Chemical Imaging for Biological Systems: Techniques, AI-Driven Processing, and Applications

Abstract

Visualizing the chemical compositions of biological samples is pivotal to advancing biological sciences, with the past two decades witnessing the emergence of innovative chemical imaging platforms such as single-molecule imaging, coherent Raman scattering microscopy, transient absorption microscopy, photothermal microscopy, ambient ionization mass spectrometry, electrochemical microscopy, and advanced chemical probes. These technologies have enabled significant breakthroughs in diagnosing pathological transitions, designing targeted therapies, and understanding drug resistance mechanisms. Recent advancements in resolution, contrast, sensitivity, and speed have transformed the field, with techniques like fluorescence, infrared absorption, and Raman scattering being widely applied across diverse biological domains. This review provides a comprehensive overview of the evolution and current state of chemical imaging technologies, coupled with systematic analyses of data processing workflows, including pre-processing, machine learning-assisted pattern extraction, and neural network-based predictions. Artificial intelligence (AI) and machine learning-assisted imaging are transforming chemical imaging through key advancements such as improved resolution and sensitivity via noise reduction techniques, enhanced data analysis (e.g., spectral unmixing, pattern recognition), automated feature extraction using neural networks, real-time processing via high-performance cluster, and data fusion across optical platforms. These innovations are significantly advancing both current applications and the future development of chemical imaging techniques in biomedical research. However, several critical challenges remain, including the scarcity of high-quality training datasets, limited generalizability across different instruments and experimental conditions, high computational costs, challenges in output interpretability and trust, and the lack of standardized validation protocols across different approaches. Looking ahead, the integration of bioimaging into cell biology, lipid research, tumor studies, microbiology, neurobiology, and developmental biology is anticipated to expand its impact, aided by interdisciplinary expertise in biochemistry, physics, and optical engineering. These developments promise unprecedented resolution and speed, facilitating high-speed, high-resolution imaging of living systems, with applications leading to discoveries such as biomarkers for cancer aggressiveness and drug resistance. Moreover, the miniaturization and commercialization of imaging platforms are broadening accessibility, enabling on-site clinical investigations and in vivo measurements, underscoring the transformative potential of chemical imaging in advancing biological science and medical research.

1. Overview of imaging techniques

In recent decades, microscopy has advanced rapidly, with significant improvements in resolution, contrast, sensitivity, and speed. While early efforts focused on visualizing specimen morphology, there is an increasing demand for deeper insights into biological functions through chemical composition analysis. This shift has spurred the development of innovative diagnostic techniques, therapeutic strategies, and drug resistance studies. Among the various chemical imaging methods, optical microscopy remains the most widely used tool in biological research, owing to its exceptional chemical specificity and spatial resolution. Chemical contrast in optical microscopy is typically achieved through fluorescence labeling, light absorption, and Raman scattering, with breakthroughs in these techniques significantly expanding the scope of biological research. Achieving chemical-specific optical manipulation requires highly selective detection of target molecules. Figure 1 summarizes the optical microscopy modalities commonly employed for mapping chemical compositions in live cells, including fluorescence microscopy, infrared absorption, and Raman scattering. In the following sections, we discuss their historical origins and recent developments.

Figure 1: Timeline of the technical advancements in imaging for biological systems.

Reproduced from reference [12] with permission from [Springer Nature], copyright [2019]. Reproduced from reference [73] with permission from [Springer Nature], copyright [2019].

1.1. Timeline of development

Fluorescence microscopy.

Introduced in the early 20th century, fluorescence microscopy revolutionized biological imaging by enabling the visualization of specific targets within cells and tissues, providing detailed chemical information. This transformative capability has significantly advanced our understanding of cellular processes and disease mechanisms. Key milestones include the development of fluorescent antibody labeling^1,2 and fluorescent protein cloning^3,4.

Fluorescent antibody labeling employs antibodies tagged with fluorescent dyes to bind specifically to target antigens, offering exceptional specificity and sensitivity in complex biological samples. Similarly, fluorescent proteins like enhanced green fluorescent protein, enabled by targeted gene editing techniques, have expanded the possibilities of fluorescence microscopy. For example, tagging end-binding protein in HeLa cells allows researchers to monitor microtubule dynamics in real time, shedding light on processes such as cell division and intracellular transport.

Advanced fluorescence techniques have further enhanced the field. Confocal microscopy⁵ provides improved spatial resolution by eliminating out-of-focus light, while light sheet microscopy⁶ enables rapid, high-resolution imaging with minimal photodamage. Techniques like total internal reflection fluorescence microscopy^7,8 focus on membrane dynamics, and fluorescence lifetime imaging microscopy⁹ measures fluorescence decay times to reveal interactions and molecular environments. Förster resonance energy transfer¹⁰ detects nanoscale molecular interactions, and fluorescence in situ hybridization¹¹ visualizes specific nucleic acid sequences for gene expression and chromosomal studies. The increased quantum yield of fluorophores combined with low background noise allows for single-molecule detection, providing good sensitivity for studying molecular activities¹².

Super-resolution fluorescence techniques, such as stimulated emission depletion¹³, photoactivatable localization microscopy¹⁴, and stochastic optical reconstruction microscopy¹⁵, and structured illumination microscopy¹⁶, have surpassed the diffraction limit, achieving nanometer-scale resolution. Additionally, multiphoton fluorescence^17,18 extends applications to deep tissue imaging, using longer wavelengths to minimize scattering and penetrate deeper into biological samples. This capability is particularly valuable in studying complex organisms and the dynamic interactions within their biological systems.

Infrared absorption microscopy.

Infrared absorption microscopy is a technique that utilizes infrared light to both visualize a sample and analyze its chemical composition. By measuring the absorption of specific infrared wavelengths at different locations on the sample, it generates a detailed chemical map of the specimen at a microscopic scale, particularly for small molecules and metabolites^19,20. While infrared absorption microscopy is valuable for probing functional groups in biological samples, its application in thick tissues is limited due to strong water absorption and lower spatial resolution compared to fluorescence microscopy.

Recent advancements, such as super-resolution far-field IR imaging and photothermal IR microscopy, have significantly improved spatial resolution. These methods^21–23 leverage thermal effects induced by mid-IR absorption to achieve sub-micron resolution. Transient absorption microscopy, which examines electronically excited states, provides insights into chemical dynamics and interactions in real time, complementing traditional absorption techniques.

Raman scattering microscopy.

Raman scattering microscopy uses the inelastic scattering of photons to analyze molecular vibrations²⁴, offering good spatial resolution without the limitations of water absorption in IR techniques. However, spontaneous Raman scattering generates weak signals, requiring long integration times. Innovations like resonance enhancement, surface-enhanced Raman spectroscopy, and tip-enhanced Raman spectroscopy have addressed this limitation by improving signal strength and spatial resolution.

Coherent Raman techniques^25–30, including coherent anti-Stokes Raman scattering and stimulated Raman scattering, provide high-speed imaging with significant signal enhancements. These methods enable Raman imaging at speeds comparable to fluorescence microscopy, making them suitable for studying dynamic biological processes. In comparison, while fluorescence microscopy may still provide certain advantages in sensitivity and selectivity for protein detection, Raman imaging excels in analyzing small molecules and labeling-free detection to avoid potential invasions with biological processes.

1.2. Imaging Chemicals: Beyond General Optics

Chemical imaging is a powerful analytical technique that extends beyond conventional imaging by integrating spatial resolution with molecular specificity. Unlike general imaging methods, which primarily capture morphological and structural features, chemical imaging combines imaging techniques with spectroscopy to provide both spatial and spectral information. This integration allows for the visualization of the chemical composition of biological and biomedical samples with high precision.

1.2.1. Fundamentals of Chemical Imaging: Integration of Spectroscopy and Imaging

The defining characteristic of chemical imaging is the acquisition of spectral data at each pixel, enabling molecular identification and spatial distribution mapping. Several spectroscopy-based techniques are widely utilized in chemical imaging, including: (1) Raman imaging^25–30 – Combines Raman spectroscopy with optical imaging to detect vibrational energy shifts, enabling label-free molecular characterization in biological tissues and cells. Advances such as coherent anti-Stokes Raman scattering and stimulated Raman scattering further enhance sensitivity and speed. (2) Fourier transform infrared imaging^19–23 – Uses infrared absorption spectroscopy to provide molecular fingerprints of biological samples, making it useful for disease diagnostics and tissue characterization. (3) Mass spectrometry imaging – Integrates mass spectrometry with spatially resolved sampling methods to map biomolecules, metabolites, and drug distributions across biological samples. (4) Hyperspectral and multispectral imaging – Captures a broad spectral range at each pixel to differentiate between biochemical components in tissues and cells, often applied in disease diagnostics and tissue engineering. To fully realize the potential of chemical imaging, it is critical to improve both spatial and spectral resolution, allowing for the precise identification of molecular structures and localization in complex biological systems.

1.2.2. Enhancing Spatial Resolution in Chemical Imaging

Achieving high-resolution spatial and spectral information remains a key challenge in chemical imaging. Various approaches have been developed to push these limits, including advanced optical techniques, computational methods, and AI-driven processing.

Spatial resolution in chemical imaging is primarily limited by diffraction, optical system performance, and detector sensitivity. Several methods have been developed to push beyond these limits. For examples, stimulated emission depletion microscopy is to utilize a secondary depletion laser to confine fluorescence emission to sub-diffraction spots, significantly enhancing spatial resolution. In addition, structured illumination microscopy is to use patterned light to extract high-frequency spatial details beyond the diffraction limit. Furthermore, single-molecule localization microscopy is a technique like stochastic optical reconstruction microscopy and photoactivated localization microscopy enable nanometer-scale imaging by precisely localizing individual fluorophores. Combining complementary imaging techniques, such as fluorescence imaging with Raman imaging or AFM with infrared spectroscopy, improves spatial details while maintaining chemical specificity.

1.2.3. Enhancing Spectral Resolution in Chemical Imaging

Spectral resolution determines how well closely spaced spectral features can be distinguished, crucial for identifying molecular structures. Several strategies enhance spectral resolution in chemical imaging. Firstly, high-performance spectrometers, include FTIR spectroscopy with high-resolution detectors, is to reduce noise and increases precision in spectral peak identification. Also, high-resolution dispersive Raman spectroscopy is another example to use advanced gratings and detectors to resolve narrow Raman bands for detailed molecular fingerprinting. Secondly, hyperspectral and ultrafast spectroscopy, include hyperspectral imaging, to capture a full spectrum at each pixel, allowing for detailed spectral differentiation of biomolecules. In addition, time-resolved and coherent Raman spectroscopy is another strategy to improve spectral resolution, techniques such as CARS and SRS use nonlinear optical effects to enhance spectral resolution while reducing background fluorescence. Furthermore, AI-driven spectral enhancement is also an effective strategy. As an example, in deep learning for spectral unmixing, AI models, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs), can separate overlapping spectral features, improving resolution and interpretability. Super-resolution spectral reconstruction is a AI-based upscaling method to reconstruct high-resolution spectra from low-resolution measurements, enhancing chemical specificity.

The combination of advanced optical techniques with AI-driven data processing is revolutionizing chemical imaging. AI-driven approaches, such as deep learning-based image reconstruction and spectral analysis, are improving data acquisition efficiency, denoising, and feature extraction. Future developments in quantum-enhanced imaging, ultrafast spectroscopy, and AI-based resolution enhancement promise further breakthroughs, leading to unprecedented levels of molecular detail in biomedical applications. These advancements will not only enhance fundamental research but also accelerate the translation of chemical imaging technologies into clinical diagnostics and drug discovery.

1.3. Workflow of AI-assisted bioimaging and diagnose

In Figure 2, we present an exemplary AI-assisted bioimaging workflow, detailing each step from sample preparation to diagnosis^10–15. Each step is critical for generating accurate and interpretable images that can be reliably used for scientific and medical purposes. The process begins with sample preparation, which is essential for preserving the sample’s natural structures and highlighting specific features of interest. This often involves fixation to prevent degradation and sectioning of thicker samples to enable deeper imaging. Staining or labeling with fluorescent dyes or antibodies is frequently required to enhance contrast or target specific molecules, enabling precise visualization of structures within tissues, cells, or subcellular components. Once prepared, the sample undergoes image acquisition using techniques tailored to the sample type and research objectives. Light microscopy is commonly used for visualizing basic structures, while advanced methods such as confocal microscopy and super-resolution techniques provide higher resolution and three-dimensional imaging. For ultrastructural details, particularly in materials or tissues, electron microscopy (either scanning or transmission) achieves nanoscale resolution. For imaging larger organisms or tissue volumes, modalities such as magnetic resonance imaging, computed tomography, and positron emission tomography scans are employed, enabling whole-organism visualization with three-dimensional depth and multi-layered information. Following acquisition, image processing refines the raw data through computational techniques designed to enhance clarity, reduce noise, and correct imaging distortions. Methods like contrast enhancement improve visibility of critical features, noise reduction algorithms minimize background interference, and segmentation isolates regions of interest. Increasingly, machine learning tools are used to automatically identify patterns and structures, saving time and enhancing reproducibility. Once processed, the data undergo analysis and diagnosis, providing both quantitative and qualitative insights essential for drawing scientific or clinical conclusions. Quantitative analysis might involve measuring parameters such as cell size, distribution, or signal intensity, while statistical methods compare experimental groups or track changes over time. With advancements in artificial intelligence, bioimaging analysis increasingly leverages machine learning algorithms to recognize subtle patterns and predict disease progression, adding a predictive dimension to traditional workflows.

Figure 2: Workflow of AI-assisted bioimaging and diagnostics.

Reproduced from reference [73] with permission from [Springer Nature], copyright [2019]. Reproduced from reference [108] with permission from [American Chemical Society], copyright [2013]. Reproduced from reference [132] with permission from [American Association for the Advancement of Science], copyright [2023].

For personalized diagnostics, bioimaging technologies can integrate molecular information from targeted imaging agents^20–24. For example, molecular imaging probes, which are often designed to bind specifically to disease biomarkers, enable the visualization of the exact molecular processes occurring in a patient’s body. This could be applied in oncology, where specific probes might target cancerous cells or blood vessels, allowing for more accurate staging and monitoring of treatment responses. Similarly, in cardiovascular diagnostics, advanced imaging agents can identify early atherosclerotic lesions or blood vessel abnormalities before they cause significant clinical symptoms. The combination of molecular imaging with traditional anatomical imaging provides clinicians with an unprecedented level of detail for decision-making. The integration of multimodal imaging also plays a pivotal role in diagnosis. This multi-layered information enriches the diagnostic process by providing a more comprehensive understanding of the patient’s condition. Together, these modalities provide a holistic view of the disease, enabling more accurate staging, treatment planning, and monitoring of therapeutic efficacy.

