|Home | About | Journals | Submit | Contact Us | Français|
Coherent Raman scattering (CRS) microscopy is a high-speed vibrational imaging platform with the ability to visualize the chemical content of a living specimen by using molecular vibrational fingerprints. We review technical advances and biological applications of CRS microscopy. The basic theory of CRS and the state-of-the-art instrumentation of a CRS microscope are presented. We further summarize and compare the algorithms that are used to separate the Raman signal from the nonresonant background, to denoise a CRS image, and to decompose a hyperspectral CRS image into concentration maps of principal components. Important applications of single-frequency and hyperspectral CRS microscopy are highlighted. Potential directions of CRS microscopy are discussed.
Thus far, the majority of our knowledge about the chemical content in cells or tissues comes from analysis of fixed samples or homogenates. These in vitro assays can only detect the presence of molecules. Without information of spatiotemporal dynamics, it still remains difficult to either probe transient/localized accumulation of metabolites or biomarkers, or determine whether the detected species is the cause or consequence of a biological event. To move from detecting the presence of molecules to fully understanding their functions (1), in situ and in vivo molecular analysis is becoming a frontier in the field of analytical chemistry. In this article we scrutinize the recent progress in developments and applications of coherent Raman scattering (CRS) microscopy, an emerging platform that allows high-speed vibrational spectroscopic imaging of biomolecules in living systems.
Raman scattering, an inelastic scattering process induced by molecular vibrations, was reported in 1928 by Raman & Krishnan (2). Since the invention of lasers in the 1960s, Raman spectroscopy has become a powerful analytical tool for label-free, quantitative analysis of molecules in gas, in solutions, or at an interface. In 1979 Abraham & Etz (3) resolved the spatial distribution of molecules in biological tissues using a Raman microscope. Now, commercially available confocal Raman microscopy providing submicron spatial resolution has reached millisecond-scale spectral acquisition speed. This speed is not yet sufficient for imaging living systems. CRS microscopy (4), based on either coherent anti-Stokes Raman scattering (CARS) or stimulated Raman scattering (SRS), has been developed to overcome the speed limitations. CARS is a third-order nonlinear optical process in which three laser fields, namely a pump field at frequency ωp, a Stokes field at frequency ωs, and a probe field at frequency ωpr, are involved. When the energy difference of the pump and Stokes beams (ωp − ωs) matches the frequency of a molecular vibration mode (Ω), the Raman resonance occurs. This resonance is then probed by the third field at ωpr, generating an anti-Stokes field at (ωp − ωs) + ωpr. The CARS signal appears at a higher-frequency end and therefore is spectrally separated from the one-photon fluorescence. Meanwhile, the CARS signal is accompanied by a vibrationally nonresonant background contributed by electronic motions.
CARS was observed first by Terhune et al. in 1965 (5). Thereafter, CARS spectroscopy was widely applied to study the dynamics of chemical reactions (6, 7). The first CARS microscope with noncollinearly overlapped pump and Stokes beams focusing on a single vibrational frequency was reported in 1982 (8). In 1999, Zumbusch et al. (9) demonstrated CARS imaging of living cells using a tightly focused femtosecond laser source with collinear beam geometry. Subsequent developments of picosecond excitation with polarization control (10), epi-detected CARS (11), time-resolved CARS (12), interferometric CARS (13), and frequency modulation (14) aimed to reduce the nonresonant background and thus enhance the vibrational contrast. The image acquisition speed of single-frequency CARS reaches video rates (15). Spectroscopic CARS imaging techniques that acquire a CARS spectrum at each pixel were developed to resolve spectrally overlapped Raman bands. Methods include the fast tuning of an optical parametric oscillator (OPO) (16), spectral focusing (17), multiplex excitation (18, 19), and broadband excitation (20). In broadband CARS microscopy, spectral coverage of ~3,000 cm−1 with an acquisition time of ~3.5 ms per pixel was recently reported (21).
SRS is a dissipative process that occurs simultaneously with CARS. When the beating frequency (ωp − ωs) is tuned to excite a molecular vibration, the energy difference between ωp and ωs pumps the molecule from a ground state to a vibrationally excited state. The laser field manifests itself as a weak decrease in the pump beam intensity, called stimulated Raman loss (SRL), and a simultaneous increase in the Stokes beam intensity, called stimulated Raman gain (SRG). Because the Raman gain or loss signal appears at the same frequency as one of the incident beams, optical modulation and demodulation are needed to extract the SRS signal.
Woodbury & Ng (22) reported SRS phenomenon in 1962. Thereafter, femtosecond SRS spectroscopy was developed to study the dynamics of chemical and biochemical reactions at femtosecond-scale temporal resolution (for a review, see 23). In 2007, Ploetz et al. (24) demonstrated SRS microscopy using a femtosecond laser with a kilohertz repetition rate to generate the narrowband pump beam and the white continuum Stokes beam. In 2008, Freudiger et al. (25) demonstrated real-time, single-frequency SRS microscopy using a laser with a repetition rate of 76 MHz. In their scheme, an electro-optic modulator modulated the Stokes beam at megahertz frequency, and a lock-in amplifier demodulated the SRL signal from the pump beam. Since then, several hyperspectral SRS techniques have been published (26–29). Slipchenko et al. (30) established an alternative scheme of demodulation using a tuned amplifier, which enabled the development of a tuned amplifier array for multiplex SRS imaging (31).
