1. Introduction

Coherent anti-Stokes Raman scattering (CARS) is a four-wave mixing (FWM) process used for probing the resonant vibrational states of materials. CARS occurs when at least two light pulses—pump/probe and Stokes, having respective frequencies of ω_P,Pr and ω_S—interact in a non-linear material such that the difference frequency, Ω_R = ω_P,Pr – ω_S, matches a vibrational mode of the material, resulting in the resonant generation of photons at the anti-Stokes frequency ω_AS = 2ω_P,Pr – ω_S [1–3]. With its combination of rapid imaging with rich spectroscopic information, CARS “hyper-microscopy” is gaining popularity in research fields such as biomedical imaging [4–6], materials science [7], and biology [8–11].

Experimental implementations of CARS microscopy fall into two categories based on the means of signal detection: spectral-detection (multiplex) and single-wavelength detection. In multiplex-CARS, one or both of the pump/probe and Stokes beams are spectrally broad and a wide range of vibrational frequencies are probed simultaneously. Anti-Stokes light is detected with a spectrometer and hyperspectral images are composed one pixel-spectrum at a time. This approach could be considered as “spectrum first”, where rapid collection of good-quality and dense CARS spectra is a dominant attribute. However, despite having spectral acquisition times as low as a few ms [6], such pixel dwell-times limit image acquisition to several minutes. With single-wavelength detection, a fixed vibrational frequency is probed at any given time providing the requisite contrast for CARS imaging [12,13]. CARS spectra can then be built by scanning the pump-Stokes frequency difference. Traditionally, this approach has utilized synchronized picosecond pulses to most efficiently excite sharp vibrational resonances, and the CARS spectrum is scanned by systematically varying the wavelengths of one of the sources. Such frequency-scanning CARS can thus be thought of as “imaging first”, where rapid—even video rate [13]—image collection enables live-cell imaging, albeit at the expense of sparse or time-consuming hyperspectral capabilities. A hybrid technique that utilizes both single-wavelength detection and broadband laser excitation is known as spectral-focusing CARS (SF-CARS) [5,14–16]. With SF-CARS, a narrowband pump pulse and broadband Stokes pulse are stretched in time such that their frequency-time relationship (i.e. their “chirps”) are made to be equal. The instantaneous frequency difference of such chirp-matched pulses has significantly narrower bandwidth, which allows for improved CARS spectral resolution. Furthermore, this instantaneous frequency difference—i.e. the probed CARS frequency—can be simply scanned by sweeping the relative temporal overlap between the two pulses. Because it utilizes single-wavelength detection, SF-CARS is predominantly an imaging-first technique; and since the spectrum can be scanned in a matter of seconds-to-minutes [5,10,16,17], it can yield good-quality hyperspectral image stacks at a rate that challenges multiplex-CARS techniques.

Among the various CARS microscopy techniques, SF-CARS offers the greatest flexibility to switch between spectroscopy and imaging or perform a mixture of both. SF-CARS has been implemented using three distinct experimental approaches: by using two synchronized laser sources to generate pump/probe and Stokes [15]; by using a single spectrally-broad sub-25-fs laser source with intra-pulse filtering to generate the pump/probe and Stokes [16,18]; or by using a fs-laser-pumped photonic crystal fibre (PCF) source to generate a supercontinuum from which some (or all) of the pump/probe/Stokes beams are derived [5,9,10,14,19]. Implementations of SF-CARS that use a single spectrally-broad femtosecond laser source have been successful in performing both CARS imaging and dense spectral acquisition, albeit with the added challenges and costs of utilizing an ultra-broadband laser system. On the other hand, SF-CARS systems using supercontinuum generating-PCFs (SCG-PCFs) to generate the Stokes excitation beam are among the most cost-effective solutions for CARS microscopy, although they generate supercontinua with spectral intensities and bandwidths that are often insufficient for spectrally-broad CARS microscopy. To date, an approach that requires only a single set of optics that can be used for broadband SF-CARS microscopy spanning both the fingerprint and the CH/OH vibrational frequencies has not been demonstrated. For example, in some recent implementations of SF-CARS, different optics [16], pump/probe wavelengths [17], or totally different experimental setups [18] have been used to separately access the fingerprint and the CH/OH frequencies.

