test accuracy was assessed on held-out data. Variants replacing 3D QPI with 2D QPI or 2D sinograms, or replacing the ANN with conventional machine learning, were evaluated for comparison (details in supplementary information). The framework also allows combining multiple independent measurements (cells or clusters) from the same specimen to boost accuracy by aggregating predictions.

- Single-measurement performance: Achieved 82.5% blind test accuracy in species identification from a single 3D RI tomogram (single cell or small cluster), comparable to MALDI-TOF MS when sufficient bacterial quantity is available. - Multi-measurement improvement: Accuracy increased with multiple independent measurements, reaching 99.9% with seven different measurements of cells or clusters. - Dataset scope: Built a database of 10,556 3D RI tomograms across 19 BSI-relevant bacterial species. - Ablation/variant analysis: Performance dropped markedly when 3D QPI was replaced by 2D QPI or 2D sinograms, and when the ANN was replaced by conventional machine learning algorithms, underscoring the importance of both 3D morphological information and the ANN architecture. - Practicality: The method is label-free and non-destructive, enabling potential integration into existing clinical workflows to provide early advisory results before culture-based identification. - Top-N consideration: The correct species inclusion rate can be increased by considering multiple top candidate species (at some cost to specificity).

The study demonstrates that combining 3D QPI with a tailored 3D CNN can accurately identify bacterial species from minute quantities—down to a single cell or small clusters—well before sufficient biomass accumulates for gold-standard MALDI-TOF MS. By capturing and leveraging detailed 3D morphological features in RI tomograms, the approach achieves high single-shot accuracy and near-perfect accuracy with a small number of measurements, addressing the clinical need for rapid, early identification to guide antibiotic therapy. The non-destructive, label-free nature of QPI facilitates integration into clinical workflows without interfering with downstream culture or molecular assays. Comparative analyses indicate that both the 3D imaging modality and the ANN-based classifier are critical to performance, highlighting the specificity contained in 3D morphology for species-level discrimination.

This work introduces a microscopy-based framework that integrates three-dimensional quantitative phase imaging with a 3D convolutional neural network to rapidly identify pathogenic bacteria from single to few cells. On a dataset of 10,556 tomograms spanning 19 major BSI-causing species, the method achieves 82.5% accuracy with a single measurement and up to 99.9% with seven measurements, comparable to MALDI-TOF MS but achievable much earlier without labeling or destructive processing. The framework shows promise as an advisory tool for early clinical decision-making in infections.