The ability to analyze biomarkers using bioimaging also greatly improves diagnostics. In diseases such as cancer, neurodegeneration, or infectious diseases, the presence, absence, or abnormal expression of certain biomarkers can be indicative of disease progression or therapeutic response. Advanced imaging tools, such as multiplexed fluorescent imaging, allow for the simultaneous visualization of multiple biomarkers within a single tissue section. This is particularly important in cancers, where the expression of several markers might influence treatment strategies. Additionally, imaging-based biomarker analysis can monitor therapeutic response in real-time, providing valuable feedback to clinicians and adjusting treatment protocols as needed.

In clinical decision-making, bioimaging advances help physicians not only diagnose but also monitor treatment responses and disease progression. High-resolution, longitudinal imaging enables the tracking of dynamic changes over time, providing insights into how a disease is evolving or how a patient is responding to therapy. For instance, in the case of cancer, serial PET scans can monitor the size and metabolic activity of tumors during chemotherapy, allowing clinicians to assess whether the treatment is effective or if alternative therapies need to be considered. The use of AI in this context allows for automated analysis of serial imaging data, improving the reproducibility and consistency of assessments and aiding in more personalized treatment decisions.

Moreover, early diagnosis is a critical benefit of bioimaging. Early detection of diseases like cancer, neurodegenerative conditions, and cardiovascular diseases significantly improves treatment outcomes. Innovations in super-resolution and molecular imaging are enabling earlier and more precise detection of pathological changes that occur long before symptoms are visible. For example, in Alzheimer’s disease, advanced imaging techniques can detect amyloid plaque formation or tau tangles in the brain at stages where patients are still asymptomatic, allowing for earlier interventions. This kind of early detection can also extend to infectious diseases, where imaging techniques are employed to monitor pathogens’ interactions with the host’s immune system or visualize the progression of an infection at the cellular level. Finally, predictive diagnostics are increasingly becoming a powerful tool in bioimaging. By combining machine learning algorithms with bioimaging data, it is now possible to predict disease trajectories, responses to treatments, and even patient outcomes. Predictive models that analyze imaging data alongside genetic, clinical, and environmental information are being developed to provide a more comprehensive risk assessment for diseases such as cancer, diabetes, and heart disease. These models can inform personalized treatment strategies, identifying patients who are likely to respond to certain therapies and those who may require alternative approaches. As these predictive tools become more sophisticated, bioimaging will play a central role in the transition toward precision medicine, where treatments are tailored to an individual’s unique biological characteristics.

In summary, the diagnostic potential of bioimaging is vast, with innovations enabling more accurate, early, and personalized diagnoses. The integration of advanced imaging modalities, molecular probes, AI, and predictive models promises to revolutionize clinical practice by providing highly detailed, real-time insights into the molecular and structural basis of diseases. This will not only enhance disease detection but also guide the development of targeted therapies, optimize patient management, and improve clinical outcomes across a wide range of medical fields.

1.4. Market landscape, research gap and market gap

Imaging plays a critical role in biological research, diagnostics, and pharmaceutical development by providing spatially resolved molecular information. The global market for chemical imaging in life sciences has been rapidly expanding, driven by the increasing demand for high-resolution, label-free analytical techniques. The chemical imaging market size was valued at USD 4.5 billion in 2023 and is projected to grow at a CAGR of 8.7% from 2024 to 2030, fueled by advancements in hyperspectral imaging, infrared spectroscopy, Raman microscopy, and mass spectrometry imaging³¹. Key drivers of this growth include the rising adoption of non-invasive diagnostic tools, AI-driven data processing, and precision medicine approaches, which rely on highly sensitive and specific chemical imaging techniques for biomolecular analysis.

The miniaturization and commercialization of imaging platforms are broadening accessibility by enabling the development of portable, cost-effective, and user-friendly chemical imaging devices^5–10. Recent advances in microfabrication, fiber optics, and chip-based spectroscopy have led to the emergence of compact and field-deployable imaging solutions that maintain high analytical performance. These innovations facilitate point-of-care diagnostics, real-time environmental monitoring, and on-site pharmaceutical quality control, reducing the reliance on bulky and expensive laboratory-based systems. Furthermore, the integration of AI and cloud-based data analysis enhances the usability of these platforms, allowing non-expert users to leverage chemical imaging technologies for diverse applications in healthcare and industry.

Despite significant progress in chemical imaging, several research gaps remain in terms of spatial resolution, real-time imaging capabilities, and data interpretation. Traditional techniques, such as Fourier Transform Infrared Spectroscopy (FTIR) and Raman Spectroscopy, face limitations in spatial resolution (diffraction limit), weak signal intensity, and complex data processing. While emerging nanoscale approaches, such as NanoIR spectroscopy and Tip-Enhanced Raman Spectroscopy (TERS), have improved imaging resolution, their integration with AI-driven data analysis remains underexplored. Additionally, challenges in interpreting high-dimensional spectral data limit the widespread application of chemical imaging in dynamic biological systems, such as real-time single-cell analysis, disease diagnostics, and biomolecular interaction studies.

From a commercial perspective, the market lacks a fully integrated chemical imaging platform that combines high-resolution imaging, AI-enhanced spectral processing, and real-time diagnostics. Current solutions are often fragmented, requiring multiple imaging modalities and extensive data post-processing, which slows down biomedical research and clinical applications. The integration of machine learning (ML) and AI-driven image processing has the potential to bridge this market gap by enabling automated feature extraction, enhanced spectral contrast, and predictive analysis for early disease detection. The next generation of AI-assisted chemical imaging can drive transformative innovations in precision medicine, biomarker discovery, and biomedical engineering, creating new commercial and research opportunities.

In the following sections, we provide a detailed exploration of the working principles, challenges, demonstrated applications, and future opportunities of various bioimaging microscopy techniques.

2. Chemical Imaging microscopies

2.1. Fluorescence microscopies

2.1.1. Confocal fluorescence microscopy.

Confocal fluorescence microscopy (CFM) ^31–37 is a powerful imaging technique widely used in biological and material sciences for acquiring high-resolution and three-dimensional (3D) images. As illustrated in Figure 3a, the working principle of CFM involves point illumination and a spatial pinhole to eliminate out-of-focus light in specimens thicker than the focal plane. This enables optical sectioning and clear 3D reconstruction of the sample. Typically, CFM uses a laser as the excitation light source. Fluorescent dyes or proteins within the sample absorb this light and emit longer-wavelength fluorescence. A system of mirrors or a moving stage directs the laser spot across the sample to build a two-dimensional (2D) image. Combining multiple 2D images produces detailed 3D reconstructions. The in-focus emitted light is collected by a photodetector, such as a photomultiplier tube or an avalanche photodiode.

Figure 3: Imaging techniques.

(a) Fluorescence microscopy. (b) Super-resolution microscopy. (c) Raman microscopy. Reproduced from reference [12] with permission from [Springer Nature], copyright [2019]. Reproduced from reference [92] with permission from [Springer Nature], copyright [2019].

CFM offers several advantages, making it a widely adopted technique for imaging living cells, intracellular organelles, and components^33–37. First, compared to wide-field fluorescence microscopy (WFFM), CFM provides higher optical resolution and contrast by eliminating background fluorescence from out-of-focus planes. Second, it enables the collection of thin optical sections, facilitating 3D imaging without physical sectioning of the sample. Third, CFM minimizes photobleaching because only the in-focus regions are illuminated at any given time. This reduces light-induced damage compared to WFFM. Additionally, CFM supports the use of multiple fluorescent dyes simultaneously, enabling colocalization studies. However, CFM also has limitations^38,39, including slow scanning speed and the use of high-intensity light, which can lead to shorter fluorescence lifetimes and phototoxicity in living cells. To mitigate these issues, imaging time should be carefully controlled during operation.

CFM is versatile and plays a crucial role in modern scientific research, particularly in high-resolution imaging of complex biological processes^38–51. For example, it is frequently used in live-cell imaging to monitor cellular behaviors, such as responses to drugs and intracellular processes like mitosis and apoptosis. In one study⁴⁰, HeLa cells imaged with CFM revealed the number of internalized vaterite particles and their crystalline structures. Additionally, CFM has been employed to track cellular development and differentiation in embryos, providing insights into gene expression patterns during early developmental stages. CFM is also invaluable for studying the distribution of proteins, organelles, and subcellular structures, such as actin filaments and microtubules, with remarkable clarity. For instance, the actin cytoskeleton architecture of leukemia cell lines, Jurkat and K562 cells, was fully visualized using CFM to investigate F-actin remodeling mechanisms⁴¹. Recent efforts have focused on improving CFM resolution beyond the diffraction limit. For example, a pinhole displacement method has been proposed, enhancing contrast by up to 80% with a diffraction-limited resolution of 250 nm. Additionally, the use of Bessel–Gauss beams, which produce a tighter focal spot than classic Gaussian beams, has achieved resolution enhancements. A CFM employing Bessel–Gauss beams⁴² improved lateral resolution to 219 nm, demonstrating the potential for further advancements in resolution and imaging capabilities. CFM continues to be an indispensable tool in scientific research, offering unparalleled insights into cellular and molecular processes.

2.1.2. Two-Photon Fluorescence Microscopy.

Two-photon fluorescence microscopy (TPFM) is an advanced imaging technique that shares similarities with confocal fluorescence microscopy but offers unique advantages, such as deeper tissue penetration and reduced photodamage. These features make TPFM particularly well-suited for live animal imaging and in-depth studies of biological samples.

As shown in Figure 3a, the working principle^17,43,44 of TPFM involves the use of longer-wavelength, near-infrared lasers to excite fluorophores through the simultaneous absorption of two lower-energy photons, rather than a single high-energy photon. This two-photon absorption is a non-linear process requiring extremely high photon densities, achieved by tightly focusing the laser beam on a small point within the specimen. Fluorophore excitation occurs exclusively at the focal point, where the photon density is sufficient for two-photon absorption, minimizing out-of-focus excitation and reducing photobleaching and phototoxicity. Similar to confocal microscopy, TPFM enables optical sectioning by scanning at different depths within the sample, allowing for 3D imaging. Its ability to achieve precise excitation only at the focal point, combined with its deeper tissue penetration, makes TPFM a powerful tool for studying biological structures and dynamic processes in thick or intact samples.

TPFM offers several advantages that make it ideal for in vivo studies. Firstly, the near-infrared wavelengths used in TPFM penetrate deeper into biological tissues (up to approximately 1 mm) compared to single-photon confocal microscopy, enabling the imaging of intact tissues or organs in living animals. Secondly, because excitation occurs only at the focal point, photodamage to out-of-focus areas is minimized, making this technique particularly suited for live-cell and in vivo imaging. Additionally, by restricting fluorescence to the focal plane, TPFM provides higher contrast and an improved signal-to-noise ratio, particularly for thick samples, compared to widefield fluorescence microscopy that may be affected by background fluorescence from out-of-focus planes.

However, TPFM has certain limitations. The low imaging speed (approximately 0.5 Hz)⁴⁵ means that acquiring a high-resolution image stack using traditional TPEM can take tens of minutes. Furthermore, visualizing biological structures smaller than hundreds of nanometers remains a significant challenge^46,47. Despite these limitations, TPFM has been widely used to advance our understanding of dynamic biological processes, particularly in real-time and within living systems. Its applications span multiple fields, including neuroscience, cancer research, and developmental biology. For example, in immunology, TPFM has been employed to monitor the movement of transferred T-cells in mice⁴³, as it is less affected by fluorescence concentration and light source interference. In neuroscience, TPFM enables precise detection of fluorescent probes to track dynamic changes in hydrogen peroxide and adenosine triphosphate (ATP) within neurons⁴⁸. Additionally, TPFM has garnered significant interest in the study of mammalian brain structure^46,47 and function in vivo. In microbiology, TPFM has been used to develop rapid bacteria localization platforms, where z-stack images distinguish intracellular from extracellular bacteria⁴⁹. These applications highlight the versatility of TPFM in addressing complex biological questions across diverse fields.

2. 1. 3. Light sheet fluorescence microscopy.

Light sheet fluorescence microscopy (LSFM) is an advanced imaging technique that facilitates rapid, high-resolution, and minimally invasive imaging of large biological specimens^50–53. As illustrated in Figure 3a, LSFM works by illuminating a thin plane (light sheet) within the sample, positioned perpendicular to the detection axis. The sample is illuminated from the side by a sheet of light, exciting only the fluorophores within this narrow plane. A camera positioned at a 90-degree angle to the light sheet captures the fluorescence emitted from the illuminated plane, enabling fast, high-resolution imaging. By moving the sample or the light sheet across different planes, LSFM can reconstruct an entire 3D volume without the need for point-by-point scanning.

LSFM offers several significant advantages. Firstly, its plane-by-plane illumination enables the rapid acquisition of large 3D datasets, making it much faster than traditional point-scanning methods, such as confocal and TPFM. Unlike CFM, which scans point-by-point and suffers from slow acquisition times for large samples, LSFM efficiently images entire planes at once, significantly reducing phototoxicity and photobleaching. This makes LSFM particularly suited for long-term live-cell imaging, where repeated exposure to high-intensity light can compromise cell viability. Furthermore, LSFM excels at imaging large and complex samples, such as whole embryos or organoids, due to its capacity to capture extensive areas at high resolution. This technique can also produce high-resolution images of an entire 3D structure, enabling detailed visualization of cellular and subcellular features within tissues. Unlike conventional wide-field microscopes, LSFM images are resistant to out-of-focus blurs, revealing more detailed features and offering better contrast.

However, LSFM is not without limitations^54,55. Its axial resolution is generally lower than that of CFM, typically not exceeding 1 μm, whereas CFM and super-resolution techniques such as stimulated emission depletion (STED) microscopy can achieve much finer axial resolution. Additionally, the orthogonal configuration of LSFM poses challenges in sample preparation, limiting the use of standard Petri dishes and multi-well plates for sample storage. This restricted setup also complicates integration with other manipulation tools, significantly limiting its applicability in intracellular micromanipulation.