Access to the broadest range of Raman frequencies with supercontinuum-based SF-CARS has been limited by both the spectral intensity and bandwidth generated by existing PCFs. The PCF most typically used for SF-CARS [20] has been selected by some groups for the stability and spectral density of its generated supercontinuum [14,21–23]. Previous work by Hilligsoe et al. highlighted this fibre’s efficiency in generating supercontinuum beyond its two zero-dispersion wavelengths (ZDWs), which are 775 and 945 nm, using sub-70 fs input pulses [19]. However, within the two ZDWs, this fibre has not been shown to generate supercontinuum with enough spectral intensity to be useful for CARS. This means that when using a standard 800-nm (12500 cm⁻¹) wavelength for both pump/probe and PCF input pulses, Stokes can only be generated at wavelengths above 945 nm (<10580 cm⁻¹). This makes it difficult to detect vibrational frequencies below 1900 cm⁻¹. While this limitation can be overcome by tuning the pump/probe wavelength to above 900 nm (<11110 cm⁻¹) [17], the resulting Stokes supercontinuum is insufficiently broad to enable access to the important CH/OH vibrational region. In this work, we show that by using conveniently-chirped (i.e. not transform-limited) pulses as input for the PCF, one can generate supercontinuum both within and beyond the two-ZDWs with sufficient spectral power for broadband CARS microscopy. In doing so, we demonstrate hyperspectral CARS microscopy that can span 630 cm⁻¹ – 3250 cm⁻¹ using >100 fs, 800 nm pump pulses without the need for wavelength tunability. Relaxing the requirements for transform-limited pulses and tunable laser oscillators significantly lowers the complexity and cost of what is already one of the least expensive and most agile approaches to implementing multimodal CARS hyperspectral microscopy, thus opening the technique to a new segment of nonlinear-optical microscopy laboratories.

2. Experiment

A schematic of the experimental setup is shown in Fig. 1. A Ti:Sapphire oscillator (Spectra-Physics Tsunami) generates a beam of 800 nm, 190 femtosecond pulses. A half-wave-plate and cube polarizer act as a variable beam splitter to create the pump/probe and Stokes-generating arms. The pump/probe (henceforth: “pump”) is routed to a computer-controlled optical delay stage before re-combining with the Stokes. The other beam is coupled into a commercial PCF-based supercontinuum generation module (FemtoWHITE-CARS, NKT Photonics) using a Newport M-40 × objective lens to generate the Stokes beam. The Stokes beam is collimated with a Mitutoyo M Plan NIR 50 × long-working distance objective. Coupled power into the PCF is ~54%. Both beams are combined using a Semrock LPF-937 filter oriented at 45 degrees incidence, making it an effective 840 nm long-pass filter. The pump and Stokes beams are further dispersed by a total of 101 mm and 158 mm of S-NPH2 glass, respectively, before entering the microscope. Not including other dispersion in the optical setup, this added glass represents 30,000 fs² of dispersion in the pump arm, and between 28,000 fs² and 43,000 fs² of dispersion for the Stokes beam spanning 1090 nm (Ω_R = 3300 cm⁻¹) to 840 nm (Ω_R = 600 cm⁻¹), respectively. The lengths of these blocks of glass are chosen to match the chirps of the pump and Stokes pulses, thereby enabling spectral-focusing CARS with an optimized spectral resolution of ~25 cm⁻¹ at the CH vibration region (~2900 cm⁻¹). As discussed below, the broadband Stokes is predominantly linearly chirped, with little higher-order dispersion effects, and thus we do not find the spectral resolution to vary significantly along the CARS spectrum. The microscope is a modified Olympus IX73 inverted laser-scanning microscope utilizing an Olympus UApo-N340 40 × objective, Thorlabs scanning galvos and x-y computerized sample stage, and a unique multimodal fibre collection system in the transmitted direction, as described previously [24]. Two-photon excitation fluorescence (TPEF) signals are collected in a non-descanned epi-channel. Second harmonic generation (SHG) and anti-Stokes (CARS) signals are collected in the forward direction and are wavelength-separated by a dichroic filter, with each of the two signals going to separate detectors. Hamamatsu H10723-01 PMTs are used in all three channels. Alternatively, the SHG and CARS signals can be sent to a modular USB spectrometer (Black Comet CXR SR-50, StellarNet) for signal monitoring and for calibration of vibrational frequency as a function of pump delay stage position. The Stokes spectra generated by the SCG module were recorded using an optical spectrum analyzer (HP 70450A). Pulse durations and chirps were measured with a home-built scanning interferometric autocorrelator [25]. Data acquisition and microscope control were implemented using a custom LabView interface. Data analysis and image processing were conducted using LabView, Python, and ImageJ [26].

