Synchronously pumped Raman laser for simultaneous degenerate and nondegenerate two-photon microscopy

: Two-photon fluorescence microscopy is a nonlinear imaging modality frequently used in deep-tissue imaging applications. A tunable-wavelength multicolor short-pulse source is usually required to excite fluorophores with a wide range of excitation wavelengths. This need is most typically met by solid-state lasers, which are bulky, expensive, and complicated systems. Here, we demonstrate a compact, robust fiber system that generates naturally synchronized femtosecond pulses at 1050 nm and 1200 nm by using a combination of gain-managed and Raman amplification. We image the brain of a mouse and view the blood vessels, neurons, and other cell-like structures using simultaneous degenerate and nondegenerate excitation.


Introduction
Two-photon microscopy (2PM) has become a standard technique used for cellular-resolution in vivo deep tissue microscopy [1]. In standard 2PM, it can be difficult to distinguish between different types of cells or other structures. One way to address this problem is to label the structures to be tracked with spectrally distinct fluorescent dyes, whose fluorescence can be demultiplexed by separating the detected signal into distinct spectral bands and applying an appropriate unmixing algorithm [2]. However, efficiently exciting fluorophores with widely separated excitation spectra usually requires exciting them at different wavelengths. For many fluorophores, this can significantly increase the complexity and cost of the pulsed excitation laser. So far, most systems targeting multicolor excitation have sequentially taken images with tunable titanium-sapphire lasers [3][4][5][6][7][8]. Unfortunately, in these cases, separate images need to be taken for each excitation wavelength. It is possible and conceptually simple to use completely separate laser systems to excite fluorophores at separate wavelengths, but doing so significantly increases the complexity and cost of the system [9,10]. The use of synchronized short pulses at separated wavelengths alleviates this issue by not requiring tuning the output laser for different wavelengths and by providing a nondegenerate excitation channel, for a total of three spectrally distinct excitation channels in a single image [11][12][13].
The requirements for a laser source for multispectral deep-tissue 2PM are similar to those for standard 2PM, except that such pulses must be synchronously generated at widely separated wavelengths. The generated pulses should be short (typically less than 200 fs full-width halfmaximum) and have an energy of tens of nanojoules for peak power in the range of hundreds of kilowatts before microscope optics. Additionally, most fluorophores used for 2PM have an excitation spectral bandwidth about 50 nm wide. Thus, for the multispectral case, the pulses that are generated should be separated by at least 100 nm in center wavelength to allow for a spectrally-distinct nondegenerate excitation channel. Laser systems that deliver synchronized short pulses with the above characteristics are usually constructed with bulk optical components, most commonly a titanium-sapphire laser or ytterbium fiber laser driving an optical parametric oscillator that integrates a second-order nonlinear crystal for wavelength conversion. Unfortunately, these oscillators are complex and expensive devices due to necessarily sensitive cavity alignment. Fiber lasers are typically much cheaper than their bulk counterparts, compact, and naturally robust. Nonlinear wavelength conversion of high-energy pulses in optical fiber is typically much more difficult than with bulk components though due to the lack of a second-order nonlinearity in glass and competing third-order nonlinear processes.
The most mature fiber gain media are based on ytterbium-, erbium-, and thulium-doped germanosilicate glasses. Of these, ytterbium-doped fiber is the only one that provides gain (∼1030 nm -1100 nm) close to the two-photon excitation wavelengths of fluorophores commonly used in 2PM. To excite several spectrally distinct fluorophores, it is highly desirable to generate another pulse outside the fiber gain band. Various nonlinear wavelength conversion schemes have been implemented in fiber for ultrashort out-of-band pulse generation including soliton self-frequency shift [14][15][16][17][18], parametric amplification [19], self-phase modulation [20,21], and Raman lasing [22][23][24][25][26][27][28][29][30]. A variant of the soliton self-frequency shift (soliton self-mode conversion) has demonstrated high-energy (80 nJ) and sub-100-fs pulses tunable from the ytterbium band around 1030 nm up to 1300 nm and beyond [17]. Unfortunately, the timing jitter of Raman solitons is likely to significantly increase noise of the measurement when the Raman soliton is used for nondegenerate 2PM with a pulse at 1030 nm [30]. Timing jitter reduction between Raman solitons generated simultaneously has been observed, but it is unclear whether it is possible to independently tune the wavelengths of the solitons [31]. Four-wave mixing in photonic crystal fiber has generated energetic pulses, but fiber integration is challenging due to loss when the photonic crystal fiber is spliced to step-index fiber [32]. Self-phase modulation has been used to generate megawatt-level pulses at 1300 nm, but requires the use of short (1 cm) fibers, which limits the potential for fiber integrability. Lastly, discrete Raman shifting in the past has been limited by the 13 THz Stokes shift in silica. While phosphosilicate fiber has been used to extend the spectral range of these systems, efforts to achieve pulse energies relevant for deep-tissue imaging have not yet proven fruitful.
Gain-managed nonlinear amplification [33] significantly extends the bandwidth achievable with a ytterbium-doped fiber amplifier. In this technique, the amplifier is seeded by a relatively narrowband pulse which then spectrally broadens approximately 100-fold while shifting to longer wavelengths. The result is a broadband pulse spanning from 1030 nm to 1180 nm which is compressible to close to the transform limit. The bandwidth and energy of the pulse are limited by stimulated Raman scattering (SRS). If the limit is exceeded, a significant fraction of the pulse energy is frequency-shifted into the Stokes wave. This process is incoherent because it is seeded by noise. The bandwidth achieved so far with gain-managed nonlinear amplification is not adequate for generating three spectrally separated excitation channels for fluorophores.
Here, we describe a synchronously-pumped Raman fiber laser that generates pulses at 1050 nm and 1200 nm. In conventional synchronously-pumped Raman lasers pumped with ytterbiumdoped fiber lasers, Raman lasing is limited to approximately 1080 nm, as the driving pulse is near 1030 nm. In this work, we use gain-managed nonlinear amplification to shift the first-order Stokes peak to 1200 nm while still using only ytterbium-doped silica fiber. Cavity feedback at 1200 nm redshifts a spectral slice of the broadened pump pulse by 13 THz, while leaving the rest of the pump unshifted. The unshifted pulse is then filtered to increase the spectral separation of the pulses, producing synchronized pulses at 1050 nm and 1200 nm with energies of ∼24 and ∼15 nJ, respectively. We demonstrate the utility of this system for three-color two-photon imaging by imaging the brain of a mouse and detecting blood vessels, neurons, and other cell-like structures. The excitation channel resulting from the nondegenerate excitation of the fluorophores by the two pulses provides information distinct from the two excitation channels provided by the degenerate excitation of the fluorophores by each pulse individually.