Despite these challenges, LSFM has revolutionized biological imaging by enabling fast, high-resolution, and minimally invasive imaging of large and complex specimens. It has been particularly instrumental in advancing developmental biology and neuroscience. In developmental biology, LSFM has been used to monitor morphological changes in Drosophila embryos at various developmental stages, including germ band retraction, segment formation, and dorsal closure⁵⁶. This capability allows researchers to track cellular dynamics and gene expression patterns in real time over extended periods, providing crucial insights into developmental processes. In neuroscience, LSFM has been utilized with dual orthogonal detection and illumination objectives to capture neural wiring during the brain development of Caenorhabditis elegans. Using LSFM, researchers have mapped entire neural networks and tracked neuronal activity in live brain tissue. For example, LSFM has been employed to observe brain function in larval zebrafish⁵⁷, capturing neural activity at high spatial and temporal resolution. These examples highlight LSFM’s transformative impact on understanding dynamic biological processes.

2.2. Super-Resolution Microscopies

2.2.1. Structured illumination microscopy.

Structured Illumination Microscopy (SIM) ^58–61 is a super-resolution technique that improves the resolution of a standard wide-field microscope by using a structured light pattern to illuminate the sample. By capturing multiple images under different illumination patterns and applying computational reconstruction, SIM achieves finer detail and doubles the resolution compared to traditional imaging methods. As illustrated in Figure 3b, SIM enhances resolution by projecting a known pattern of light, such as stripes or grids, onto the sample. This patterned illumination interacts with the sample’s fluorescent structures, generating moiré patterns that encode high-frequency information beyond the system’s diffraction limit. Multiple images are acquired with varying pattern orientations and phases. These raw images are computationally processed and combined to reconstruct a super-resolved image, revealing details that exceed the resolution of standard optical microscopy.

SIM achieves approximately twice the resolution of conventional light microscopy, with lateral resolution typically around 100 nm in standard setups. This improvement allows researchers to visualize finer details within cells and tissues. Additionally, SIM uses relatively low light intensity, making it less phototoxic and suitable for live-cell imaging over extended periods. Unlike STED, which requires high-intensity laser depletion, or SMLM (e.g., PALM/STORM), which relies on sparse activation of fluorophores and long acquisition times, SIM does not require specialized fluorophores, allowing compatibility with standard fluorescent dyes and proteins. Furthermore, SIM is relatively fast compared to other super-resolution methods, making it ideal for dynamic studies in live samples.

However, SIM has limitations. Its high-resolution capability is sensitive to slight positional movements and fluctuations in light intensity, making it prone to artifacts if the sample is not stable. In contrast, techniques like SMLM and STED, though also sensitive to sample drift, can sometimes tolerate minor fluctuations through post-processing corrections. When the signal-to-noise ratio is low, samples often need to be fixed during SIM imaging, limiting its utility for live-cell or dynamic intracellular observations. Additionally, while SIM surpasses confocal microscopy in resolution, it does not offer the same optical sectioning capability as multiphoton imaging techniques like TPFM, which provide deeper tissue penetration.

Despite these challenges, SIM remains a powerful tool for super-resolution imaging across various biological applications. In cell biology, SIM has been widely used to study cellular structures such as microtubules, actin filaments, and mitochondria in high detail. It has provided greater clarity in observing interactions between cytoskeletal components and organelles⁶⁰. In virology, SIM has enabled detailed studies of virus structure and assembly within host cells. For example, it has been used to investigate the entry and replication mechanisms of influenza and HIV, offering critical insights for developing antiviral strategies. In cancer research, SIM has been utilized to examine interactions between cancer cells and the extracellular matrix, as well as cell-cell junctions in tumor tissues. These studies have advanced our understanding of metastasis and tumor cell behavior⁶⁸.

SIM’s ability to combine high resolution, compatibility with standard fluorophores, and relatively rapid imaging makes it an essential tool for super-resolution microscopy, particularly in fields requiring detailed visualization of biological processes.

2.2.2. Stimulated emission depletion microscopy.

Stimulated Emission Depletion (STED) Microscopy is an advanced fluorescence microscopy technique that uses two laser beams to improve resolution beyond the limits of confocal microscopes, enabling researchers to visualize structures at the nanoscale^3,69–72. As shown in Figure 3b, the working principle of STED microscopy employs two synchronized laser beams to achieve super-resolution imaging. The first laser, the excitation beam, excites fluorescent molecules within the sample, while the second laser, the depletion (or STED) beam, selectively suppresses fluorescence from areas outside a small, focused spot. The depletion beam is donut-shaped, with zero intensity at its center. Consequently, only the fluorophores located at the very center of the donut remain fluorescent and are detected, resulting in a highly localized fluorescent spot. By scanning this focused spot across the sample, STED microscopy generates high-resolution images of the entire field of view. This technique typically achieves resolutions of 20–50 nm, significantly surpassing the diffraction limit.

STED microscopy is particularly well-suited for studying subcellular structures due to its ability to provide resolution far below the diffraction limit. Unlike SMLM, which requires stochastic fluorophore activation and extensive post-processing, STED provides real-time super-resolution imaging, making it advantageous for studying dynamic cellular processes. Additionally, it can be used with a broad range of conventional fluorescent dyes, eliminating the need for specialized fluorophores, which is a limitation in techniques such as SMLM, making it highly adaptable to various biological applications. Although STED requires high-intensity light, it is compatible with live-cell imaging under certain conditions, enabling the study of dynamic processes that demand high spatial resolution. Moreover, the high-resolution images produced by STED microscopy allow for quantitative measurements of molecular distributions and distances within cells, offering critical insights into molecular interactions and structural organization. However, the high-powered depletion laser required to achieve super-resolution introduces a risk of photobleaching, making photodamage a significant concern for STED microscopy.

STED has become an indispensable tool across a range of scientific disciplines, enabling researchers to investigate biological processes and molecular architectures previously beyond reach. In neuroscience, STED microscopy has been extensively employed to study synaptic structures at the nanoscale. For instance, it has been used to capture high-resolution images of living hippocampal primary neurons, demonstrating its ability to provide significantly greater detail than confocal fluorescence microscopy. In cell biology, STED has facilitated studies of cytoskeletal components, nuclear pores, and membrane-bound proteins with unprecedented clarity. It has enabled researchers to observe the organization of microtubules, actin filaments, and other intricate cellular structures that are challenging to resolve with conventional microscopy. In virology, STED has advanced our understanding of viral particles, their mechanisms of entry into host cells, and their interactions with cellular machinery. This detailed imaging has been crucial for exploring the structural dynamics of viruses, such as HIV and influenza, during the infection process.

By providing super-resolution images, STED microscopy continues to be a transformative tool for exploring the complexities of biological systems at the nanoscale.

2.2.3. Stochastic optical reconstruction microscopy.

By precisely localizing individual fluorophores within a sample, Stochastic Optical Reconstruction Microscopy (STORM)^15,62–66 is a super-resolution imaging technique that enables researchers to surpass the diffraction limit. As illustrated in Figure 3b, STORM achieves super-resolution by stochastically activating and imaging individual fluorescent molecules in a sample, one at a time, and reconstructing a high-resolution image based on the precise localization of these molecules.

STORM relies on fluorescent probes capable of switching between “on” and “off” states. In each imaging cycle, only a sparse subset of fluorophores is activated to avoid overlap, enabling precise localization of individual molecules. When activated, each molecule emits light, which is detected as a diffraction-limited spot on the detector. The center of this spot is calculated to determine the molecule’s exact location. This process is repeated across thousands of imaging cycles, with different subsets of fluorophores activated each time. By combining the localizations from these cycles, STORM reconstructs a high-resolution image, achieving a resolution of approximately 20–30 nm, about 10 times better than conventional light microscopy.

STORM offers several advantages. One of its primary strengths is its ultrahigh resolution, reaching as fine as 20 nm — comparable to STED and SIM, which is typically limited to ~100 nm. This nanoscale resolution makes STORM suited for studying molecular structures and interactions beyond the reach of diffraction-limited optical imaging. In terms of versatility, STORM is compatible with standard fluorophores, particularly photo-switchable ones, making it suitable for a variety of biological applications. Additionally, STORM enables multicolor imaging with different fluorophore pairs, which facilitates the study of multiple molecular targets within the same sample.

STORM’s ability to achieve nanoscale resolution has significant applications across various fields: In neuroscience, STORM has been extensively used to study synaptic structures. It enables mapping the organization of proteins at synaptic sites, revealing the nanoscale architecture of neuronal synapses. These insights into the spatial distribution of synaptic proteins and their interactions are essential for understanding neural signaling. In cellular research, STORM has been used to visualize fine cellular structures such as the cytoskeleton, nuclear pores, and cell membrane proteins. Researchers have mapped the arrangement of actin, microtubules, and other cytoskeletal elements, which are critical for maintaining cell shape and function. In membrane biology, STORM has provided valuable insights into the organization of proteins and lipids within the plasma membrane. It has been used to study receptor clustering, protein distribution, and nanoscale membrane domains like lipid rafts, deepening our understanding of membrane organization and function.

By offering unprecedented resolution and the ability to track individual molecules, STORM has transformed our understanding of cellular processes and molecular mechanisms, making it a cornerstone technique in super-resolution microscopy.

2.2.4. Photoactivated localization microscopy.

By leveraging photoactivatable fluorescent molecules, photoactivated Localization Microscopy (PALM)^14,67–69 is another super-resolution imaging technique that surpasses the diffraction limit of conventional microscopy. This approach allows researchers to visualize cellular structures with a resolution of approximately 10–20 nm, revealing details that are not accessible under standard light microscopy. Unlike STED microscopy, which requires high-powered depletion lasers that can lead to phototoxicity, PALM relies on lower-intensity activation lasers, reducing photodamage and making it more suitable for live-cell imaging under controlled conditions.

PALM achieves high-resolution imaging by utilizing photoactivatable fluorescent proteins or dyes that can switch between “on” and “off” states. A low-intensity activation laser randomly activates a sparse subset of fluorophores, ensuring that only a few isolated molecules emit light at any given time. This controlled activation prevents overlap between fluorescent spots, enabling accurate localization of each molecule. Once activated, each molecule emits light that forms a diffraction-limited spot on the detector. By fitting these spots to Gaussian curves, the precise center of each spot is calculated to determine the molecule’s exact position. This process is repeated over thousands of frames, with different subsets of fluorophores activated in each frame. The positions of all localized molecules are then combined to reconstruct a super-resolved image with nanometer-scale resolution.

PALM is particularly valuable for single-molecule tracking and analysis, making it an ideal tool for studying the behavior of individual proteins or nucleic acids within cells. In neuroscience, PALM has been extensively used to map synaptic proteins and their arrangements within neuronal structures, providing insights into the molecular architecture of synapses. These studies have advanced our understanding of the spatial organization of neurotransmitter receptors and their roles in neural signaling. In cell biology, researchers have applied PALM to visualize cytoskeletal components⁷⁰, such as actin filaments and microtubules, with nanometer precision. This has been critical for exploring cell shape, mechanics, and migration at a molecular level. Furthermore, PALM has been employed to study the clustering and distribution of membrane proteins, such as receptors and ion channels, on the cell surface. These investigations have shed light on how proteins are organized within the plasma membrane and how they interact with signaling molecules¹⁴.

By comparison, both PALM and STORM achieve super-resolution imaging by precisely localizing single fluorescent molecules, but they employ distinct strategies for fluorophore activation, labeling, and localization, making each uniquely suited to specific applications.

STORM, relies on organic fluorophores conjugated to antibodies that label endogenous molecules within fixed samples. Instead of relying on photobleaching, STORM takes advantage of the stochastic blinking behavior (photoblinking) of organic fluorophores, which repeatedly switch between “on” and “off” states. This blinking behavior enables neighboring molecules to be spatially separated over time, as only a subset of fluorophores is active in each imaging cycle. The use of brighter and more stable organic dyes allows STORM to achieve high-resolution imaging with reduced photobleaching, making it particularly suitable for fixed-cell imaging of endogenous proteins. However, the labeling and sample preparation requirements make STORM less suited for live-cell imaging.

PALM’s reliance on genetically encoded proteins makes it a powerful tool for live-cell imaging and dynamic studies, while STORM’s use of stable organic dyes makes it better suited for high-resolution imaging in fixed samples. Together, these techniques provide complementary approaches for studying nanoscale structures and interactions in biological systems, enabling breakthroughs in understanding molecular processes at unprecedented resolution.

2.3. Raman Microscopies

2.3.1. Coherent anti-stokes Raman spectroscopy.

Coherent Anti-Stokes Raman Spectroscopy (CARS)^25–30 is a nonlinear optical technique that enhances the sensitivity of Raman spectroscopy by leveraging coherent light interactions. In CARS, two laser beams—the pump beam and the Stokes beam—simultaneously interact with the sample. The pump beam excites molecular vibrations by promoting the molecules to a virtual energy state, while the Stokes beam stimulates these vibrations, resulting in the generation of a coherent anti-Stokes signal. This anti-Stokes signal is produced at a higher energy than the pump and Stokes photons and is detected as a Raman scattering signal. The coherent nature of CARS significantly amplifies the Raman signal, enabling the detection of lower concentrations of analytes compared to traditional Raman spectroscopy.

CARS offers several advantages, including high sensitivity, label-free imaging, and chemical specificity. As a coherent process, CARS produces intense signals, allowing for rapid and high-resolution imaging, which is particularly suitable for real-time observations. Since CARS targets intrinsic molecular vibrations directly, it eliminates the need for external labels, making it an excellent choice for minimally invasive studies of biological tissues and live cells. Additionally, CARS selectively visualizes specific bonds, such as C-H or O-H, enabling the creation of detailed chemical maps within complex samples. Its near-infrared operating range further reduces photodamage, making it ideal for live imaging and sensitive biological applications.

In practice, CARS has found advanced applications in biomedical research, cancer diagnostics, and materials science. In cellular biology, CARS has been widely used to study lipid storage dynamics, providing insights into lipid metabolism without the use of dyes. For example, Zumbusch et al. demonstrated the utility of CARS in live-cell imaging by observing lipid droplets in living cells without labeling⁷¹. This work offered significant contributions to understanding cellular metabolism and lipid storage mechanisms. In oncology, CARS has shown promise as a non-invasive diagnostic tool. For instance, Tao et al. employed CARS to detect cancerous tissues by analyzing the biochemical composition of tissues²⁵. This approach highlighted the potential of CARS in cancer diagnostics, offering detailed molecular insights without requiring invasive procedures.