Along with the efforts to improve instrumentation, various chemometrics approaches have been adopted to decompose the three-dimensional data (x-y-Ω) in a hyperspectral CRS image into chemical maps (21, 27, 32, 33). The spectral fidelity and image quality were improved by algorithms that remove the noise in spatial and spectral domains (34–37). For CARS microscopy, researchers developed phase retrieval algorithms (38–42) to extract the Raman-resonant signals from the raw CARS spectra.
We review the basic theory of CRS and the most widely adopted instrumentation strategies for CRS microscopy. We discuss methods for hyperspectral data analysis and present different modalities of CRS microscopy and their applications, including single-frequency CRS, frame-by-frame-based hyperspectral CRS, and multiplex CRS. Because of space limitations, we reviewed only part of the literature relevant to CRS microscopy. We recommend other reviews that can help readers achieve a broader coverage of this field (43–51).
CRS belongs to a nonlinear optical process termed four-wave mixing that involves four waves at frequencies ωsig, ω0, ωp, and ωs. Here, p and s stand for pump and Stokes analogous to spontaneous Raman. In most experiments, only the pump and Stokes waves are used. When these waves are spatially and temporally overlapped into a medium, four CRS processes occur simultaneously, namely CARS at frequency (ωp −ωs)+ωp, coherent Stoke Raman scattering (CSRS) at frequency ωs − (ωp − ωs), SRG at frequency ωs, and SRL at frequency ωp. In SRG the Stokes field experiences a gain in intensity, whereas in SRL the pump field experiences a loss. There also exists a Raman-induced Kerr effect (RIKE), where a new polarization component is produced. Detailed descriptions of these processes can be found elsewhere (52). Figure 1 illustrates the input and output fields in CARS, SRG, and SRL.
The CRS signal arises from an induced third-order nonlinear polarization expressed by
where D is the number of distinct permutations of the fields. D = 6 for three distinct fields, and D = 3 for ω0 = ωp or ωs (53). χ(3) is the third-order susceptibility containing an electronic and a vibrational contribution,
where NR is the electronic contribution that is vibrationally nonresonant, is related to the Raman scattering cross-section, Γ is the Raman line width, and Δ = Ω − (ωp − ωs) is the detuning from a vibrational transition at Ω. The imaginary part of χ(3),
exhibits a Lorentzian profile as the spontaneous Raman line shape. Under the slowly varying amplitude approximation, the CRS field has the following form,
Here we neglect the coefficients before χ(3) for simplicity. The CARS signal appears at the anti-Stokes frequency. Under the tight focusing condition, the phase-matching condition is relaxed, and the forward-detected CARS intensity can be written as
As shown in Equation 5, the CARS spectral profile is dispersed owing to the interference between the resonant and nonresonant contributions, resulting in a red shift of the peak and the appearance of a dip at the higher wavenumber side. An effective way to remove the nonresonant background is to measure the CARS signal at the peak and dip of a Raman band. For a weak Raman scatterer, the difference of the CARS intensities at the peak and dip positions is written as
Thus, modulation between the peak and dip frequencies is a straightforward way to remove the nonresonant background. Under the shot noise limit, the noise in frequency modulating CARS equals the square root of the nonresonant background, and the signal-to-noise ratio (SNR) is written as
Unlike CARS that appears at a new frequency, the SRL field appears at the same frequency as the pump wave,
The total intensity, which is the mixing of the SRL field and the local oscillator (Ep), is expressed as
Being heterodyne-detected via mixing with the modulated Stokes wave, the Raman-induced modulation of the pump laser intensity is dominated by the interference term between the SRL field and the local oscillator Ep,
where the negative sign indicates a loss of intensity in the pump wave. Similarly, for SRG the Stokes laser intensity is modulated with a gain of
Both SRL and SRG have the same spectral profile as spontaneous Raman. A synergistic effect of SRL and SRG is the transfer of photon energy h(ωp − ωs) from the incident waves to the sample, resulting in a real vibrational excitation. Under the shot noise limit and resonant condition (Δ = 0), the noise in SRG equals the square root of Is, and the SNR is written as
In a most widely adopted CRS microscope (Figure 2a), two synchronized laser pulse trains are spatially overlapped and then sent to a laser scanning microscope. In this section, we discuss various laser sources and detection schemes used for single-frequency and multiplex CRS imaging.