 

Fig. 1 A simplified schematic of the multimodal CARS microscopy setup. A Ti:Sapphire oscillator generates 190 fs pulses at 800 nm, which are split into pump and Stokes-generating beams. The Stokes-generating beam passes through a Faraday isolator (FI) before being coupled into a FemtoWHITE CARS (NKT photonics) microstructured-fibre module that generates the Stokes supercontinuum. Blocks of high-dispersion S-NPH2 glass are used to disperse and match the chirps of the pump and Stokes pulses. The pump and Stokes beams are recombined using a long-pass filter (LPF) and routed to the laser-scanning microscope. Isotropically-generated TPEF is collected in the backwards direction, reflected by a dichroic and detected by a PMT. Forward-generated SHG and CARS signals are isolated with a short-pass filter (SPF) and collected using a customized multimode-fibre assembly and routed off-board, where they are wavelength-separated en route to separate PMTs.

Download Full Size | PDF

The samples used in this study were algae-derived astaxanthin-rich extract (AstaReal L10) diluted as needed in food-grade canola oil, author-collected ‘stargazer’ lily pollen dispersed in tap water, reagent-grade benzonitrile, dimethyl sulfoxide (DMSO), and nitrobenzene. DMSO was diluted 50:50 with water and then dispersed in benzonitrile together with cotton fibres from Puritan Brand cotton swabs, and placed between a standard glass slide and coverslip to produce a heterogeneous sample mixture. The astaxanthin extract served as a calibration sample, both the solvent mixture and lily pollen samples were used for demonstrating multimodal hyperspectral CARS imaging. Nitrobenzene served as a reference sample to demonstrate the efficacy of Raman-retrieval algorithms in SF-CARS. The astaxanthin, pollen, solvent mixture, and nitrobenzene samples were characterized using respective pump powers of 3.7, 7.4, 74, and 300 mW, with a fixed Stokes power of 4.2 mW; all powers measured at the sample plane.

3. Results and discussion

The FemtoWHITE CARS SCG module houses a microstructured fibre with two zero-dispersion-wavelengths (ZDWs) at 775 nm and 945 nm [19,20]. Originally, the SCG spectra from the FemtoWHITE CARS were modeled using transform-limited 40 fs input pulses, and were studied as a function of pulse energy rather than peak intensity [19]. With such short pulses, cascaded nonlinearities provide efficient wavelength shifting and broadening, rapidly evolving spectral weight beyond the ZDWs, and leaving the region between the pumping wavelength (790 nm) and the ZDWs largely devoid of spectral power. One consequence of efficient supercontinuum generation is that access to frequencies near that of the pump is limited. For example, in a prior demonstration of CARS microscopy at fingerprint (<1800 cm⁻¹) frequencies using the same PCF [17], the Stokes generation efficiency was reduced by lowering the pump power into the PCF, while concurrently shifting the pump wavelength above 900 nm, closer to the 945 nm ZDW. While this approach works to some degree, the need to change the pump wavelength significantly away from 800 nm is problematic in that it requires changing various routing and collection filters; requires detection of anti-Stokes wavelengths above 820 nm where alkali PMTs are precipitously inefficient; and requires a broadly-tunable laser oscillator source. Furthermore, the need to reduce the input power for the SCG module meant a corresponding reduction in generated Stokes power available for imaging [17]. Here, we find that by using longer pulses (or ones that are already chirped) we can generate significant Stokes power in the spectral region between 800 nm and 945 nm. Using longer pulses means that the peak intensity in the fibre is reduced, thereby significantly reducing the efficiency of the SCG process. This is an advantageous inefficiency that allows us to access CARS frequencies across the fingerprint, silent, and CH/OH regions. This principle is demonstrated in Fig. 2, where it is shown that pumping the PCF using 190 fs, 800 nm, pulses generates supercontinuum close to the pump wavelength, which is useful for accessing the fingerprint vibrational frequencies. Introducing additional dispersion to the input pulse generates a supercontinuum with greater spectral intensity closer to the pump wavelength. In this case, as shown in Fig. 2(b), the use of a Faraday isolator is doubly-advantageous as it adds dispersion to the SCG input pulse for improved Stokes generation while also eliminating feedback to the oscillator. Such feedback is a common nuisance when using SCG-PCF modules [27]. However, because conventional wisdom dictates that for low-noise supercontinuum generation the SCG module must be pumped with short transform-limited pulses, any use of an isolator is invariably complicated by the added need for a pulse compressor [28]. As shown in Fig. 2(b), the added dispersion offered by the isolator provides a more useful range of supercontinuum wavelengths for probing a wider span of vibrational frequencies, obviating the need for a pulse compressor. For any given pulse energy, the spectrum of the generated supercontinuum ultimately depends on the peak power in the fibre, and thus adding too much dispersion before the SCG module leads to insufficient broadening beyond the 945 nm ZDW, and thus to a diminished utility of the generated continuum. For example, adding an extra 25 mm of high-dispersion S-NPH2 glass before the PCF yields a supercontinuum with more power close to the pump wavelength of 800nm, but with insufficient power above 885 nm to probe vibrational frequencies above 1200 cm⁻¹, as shown in Fig. 2(c).