With ongoing technological advancements, CARS continues to expand its impact across multiple disciplines, providing researchers with a powerful tool for investigating molecular-level phenomena and driving innovation in fields such as biology, medicine, and materials science.

2.3.2. Stimulated Raman scattering microscopy.

Stimulated Raman Scattering (SRS) ^{24,26,27,30,72–76} is a highly powerful and rapidly evolving vibrational imaging technique that enables real-time mapping of chemical bond distributions in three-dimensional space. As a label-free imaging method, SRS enhances Raman scattering by exploiting the stimulated emission of molecular vibrations, offering high sensitivity and chemical specificity.

The working principle of SRS is illustrated in Figure 3c. Each vibrational excitation event results in the loss of one photon from the pump beam (stimulated Raman loss) and the gain of one photon by the Stokes beam (stimulated Raman gain), in accordance with the principle of energy conservation. The SRS signal is detected as a relative intensity change in the incident laser, with sensitivity levels as low as ΔI/I ~ 10⁻³ to 10⁻⁷. To achieve precise detection above the low-frequency noise of the laser, one of the laser beams is modulated at a high frequency (megahertz), and the modulation transfer to the other beam is measured using a radio-frequency lock-in technique. This approach effectively eliminates noise caused by slow laser intensity fluctuations, achieving near shot-noise-limited sensitivity. Three-dimensional images are generated by raster-scanning the laser focus across the sample using a laser-scanning microscope configured for either transmitted or reflected light detection.

SRS microscopy offers several key advantages. Its high sensitivity enables rapid imaging, facilitating real-time visualization of molecular distributions. Unlike TPFM and CFM, which rely on exogenous fluorophores that may alter biological function, SRS provides label-free contrast, making it ideal for studying native biomolecular compositions. SRS detects intrinsic vibrational contrasts of molecules, such as lipids and proteins, without requiring fluorescent labels. This preserves sample integrity and reduces preparation time. Compared to CARS microscopy, SRS eliminates non-resonant background signals, resulting in clearer, more quantitative imaging with improved spectral fidelity. Furthermore, unlike STED or STORM, which achieve nanometer-scale resolution at the cost of long acquisition times and photobleaching risks, SRS enables rapid imaging at video-rate speeds, facilitating real-time biological observations.

The applications of SRS microscopy are diverse, spanning biomedical research, cancer diagnostics, and materials science. In cellular biology, SRS has been extensively employed to study lipid dynamics in live cells, providing critical insights into metabolic processes such as lipid storage and mobilization. For instance, researchers have used SRS microscopy to visualize lipid droplets in live cells, mapping their distribution and quantifying lipid storage and mobilization in real time⁷⁷. This research has advanced understanding of lipid metabolism under various physiological conditions, including obesity and metabolic disorders, by tracking lipid accumulation and release in individual cells without fluorescent tagging.

In cancer research, SRS has been applied to differentiate tumor tissues from healthy ones by revealing biochemical composition differences that are valuable for diagnostic purposes⁷⁸. Additionally, SRS has been used to study polymers, pharmaceuticals, and complex chemical environments, achieving high-resolution chemical mapping of diverse materials.

Ongoing advancements in SRS systems continue to enhance its capabilities, further solidifying its role as a versatile and indispensable tool for scientific research and medical diagnostics. SRS microscopy’s combination of speed, sensitivity, and label-free imaging makes it a valuable technique for exploring molecular-level processes across a wide range of disciplines.

2.3.3. Scanning probe microscopies and tip-enhanced Raman spectroscopy.

Scanning probe microscopies, particularly Tip-Enhanced Raman Spectroscopy (TERS)^79–84, represent advanced techniques for nanoscale bioimaging, combining the molecular specificity of Raman spectroscopy with the spatial resolution of scanning probes. TERS enhances traditional Raman spectroscopy by utilizing a sharp metallic tip, typically made of gold or silver, positioned extremely close to the sample surface. This setup, as illustrated in Figure 3c, capitalizes on the tip’s plasmonic properties to amplify the local electromagnetic field. When laser light is focused onto the sample, the enhanced electromagnetic field around the tip significantly increases the Raman scattering signal from nearby molecules. This coupling of light with surface plasmons improves the Raman signal, enabling the detection of single molecules or nanostructures. TERS achieves spatial resolutions down to the nanometer scale, far surpassing the capabilities of conventional Raman spectroscopy.

TERS offers several key advantages over traditional Raman spectroscopy. Unlike traditional Raman spectroscopy, which is often limited by weak signals and diffraction-limited spatial resolution (~200–300 nm), TERS enhances signal intensity by several orders of magnitude, achieving single-molecule sensitivity. This allows researchers to analyze the molecular composition of materials with unprecedented detail, making it essential for studying nanoscale materials, biomolecules, and structural heterogeneities in complex systems. Furthermore, TERS provides chemical specificity through vibrational modes, enabling the identification of molecular species and structural features in heterogeneous samples. The technique also boasts a low detection limit, facilitating the identification of trace amounts of substances. In contrast, while SRS and CARS also offer chemical contrast, their spatial resolution is diffraction-limited, making them less effective for imaging nanoscale structures. Additionally, TERS can be conducted in various environments, including aqueous solutions and ambient air, broadening its applications across materials science, biology, and nanotechnology.

Numerous studies have highlighted the power and versatility of TERS across diverse research domains. In biological and medical research, TERS has shown great promise. For example, TERS has successfully probed single DNA molecules⁸², providing critical information about molecular conformations and interactions. Similarly, TERS has been utilized to study catalytic processes on metal surfaces, revealing mechanisms underlying heterogeneous catalysis with atomic-level precision. TERS has also been applied to distinguish cancerous cells from normal cells in biological tissues by analyzing their vibrational signatures. This approach has enabled the identification of specific biochemical markers associated with cancer, demonstrating the potential of TERS in medical diagnostics. By combining nanoscale spatial resolution with chemical specificity, TERS continues to push the boundaries of molecular imaging and analysis. Its ability to operate in diverse environments and its sensitivity to trace-level substances make it an indispensable tool for advancing research in fields ranging from materials science to biology and medicine.

2.3.4. Surface-enhanced Raman spectroscopy.

Surface-Enhanced Raman Spectroscopy (SERS) ^85–88 is a powerful analytical technique that enhances traditional Raman spectroscopy by leveraging roughened metal surfaces, typically composed of gold or silver, to amplify the Raman signal of molecules adsorbed onto these surfaces. As illustrated in Figure 3c, SERS operates on the fundamental principle of the interaction between incident laser light and the surface plasmons of metallic nanostructures, resulting in a significant enhancement of the electric field near the surface. This enhancement arises from two primary effects: electromagnetic enhancement, which is attributed to the localized surface plasmon resonance of the metal, and chemical enhancement, which stems from charge transfer interactions between the adsorbed molecules and the metal surface. These combined effects enable even low concentrations of analytes to produce highly intensified Raman signals, allowing for ultrasensitive detection and characterization of various chemical species.

SERS offers several advantages over traditional Raman spectroscopy, making it a preferred choice for many analytical applications. Its exceptional sensitivity, enabling the detection of molecules at femtomolar concentrations, is particularly valuable in fields such as biomedical diagnostics, environmental monitoring, and food safety, where detecting low levels of analytes is critical. While SRS and CARS offer label-free chemical imaging with improved speed and reduced photodamage, they are diffraction-limited and lack the extreme sensitivity of SERS. Additionally, SERS provides molecular-specific information by probing vibrational modes, facilitating the identification of chemical structures and functional groups. Its versatility extends to a wide range of samples, including solids, liquids, and gases, and it can be performed in situ or ex-situ, broadening its applicability. On the other hand, TERS provides higher spatial resolution (~10 nm), but it requires precise tip alignment and is limited in its ability to analyze large sample areas. Furthermore, SERS is a rapid technique requiring minimal sample preparation, making it ideal for high-throughput screening applications.

Studies have demonstrated the efficacy of SERS in various fields, underscoring its potential for groundbreaking research and practical applications. In biomedical diagnostics, for example, a study by Chen et al. utilized SERS to detect cancer biomarkers in serum samples, achieving detection limits that surpassed those of traditional diagnostic techniques⁸⁹. This work emphasized the utility of SERS in early cancer detection and personalized medicine. In environmental science, SERS has proven effective in monitoring pollutants, such as detecting pesticide residues in agricultural products. This capability supports environmental safety and regulatory compliance. Additionally, in pathogen detection, SERS has been instrumental in identifying harmful bacteria. For instance, Zhao et al. employed SERS to detect pathogenic bacteria⁹⁰, such as Escherichia coli and Salmonella, in food samples. By using gold nanoparticles to enhance the Raman signal of specific bacterial biomarkers, the study demonstrated rapid and sensitive detection, highlighting the technique’s potential in ensuring food safety.

SERS’s combination of high sensitivity, molecular specificity, and versatility continues to make it a transformative tool across diverse fields, driving advancements in research, diagnostics, and industrial applications. Its ability to provide detailed chemical insights while enabling rapid, low-concentration detection ensures its continued relevance in addressing critical scientific and societal challenges.

2.4. NanoIR Spectroscopy

NanoIR spectroscopy, also known as infrared nanospectroscopy, is an advanced technique that combines atomic force microscopy with infrared spectroscopy to achieve high-resolution chemical characterization beyond the diffraction limit of traditional IR spectroscopy^91–97. The core principle of NanoIR is based on either photothermal-induced resonance (AFM-IR) or scattering-type near-field optical microscopy (s-SNOM). In the AFM-IR approach, an AFM tip detects the localized thermal expansion of a sample upon infrared absorption, translating molecular vibrational signatures into high-resolution chemical maps. Alternatively, in s-SNOM, an AFM tip interacts with the scattered IR light to obtain phase and amplitude information, enhancing sensitivity to surface chemistry and material composition. These approaches enable NanoIR to probe nanoscale variations in molecular structures and chemical environments, making it a powerful tool for studying heterogeneous and hierarchical systems.

One of the major advantages of NanoIR spectroscopy is its nanoscale spatial resolution, which can reach 10–20 nm, far surpassing conventional FTIR and standard Infrared Absorption Microscopy^93–94. This high resolution allows for precise chemical characterization of individual nanoparticles, thin films, and biological structures such as proteins, lipids, and membranes. Additionally, NanoIR provides label-free chemical identification, which is particularly useful in biomedical research where fluorescence labeling can alter biological functions. The ability to correlate nanoscale chemical and mechanical properties using AFM further enhances its utility in material science and biomedical applications. However, NanoIR has limitations, including the requirement for samples with strong IR absorption, potential artifacts from AFM tip-sample interactions, and longer acquisition times compared to conventional spectroscopic techniques. Despite these challenges, continuous advancements in detector sensitivity and data processing have significantly improved the technique’s applicability.

In biological applications, NanoIR spectroscopy has been employed to study intracellular structures, biomembrane composition, protein aggregation, and disease-related biomolecular changes. For example, NanoIR has been used to analyze amyloid fibrils associated with neurodegenerative diseases such as Alzheimer’s and Parkinson’s, providing nanoscale insights into protein misfolding mechanisms⁹⁵. Additionally, researchers have utilized NanoIR to investigate lipid-protein interactions in cell membranes, revealing compositional variations that impact membrane fluidity and function⁹⁶. Another emerging application is in single-cell analysis, where NanoIR enables spatial mapping of biomolecular heterogeneity within individual cells, aiding in cancer diagnostics and personalized medicine. These capabilities highlight the potential of NanoIR to advance our understanding of complex biological systems at the molecular level.

In together, we summarized the key features of above-mentioned different techniques in the Table 1.

Table 1.

Summary of different techniques features.

Technique	Activation/Labeling Method	Localization Mechanisms	Applications
Confocal Fluorescence Microscopy	Fluorophores with laser scanning	Pinhole-based optical sectioning	High-resolution imaging of cellular structures
Two-Photon Fluorescence Microscopy	Fluorophores with near-infrared lasers	Two-photon excitation for deeper imaging	Deep-tissue and in vivo imaging
Light Sheet Fluorescence Microscopy	Selective plane illumination	Optical sectioning with a thin light sheet	Fast imaging of large samples, developmental biology
Structured Illumination Microscopy (SIM)	Fluorophores (various dyes, proteins)	Moiré pattern analysis (structured light)	Live-cell and deep-tissue imaging with moderate resolution enhancement
Stimulated Emission Depletion Microscopy (STED)	Fluorophores with depletion lasers	Point-spread function reduction by stimulated emission	High-resolution imaging of subcellular structures
Stochastic Optical Reconstruction Microscopy (STORM)	Organic fluorophores on antibodies (immunolabeling)	Stochastic blinking (photoblinking)	Fixed-cell imaging, endogenous protein studies
Photoactivated Localization Microscopy (PALM)	Photoactivatable fluorescent proteins (genetically encoded)	Stochastic activation and photobleaching	Live-cell imaging, protein-specific localization
Coherent Anti-Stokes Raman Spectroscopy (CARS)	Molecular vibrations excited by two lasers	Nonlinear Raman scattering	Label-free imaging of lipids, cells, and tissues
Stimulated Raman Scattering (SRS) Microscopy	Molecular vibrations excited by stimulated Raman processes	Signal amplification for higher sensitivity	Chemical imaging of biomolecules and metabolites
Tip-Enhanced Raman Spectroscopy (TERS)	Nanometer-scale metallic tip with Raman excitation	Localized surface plasmon enhancement	Nanoscale chemical imaging, catalysis studies
Surface-Enhanced Raman Spectroscopy (SERS)	Raman enhancement using metallic nanoparticles	Localized surface plasmon resonance	Ultra-sensitive detection of biomolecules
NanoIR Spectroscopy	Infrared absorption combined with atomic force microscopy	Nanometer-scale IR absorption mapping	Chemical characterization of nanomaterials and biological samples

3. Data processing and AI-assisted bioimaging

The primary goal of chemical imaging in biological systems is to identify and map local elements, molecules, compositions, cells, tissues, organics, and other critical components. However, the inherent heterogeneity and complexity of cellular environments generate massive amounts of raw data, often accompanied by noise and interference. A single experiment can produce gigabytes or even terabytes of data, typically organized into multidimensional, multivariable matrices. Repeated measurements across diverse samples under varying controlled conditions further expand these datasets. Additionally, advancements in spatial and temporal resolution increase the dimensionality of data from the same sample regions. As a result, efficient data analysis techniques have become as crucial as experimental advancements in managing and interpreting this “big data.”