A laser source that outputs either dual-color, synchronized pulse trains or a broadband pulse train is needed for CRS imaging. Moreover, because both CARS and SRS are nonlinear processes, pulsed lasers are required to utilize pulsed lasers to generate a strong signal. Several sources were developed to meet these requirements. The first approach uses two electronically synchronized, mode-locked Ti:sapphire lasers that produce mode-locked pulse trains with a repetition rate of ~80 MHz, watt-level average power, and pulse duration from several picoseconds to 100 fs (54). The second approach uses a mode-locked picosecond Nd:YVO4 laser at 1,064 nm that synchronously pumps an OPO by its second harmonic (55). The third approach deploys a femtosecond Ti:sapphire laser to pump an OPO (56). For a single-frequency CRS microscope, high spectral resolution can be achieved with two transform-limited picosecond lasers or two spectrally focused femtosecond lasers. For a hyperspectral CRS microscope, the laser sources can be two tunable lasers with picosecond pulse duration, or two femtosecond lasers by the aid of linear chirping (Figure 2b) or pulse shaping (Figure 2c) techniques. The combination of a picosecond laser with a femtosecond laser or supercontinuum for multiplex CRS imaging has also been reported (19, 20, 31, 57–61).
In a CRS microscope, an objective lens with high numeric aperture is commonly used to fulfill the tight focusing condition and maximize the power density on the sample. A photomultiplier tube (PMT) collects CARS signals with a bandpass filter in the front to block the excitation laser beams. The SRS imaging modality can be built on a CARS microscope with a few modifications. An acousto-optic modulator (AOM) or electro-optic modulator (EOM) is added to modulate the pump or Stokes laser intensity at megahertz frequency. For SRS imaging, the use of an oil condenser is critical for suppressing the background induced by cross-phase modulation. Unlike a CARS microscope, which uses a PMT as the photodetector, an SRS microscope uses photodiodes that can handle high laser power at the milliwatt level. A demodulator then extracts the SRL or SRG signals at the modulation frequency from the local oscillator. For high-speed SRS imaging, demodulation can be performed by a lock-in amplifier (25). A tuned amplifier that resonates at the modulation frequency provides a cost-effective way to demodulate the SRS signals (30). Hyperspectral CARS or SRS imaging is performed by tuning the wavelength of a picosecond laser or tuning the delay between two chirped femtosecond pulses.
Advanced detection schemes for multiplex CRS imaging have been developed. In multiplex or broadband CARS, a charge-coupled device (CCD) collects the dispersed CARS signals. For multiplex SRS, a photodiode array (62) or complementary metal-oxide semiconductor array (63) is used to collect dispersed SRS signals. The demodulation can be performed by either subtracting the background from the signal or using demodulators such as a multichannel lock-in amplifier (64) or a tuned amplifier array (31).
Providing spectral information at each pixel, hyperspectral CRS images are three-dimensional, including x, y, and the Raman shift Ω. The primary goal of hyperspectral data analysis is to generate concentration maps of individual species based on their Raman spectral profiles. A panel of image-processing methods has been adopted to convert hyperspectral CRS image data to chemical maps. These methods include denoising, phase retrieval, principal component analysis, and multivariate curve resolution. The denoising algorithms, based on either singular value decomposition (SVD) or spectral total variation (STV), remove the noise in spatial and spectral domains to improve the SNR and spectral fidelity. Based on either time-domain Kramers-Kronig (TDKK) transform or the maximum entropy model (MEM), the phase retrieval algorithms extract the Raman-resonant signals from raw CARS signals mixed with the nonresonant background. Principal component analysis (PCA) determines the number of major components in the hyperspectral image data. Finally, multivariate curve resolution (MCR) generates a concentration map for every principal component.
Under the optimized condition, shot noise is the main noise source that degrades the image quality and the spectral fidelity. Longer integration time or frame averaging could increase the SNR but at the price of sacrificing the imaging speed. An image-processing algorithm that reduces the noise in both spectral and spatial dimensions is important for in vivo imaging applications. Denoising two-dimensional images is well studied in the literature (65, 66). Two methods denoise hyperspectral image data that contain three-dimensional information. One method uses SVD to remove components that are dominated by random noise. The other method utilizes the spatial-spectral correlation of signals and the statistical independence of the noise.
SVD is widely used for noise reduction in spectroscopic images (32, 34–36). This method first factorizes the data matrix D into three matrix factors, D = USVT, where the unitary matrix U corresponds to an array of spectral vectors; S is a diagonal matrix composed of singular values; and V corresponds to an array of spatial vectors. Then the number of significant singular values in S is objectively determined, and the rest of the singular values are considered to be noise dominated and set to zeros to generate a new diagonal matrix S′. The noise-reduced data D′ can be reconstructed using D′ = US′VT. The determination of significant singular values is the key to denoising. Several criteria have been reported, including the decline in the slope of singular values, the first-order autocorrelation function of spectral and spatial matrices, and the randomness of residual plots for the difference between the original and reconstructed spectroscopic image data (67, 68).
