For more than a century, the workhorse of bacterial detection has been essentially unchanged: smear a sample onto a gel-like plate, wait one to three days for colonies to grow, then run a battery of biochemical and serological tests to work out what has emerged. Researchers at the Korea Food Research Institute (KFRI) argue that this bottleneck is finally being dismantled—by light and by algorithms. In a comprehensive review published on 29 August 2026 in the journal Food Science and Biotechnology, a team led by corresponding author Min-Cheol Lim surveys the accelerating convergence of spectroscopy, optical imaging and artificial intelligence into tools that can identify pathogenic bacteria in food and clinical samples in minutes to hours, often without dyes, antibodies or any other label. The technologies read like a photonics catalogue—Fourier transform infrared spectroscopy, hyperspectral imaging, laser speckle imaging, time-lapse shadow image analysis, holography and terahertz sensing—but their common thread, the authors contend, is speed without the sacrifices of conventional microbiology.

The stakes are enormous. Foodborne pathogens such as Salmonella, pathogenic Escherichia coli O157:H7, Listeria and multidrug-resistant Vibrio strains sicken hundreds of millions of people each year, while in hospitals the delay between sampling and a definitive bacterial identification forces physicians to prescribe broad-spectrum antibiotics blindly—a practice that feeds the resistance crisis. Conventional culture-based methods, though inexpensive and sensitive, are slow because they depend on the biology of the pathogen itself: colonies must reach a visible size before they can be typed. Polymerase chain reaction shortens the wait but demands DNA extraction, thermal cycling equipment and trained personnel, targets only organisms already suspected, and cannot easily distinguish living cells from residual DNA. Antibody-based immunoassays are faster still but require bespoke reagents for every target. What food safety inspectors and clinical microbiologists alike need, the KFRI team argues, is detection that is rapid, real-time, label-free and non-invasive—a combination that only optical methods, supercharged by machine learning, appear poised to deliver.

The most mature of these approaches is Fourier transform infrared (FTIR) spectroscopy. When mid-infrared light, spanning roughly 4,000 to 400 wavenumbers, interacts with a bacterial sample, its chemical bonds absorb radiation at characteristic frequencies: proteins dominate the spectrum through the amide I and amide II bands near 1,650 and 1,540 cm⁻¹, lipids contribute peaks around 1,740 and between 3,000 and 2,800 cm⁻¹, while polysaccharides and nucleic acids leave their marks at lower wavenumbers. The result is a whole-cell biochemical fingerprint, reproducible enough to distinguish species and even strains. The technique’s microbiological pedigree stretches back to 1985, when researchers first applied it to bacteria and biofilms, and it has since been shown to detect E. coli O157:H7 and spoilage-causing Alicyclobacillus strains in apple juice, to identify the same pathogen in ground beef when coupled with filtration and immunomagnetic separation, and to type the gram-negative bacilli behind hospital outbreaks in a fraction of the time of classical typing. Handheld and portable FTIR instruments now promise on-site quality control from farm to fork.

Near-infrared (NIR) spectroscopy, operating between roughly 780 and 2,500 nanometres, trades some of FTIR’s chemical specificity for practical advantages: light penetrates deeper into samples, instruments are cheaper and more robust, and measurements can be made with minimal preparation. The price is that NIR spectra consist of broad, heavily overlapping overtone and combination bands rather than sharp fundamental absorptions, so the technique leans heavily on spectral preprocessing—standard normal variate correction and multiplicative scatter correction among them—and on chemometric models such as principal component and partial least squares regression. Applied to microbiology, the approach has differentiated carbapenem-resistant from susceptible Enterobacteriaceae strains, detected bacteria in the dairy industry, identified poultry-meat-associated species, and even spotted Staphylococcus aureus biofilms on food-contact surfaces using portable instruments. More recent work extends NIR to bacterial biofilm characterisation more broadly, and machine-learning classifiers trained on NIR data have enabled rapid, reagent-free classification of bacterial isolates—no stains, no cultures, no consumables.

Raman spectroscopy takes the complementary route: rather than measuring absorption, it reads the tiny fraction of photons—roughly one in a million to one in a hundred million—that scatter off a molecule with shifted energy, revealing its vibrational structure. Because water, ubiquitous in biological samples, is nearly Raman-silent, the method suits aqueous specimens that defeat infrared approaches. Its principal weakness, a feeble signal, has been answered by surface-enhanced Raman scattering (SERS), in which molecules adsorbed onto nanostructured gold or silver surfaces experience electromagnetic and chemical enhancements that can amplify signals by factors of a million or more. The combination is proving potent: Raman spectra have rapidly differentiated the microbes responsible for urinary tract infections, machine learning has paired with Raman to identify Salmonella serovars, and single-cell Raman spectra coupled to self-transfer deep learning with ensemble prediction have classified foodborne pathogens at the level of individual cells. SERS-based immunosensors built on covalent organic framework Raman tags, and SERS–CRISPR assays on microfluidic paper devices, now push foodborne pathogen detection toward supersensitivity.