With the rising prominence of artificial intelligence (AI) approaches such as data mining and machine learning, a variety of algorithms and models are now being employed to process such complex datasets. Unlike traditional methods, these advanced techniques excel at handling high-dimensional and nonlinear data. While the specifics of data analysis depend on the objectives, the process typically includes data pre-processing, pattern recognition or visualization, clustering, classification, and prediction. Selecting the appropriate algorithms and models is highly context-dependent and relies on researchers’ expertise and the nature of the data. Thus, reviewing state-of-the-art data processing methods is essential to provide researchers with a comprehensive reference for navigating this field.

The overall workflow of chemical imaging involves several distinct steps, including data acquisition, data pre-processing (optional), data processing (target-based), and the final application of the results, as shown in Figure 2b. Each step plays a crucial role in ensuring that high-quality, accurate chemical information is obtained from the imaging data. Data acquisition typically involves capturing spectral and spatial information from chemical samples, which can be achieved using various imaging techniques such as Raman spectroscopy, hyperspectral imaging, or mass spectrometry. Depending on the nature of the dataset and the specific technique employed, data pre-processing may be necessary to clean or enhance the raw data, including noise reduction or spectral unmixing, etc. The data processing stage is where AI and machine learning methods come into play, with approaches varying based on the target application. For instance, supervised learning methods, such as neural networks, require labeled training sets to model relationships within the data and make predictions. On the other hand, unsupervised techniques, such as clustering or dimensionality reduction, may not need predefined labels but instead extract inherent patterns or features from the data. The ultimate goal of the data processing stage is to enable accurate classification, segmentation, or quantification of chemical information. Finally, the processed data is applied to real-world scenarios, whether in biomedical research, industrial quality control, or environmental monitoring. Given the diversity of imaging methods and applications, the workflow must remain flexible and adapt to the specific demands of each process.

It is important to distinguish between “data mining” and “machine learning,” as these terms are often used interchangeably but represent distinct concepts. While they share some overlap, they are not synonymous. To maintain clarity and organization in this review, the discussion is structured around the general stages of data analysis rather than specific algorithms. The subsequent sections begin with a brief overview of the fundamental concepts of data mining and machine learning. This is followed by dedicated sections on data pre-processing, pattern extraction, and prediction methods, as illustrated in Figure 4. While not all stages apply to every study, their inclusion depends on the specific goals and requirements of the research.

Figure 4: Summary of data-processing methods in imaging.

(a) Denoise: Comparison of Raman spectra with Savitzky-Golay filtering and detrending under low laser excitation conditions. Reproduced from reference [98] with permission from [Elsevier], copyright [2021]. (b) Spike removal: Raman spectra from a cellular system. Reproduced from reference [92] with permission from [Springer Nature], copyright [2019]. (c) Normalization: Normalized Raman spectra with different PQT-12 polymer layers. Reproduced from reference [104] with permission from [AIP Publishing], copyright [2014]. (d) PCA: Score plot of IMS data from normal and the tumor tissue sections, showing complete and reduced data sets in blue and red, respectively. Reproduced from reference [108] with permission from [American Chemical Society], copyright [2013]. (e) VCA: Retrieved VCA image with three endmembers from atherosclerotic rabbit aorta tissue samples. Reproduced from reference [116] with permission from [WILEY], copyright [2015]. (f) Pipeline for Identifying Cancer Driver Modules by Graph Embedding and Hierarchical Clustering (ICDM-GEHC). Reproduced from reference [130] with permission from [Springer Nature], copyright [2024]. (g) Regression: Schematic illustration of pixel-wise LASSO spectral unmixing for chemical mapping generation. Reproduced from reference [132] with permission from [American Association for the Advancement of Science], copyright [2023]. (h) Architecture of a hybrid HCNN-KNN Model for age estimation in orthopantomography. Reproduced from reference [144] with permission from [Frontiers], copyright [2022]. (i) t-SNE visualization of top-ranked analogue of Palbociclib and analogues of Ribociclib. Reproduced from reference [152] with permission from [MIT Press, Microtome Publishing, JMLR, Inc.], copyright [2008]. (j) Neural network architecture with multiple hidden layers, each adopting potentially different activation functions. Reproduced from reference [155] with permission from [Taylor & Francis Online], copyright [2024]. (k) Improved CNN architecture for recognizing and classifying biological images. Reproduced from reference [159] with permission from [Innovative Information Science & Technology Research Group], copyright [2025]. (l) Optimized deep neural network architecture for gene expression analysis. Reproduced from reference [163] with permission from [American Chemical Society], copyright [2022].

3.1. General concept of data mining and machine learning

Data mining involves uncovering patterns, relationships, anomalies, and insights from large datasets. It focuses on extracting valuable and previously unknown information from raw data through a combination of statistical analysis and computational algorithms. The insights derived from data mining are instrumental in informed decision-making, problem-solving, process optimization, and strategy development.

Machine learning (ML), a subset of AI, is dedicated to developing algorithms and statistical models that enable computers to learn from data and make predictions or decisions without explicit task-specific programming. Machine learning methods are generally categorized into four types: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning.

Although data mining and ML share common ground—especially in pattern recognition—they serve different purposes. For instance, Principal Component Analysis (PCA) is used in both fields but with distinct objectives. In data mining, PCA aids in dimensionality reduction, anomaly detection, and data visualization, revealing inherent structures within complex datasets. In ML, PCA serves as a preprocessing step, transforming data into a lower-dimensional space to enhance feature extraction, reduce overfitting, and improve computational efficiency.

Despite their overlap, the two fields emphasize different goals and approaches. The strength of data mining lies in its ability to uncover patterns and insights without prior assumptions, though interpreting these findings requires the researcher’s domain knowledge and understanding of the underlying physical phenomena. In contrast, machine learning emphasizes enabling computers to “learn” from data and independently generate predictive or prescriptive insights. Once an ML model is trained, it typically operates autonomously, producing outputs based on the data it has been exposed to, without further human intervention.

3.2. Data Pre-processing

Raw data obtained from chemical imaging experiments are typically represented as multidimensional matrices, with two or three dimensions corresponding to spatial locations and at least one dimension reflecting intensity-related information, such as Raman or fluorescence signals. Before advanced data analysis can proceed, fundamental preprocessing steps are required to refine this raw data. These steps include denoising, baseline correction, data unfolding, spike removal, scaling, and normalization, among others. The choice of preprocessing techniques depends on the specific experimental method. This section introduces several classical methods, as summarized in the first row of Figure 4.

Denoising.

Noise, representing random disturbances unrelated to the variables of interest, can obscure underlying patterns in data. In fluorescence microscopy, noise primarily arises from shot noise, which is inherent to the properties of light, and detector noise, stemming from the electronics used for detection⁹¹. In Raman spectroscopy, the two primary types of noise are heteroscedastic and homoscedastic. Heteroscedastic noise often appears as a baseline signal due to fluorescent contamination, while homoscedastic noise, or shot noise, is attributed to the instrument itself ⁹². Common denoising methods include smoothing filters (Gaussian, median, and mean filleter), frequency domain filtering, more sophisticated Non-local means (NLM) ⁹³, wavelet transform ⁹⁴, and deep learning-based denoising methods (e.g., 3D-RCAN ⁹⁵ and DenoSeg ⁹⁶). Methods such as the Savitzky-Golay smoothing and differentiation filter ⁹⁷, are often used as a preprocessing in spectroscopy and signal processing. For example, Zhao et al.⁹⁸ applied a Savitzky-Golay filter with a smoothing number of 31 and a polynomial order of 3 to Raman spectra, which can significantly smooth raw data while maintaining the integrity of the main peak, as shown in Figure 4a. The Signal-to-Noise (SNR) and mean square error (MSE) are significantly augmented and attenuated by 35 times and 15 times, respectively. However, despite this improvement, only the main peak at 672 cm⁻¹ is clearly visible, while other peaks remain obscured by background interference. This limitation highlights the impact of low laser excitation intensity in the experimental setup. By selecting appropriate denoising techniques, researchers can significantly enhance signal clarity, enabling better interpretation and subsequent analysis.

Baseline Correction.

The baseline refers to non-specific background signals or interference that can obscure the true spectral features of interest. Removing the baseline is essential for the accurate analysis and interpretation of chemical imaging data. Studies have been developed to improve baseline correction techniques. For example, Zhan et al.⁹⁹ introduced the Morphologically Weighted Penalized Least Squares algorithm, which uses a penalized least square function to estimate a rough background. This method has been applied to chromatographic data and Raman spectra, producing competitive results that are relatively insensitive to parameter settings, though it can be time-consuming. Hu et al.¹⁰⁰ proposed an improved piecewise baseline correction method for Raman spectroscopy using polynomial fitting, incorporating iterative optimization to eliminate discontinuities between segments. By selecting appropriate polynomial orders for each section, their approach mitigates the Runge Phenomenon and improves accuracy, demonstrating superior adaptability and robustness compared to traditional methods.

Spike Removal.

Spike removal refers to the process of identifying and eliminating sudden, isolated peaks or spikes in spectral data that arise due to instrumental artifacts, outliers, or other sources of interference. For example, a direct comparison of spectra⁹² after spike removal is presented in Figure 4b, where two outstanding outliers at 1500 and 1580 cm⁻¹ are eliminated and the baseline is corrected at the same time, making a “cleaner” spectra. These spikes are a form of short noise that can be addressed using the aforementioned filters or smoothing techniques. Additionally, statistical tests such as Dixon’s Q test, Grubbs’ test, and Tukey’s fences can be applied to detect outliers. More advanced techniques like the Minimum Volume Ellipsoid (MVE), Multivariate Trimming (MVT), and Minimum Covariance Determinant (MCD) employ complex algorithms, which can be challenging to implement and may require significant processing time ¹⁰¹. Very recently, Liu et al. ¹⁰² introduced a novel outlier removal technique named PCA-DBSCAN, designed to enhance blood-based Surface-Enhanced Raman Spectroscopy (SERS) data analysis for cancer screening. This method integrates Principal Component Analysis (PCA) for dimensionality reduction with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for clustering, effectively identifying and eliminating outliers in spectral data. The approach incorporates adjustable parameters, Eps and MinPts, to fine-tune the clustering process. Validation of the PCA-DBSCAN method demonstrated significant improvements in the performance of SERS-based cancer screening models, achieving a macro-average recall of 97.41% and a macro-average F1-score of 97.74%. Additionally, Rhyu et al. ¹⁰³ developed new software for outlier detection and missing data estimation, based on T² and Q contributions, along with several general-purpose algorithms such as PCADA, PPCA, and PPCA-M. Their approach was tested on data from a continuous biomanufacturing pilot facility, particularly monoclonal antibody production in a perfusion bioreactor. The results confirmed that their method improved data quality and predictive accuracy.

Scaling and Normalization.

Scaling and normalization are essential preprocessing steps aimed at reducing systematic variations between spectra or images, enhancing the comparability of spectral features, and improving the performance of analytical models. The fundamental idea is to identify an appropriate scaling factor (denominator) to ensure variables fall within a scalable and comparable range. For example, as demonstrated in Figure 4c, Raman spectra of different layers of PQT-12 polymer can be integrated together after normalization¹⁰⁴, which is convenient for the observation of intensity change with main peaks at 1393 and 1479 cm⁻¹ and new peaks at 1057 cm^-1. Depending on the processing goal, scaling can be applied to rows or columns of data. Common scaling factors include peak value, standard deviation, Manhattan distance, and Euclidean distance. For instance, Min-Max scaling rescales each feature to a range between 0 and 1 by subtracting the minimum value from each data point and dividing by the range (maximum value minus minimum value). Similarly, Z-score scaling involves subtracting the mean value of the data and then dividing by the standard deviation, effectively centering the data around zero with a unit variance.

3.3. Pattern Extraction.

The term pattern refers to inherent trends, relationships, distributions, modes, and features within a dataset that exist prior to analysis. Sometimes, the term “explorative methods” is used to describe pattern recognition algorithms in data mining, while in machine learning, these are often covered by unsupervised methods, including clustering. In this context, any algorithm that helps uncover deeper insights from the original data matrix can be considered a pattern extraction method. These methods are typically categorized into unsupervised and supervised approaches, with unsupervised methods encompassing both explorative techniques and clustering algorithms.

3.3.1. Unsupervised Method.

Several explorative methods are introduced below, along with their respective applications, including Principal Component Analysis (PCA), Independent Component Analysis (ICA), Vertex Component Analysis (VCA), and clustering methods such as K-means and Hierarchical Clustering Analysis (HCA).

Principal Component Analysis.

PCA is one of the simplest techniques for reducing dimensionality and visualizing multidimensional matrices. It is widely used in various applications, particularly in electron and force-based scanning probe microscopy data^105,106. PCA requires no additional parameters beyond the raw data and produces three main outputs: eigenvectors (ranked by information density), corresponding weight (or score) maps, and a Scree plot showing information content versus eigenvector number. Typically, the first two outputs are sufficient to visualize the principal features of the dataset. However, interpreting higher-order eigenvectors becomes increasingly difficult due to their reduced information content, and processing large datasets can be computationally demanding¹⁰⁷.

In PCA, a spectroscopic or imaging dataset is transformed into a linear combination of orthogonal eigenvectors. The covariance matrix, computed from the product of the experimental matrix and its transpose, is used in singular value decomposition to generate the eigenvectors and their associated eigenvalues. These eigenvectors are ordered by their eigenvalues, with the first capturing the greatest variance, followed by the second and so on. This hierarchy allows for efficient data visualization and quantification based on the most significant features of the dataset.