STV denoises a hyperspectral image by utilizing the spatial-spectral correlation of the signal and the statistical independence of the noise (37). This method is a numerical optimization algorithm that solves the problem
where g is a vector of the observed (noisy) image; f is a vector of the desired (clean) image; μ is a parameter to determine the relative emphasis of the first term and the second term; and D is a matrix representing the first-order forward finite-difference operators along the horizontal, vertical, and wavelength directions, defined as
The first term of Equation 13 is a quadratic term for the residue between the observed image and the solution, and the second term is a total variation term for the solution. The parameter μ is a function depending on the noise level σ of each spectral frame. An image with a large noise level requires more smoothing, and therefore μ should be small. The relationship between μ and σ can be determined empirically:
where 1 ≤ α ≤ 1.2. Therefore, by estimating the noise level σ of each spectral frame, Liao et al. (37) denoised hyperspectral SRS images of diluted dimethyl sulfoxide solutions and Caenorhabditis elegans and improved SNR by up to 57 times.
A quantitative comparison of denoising performance between SVD and STV was given by Liao et al. (37). In general, SVD eliminates the noise contributed by insignificant components, and therefore, it requires prior knowledge of the sample, such as the number of major species and the spectral features. STV reduces the noise level on the basis of the spatial-spectral correlation of signals. The knowledge of the optical system, for example, the spatial and spectral resolutions, determines the denoising parameter and the performance of the STV algorithm.
The CARS signal contains a Raman-resonant and a nonresonant contribution to the third-order nonlinear susceptibility, which can be expressed as
where the resonant component is composed of real and imaginary parts, and the nonresonant contribution contains only a real part. The imaginary part of representing the Raman line shapes, can be calculated by the following equation if the spectral phase () is known,
The Kramers-Kronig relation is widely used to obtain phase information from the modulus of susceptibility under the assumption of a causal and holomorphic response,
where χ(ω) = |χ(ω)|exp[i(ω)], and P is the Cauchy principal value. For CARS, there are two major difficulties when applying the Kramers-Kronig relation in a frequency domain. One is that the CARS signal is a mixture of nonresonant background and Raman resonance. This mixture leads to the inequality in the magnitudes of real and imaginary parts of the spectral phase, which does not agree with the assumption of causality. Moreover, in the case that ω goes to infinity and χ(ω) approaches zero, ln|χ(ω)| goes to negative infinity. To resolve these problems, an alternative form of TDKK using the Fourier series approach was reported (69). The spectral phase as a function of the signal modulus is represented in the time domain as
where the operator ψ(f(ω)) is defined as
and u(t) is the Heaviside step function (41). Based on the above Kramers-Kronig relation, one can represent the nonresonant background as a finite amplitude negative-time component. The Fourier transform of the nonresonant background is a symmetric function centered at time zero with a finite width. Assuming that the CARS signal has a negative-time component only from the nonresonant background and the positive-time response contains both Raman-resonant and nonresonant components, one can extract the spectral phase by replacing the Heaviside step function with the following function (41):
The MEM approach assumes that the data are consistent with the maximized entropy of the associated probability distribution. Therefore, the CARS spectrum can be fitted with the model by MEM and the spectral phase function can be estimated. In MEM, the CARS line shape function is represented as (70)
where υ is the normalized frequency defined as
The maximum entropy coefficients ak and β can be formulated as a Toeplitz matrix,
where CM is the discrete Fourier transform of the CARS line shape at normalized frequencies, represented as
Therefore, one can estimate CM, resolve the Toeplitz matrix for the maximum entropy coefficients, and then obtain the CARS line shape function. The spectral phase can be retrieved by
This approach has been explicitly discussed in References 38 and 39 and used in CARS microscopy to extract the Raman signal.
Principal component analysis (PCA) is widely applied for both hyperspectral CARS and SRS microscopy (71, 72) to characterize the principal components in a hyperspectral image. PCA rotates the coordinate of the dataset to a new space to reduce the number of dimensions. This can be done by SVD, in which the rank of eigenvalue and eigenvector is provided to determine the major components. As discussed above, the data matrix D is factorized into three matrix factors, D = USVT, where the unitary matrix U corresponds to an array of spectral vectors, S is a diagonal matrix composed of singular values, and V corresponds to an array of spatial vectors. The number of major components can thus be assessed by the singular values, given that the nonmajor components have a singular value close to zero.
Hyperspectral images can be decomposed into chemical maps of major components by methods including linear unmixing (21), multivariate curve resolution (MCR) (27), independent component analysis (32), spectral phasor analysis (33), or factorization into spectra and concentrations (73). MCR analysis has been applied to hyperspectral SRS microscopy (27, 74, 75). MCR is a bilinear model (76), capable of decomposing a measured spectral data set D into concentration profiles and spectra of chemical components, represented by matrices C and ST, D = C · ST + E. Here, T is the transpose of matrix S. E is the residual matrix or experimental error. The inputs to MCR are the dataset D and the reference spectra of each component. S contains the output spectra of all fitted components. The output concentration of a chemical component at each pixel is expressed as a percentage relative to the intensity of the MCR-optimized spectrum. Given an initial estimate of pure spectra either from PCA or prior knowledge of the sample, an alternating least squares algorithm calculates C and S iteratively until the results optimally fit the data matrix D. Non-negativity on both concentration maps and spectral profiles can be applied as a constraint during the alternating least squares iteration. To reduce ambiguity associated with MCR decomposition (77), a data augmentation matrix composed of repeating reference spectra can be added to the spectral dataset D. The enhanced weight on pure reference spectra ensures that the MCR algorithm selectively recovers concentration profiles for the corresponding Raman bands from dataset D.