 

Fig. 2 Spectra of the FemtoWHITE-CARS output vs. coupled pump power plotted in wavelength (left scale) and the difference frequency (right scale), Ω_R = ω_P – ω_S. Generated supercontinuum (a) from transform-limited 190 fs input pulses; (b) when the 190 fs input pulses are dispersed to 200 fs by a faraday isolator; (c) when the input pulses are further dispersed to 315 fs with a block of high-dispersion glass. The dashed horizontal lines are guides to the eye that correspond to Stokes wavelengths used to probe some relevant fingerprint and CH/OH vibrational frequencies. The dashed vertical line in (b) represents the slice of the Stokes spectrum (coupled power of 110 mW) used for the proceeding hyperspectral imaging experiments.

Download Full Size | PDF

In SF-CARS, the vibrational spectrum is scanned by varying the temporal overlap of the pump and Stokes beams. In order to calibrate the frequency scale, we route the CARS signal to a spectrometer and collect the “anti-Stokes” signal (ω_AS = 2ω_P – ω_S) generated by the FWM “nonresonant background” (NRB) in a nonlinear sample. This allows us to calibrate the stage delay with respect to the vibrational frequency (Ω_R = ω_P – ω_S) being probed. An astaxanthin-rich sample (AstaReal L10) exhibits very strong FWM at low pump powers [9,10], and is used as the nonlinear sample for frequency calibration. Alternatively, calibration can take place via FWM in the coverslip glass or in water, but this requires nearlytwo orders of magnitude more pump power. Figure 3(a) shows the FWM spectrogram of astaxanthin as a function of the pump delay-stage position (and corresponding pump-Stokes time delay). The FWM signal in astaxanthin is seen to be sufficiently strong for frequency calibration across a broad spectral range, with some prominent and narrow peaks in the spectrogram attributed to vibrational resonances (CARS); most notably around 1520 cm⁻¹, 1150 cm⁻¹, and 1000 cm⁻¹ [29]. As seen from the spectrogram, the Stokes spectrum is sufficiently broad to allow probing of vibrational frequencies from 630 cm⁻¹ to 3250 cm⁻¹. Previous studies [17] measured the Stokes pulses to be essentially linearly chirped, with little indication of higher-order dispersion effects coming from either the SCG module or the added high-dispersion glass. Despite appearing linear at any given 1000 cm⁻¹ span, close inspection of Fig. 3(a) shows that when sampling a wider spectral span, the spectrogram has a distinct positive curvature. With our single-scan access to a much wider range, nonlinear chirp effects become more pronounced. We find that a simple quadratic polynomial fit suffices for calibrating stage position to vibrational frequency. A CARS spectrum of astaxanthin obtained through a point scan is shown in Fig. 3(b). This “raw” CARS spectrum contains both narrow peaks (as labeled) assigned to the expected vibrational resonances of astaxanthin across the fingerprint region, as well as broader features that manifest between 2500 cm⁻¹ and 3000 cm⁻¹ which represents the non-resonant FWM signal that follows variations in the Stokes spectrum.