Thomas et al. ¹⁰⁸ introduced a histology-driven imaging mass spectrometry (IMS) approach to identify distinct lipid signatures in human colorectal cancer liver metastasis (CRCLM) biopsies. Utilizing matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF/TOF) mass spectrometry in both negative and positive ionization modes, the researchers analyzed tissue sections after applying a 1,5-diaminonaphthalene matrix via sublimation. To manage the extensive data generated, they implemented a data reduction strategy by randomly selecting spectra from histologically defined regions of interest, achieving a tenfold decrease in data volume while preserving molecular specificity. After PCA extraction, normal and tumor tissue show distinctive patterns as compared in Figure 4d. Their findings confirmed that molecular selectivity in the regions of interest was preserved even after data reduction (blue: complete; red: reduced data). This methodology offers a comprehensive framework for deriving disease-specific lipid signatures directly from IMS datasets, potentially advancing clinical diagnostics and therapeutic strategies. In another study aimed at extracting chemical information from a formalin-fixed and paraffin-embedded rat colon tissue section via a label-free Raman spectroscopic imaging approach, the new variables derived from PCA were linked to spectral information related to the paraffin and specific biochemicals (muscle, mucin, and nuclei). By utilizing the weighting scores, the researchers produced images with improved contrast and sharpness compared to conventional hematoxylin and eosin (H&E) staining, effectively displaying the relative concentration distribution of these biochemicals ¹⁰⁹.This method enables simultaneous extraction of both morphological and chemical information, providing insights into tissue composition without the need for stains or dyes.

Independent Component Analysis and Vertex Component Analysis.

ICA and VCA are similar to PCA in that they generate two matrices: loading variables and weighting scores. However, there are some key differences between these methods. Unlike PCA, the order of the components generated by ICA is not significant, and it typically requires additional input parameters, which generally results in longer processing times compared to PCA ¹⁰⁷. Before applying ICA, the number of independent components must be determined, which is a highly non-trivial task. While the number of principal components to retain in PCA is often overlooked, it plays a crucial role in ICA. ICA aims to decompose the raw dataset into a new linear representation of statistically independent variables, focusing on achieving independence from non-Gaussian data. Consequently, the new variables are designed to minimize the mutual dependence that may exist between them. ICA has been employed in various studies, including investigations of brain function using functional magnetic resonance imaging¹¹⁰ and electroencephalograph¹¹¹, the examination of the distribution of actives and major excipients within tablets ¹¹², and separation of unknown sample mixtures ¹¹³.

The VCA exploits two facts: (1) the reference substances are the vertices of a simplex and (2) the affine transformation of a simplex is also a simplex, and this algorithm is first proposed by Nascimento et al. ¹¹⁴. It is a very useful method to decompose hyperspectral data, with the new variables expected to be pure spectra or components. In their 2017 study, Kallepitis et al. ¹¹⁵introduced a computational framework for label-free quantitative volumetric Raman imaging (qVRI) to analyze three-dimensional (3D) cell cultures. This method combines confocal Raman micro-spectroscopy with VCA to visualize and quantify biomolecular structures within 3D environments without the need for labels or dyes. The researchers applied qVRI to various biological systems, including human pluripotent stem cells and mesenchymal stem cells within biomimetic hydrogels, successfully identifying fine details in cell morphology, such as cytoplasm, nuclei, lipid bodies, and cytoskeletal structures. This approach offers a non-destructive means to obtain high-resolution, biomolecular-specific information, enhancing the understanding of complex cellular mechanisms and cell-material interactions in 3D culture systems. Tabarangao et al. ¹¹⁶ demonstrated the effectiveness of Vertex Component Analysis (VCA) in generating multimodal-like contrast from hyperspectral Coherent Anti-Stokes Raman Scattering (CARS) images. Using an atherosclerotic rabbit aorta test image, they showed that VCA could replicate contrast achieved through multimodal imaging, which typically relies on signals from three separate nonlinear optical techniques. Their results, particularly Figure 4e, indicate that VCA can effectively extract meaningful image contrast from hyperspectral data. Additionally, their findings highlight the utility of Principal Component Analysis (PCA) as an unsupervised method for contrast enhancement in hyperspectral CARS imaging, preserving essential structural and biochemical information.

Cluster Analysis.

In addition to dimensionality reduction and pattern extraction methods, cluster analysis— such as K-means and HCA — is widely used to visualize raw data. Unlike decomposition techniques that create new variables, cluster analysis groups object or pixels based on shared features. A key concept in clustering is the distance measure that determines similarity, with common metrics including Euclidean, Manhattan, Chebyshev, and Mahalanobis distances ¹¹⁷. Since cluster analysis does not provide concentration information, it is useful when relative concentrations are not the main focus.

K-means is a common partitional clustering method that partitions a dataset into K clusters while minimizing the within-group sum of squares. The primary input is the number of clusters, and the algorithm assigns each observation to a cluster. The results can be influenced by parameters like distance metric, iteration count, initial sample selection, and handling of outliers ¹¹⁸.

Various extensions of K-means can be found in the literature ¹¹⁹. and algorithm consists of two steps: first, randomly selecting K initial clusters; second, assigning samples to the nearest cluster centroid and recalculating centroids. Iteration continues until moving an object to a different cluster does not reduce the within-group sum of squares. A notable drawback is that K-means is sensitive to initial cluster selection, which can lead to different results across runs. Insights into cluster stability can be gained by examining the average distance between clusters and the number of points in each cluster.

In their study, Kochan et al.¹²⁰employed K-means cluster analysis with the Euclidean distance metric to analyze the biochemical content of murine liver sinusoidal endothelial cells (LSECs). This analysis was combined with label-free confocal Raman imaging, allowing for a detailed, subcellular-level characterization of LSECs’ chemical composition. The study provided valuable insights into the unique biochemical signatures of LSECs, helping to distinguish them from other liver cells and improving the understanding of their role in liver function. Their approach demonstrated the potential of Raman spectroscopy coupled with chemometric techniques for studying live cells in a non-invasive manner. Similarly, Saxena et al. ¹²¹ utilized computer-assisted K-means analysis to examine individual cells within a tissue sample, exploring their molecular, phenotypic, and functional attributes. This research aids medical science by identifying expressed and non-expressed cells in tissue samples and their spatial locations.

While K-means excels with large datasets, HCA is more effective for smaller datasets. HCA uncovers the underlying structure of objects and outputs a nested, hierarchical set of partitions represented by a dendrogram. This tree diagram displays individual samples at the bottom and a single cluster encompassing all elements at the top ¹¹⁷.

The whole clustering is accomplished through an iterative process that associates (agglomerative methods) or dissociates (divisive methods) object by object, and that is halted when all objects have been processed ¹²². The agglomerative method starts with each object in a separate cluster and then combines the clusters sequentially, reducing the number of clusters at each step until all objects belong to only one cluster; the divisive methods start with all of the objects in one cluster, and then proceed to their partition into smaller clusters until there is one object per cluster ¹²³. Therefore, for N objects, the process involves N ₋ 1 clustering steps.

In HCA, the type of similarity measure between objects and/or groups, and the linkage technique ¹²⁴ are two important choices when defining a method. Euclidean distance and the correlation coefficient are the most common way to determine the similarity between objects, but there are many alternatives for similarity indicators ¹²⁵. Some of the most used linkage algorithms are single linkage (the distance between two groups is the distance between their closest members), complete-linkage (defined as the distance between the two farthest points), Ward’s hierarchical clustering method (at each stage of the algorithm, the two groups that produce the smallest increase in the total within-group sum of squares are amalgamated), centroid distance (defined as the distance between the cluster means or centroids), and so on ¹²⁶. A comprehensive examination of eight widely employed linkage algorithms in HCA resonance Raman spectra of bacteria has been reported by Kniggendorf et al.¹²⁷.

Papin et al. ¹²⁸ utilized hierarchical clustering to identify distinct clinical and biological clusters among sepsis patients. By analyzing data from 6,046 patients in the OUTCOMEREA cohort, they identified six clusters based on characteristics available at ICU admission. These clusters exhibited distinct mortality outcomes, with variations observed at 28-day, 90-day, and one-year intervals. Notably, the study found that these clusters maintained significant differences in mortality rates even after adjusting for the Sequential Organ Failure Assessment (SOFA) score and the year of ICU admission. At the same time, Espinoza et al. ¹²⁹ developed a novel HCA-based clustering approach to investigate the fine-scale structure of cell membranes, aiming to elucidate the role of spatial organization in regulating transmembrane signaling. More recently, Deng et al. ¹³⁰ applied hierarchical clustering analysis (HCA) to subgroup vertices in a weighted Protein-Protein Interaction (PPI) network using Mahalanobis distance. Their method, Cancer Driver Modules based on Graph Embedding and Hierarchical Clustering (ICDM-GEHC), demonstrated superior performance in identifying cancer-related genes. The proposed pipeline, as illustrated in Figure 4f, successfully generated modules with high coverage and mutual exclusivity. These modules were significantly enriched for specific cancer types, showcasing the method’s potential for identifying critical cancer driver modules and advancing cancer research.

3.3.2. Supervised Method.

Regression and classification are both supervised learning tasks but differ in their target variables. Regression predicts continuous numerical values, while classification assigns data points to predefined categories. The choice between them depends on the problem and desired output. Both techniques are widely used in bioinformatics, including tasks like chemical composition decomposition, protein secondary structure prediction, gene expression-based diagnosis, and splice site prediction.

Linear Regression.

Linear regression predicts the value of a dependent variable based on one or more independent variables (predictors). It estimates coefficients for a linear equation that best fits the data, minimizing discrepancies between predicted and actual values using the least squares method. To reduce overfitting and enhance model performance, regularization techniques like LASSO, Ridge regression, and Elastic Net are employed, each incorporating different penalty terms into the objective function ¹³¹.

Least Absolute Shrinkage and Selection Operator (LASSO) adds a penalty term proportional to the absolute values of the coefficients to the linear regression objective function. This L1 regularization term promotes model sparsity by setting some coefficients exactly to zero. In their 2023 study, Tan et al. ¹³² applied LASSO regression to simultaneously map five major biomolecules—proteins, carbohydrates, fatty acids, cholesterol, and nucleic acids—at the single-cell level using hyperspectral Stimulated Raman Scattering (SRS) imaging. The schematic in Figure 4g shows the pixel-wise LASSO regression process, which allowed for improved chemical mapping results. This method outperformed traditional techniques like least squares (LS) fitting and Multivariate Curve Resolution (MCR), offering more accurate and comprehensive results by analyzing multiple species simultaneously within individual cells.

Ridge regression introduces a penalty term proportional to the square of the coefficients’ magnitudes (L2 regularization) to the linear regression objective function. This term shrinks the coefficients toward zero but does not set them exactly to zero. In a study by Guo et al.¹³³, the researchers introduced a novel approach for multi-step influenza forecasting by integrating Singular Value Decomposition (SVD) with Kernel Ridge Regression (KRR), enhanced by MARCOS-guided gradient-based optimization. This methodology effectively captures complex temporal patterns in influenza data, improving the accuracy of long-term forecasts. The authors validated their approach using historical influenza data, which reduced the Root Mean Square Error (RMSE) by 15–20% and Mean Absolute Error (MAE) by 10–15%, demonstrating its superior performance over traditional forecasting models and highlighting its enhanced predictive capability.

Elastic Net merges the penalties of Ridge regression and LASSO, incorporating both L1 and L2 regularization terms into the linear regression objective function. This hybrid approach addresses the limitations of each method and balances variable selection with coefficient shrinkage. Recently, Jiang et al. ¹³⁴ developed a novel approach using single-site-based tissue-specific Elastic Net models to predict tissue-specific methylation levels from gene expression data. The model achieved a 20–30% reduction in Mean Absolute Error (MAE) compared to traditional methods, showcasing enhanced prediction accuracy. Additionally, the model yielded R-squared (R²) values between 0.75 and 0.85, indicating strong predictive power across different tissues. The model also performed consistently well during 10-fold cross-validation, further validating its robustness and reliability for RNA methylation prediction in various tissue types.

Decision Tree.

The decision tree is a classifier trained through iterative selection of the most significant features at each node. Its advantages include simplicity, speed, and easily interpretable graphical outputs. Trees are typically constructed top-down, starting from a root node with all training samples, and partitioning the feature space until reaching terminal nodes or leaves¹³⁵. The main construction steps are establishing a dividing rule for internal nodes, determining terminal nodes, and assigning class labels to minimize estimated error ¹³⁶. Binary decision trees, which use a single feature at each node, are the most common. However, they may not be suitable for limited sample sizes, as they can produce overly simple trees. Additionally, small changes in training samples can lead to significant output variations, reflecting the algorithm’s intrinsic instability ¹³⁷. As a result, decision tree applications in biology are somewhat limited ^138,139.

k-nearest-neighbor

The KNN algorithm classifies unlabeled data by measuring similarity among nearby data points using a distance function. It requires no parameters and makes no assumptions about data distribution, except that features are continuous. The most common distance function is Euclidean distance, along with variants like Mahalanobis and modified value difference metrics ¹⁴⁰. To classify a new object, the algorithm calculates distances from it to all training samples, orders these distances, and retains the top k closest samples. The new object is classified based on the most frequent class among its k nearest neighbors ¹⁴¹. KNN offers good interpretability and outlier detection, but it requires domain knowledge to determine the optimal value of k ¹⁴². Ramteke and Khachane¹⁴³ proposed an automatic medical image classification and abnormality detection system using the K-Nearest Neighbour (KNN) algorithm. The system achieved an 80% classification rate, outperforming kernel-based Support Vector Machine (SVM) classifiers. The methodology involved four phases: preprocessing, feature extraction, classification, and post-processing. Sharifonnasabi et al.¹⁴⁴ introduced a hybrid model combining Convolutional Neural Networks (CNN) and K-Nearest Neighbors (KNN) to enhance age estimation accuracy in dental radiographs, specifically Orthopantomography (OPG). The model, as shown in Figure 4h, achieved impressive accuracy of 99.98% for 1-year-old cases, 99.96% for 6-month-old cases, 99.87% for 3-month-old cases, and 98.78% for 1-month-old cases. These results were obtained by analyzing a dataset of 1,922 panoramic dental radiographs from patients aged 15 to 23. The study demonstrated that integrating CNN and KNN significantly improved age estimation precision compared to conventional methods.