Single-frequency CRS microscopy focuses all laser energy onto a specific Raman band to maximize the imaging speed until video rate (15, 78). The spectral bandwidth of a typical Raman band is ~10 cm−1. Therefore, to achieve high spectral resolution, picosecond excitation is preferred. By contrast, owing to the nature of the nonlinear interaction of CRS, a high-peak laser power is required to generate signals. As a result, femtosecond excitation can generate a larger signal level. Moreover, to couple CRS with other widely used imaging modalities, including two-photon excitation fluorescence, second harmonic generation, and third harmonic generation, femtosecond pulse excitation is preferred (79, 80). The signal-to-background ratios of CARS and SRS as the function of the pulse spectral width have been extensively studied (11, 81, 82). For the CARS signal, the use of femtosecond pulse as the excitation increases not only the Raman-resonant signal, but also the nonresonant background. The ratio of resonant signal to the nonresonant background can reach the maximum in the case that the width of the excitation pulses equals the spectral width of the Raman band (11, 81). For SRS that is not interrupted by the nonresonant background, the signal level can be increased by more than one order using femtosecond excitation. For a broad Raman line with a bandwidth of 25 cm−1, the SRG signal is amplified 30 times when the pulse spectral width increases from 1 cm−1 (~14 ps) to 120 cm−1 (~120 fs) (82). For narrow Raman bands, picosecond or spectrally focused femtosecond pulses are preferred to minimize the cross-phase modulation background in SRS microscopy.
In addition to optimizing the pulse duration for higher signal level, the nonresonant background in the CARS signal can be removed by techniques including polarization-sensitive CARS (10, 83), time-resolved CARS (12), interferometric CARS (13), and frequency-modulated CARS (14). In polarization-sensitive CARS, a polarizer is used before the detector to block the non-resonant background, which is linearly polarized but not parallel to the Raman-resonant signal (10). In time-resolved CARS, the Raman free-induction decay of molecular vibrations is recorded. The instantaneous nonresonant background can thus be separated in the time domain (12). In interferometric CARS, a third beam at the CARS wavelength is introduced as the local oscillator and the interferometric mixing term are detected, in which the phase of the local oscillator is adjusted to minimize the nonresonant background (13). In frequency-modulated CARS, the CARS intensities at the peak and dip positions are compared at each pixel (14). The intensity difference is free from the nonresonant background and is linearly proportional to the concentration of the target species.
SRS microscopy is not completely background free. In an intensity-modulated single-frequency SRS microscope, pump-probe signals including cross-phase modulation, two-photon absorption, and photothermal effects can be generated by the same laser beams and they contribute to the non-Raman background in the SRS image. These backgrounds can be removed by harnessing their difference from the SRS signal in the spectral domain (84–86). For example, Zhang et al. (86) developed a spectral modulation scheme for SRS microscopy in which an AOM in a pulse shaper switched between two excitation wavelengths within a femtosecond pulse at megahertz rate. Another way to remove backgrounds is to detect the SRL and SRG signals simultaneously. Berto et al. (87) demonstrated three-beam SRS microscopy with frequencies at ω0, (ω0−Ω), and (ω0+ Ω), where Ω is the molecular vibrational mode. In this scheme, lasers at (ω0−Ω) and (ω0+Ω) were modulated by two AOMs and with a 180° phase delay. Therefore, SRL and SRG signals were detected simultaneously from the local oscillator at ω0, and thus the background was canceled out (87).
In the case that the excitation wavelengths cover Raman bands that are spectrally overlapped, single-frequency excitation provides poor chemical specificity. To address this issue, Freudiger et al. (88) utilized a spatial light modulator to code two excitation patterns in a femtosecond pulse, and used an EOM to fast switch these two patterns at megahertz rate. The excitation pattern can be programmed to match a known, specific molecular vibration mode. An easier way is to record a stack of CRS images sequentially at multiple Raman shifts (89). To this end, hyperspectral CRS microscopy has been extensively developed, which is discussed below.
