Virtual staining of unstained bacteria via deep learning

Introduction: : For over a century, Gram staining has been a cornerstone technique in microbiology, fundamental for differentiating bacteria based on their cell wall characteristics. This classification is indispensable across various domains, most critically in clinical diagnostics. It provides rapid bacterial identification directly from patient samples, which is crucial for guiding initial antibiotic selections in potentially severe infections. Furthermore, it serves a vital function in monitoring microbial presence in food safety and environmental contexts. Despite this enduring relevance, the conventional Gram staining process is hampered by considerable drawbacks, such as its reliance on manual execution, lengthy procedures, susceptibility to operator variability, and inherent safety concerns for laboratory workers. To address these limitations, we propose a virtual Gram staining method powered by deep learning. This approach transforms 3D darkfield axial image stacks obtained from label-free, unstained bacteria into their equivalent brightfield images, faithfully replicating the appearance of chemically Gram-stained cells without any wet-lab processing (see Fig. 1a). The successful performance of this digital staining technique has been confirmed using E. coli (Gram-negative) and L. innocua (Gram-positive).

Materials and

Methods: : Our methodology leverages a conditional generative adversarial network (cGAN). For training data, darkfield axial image stacks (Z = -2 to +2 μm, 1 μm steps) of bacterial smears were captured using an inverted microscope (IX-83, Olympus) with a 100x oil-immersion objective. These darkfield images were precisely registered with corresponding brightfield images taken from the same smears after conventional Gram staining—a one-time effort required only for training. The cGAN model, implemented using Python and the TensorFlow library, was subsequently trained to establish a physically consistent virtual Gram staining transformation. For training and subsequent testing, the dataset encompassed 620 images containing E. coli, 620 images of L. innocua, and 1000 images featuring mixtures of both species.

Results, Conclusions, and Discussions:: Initially, the virtual Gram staining model was trained using carefully registered pairs of darkfield axial image stacks from label-free bacteria and their corresponding brightfield images after conventional Gram staining. Subsequently, the trained model's generalization capability was assessed using test fields-of-view (FOVs) entirely excluded from the training phase. Figure 1a-b visually demonstrates the method's microscopic imaging fidelity, showcasing a virtually stained FOV next to its chemically stained brightfield equivalent. The effectiveness of this computational staining framework is quantitatively supported by high average image quality metrics (see Fig. 1c): a Peak-Signal-to-Noise Ratio (PSNR) of 38.35 dB and a Structural Similarity Index Measure (SSIM) of 0.9685, calculated over a set of 200 test images. To further evaluate performance, bacteria detection metrics were computed, yielding a precision of 95.5%, recall of 96.5%, and an F1-score of 96%. Low error rates, specifically a false staining (FS) rate of ~3% and a bacteria hallucination (BH) rate of ~1.5%, were also recorded, collectively demonstrating the high accuracy achieved by the virtual Gram staining technique.
In summary, this work introduces the demonstration of virtual Gram staining, achieving accurate conversion of darkfield images derived from label-free bacteria into corresponding brightfield images mimicking conventional Gram-stained bacteria. This computational staining methodology effectively translates darkfield microscopy data of unstained bacteria into their virtual Gram-stained brightfield equivalents. Such an approach promotes repeatable results, thereby circumventing the variability associated with operator technique and chemical inconsistencies prevalent in standard Gram staining protocols. Moreover, the fundamental principles employed in this virtual staining framework show potential for extension to other significant microbiological staining methods, such as those used for acid-fast and endospore detection.

Acknowledgements and/or References (Optional)::