Support Vector Machines

SVMs is popular classification techniques known for their robust mathematical foundation and strong accuracy, particularly in biomedical applications. They are parameter-free and perform well in high-dimensional and non-linear problems. However, Vabalas et al. ¹⁴⁵ suggest that SVMs may not be suitable for certain limited cases. SVMs identify the optimal boundary that maximizes the margin between two classes by transforming samples into a higher-dimensional space and establishing a hyperplane for separation. Two parallel hyperplanes are created on either side, and a wider separation implies better predictive accuracy. The support vectors—samples closest to the parallel hyperplanes—determine the width of this margin. Detailed mathematical descriptions can be found in Tarca et al.’s work ¹⁴¹ and it can easily be established in R ¹⁴⁶. Orrù et al ¹⁴⁷ provide a comprehensive review of SVMs for identifying imaging biomarkers in neurological and psychiatric diseases, while applications in biomedical and biometrical fields are discussed in the book chapter “Emerging Paradigms in Machine Learning”¹⁴⁸.

t-distributed Stochastic Neighbor Embedding

t-SNE is a widely used technique for dimensionality reduction and visualization, particularly effective for revealing complex and nonlinear relationships in data. It maps high-dimensional data points to a lower-dimensional space while preserving pairwise similarities. Unlike PCA, t-SNE can manage non-linear separations where data points cannot be divided by a straight line. However, the choice of hyperparameters, such as the number of neighbors, significantly impacts the final clusters and relies on the user’s expertise and prior knowledge.

Proposed by Maaten et al¹⁴⁹, t-SNE begins by creating a probability function to describe the similarity between neighbors, typically using Euclidean distance. This distance is proportional to the probability density under a Gaussian centered at the sample data. Next, Student’s t-distribution defines similarity in the lower-dimensional space projected from the higher-dimensional domain. A key hyperparameter, perplexity, controls the variance of the t-distribution and the effective number of neighbors for each data point. Using gradient descent optimization, t-SNE minimizes the Kullback-Leibler divergence between pairwise similarity distributions in both spaces, aiming to preserve pairwise similarities in the low-dimensional representation. Once convergence is achieved, t-SNE generates a low-dimensional embedding for visualization in two or three dimensions, with similar data points in the original space appearing close together in the reduced space, facilitating the visualization of clusters, patterns, and structures.

In their analysis of human embryo genome data, Jia et al.¹⁵⁰ used Pearson’s correlation coefficient of multiple-cumulative probabilities (PCC-MCP) to define gene expression similarities. They then employed t-SNE to create optimal maps for clustering the data, resulting in six distinct clusters for visualization. By applying this method to several transcriptome datasets, they demonstrated its effectiveness in identifying distinct gene clusters and providing clear visual boundaries between them. This approach enhances the understanding of gene function and regulation, making it a valuable tool for analyzing gene expression data. Similarly, Emadi et al. ¹⁵¹ improve gene regulation network analysis by developing a rotation forecast method based on t-SNE to identify regulatory interactions among genes. Their approach, which incorporates biological information during modeling, outperforms leading methods in the field. The intrinsic features of t-SNE enable its broad applications in Big (Chemical Data challenges ^152–154.

3.4. Data Learning with Neural Network.

Although previously discussed algorithms, such as linear regression, can achieve basic data predictions based on intrinsic assumptions, they are limited in their ability to “learn” complex features from raw data and provide precise predictions. In this section, we focus on Artificial Neural Networks (ANNs) and their related architectures, which represent more advanced models in machine learning. While ANNs can also perform functions like regression and classification, their advantage lies in their capacity to learn intricate patterns and make more accurate predictions.

Artificial Neural Networks.

The appeal of ANNs stems from their ability to learn. In this context, learning refers to the process of using a set of observations to find solutions that align with a specific task and class of functions¹⁰⁷. Inspired by the biological brain, ANNs—also referred to as Simulated Neural Networks (SNNs)—are mathematical or computational models designed to mimic the neural processing capabilities of biological systems. These models are highly adaptive and capable of learning the relationships between inputs and outputs through a process of training.

Typically, an ANN is composed of an input layer, one or more hidden layers, and an output layer. A typical structure of ANN can be found in Figure 5j, while different activation functions are employed in each layer¹⁵⁵. The network comprises numerous nodes (similar to neurons) that process and learn from large amounts of input data, either directly from the dataset or from other nodes. Each connection between nodes has a weight, which controls the strength of the signal. A neuron is activated only if the weighted sum of its inputs exceeds a predefined threshold; otherwise, no information is passed to the next layer. The architecture, defined by how neurons are connected and how inputs are processed, ultimately determines the network’s effectiveness in tasks like classification, regression, or complex data predictions.

Figure 5: Imaging for Biological Applications.

(a) Cell division visualized by fluorescence microscopy. Reproduced from reference [175] with permission from [Springer Nature], copyright [2014]. (b) Lipid organelles recorded with fluorescence microscopy. Reproduced from reference [180] with permission from [Springer Nature], copyright [2020]. (c) Three-dimensional microscopy of the tumor microenvironment in vivo using optical frequency domain imaging. Reproduced from reference [190] with permission from [Springer Nature], copyright [2009]. (d) Visualization of individual type III protein secretion machines in live bacteria. Reproduced from reference [201] with permission from [Proceedings of the National Academy of Sciences of the United States of America], copyright [2017]. (e) Whole-brain neuronal activity of a larval zebrafish recorded with a light-sheet microscope. Reproduced from reference [209] with permission from [Springer Nature], copyright [2013]. (f) Photoacoustic imaging of thrombosis via fibrin-specific homopolymer nanoparticles. Reproduced from reference [218] with permission from [Springer Nature], copyright [2023].

ANNs can incorporate various learning paradigms, including supervised, unsupervised, and reinforcement learning. In supervised learning, the network is trained on labeled data, where both the inputs and their corresponding outputs are provided in the training samples. The network learns to map the relationship between the two, making it suitable for tasks like classification or regression. In unsupervised learning, the networks, often referred to as competitive neural networks, typically consist of a single layer. Here, the training algorithm adjusts the weights between input and output nodes based solely on input data, with no labeled outputs provided. These networks excel in pattern recognition, feature extraction, and cluster analysis, functioning similarly to the algorithms discussed in the previous section. They are particularly effective in identifying hidden structures in data, which is valuable for chemical imaging when exploring intricate datasets without predefined labels. Reinforcement learning, on the other hand, enables ANNs to learn by interacting with an environment. The network makes decisions and receives feedback in the form of rewards or penalties, guiding its learning process. This approach is often used in more dynamic applications such as game playing or robotics, where the system must continuously adapt its strategy based on the outcomes of its actions.

Neural networks can be classified into various types, depending on their architecture and intended application. One of the most fundamental types is the feedforward neural network, often referred to as a multi-layer perceptron (MLP). These networks are the backbone of many machine learning applications. Unlike simple perceptron, which struggle with nonlinear problems, MLPs typically use sigmoid neurons to handle nonlinearity, making them suitable for complex tasks. They form the basis for advanced applications such as computer vision, natural language processing, and other neural network models. For instance, Mansouri¹⁵⁶ provides an in-depth review of the application of neural networks in the medical field, demonstrating their versatility. De Fauw et al ¹⁵⁷ further highlight their potential by applying neural networks to a heterogeneous dataset of three-dimensional optical coherence tomography (OCT) scans from patients at a major eye hospital. Remarkably, after training on just 14,884 scans, the model’s diagnostic recommendations for sight-threatening retinal diseases surpassed those of experts. Specifically, the model demonstrated high accuracy in making referral recommendations, effectively identifying patients requiring urgent care. Additionally, the neural network was deviceindependent, broadening the potential of deep learning in enhancing diagnostic processes and patient management in ophthalmology.

Convolutional Neural Networks.

CNNs share similarities with Multi-Layer Perceptrons (MLPs) but are particularly well-suited for tasks involving spatial structures, such as image recognition, pattern recognition, and computer vision. CNNs are designed with convolutional layers that automatically learn hierarchical patterns from raw data, making them highly effective for visual data analysis. These layers leverage principles from linear algebra, especially matrix multiplication, to detect features and patterns within images, allowing CNNs to excel in tasks like chemical and biological image processing. Kumar et al.¹⁵⁸ introduced a method for classifying medical images using an ensemble of fine-tuned convolutional neural networks (CNNs). This approach leverages multiple CNN architectures, each fine-tuned on a large dataset of natural images, to extract diverse features from medical images. The fine-tuning process adapts the generic features from natural images to the specific characteristics of medical imaging modalities. The extracted features are then used to train multiclass classifiers, whose posterior probabilities are combined to predict the modalities of unseen images. Experiments on the ImageCLEF 2016 medical image dataset demonstrated that this ensemble method achieved higher accuracy than established CNNs. It also outperformed other methods evaluated on the same benchmark dataset, with the exception of those that utilized additional training data. Similarly, Qin et al ¹⁵⁹ proposed an improved convolutional neural network (CNN) method for classifying biological images. They evaluated the computational cost and classification accuracy on five well-known benchmark datasets, demonstrating that their method achieved higher accuracy compared to traditional CNNs. For example, their classification accuracies are improved to different degrees (usually several percentages) compared with InceptionV3 and DenseNet201. Lin et al. presents an AI-assisted approach for the fusion of scanning electrochemical microscopy (SECM) images using a novel soft probe to enhance image quality and analytical precision. They introduced a CNN-based framework for merging multiple SECM images and the proposed method integrates information from different scans, improving spatial resolution and providing more accurate data on electrochemical processes at the nanoscale. Performance metrics such as root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM), demonstrate significant improvements in image clarity and data consistency compared to traditional image fusion techniques. The authors highlight the potential of the AI-assisted fusion method to advance research in electrochemistry and surface analysis, offering more reliable results for both qualitative and quantitative studies¹⁶³.

Recurrent Neural Networks.

RNNs are distinguished by their feedback loops and ability to capture dynamic temporal behavior, making them particularly effective for time-series or sequence-based data. These networks are commonly used to predict future outcomes in fields that rely on sequential information. In biological sequence analysis, RNNs have demonstrated an exploitable intrinsic bias, as reported by Hawkin et al¹⁶⁰. Their study, which utilized a dataset of protein sequences across various subcellular localization classes, revealed that RNNs outperform feedforward networks, especially when sequence patterns become ambiguous. In such cases, the choice of specific recurrent architecture becomes crucial for accurate predictions. Furthermore, long-short-term memory (LSTM) RNNs, a variant designed to mitigate issues like vanishing gradients, have been successfully applied to the analysis of cell motility data, as reported by Kimmel et al ¹⁶¹. In their study, LSTM RNNs provided a highly accurate classification of both simulated and experimentally measured cell motility across multiple cell types, outperforming models that relied on hand-engineered features. Specifically, mean test accuracy reached 85.47%±1.96% for RNN baseline models, 93.11%±0.65% for RNN models with convolutional layers, and 95.31%±0.32 for their heuristic baseline. This underscores the power of RNNs in capturing complex temporal patterns in biological data.

Deep Neural Networks.

DNNs are characterized by multiple hidden layers, allowing them to learn hierarchical representations of input data. The term “deep” in deep learning simply refers to the depth of these layers. While basic neural networks may have only two or three layers, DNNs extend this architecture, enabling them to capture complex patterns. These networks are widely applied in areas such as image recognition, speech recognition, and natural language processing. In the biomedical field, DNNs are particularly adept at modeling intricate dependencies in the genome’s regulatory landscape, making them valuable for tasks like genetic variant prediction and interpretation¹⁶². For example, Zrimec et al. ¹⁶³ applied deep learning to investigate the relationship between gene regulatory code and mRNA abundance using over 20,000 mRNA datasets from bacteria to humans. The improved architecture is presented in Figure 4l, which consists of CNN layers, RNNS layers, and fully connected layers (FC). Their model successfully predicted mRNA abundance directly from DNA sequences, explaining up to 82% of the variance in transcript levels encoded by the gene’s regulatory structure. This study demonstrated that gene expression levels are influenced by the entire regulatory structure and specific combinations of regulatory components. Yao et al. ¹⁶⁴ developed DNNs to enhance the standoff detection of dangerous chemicals. They introduced a data accentuation technique to improve the model’s sensitivity to chemical signatures. The study demonstrated that the DNNs, combined with data accentuation, effectively identified and classified various hazardous substances from standoff distances. This approach offers a promising avenue for remote detection and identification of dangerous chemicals, potentially enhancing safety protocols in various applications and highlighting the versatility of DNNs in analyzing complex data across diverse fields. Zhao et al. ¹⁶⁹demonstrate an AI-assisted mass spectrometry imaging (AI-SMSI) strategy combined with in situ image segmentation to enhance subcellular metabolomics analysis. This approach leverages deep learning algorithms to accurately segment and analyze mass spectrometry data, enabling detailed visualization of metabolite distributions at a subcellular level. The AI-driven method significantly improves spatial resolution, allowing for precise localization of metabolites within different cellular compartments. Performance metrics, such as Dice similarity coefficient and intersection over union (IoU), show high segmentation accuracy, while resolution improvements are quantified using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). This AI-enhanced method offers a powerful tool for investigating intracellular biochemical processes and has potential applications in drug discovery and cellular research.

4. Applications in biology and biomedicine.

The development of these chemical imaging techniques has provided valuable insights across various fields. Here, we highlight notable applications in cell biology, lipid biology, tumor biology, microbiology, neurobiology, developmental and pharmaceutical biology.

4.1. Cell Biology.

Bioimaging microscopy in cell biology has become an essential tool for exploring the intricate structures, dynamics, and functions of cells. By providing high spatial resolution and compatibility with live-cell imaging, bioimaging techniques enable researchers to observe cellular processes in real time, shedding light on the composition and metabolism of cells. These technologies have revealed essential insights into the behavior of genetic materials (DNA and RNA), proteins, and metabolic activities such as glucose utilization, thereby significantly advancing our understanding of cellular biology.

For genetic material imaging, such as DNA and RNA, various bioimaging techniques have been instrumental in studying genetic material such as DNA and RNA within live cells. SRS microscopy allows researchers to visualize chromosome dynamics across different stages of the cell cycle without the need for fluorescent labeling, preserving the natural state of the sample. For instance, by incorporating ethynyl deoxy uridine, a nucleoside analog, researchers can directly image newly synthesized DNA as a marker for cell proliferation and division^170–175. This approach has been applied effectively to study tumor cells and neurons, providing critical information about cell cycle progression, genetic replication, and cell proliferation in both normal and pathological conditions, such as cancer.

For Protein imaging and dynamics, proteins are fundamental to cellular structure and function, and their synthesis, modification, and degradation are essential processes that regulate cellular behavior. Bioimaging microscopy has advanced our ability to observe these processes directly within live cells. By using deuterated amino acids, researchers can label active sites of protein synthesis and degradation¹⁶⁵. This approach enables the identification and tracking of proteins in their native environment, providing a deeper understanding of how proteins are synthesized, folded, and eventually degraded or recycled within cells.