Single-frequency CRS microscopy has been applied to the fields of chemistry, materials science, biology, and medicine. For single-cell analysis, the C-H stretching region with Raman peaks at 2,850 and 2,940 cm−1 is commonly used to image lipid and protein at submicron resolution. For example, lipid droplet dynamics was monitored during differentiation in 3T3-L1 cells (90), lipolysis in adipocytes (91), accumulation induced by excess free fatty acid in cancer cells (92), and early embryo development (93). The accumulation of lipid was used to differentiate the circulating tumor cells from leukocytes (94). Besides lipid droplets, SRS microscopy can image polytene DNA in the nucleus, which contains multiple copies of tightly bound chromosomes and exhibits Raman peaks at 785 and 1,090 cm−1 (95). Quantitation of lipid storage influenced by metabolic pathways (96) was studied in C. elegans using a label-free method (97). The quantitative measurement of lipid storage has also been applied to identify 9 out of 272 genes whose inactivation increased fat content by more than 25% (98). Studies of myelin disease and spinal cord injury used CRS microscopy at the tissue level to visualize the myelin sheath in the central (99–103) and peripheral (104, 105) nervous systems. For medical imaging, Freudiger et al. (89) and Ji et al. (106) showed that, by the linear combination of SRS images at peaks of 2,850 and 2,940 cm−1 for CH2 and CH3 vibration modes, a stain-free histopathological image comparable to a hematoxylin and eosin (H&E) stained image can be achieved (Figure 3a).
To improve the chemical selectivity of single-frequency CRS microscopy, researchers deployed deuterated molecules (82, 107) or alkyne-tagged molecules (108–110) to study specific molecules from the single-cell level to the tissue level. For example, Zhang et al. (82) visualized the cellular uptake of deuterated fatty acid, which exhibited a C–D Raman signal at 2,133 cm−1 and showed that oleic fatty acid facilitated the conversion of palmitic fatty acid into lipid bodies. Li & Cheng (111) used isotope-labeled glucose to track its metabolism in pancreatic cancer cells and discovered that glucose was largely utilized for lipid synthesis (Figure 3b). Wei et al. (107) used deuterium-labeled amino acids and imaged newly synthesized proteins in live mammalian cells with quantitative ratios between new and total proteomes. To increase SRS signals from the tag molecules, Wei et al. (108) reported alkyne-tagged molecules display a signal at 2,125 cm−1 by a CC chemical bond and visualized the de novo synthesis of DNA, RNA, proteins, phospholipids, and triglycerides through metabolic incorporation of alkyne-tagged precursors (Figure 3c). For imaging cholesterol storage in living cells and C. elegans, Lee et al. (110) reported the phenyl-diyne tag exhibited a Raman scattering cross-section that was 15 times larger than the alkyne group, and 88 times larger than the endogenous C = O group.
CRS microscopy has also been applied to monitor drug delivery and biomass conversion. Saar et al. (112) showed quantitative, selective imaging of the deuterated nonsteroidal anti-inflammatory drug ibuprofen in propylene glycol solutions penetrating skin tissue using the Raman band at 2,133 cm−1 (Figure 3d). Saar et al. (113) also visualized delignification in corn stover stem by imaging the 1,600 cm−1 Raman peak from aromatic stretching in lignin.
Although single-frequency CRS uses all the laser energy in a specific Raman band to maximize the image speed, it is accompanied by loss of spectral information. To address this issue, Slipchenko et al. (114) coupled a single-frequency CRS microscope with a spontaneous Raman spectrometer. Thus, the single-color CRS image of a specific Raman band served as guidance and Raman spectroscopy was performed at the area of interest. Slipchenko et al. (114) reported that the fat in adipocytes in living animals was composed of more unsaturated fatty acid than cultured adipocyte 3T3-L1 cells (Figure 3e). This method enabled Le et al. (115) to quantify the unsaturation level of lipids in wild-type and mutant C. elegans. Yue et al. (116) further used this platform to quantitatively analyze the lipid storage in prostate cancerous tissues and discovered the accumulation of esterified cholesterol in lipid bodies in aggressive prostate cancer cells.
Frame-by-frame hyperspectral CRS imaging is performed by tuning the excitation wavelength over a defined spectral window. The wavelength tuning can be achieved by using a mode-locked and frequency-doubled picosecond Nd:YVO laser that synchronously pumps an OPO to generate a signal beam and an idler beam (16). Hyperspectral CARS and SRS imaging were generally done by tuning the temperature of the periodically poled potassium titanyl phosphate or the lithium triborate crystal (LiB3O5) inside the OPO and adjusting the etalon setting, with a scanning range of ~200 cm−1 (26, 72, 117). A limitation of this method is that it takes seconds for each step of tuning. Alternatively, Garbacik et al. (118) adjusted the angle of the intracavity Lyot filter that was synchronized to the galvano mirrors and reached the scanning range of 2,880–3,020 cm−1 at 2 cm−1 resolution and subsecond tuning speed. Kong et al. (119) developed an electro-optical tunable Lyot filter and achieved a tuning range of ~115 cm−1 at 100 μs per step.