 

Fig. 3 Frequency-calibrating spectrogram and CARS spectrum of astaxanthin. (a) Spectrogram with a sampling spatial resolution of 50 μm (i.e. temporal resolution of 333 fs). The duration of the highly-chirped Stokes supercontinuum at the sample is approximately 25 ps. (b) CARS spectrum of astaxanthin obtained through point scan having 900 data points obtained in 100 s, and the corresponding Stokes spectrum plotted as a function of both wavelength and the vibrational frequency probed. The power densities of the Stokes at the 897 nm and 1001 nm peaks are estimated to be 43 µW/nm and 24 µW/nm, respectively. The pump power and the integrated Stokes power were measured to be 3.7 mW and 4.2 mW at the sample plane, respectively.

Download Full Size | PDF

To demonstrate the multimodal and hyperspectral capabilities of our technique, imaging was performed on two samples: an artificial sample consisting of a mixture of benzonitrile and DMSO along with a cotton (cellulose) fibre (Fig. 4(a)), and a grain of lily pollen (Fig. 4(c)). For the artificial sample, benzonitrile was imaged by the strong CARS peak centered at 3074 cm⁻¹; DMSO using a strong CARS peak centered at 2934 cm⁻¹; and cellulose using SHG. Additionally, a nearly-identical image is obtained via contrast for benzonitrile in the “silent region” at 2240 cm⁻¹ (not shown), however, we chose 3074 cm⁻¹ and 2934 cm⁻¹ as the two CARS contrasts to highlight the spectral sensitivity of our SF-CARS system. Stimulated vibrational imaging of cellulose, both at the fingerprint [17, 30] and CH-OH frequency ranges [31–33], has been previously demonstrated. In our sample, the CARS signal from cellulose at 1100 cm⁻¹ is detected, but the signal is largely swamped by the large NRB from the solvents that soak the fibre. In practice, the cotton fibre is included for SHG contrast, to demonstrate compatibility of our CARS technique with other nonlinear optical imaging modalities. For the lily pollen, carotenoid was imaged using the peak centered at 1154 cm⁻¹; and muri channels are imaged via endogenous TPEF in the epi-channel. We find carotenoid to accumulate in lumina, decorated by the muri [34]. Interestingly, we also find strong endogenous TPEF signal from an unidentified substance that is co-localized with the carotenoid CARS signal.

 

Fig. 4 A demonstration of the hyperspectral imaging capabilities of the multimodal CARS setup. (a) A 200 x 200 pixel multimodal image of benzonitrile (CARS; blue contrast; 12 frames centred at 3074 cm⁻¹), DMSO (CARS; green contrast; 13 frames centred at 2910 cm⁻¹), and cellulose fibre (SHG; red contrast). (b) CARS spectra from two 10 x 10 pixel regions of interest (ROIs) showing the CARS spectrum from benzonitrile and DMSO. (c) A 200 x 140 pixel (cropped) multimodal image of lily pollen containing carotenoids in the luminae (CARS; red contrast; 43 frames centered at 1154 cm⁻¹), and muri (TPEF; white contrast). (d) Raw CARS spectrum averaged over 10 x 10 pixel ROI in (c). The shaded spectral regions in (b) and (d) correspond to the image stacks averaged to produce the multimodal image. Each image stack was taken at a duration of 0.8 seconds with a 13-μs pixel dwell time using pump powers of 74 mW for (a) and 7.4 mW for (c) measured at the sample plane. The input power in the PCF was maintained at 110 mW (200 fs pulse duration) which generates supercontinuum with a power of 4.2 mW as measured at the sample plane.