For metabolic imaging, as glucose is a primary energy source for cellular functions, understanding its uptake and utilization is crucial in studying cell metabolism, especially in the context of cancer biology. SRS microscopy has been employed to study glucose metabolism by using alkyne-tagged glucose analogues, such as 3-O-propargyl-glucose¹⁶⁵. This allows researchers to visualize the uptake of glucose at the single-cell level, providing insights into metabolic activities in various cell types, including tumor cells and neurons.

For multiplexed imaging of subcellular structures^166–170, such as the use of super-multiplexed probes, which enable 10-color imaging of various organelles within live cells, including mitochondria, lysosomes, endoplasmic reticulum, lipid droplets, plasma membrane, Golgi apparatus, microtubules, actin filaments, and the nucleus. Such detailed imaging facilitates the study of the organization, dynamics, and interactions between different organelles, which are critical for understanding cellular physiology and pathology. By visualizing these structures simultaneously, researchers can investigate the physical and functional interactions between organelles under various conditions. For example, fluorescence imaging has been used to observe nuclear envelope dynamics through the cell division, as shown in Figure 5a¹⁷¹. BY2 cell expressing the SUN1–YFP (green) and the chromatin marker histone H2B–CFP (magenta) through the cell cycle to visualize the key features during plant cell division. This technical capability supports the study of inter-organelle contacts and interactions under both physiological and pathological conditions, enhancing our understanding of cellular function and disease mechanisms.

4.2. Lipid Biology.

Lipids are among the most abundant molecules in cells, encompassing fatty acids, sterols, and their derivatives. Due to their small sizes, lipid molecules are challenging to image using traditional fluorescence methods. However, they exhibit strong and distinctive Raman signals, making them ideal candidates for label-free and bio-orthogonal chemical imaging approaches. As a result, bioimaging technologies have found early and rich applications in lipid research, providing valuable insights into lipid metabolism and its association with various metabolic diseases, such as atherosclerosis, obesity, and diabetes^172–177.

Lipid membranes are found in most intracellular organelles, and their heterogeneities play an essential role in regulating the organelles’ biochemical functionalities. For example, heterogeneity and dynamics of subcellular lipid membranes were visualized by high-dimensional super-resolution imaging. By employing Nile Red, a common intracellular lipid dye, to stain lipid membranes universally in live cells (Figure 5b)¹⁷⁶. From the GFP colocalization images and the Nile Red membrane morphology images, Nile Red successfully labels at least ten subcellular compartments, including mitochondria-GFP, Golgi-GFP, ER-GFP, lysosome-GFP, early endosome-GFP, and late endosome-GFP colocalize the corresponding organelles.

Hyperspectral imaging of lipid-specific bonds, such as CH2 and C=C–H, enables the quantification of neutral lipids, revealing differential storage patterns of triglycerides and cholesteryl esters in various cells, tissues, and metabolic disorders. By tracking the incorporation of deuterated fatty acids, researchers observed that unsaturated fatty acids tend to accumulate preferentially in lipid droplets, while saturated fatty acids are found in both membrane structures and lipid droplets¹⁷⁸. Imaging of deuterated fatty acid metabolism revealed that saturated fatty acids, such as palmitic acid, induce phase separation within the fluidic ER membrane, forming solid-like domains. This phenomenon, observed for the first time in living cells, suggests that ER membrane phase behavior plays a role in the onset of biotoxicity, highlighting a potential target for therapeutic intervention¹⁷⁹.

Imaging lipid metabolism has provided further insights, particularly in the context of cancer. Label-free imaging of cholesterol ring vibrations revealed a high accumulation of cholesteryl ester in advanced human prostate cancer tissues. This accumulation was associated with the loss of the tumor suppressor PTEN, activation of the PI3K–AKT pathway, and increased uptake of lipoproteins¹⁸⁰. Notably, inhibiting cholesterol esterification was shown to reduce cancer aggressiveness.

Various bioimaging techniques have significantly advanced lipid biology research, offering a powerful means to visualize and quantify lipid dynamics and interactions in health and disease. By exploring the molecular mechanisms underlying lipid metabolism, researchers can develop targeted therapies for metabolic diseases and cancers, contributing to the broader understanding of cellular lipid regulation.

4.3. Tumor Biology.

Tumors are characterized by uncontrolled proliferation and abnormal metabolism, with tumor heterogeneity contributing to challenges such as metastasis and resistance to cancer therapies^{20,166,181–186}.Understanding tumor heterogeneity is crucial for developing more effective diagnostic and treatment strategies. Bioimaging technologies have become vital tools in characterizing tumors at early stages, particularly for visualizing tumor boundaries and heterogeneity.

For the label-free imaging for tumor characterization, various bioimaging techniques have enabled the label-free differentiation of tumor tissues from healthy tissues using intrinsic protein (CH3) and lipid (CH2) vibrations¹⁸⁷. These techniques were first applied in mouse models and human glioblastoma tissues, showing a strong correlation with standard hematoxylin and eosin (H&E) staining results. This approach allows for in situ characterization of tumors without the need for staining or labeling, providing a real-time, non-invasive option for tumor diagnosis.

In addition, the application of multiphoton microscopy to the study of solid tumor biology in vivo has elucidated pathways and mechanisms of cancer progression and has led to new therapeutic strategi. For example, as shown in Figure 5c, Benjamin et al¹⁸⁶ introduced optical frequency domain imaging (OFDI) as an intravital microscopy technique that overcomes the technical limitations of multiphoton microscopy, granting exceptional access to critical, previously unexplored aspects of tissue biology. By utilizing specialized OFDI-based methods and fully intrinsic contrast mechanisms, we achieve rapid and repeated imaging of tumor angiogenesis, lymph angiogenesis, tissue viability, and both vascular and cellular responses to treatment. This demonstrates OFDI’s potential to advance the understanding of physiological and pathological processes and enhance the assessment of therapeutic strategies.

Raman imaging of vibrational probes has also been used to investigate tumor metabolism¹⁸⁸. By tracking the metabolic incorporation of deuterated molecules, researchers have observed increased lipid synthesis at the invasion front of breast cancer cell spheroids and in infiltrating tumor cells within a mouse xenograft model of glioblastoma¹⁸⁹. This enhanced lipid incorporation is likely linked to the epithelial-to-mesenchymal transition, where mesenchymal cells exhibit a higher accumulation of lipid droplets, as shown through biorthogonal chemical imaging techniques.

Bioimaging technologies are transforming the study of tumor biology, providing insights into tumor heterogeneity, metabolism, and structural characteristics. These advancements support the development of more effective diagnostic and therapeutic strategies, particularly in real-time and intraoperative settings.

4.4. Microbiology.

Microbiology has increasingly garnered attention for its critical roles in areas such as human gut microbiota, antibiotic resistance, and biomaterials, including biofuels^190–196. Super-resolution microscopy is instrumental in focusing on individual molecules and accessing previously unresolvable subcellular structures, such as the bacterial cytoskeleton, cell wall, and nucleoid, as well as the highly structured and compartmentalized bacterial cytoplasm. Super-resolution microscopy offered fundamental insights into the secretion of bacterial effectors and growth of bacteria in broth, and in the future can be used to study bacterial pathogens inside cells. Both 2D and 3D single-molecule switching nanoscopy were used to investigate the ‘Type III Secretion System’ of Salmonella enterica serovar Typhimurium¹⁹⁷, an important zoonotic pathogen causing gastroenteritis and inflammation of the intestinal mucosa. In a recent study¹⁹⁷, authors revealed the ultrastructure, protein stoichiometry, assembly and distribution in live bacterial cells, as well as the localization of bacterial effectors prior to secretion (Figure 5d).

Bioimaging has also been utilized to investigate antibiotic responses in bacterial biofilms, which are responsible for drug-resistant infections in clinical settings. By tagging the antibiotic vancomycin with an aryl-alkyne tag¹⁹⁸, researchers quantitatively monitored its penetration kinetics into biofilms. This revealed a non-uniform diffusion pattern with shallow depths, primarily due to preferential binding to bacterial cells over the extracellular matrix. Such insights contribute to understanding the pharmacokinetics of antibiotic treatments in biofilms.

Moreover, bioimaging techniques have shed light on antibiotic tolerance and metabolic heterogeneity within Pseudomonas aeruginosa biofilms¹⁹⁹, which are a significant cause of chronic lung infections in cystic fibrosis patients. Through the imaging of D2O incorporation, deep regions of active metabolism were observed for the first time in biofilms grown in glucose media. Particularly, metabolic activity in the hypoxic regions was linked to small metabolites such as phenazines and Cco terminal oxidases, which play crucial roles in maintaining reduction-oxidation balance in oxygen-depleted environments, thereby enhancing biofilm tolerance to antibiotics like ciprofloxacin.

4.5. Neurobiology.

The nervous system is essential for various functions, including learning, memory, movement, and behavior. Investigating complex nervous systems in both physiological and degenerative states can provide a comprehensive understanding of neural functions and help uncover the underlying causes of neural disorders. Advanced imaging techniques are crucial for this purpose^{3,48,156,158,160,161,164,200–204}.

Brain function relies on the communication between extensive populations of neurons across various brain regions. To achieve a comprehensive understanding of these interactions, it is essential to track the time-varying activity of all neurons within the central nervous system. In this study²⁰⁵, authors employed light-sheet microscopy to monitor neural activity, using the genetically encoded calcium indicator GCaMP5G, throughout the entire brain volume of larval zebrafish in vivo at a frequency of 0.8 Hz (Figure 5e). This approach enabled us to capture activity from over 80% of all neurons at single-cell resolution, which demonstrate the utility of this technique in revealing functionally defined neural circuits across the brain.

Raman microscopy has been effectively utilized for label-free imaging of neurotransmitters and membrane potentials. By employing frequency-modulated spectral-focusing SRS, researchers directly measured the local concentration of acetylcholine at approximately 10 mM at the neuromuscular junction of the frog cutaneous pectoris muscle.²⁰⁶ Additionally, label-free SRS imaging of the protein CH3 Fermi resonance peak allowed for the visualization of puff-induced depolarizations in multiple neurons within live mouse brain tissues, even enabling the detection of single action potentials in patched neurons²⁰⁷. SRS microscopy has also been applied to investigate brain pathologies, including myelin degeneration in amyotrophic lateral sclerosis (ALS) and amyloid plaques in Alzheimer’s disease²⁰⁸. Long-term SRS imaging of peripheral nerve degeneration in living ALS mouse models revealed the formation of lipid-rich ovoids at early time points. This capability allows for the monitoring of drug effects, facilitating early detection of ALS and the potential for timely interventions²⁰⁹.

4.6. Pharmaceutical.

Pharmaceutical bioimaging is a powerful tool that provides critical insights into the behavior of drugs within living organisms. By utilizing advanced imaging techniques, researchers can enhance their understanding of drug mechanisms, improve therapeutic outcomes, and advance the field of drug discovery, ultimately leading to more effective treatments for various diseases^{20,26,76,112,132,181,210–213}.

By linking the structural characteristics of pharmaceutical compounds to their distribution patterns and functional outcomes, researchers can optimize drug design to improve therapeutic efficacy and reduce side effects. Bioimaging can aid in the development of personalized treatment regimens by visualizing individual responses to drugs, which could guide dosage adjustments and treatment strategies tailored to specific patient needs. For example, researchers developed a polymer nanoplatform that integrates long-wavelength second near-infrared photoacoustic imaging-based thrombosis detection and antithrombotic activity (Figure 5f)²¹⁴.

High-throughput imaging techniques enable rapid screening of potential drug candidates, facilitating the identification of promising compounds early in the development process. Validating the action of these compounds within living systems can accelerate their progression to clinical trials. For example, Epi-SRS microscopy²¹⁵ has been utilized to investigate amlodipine besylate tablets sourced different commercial providers, revealing variations in the compositions and distribution of excipients. This information is invaluable for formulation screening, aiding in the optimization of drug release profiles.

5. Summary and Outlook.

In summary, chemical imaging has emerged as a transformative tool in biological research, enabling the visualization and analysis of complex biochemical processes with unprecedented resolution. Techniques such as fluorescence microscopy, super-resolution microscopy, and Raman microscopy, have introduced powerful approaches for studying a wide range of biological phenomena, from cellular metabolism to tissue pathology. As chemical imaging technologies continue to evolve, several exciting developments are anticipated. These include the integration of chemical imaging with other imaging modalities, contributions to personalized therapeutic strategies by tailoring treatments to individual patient profiles, and the expansion of chemical imaging applications into fields such as bioengineering, materials science, and clinical research. These advancements will foster interdisciplinary collaborations, driving innovation and discovery. Moreover, the resulting expansion of big data and the growing demand for rapid, high-throughput analysis underscore the importance of combining imaging experiments with advanced computational methods. Recent advances in AI and machine learning offer powerful tools for analyzing complex datasets, accelerating discoveries, and enhancing the utility of chemical imaging in biological research. AI-based chemical imaging has shown potential in areas like chemical analysis, biomedical detection, and environmental monitoring, where deep learning and machine learning algorithms such as convolutional neural networks and recurrent neural networks are applied to extract meaningful insights from complex data. Moving forward, several aspects of future research directions are critical in promoting and verifying the AI’s capabilities in chemical imaging. Firstly, the integration of multimodal data combines imaging techniques (e.g., hyperspectral imaging, Raman spectroscopy, and mass spectrometry) with AI can provide richer, more detailed chemical and structural information. This fusion could enable more precise characterization of biological tissues, materials, and environmental samples. Secondly, the development of unsupervised or semi-supervised learning models, which require less labeled data, could make AI models more adaptable and scalable in real-world applications. Another exciting research avenue lies in real-time chemical imaging and analysis. With advancements in computational power and model optimization, AI models could be designed to process large datasets in real time, providing immediate feedback for applications such as in vivo diagnostics, industrial quality control, or environmental hazard detection. Finally, AI’s role in automating feature extraction and data processing will likely expand, reducing the reliance on manual intervention and improving the reproducibility of results. The combination of explainable AI with chemical imaging, where models not only provide predictions but also offer insights into the underlying processes, could be another key development. This transparency could improve trust in AI-based systems, especially in critical applications like medical diagnostics and regulatory compliance. As these technologies progress, they are expected to provide deeper insights into the molecular mechanisms that underpin life.