Hyperspectral CRS imaging was also achieved with femtosecond pulses. One approach for spectral tuning is to utilize the wavelength components in a broadband femtosecond pulse. Ozeki et al. (120) used an ytterbium (Yb) fiber oscillator to generate a broadband pulse with a bandwidth of ~30 nm and sent the broadband pulse to a high-resolution, tunable bandpass filter in which a galvano mirror was used to tune the angle between the grating and the laser beam. The authors reported a spectral tuning range close to 300 cm−1, a resolution of 3 cm−1, and video-rate spectral acquisition speed (29, 120). Zhang et al. (27) utilized two femtosecond pulses from a synchronously pumped OPO system. Two pulse shapers transformed the pulses to picosecond, and an automated slit on the Fourier plane in one pulse shaper selected the excitation wavelength sequentially (27). Another approach is based on spectral focusing, in which a linear positive or negative chirp is introduced to femtosecond pulses. Therefore, Raman shift tuning can be achieved by scanning the temporal overlap of the two chirped pulses. By this method, hyperspectral CARS (17) and SRS (28) with a tuning range of ~280 cm−1 have been demonstrated. Using this concept, Beier et al. (121) used a part of a femtosecond oscillator to pump a photonics crystal fiber and generated a supercontinuum as the Stokes beam. By introducing the positive chirp to the other part of the femtosecond oscillator and scanning the temporal overlap, the authors demonstrated a spectral coverage of ~800 cm−1 (121). Andresen et al. (122) generated a soliton, which is red-shifted due to a self-frequency shift, as the Stokes beam. By introducing a negative chirp to both beams and scanning the temporal overlap, the authors reported a spectral coverage of ~280 cm−1 (122).
Frame-by-frame hyperspectral CRS microscopy has been applied to chemical imaging of single cells and tissues by resolving overlapped Raman bands. Using this approach, Fu et al. (71) demonstrated decomposition of saturated lipid, unsaturated lipid, cholesterol, protein, and water in fixed cells; Di Napoli et al. (123) showed quantitative imaging of saturated and unsaturated lipids in human adipose-derived stem cells (Figure 4a). By culturing C. elegans, a live test subject, with deuterated palmitic and oleic acid, Fu et al. (124) reported that unsaturated fatty acids are utilized preferentially for triolein synthesis incorporated into lipid droplets. Using fingerprint C = C vibrations, Wang et al. (75) discovered that lysosome-related organelles are the sites of cholesterol storage in C. elegans (Figure 4b).
At the tissue level, Ozeki et al. (29) showed the potential of hyperspectral SRS for medical imaging, in which lipid, cytoplasm, filament, nucleus, and water were clearly separated (Figure 4c). Hyperspectral CARS was used to study lipid storage in muscle tissue (125) and identify cholesterol crystals from condensed aliphatic lipids in atherosclerotic plaques in mouse model (126). Suhalim et al. (26) showed that the structure of crystalized cholesterol in plaques can be analyzed by integrating SRS microscopy with second harmonic generation modality (Figure 4d); Wang et al. (74) discriminated cholesterol crystals from lipids in the pig model, and further quantified the cholesterol ester level and the degree of unsaturation in lipid droplets, by hyperspectral SRS. For applications to drug delivery, Fussell et al. (127) studied the dissolution rate of oral dosage containing theophylline anhydrate and the corresponding change of the solid state. Fu et al. (128) further reported the accumulation of the Abelson tyrosine kinase inhibitor in the lysosome, which is one of the frontline therapies for chronic myelogenous leukemia inside single cells (Figure 4e).
For plant science and agrochemical research, Mansfield et al. (117) reported in situ chemical analysis of plant cell walls and epicuticular waxes; Liu et al. (129) mapped the aromatic ring of lignin, aldehyde, and alcohol groups in lignified plant cell walls and uncovered the spatially distinct distribution of aldehyde and alcohol groups in the thickened secondary cell wall. These studies collectively showed the potential of hyperspectral CRS microscopy with its high imaging speed and intrinsic three-dimensional spatial resolution.
In multiplex CRS, the acquisition of spectra is performed on a pixel-by-pixel base. Compared to frame-by-frame hyperspectral imaging, the advantage of multiplex acquisition is that it avoids motion-induced spectral distortion. Therefore, high-speed pixel-by-pixel spectral acquisition is preferred for in vivo applications.
The first approach for multiplex CRS microscopy is to simultaneously detect multiple Raman bands in the spectral domain. Multiplex CARS microscopy was initially demonstrated (57, 58) with synchronized femtosecond and picosecond pulses as the broadband Stokes beam and the narrowband pump beam, respectively. A prism or grating dispersed the signal light and a CCD array collected the spectrally separated CARS signals. The use of a supercontinuum source further broadened the spectral window to ~3,000 cm−1 (19, 20, 59–61). Heterodyne spectral interferometry retrieved the Raman-resonant spectral information from the nonresonant background (130–132). To date, the fastest spectral acquisition time is 3.5 ms, covering a broad Raman spectral window ranging from 500 to 3,500 cm−1 (21). The spectral acquisition speed of multiplex or broadband CARS is limited mainly by the CCD readout rate.