Where spectroscopy interrogates a bulk average, hyperspectral imaging maps chemistry across space, capturing a full spectrum at every pixel to build a three-dimensional spectral cube spanning the visible, near-infrared and short-wavelength-infrared ranges. First deployed around the turn of the millennium to spot contaminants on poultry carcasses, the technology has matured into a non-invasive workhorse for microbial quality: it has tracked fish spoilage and predicted microbial loads without touching the fillet, mapped contamination on meat, and, at the laboratory bench, differentiated common foodborne pathogens at both colony and single-cell levels in dairy products. Multi-scale spectral imaging—stitching microscopic and macroscopic views—has even identified multiple bacterial species growing on stainless steel, the workhorse surface of food processing plants, suggesting a future in which a hyperspectral camera routinely audits hygiene across an entire production line.

Perhaps the most counterintuitive entry is laser speckle imaging. Illuminate any rough or turbid object with coherent laser light and it appears covered in a grainy interference pattern—a speckle field. When the illuminated matter is alive and moving, from the Brownian jostling of molecules to swimming cells, the speckle fluctuates, and the statistics of that fluctuation encode biological activity. Reviews of the biospeckle method trace its agricultural roots, but microbiology has embraced it: speckle patterns shift measurably as bacterial suspensions grow, allowing real-time monitoring of growth kinetics without sampling; growing colonies can be told apart from non-growing ones well before they become visible to the eye; and the approach has exposed hidden inhibition zones in antibiotic disc-diffusion tests long before they are apparent. Machine learning has turned the physics into diagnostics: the DyRAST platform couples dynamic laser speckle imaging to algorithms for rapid antibacterial susceptibility testing, while related work has pinpointed the minimum lethal concentration of ampicillin in E. coli liquid cultures.

Removing the lens altogether pushes detection toward the factory floor and the clinic. Time-lapse shadow image analysis films the growing shadows of microcolonies at high frame rates, achieving noise-free colony counts and retrievable three-dimensional time-lapse datasets, and has already served environmental monitoring in biological manufacturing. Digital inline holography goes further: a sample held a fraction of a millimetre from an image sensor casts an interference hologram that computers reconstruct into quantitative phase and amplitude images—lensless, compact and fieldable. Deep-learning classifiers reading such holograms have detected and categorised live bacteria in water, flagged contamination in sterile liquid products in real time and analysed particulates and bacteria in peritoneal dialysis fluid. At the far end of the spectrum, terahertz radiation—lying between microwaves and infrared—probes collective vibrational modes of whole biomolecules: terahertz thermal curve analysis has identified pathogens label-free, near-field terahertz imaging has resolved single bacteria, and terahertz attenuated-total-reflection spectroscopy fused with multi-classifier voting has automated clinical microbial identification.

None of these optical modalities reaches its potential without the final ingredient: artificial intelligence. Spectral data are high-dimensional and subtly non-linear, precisely the regime in which convolutional neural networks excel; unified one-dimensional CNN architectures now routinely recognise Raman spectra, and CNN-based pipelines have become standard in food spectral analysis. Transformers, whose self-attention mechanisms capture long-range dependencies, are supplanting CNNs for hyperspectral data—spatial-spectral transformers classify hyperspectral images, and a hyperspectral microscope upgraded with a transformer network has classified infectious bacteria directly. Object-detection frameworks such as YOLOv8 have identified bacteria from optical scattering patterns generated by simple light-emitting diodes, dispensing with sample alignment, while transfer learning with ResNet-50 has classified bacterial colony images, and colony morphology alone has sufficed for machine-learning identification of Pseudomonas aeruginosa strains. The authors, however, sound a cautionary note: machine-learning science is plagued by data leakage, in which information from test sets contaminates training and inflates reported accuracy; rigorous external validation across instruments, laboratories and sample matrices must precede any regulatory deployment.

The KFRI review’s verdict is that no single technique will conquer the Petri dish alone. Mid-infrared and NIR spectroscopy offer rich fingerprints but struggle with complex matrices; Raman and SERS reach single cells but need nanostructured substrates; speckle, shadow and holographic methods deliver speed and simplicity but still face detection-threshold hurdles in real foods and bodily fluids. The future the authors propose is deliberately integrative: multimodal platforms that fuse spectroscopy with imaging, wrapped in AI models trained on standardised, well-curated datasets, miniaturised with LEDs, smartphone cameras and lensless optics, and validated for the food chain and the clinic alike. Such systems, they conclude, would realise the long-sought goal of non-invasive, label-free bacterial detection—turning invisible killers visible in minutes rather than days, catching contamination before shipment rather than after recall, and telling a physician which antibiotic will work while the patient still waits. The research was supported by the Ministry of Science and ICT of South Korea through the Korea Food Research Institute’s main research program.