Download Full Size | PDF

In terms of image and spectral acquisition speeds, each 200 x 200 pixel frame in the hyperspectral stack was obtained in 0.80 s, corresponding to a 13-μs pixel dwell time. The pump delay stage—and thus instantaneous CARS frequency—is continually scanned throughout the hyperspectral stack acquisition. This procedure can lead to a “smearing” effect of the vibrational frequencies probed in different portions of an image frame, but depending on the delay-stage scan speed and image acquisition speed the effect is negligible. For example, with vibrational frequencies scanned at an average rate of 4.3 cm⁻¹/s, each frame is calculated to have an average frequency bandwidth of 3.4 cm⁻¹. Of course, this smearing can be minimized by slowing down the scan speed. Ultimately, this reduction in sampling resolution negligibly impacts the obtained spectral resolution of our system, which is largely determined by the amount and similarity of chirps (dispersion) imparted to the two beams by the high-dispersion glass blocks [16,17]. By comparison, the hyperspectral acquisition capability of our system is about four times slower, and the contrast-based imaging speeds are over 100 times faster than the current state-of-the-art multiplex CARS approach [6]. On the other hand, our system has a significantly slower imaging frame-rate but significantly higher spectral acquisition capabilities compared to the commonly utilized picosecond-pulse CARS systems [12,13]. Our current approach to SF-CARS has the ability to image live cells and molecular dynamics at 2 fps, and can perform a 2500-cm⁻¹-wide hyperspectral image acquisition in under ten minutes or microspectroscopy with point scans in under two minutes. Thus, this version of SF-CARS is most advantageous for research that requires both rapidcontrast-based imaging and rich label-free molecular fingerprinting where chemical/structural dynamics can be monitored on the second-to-minute timescales, such as live cell imaging and materials science applications.

Retrieving Raman-like signals from CARS spectra has been a challenge in CARS hyperspectral microscopy given that non-resonant background signals, such as FWM, affect the overall CARS spectral line shape. Despite the wide spectral range of our scans, we find that our technique is compatible with common algorithms for removing the effects of the nonresonant background. To demonstrate this, we collected a CARS spectrum from nitrobenzene, followed by a reference nonresonant-background scan from the sample’s coverslip. For retrieval, we employ an algorithm based on the time-domain Kramers-Kronig transform, closely following the approach presented by Liu et al [35]. As demonstrated in Fig. 5, such retrieval algorithms largely remove the effects of the highly nonuniform Stokes spectrum on the NRB, and provide a CARS-retrieved spectrum that closely resembles the spontaneous Raman spectrum.

 

Fig. 5 Demonstration of the efficacy of a Kramers-Kronig-based Raman-retrieval algorithm [35] in spectrally-broad SF-CARS. The Raman-like spectrum (blue) was retrieved from the raw CARS signal (black) and a nonresonant background spectrum collected in the sample coverslip. A comparison to the spontaneous Raman spectrum (red) is shown as a reference.

Download Full Size | PDF

4. Conclusion

In this work we have demonstrated spectral-focusing-based hyperspectral CARS microscopy that spans over 2500 cm⁻¹ of the vibrational spectrum using a single 800 nm femtosecond laser, where the Stokes is generated by a commercially-available PCF. Despite typically being used to produce a stable supercontinuum outside of its two ZDWs, the spectral output of this PCF can be manipulated via power-tuning its input pulses to yield spectra with significant power densities in the wavelengths between the two ZDWs, thus allowing better access to lower vibrational frequencies deep in the fingerprint. Furthermore, the spectral power produced within the two ZDWs can be increased by concurrently increasing the seed pulse power and pulse duration. This atypical use of the PCF opens up economical and highly versatile single-laser coherent microscope systems capable of broadband CARS microscopy that integrate well with existing nonlinear optical modalities such as TPEF and SHG. Overcoming the prior need to tune the pump wavelength away from 800 nm to be able to access the fingerprint region, together with the ability to operate with lasers generating 100 fs to 200 fs pulses significantly reduces the technical requirements of the experimental system and opens the addition of hyperspectral-CARS microscopy to a new range of pre-existing nonlinear optical microscopy laboratories. In the current work, we demonstrated access to vibrational frequencies as low as 630 cm⁻¹. In the future, careful power-tuning of the Stokes spectrum and a judicious choice of filters/combiners will allow access to lower frequencies and a broader span of the CARS spectrum. Furthermore, while the current work focused on power-tuning the supercontinuum generation in a commercial PCF module with a 12-cm fibre, studying the possibility of utilizing smaller lengths of the same fibre [19] to obtain higher spectral powers below 945 nm is a promising avenue for future research.