Multiplex SRS microscopy requires parallel detection of dispersed SRS signals. Lu et al. (133) demonstrated multicolor SRS using three independent lock-in amplifiers. Using a multichannel lock-in amplifier with a frequency of hundreds of kilohertz, Seto et al. (64) developed 128-channel multiplex SRS microscope. The use of different detector arrays for multiplex SRS was discussed by Marx et al. (62). With a complementary metal-oxide semiconductor array, Rock et al. (63) recorded an SRS spectrum within 20 ms. A 32-channel tuned-amplifier array integrated with a photodiode array significantly reduced the spectral acquisition time to 32 μs, covering a spectral window of ~200 cm−1 (31).
The second approach is to detect the simultaneously excited Raman bands in the time domain. Weiner et al. (134) reported impulsive SRS in which a sudden impulse drove multiple Raman vibration modes simultaneously and Fourier transformation retrieved the spectral information from the time-dependent phonon oscillations on the femtosecond timescale. Fourier-transform CARS microscopy was demonstrated by scanning the time delay between the impulsive excitation pulse and the probe pulse through either a Michelson interferometer (135) or frequency combs (136).
The third approach is to code the excitation photons in the spectral domain, collect the photons with a single detector, and decode the signals thereafter. Fu et al. (137) reported three-color SRS by modulating each color of the excitation laser with an acousto-optical tunable filter at a specific frequency at kilohertz rate. By scanning a dispersed excitation beam on a spatial pattern inside a femtosecond pulse shaper, Liao et al. (138) demonstrated multiplexed modulation at megahertz rate and spectral acquisition within 60 μs per pixel.
Multiplex CRS microscopy has enabled multiple applications in imaging a living system. Multiplex CARS was used to study the fluidity and density of lipid in vesicle and supported bilayers (57, 139–142), and to monitor the transition to lipid phase (40). Parekh et al. (19) used broadband CARS to discriminate between nuclear and cytoplasmic regions of single living cells. Okuno et al. (143) observed the spectral profile of mitochondria in living budding yeast cells in the fingerprint region. Kano & Hamaguchi (144) visualized multiple organelles including mitochondria, septum, and nucleus in yeast cells. Rinia et al. (145) used multiplex CARS to map the lipid unsaturation and acyl-chain order of different lipid droplets in 3T3-L1 cells. Liao et al. (31) used multiplex SRS to measure the percentage of cholesterol ester storage in lipid droplets in live cancer cells, which is a potential indicator of cancer aggressiveness (Figure 5a). Broadband CARS microscopy showed its potential in histopathology. Pohling et al. (146) demonstrated imaging contrast of gray and white matter as well as layers of granule and Purkinje cells in fresh mouse brain tissue. Camp et al. (21) presented CARS images of healthy and neoplastic tissues, including xenograft glioblastoma in mouse brain, pancreatic ducts, and liver tissues, and showed clear contrast of nucleus, collagen, and lipid (Figure 5b). Liao et al. (138) demonstrated in situ SRS imaging of human breast cancerous tissue that provided the same information as the H&E stained histologic result (Figure 5c). Additional applications include quantitative and qualitative assessments of multicomponent pharmaceuticals (36), decomposition of individual component in polymer blends (60, 68), and discrimination of fat stores from protein-rich organelles in C. elegans (31).
Both single-frequency and hyperspectral CRS microscopy have generated promising results in compositional analysis of single cells and tissues. Looking into the future, we expect several advancements in developments and applications. First, microsecond-scale SRS spectroscopic imaging, based on parallel detection (31), modulation frequency multiplexing (138), or methods yet to be developed, will enable fast chemical analysis of living systems and high-throughput analysis of single cells in a flow cytometry setting. Second, integration of CRS microscopy with tailor-made Raman tags will enable broad functional imaging applications such as proteins synthesis and lipid metabolism. Third, device miniaturization and reductions in laser costs will transform CRS microscopy into a mobile analytical tool. The solid-state laser currently used for CRS is generally bulky, costly, and sensitive to the environment. Thus, development of reliable and cost-effective fiber laser sources will be crucial for bringing CRS microscopy to clinic. To date, several proof-of-concept works based on lab-built fiber laser systems have shown the potential clinical applications for CARS, such as an in vivo endoscope and a miniaturized probe (147–149). For SRS microscopy that demodulates the Raman resonant signals from the local oscillator, a low-noise fiber laser source as the excitation and an auto-balanced detector were demonstrated (150). A mobile system with such a reliable fiber laser source is expected to produce insightful clinical data in the near future. In conclusion, with continuous efforts to improve detection sensitivity, reduce costs, and promote commercialization, we anticipate CRS microscopy will have a broad impact on both discovery-driven biological sciences and label-free molecular-guided diagnoses.
This work was supported by National Institutes of Health grants R01HL125385, R21CA182608, and R21GM114853 and a grant from the W.M. Keck Foundation to J.-X.C. The authors thank Hyeon Jeong Lee, Wei Chen, Yin-Xin Zhang, and Caroline Karanja for proofreading the manuscript.
J.-X.C. has a financial interest in Vibronix Inc. C.-S.L. is not aware of any affiliations, memberships, funding, or financial holdings that might be perceived as affecting the objectivity of this review.