Microbial infections pose a significant global health challenge, leading to high mortality and healthcare costs. Early, appropriate antibiotic treatment is crucial for reducing mortality, but current microbial identification methods, such as culture tests and molecular diagnostics, suffer from long turnaround times (often exceeding 24 hours). Culture tests are time-consuming and nonspecific; molecular methods are faster but lack scalability. MALDI-TOF MS, while considered the gold standard, requires sufficient sample growth, delaying results. Image-based methods, including fluorescence and autofluorescence microscopy, offer speed but may be destructive or lack specificity. This study proposes a novel framework that utilizes 3D QPI, a non-destructive technique measuring the refractive index tomogram of live cells, coupled with an ANN for rapid and accurate bacterial identification. The goal is to achieve species identification from a limited number of cells, enabling earlier and more effective antibiotic treatment.

The literature review highlights the limitations of existing microbial identification techniques. Culture tests are slow and nonspecific; molecular methods offer speed but limited scalability. MALDI-TOF MS, the current gold standard, requires sufficient bacterial growth before analysis. Image-based methods, such as fluorescence microscopy and its label-free counterparts, have been explored but face challenges in terms of invasiveness, specificity, and applicability to diverse pathogens. The authors identify a gap in rapid and accurate identification of pathogens from limited samples, which this study aims to address through 3D QPI and ANN.

The proposed framework integrates 3D QPI and an ANN for bacterial species identification. 3D QPI is performed using a commercialized holotomography system (HT-2H, Tomocube Inc.) employing Mach-Zehnder laser interferometry and a digital micromirror device (DMD) to reconstruct 3D refractive index tomograms. The ANN, specifically designed for 3D data, utilizes 3D convolutional operations and dense connections to effectively learn the morphologies of different bacterial species from the 3D tomograms. The ANN was trained on a database of 10,556 3D RI tomograms from 19 bloodstream infection (BSI)-related bacterial species. The performance of the framework was evaluated using a blind test, analyzing the accuracy of species identification based on single and multiple 3D QPI measurements. The methodology also involved comparing the performance with variations in imaging strategies and algorithms (2D QPI, different machine learning algorithms) to demonstrate the contributions of each component.

The framework demonstrated remarkable accuracy in bacterial species identification. Using a single 3D RI tomogram, the ANN achieved an 82.5% accuracy in identifying 19 BSI-related bacterial species. This is comparable to the performance of MALDI-TOF MS with a sufficient sample amount. Importantly, this accuracy was achieved from single cells or small clusters, significantly reducing the time required for identification. The accuracy dramatically increased with multiple measurements, reaching 99.9% with seven measurements. The superior performance of the 3D QPI and ANN-based framework, compared to variants using 2D QPI or traditional machine learning algorithms, highlights the importance of both 3D morphological information and the specific architecture of the ANN. The study demonstrates that the framework can potentially guide early antibiotic treatment before the results of conventional methods are available.

The high accuracy achieved by the proposed framework from a minute sample quantity addresses the critical need for rapid microbial identification in clinical settings. The results significantly improve upon current methods by reducing the time required for diagnosis, enabling earlier and more targeted antibiotic treatment. The comparable accuracy to MALDI-TOF MS, obtained from a drastically reduced sample size, underscores the potential of this method for clinical applications. The successful integration of 3D QPI and an ANN provides a powerful tool for rapid pathogen identification, with potential for broader applications beyond BSI-related pathogens.

This study presents a novel framework for rapid bacterial identification using 3D QPI and ANN, achieving high accuracy from a minute quantity of bacteria. This method offers significant advantages over existing techniques by reducing turnaround time and enabling earlier, more informed clinical decisions. Future work should focus on expanding the database to encompass a wider range of bacterial species and exploring the feasibility of integrating this technology into point-of-care diagnostics.

The study's accuracy was primarily evaluated on 19 BSI-related bacterial species. Further testing with a broader range of bacterial species is necessary to assess the generalizability of the method. The reliance on a specific 3D QPI system might limit the accessibility and reproducibility of the findings. The complexity of the ANN might require specialized expertise for implementation and maintenance in clinical settings.