Laser design and numerical simulations
The conceptual design of the Raman laser can be seen in Fig. 1 (some components required in a real device are omitted for clarity). The short-pulse seed laser and the ytterbium-doped fiber together act as a gain-managed nonlinear amplifier, which is known to generate spontaneous Raman scattering around 1200 nm. The other optics are intended to provide positive feedback at 1200 nm by isolating the Stokes wave generated by the gain-managed pulse, synchronizing its delay relative to the repetition rate of the seed, and recombining it with the seed at the beginning of the amplifier. Without feedback present, the Stokes wave has very low energy and is incoherent. However, with the feedback present, there is the possibility of both improving the coherence of the Stokes wave and significantly increasing its energy. The behavior of this system depends on a complex interplay between the ytterbium population inversion, self-and cross-phase modulation, and Raman scattering. We performed numerical simulations to better understand the physics of the pulse propagation and thus the behavior that we would expect to observe in the laboratory setting. We solved the generalized nonlinear Schrödinger equation with Kerr nonlinearity, Raman scattering, second-and third-order dispersion, and self-steepening along with the equation for the evolution of the population inversion in a single spatial and polarization mode using a model developed recently [34]. We model a 4 m long ytterbium-doped fiber amplifier with the parameters of Nufern PLMA-YDF-10/125-VIII and a 22 m fiber dispersive delay line with the parameters of Nufern PM980-XP. An 1150-nm dichroic mirror couples the unshifted part of the pulse out of the cavity after the gain fiber. A half-wave plate and polarizing beam splitter are used as a variable output coupler. A 1200-nm long-pass filter further isolates the Stokes wave feedback, which propagates through the delay line and back into the gain fiber.
With the feedback blocked, the pulse evolution exhibits the features of gain-managed nonlinear amplification (Figs. 2(a), 2(b)). An initially narrowband pulse at 1030 nm rapidly spectrally broadens as it is amplified. The broadening eventually becomes asymmetric and the spectrum shifts to the red. When the Stokes feedback is engaged (Figs. 2(c), 2(d)), the evolution of the main pulse is similar for about 2.5 m, after which the reddest part is efficiently shifted into the Stokes wave. The properties of the output-coupled Stokes pulse are parameter-dependent, but typically the output-coupled Stokes energy is between 20 nJ and 50 nJ with a transform limited duration of approximately 80 fs. The Stokes output pulse shown in Fig. 2(d) (corresponding to λ > 1150 nm) has 26 nJ of energy with a transform-limited duration of 73 fs. The unshifted remaining pulse shown in Fig. 2(d) (corresponding to λ < 1150 nm) has 73 nJ of energy and could support a 31-fs transform-limited pulse. We then apply a 1070 nm short-pass filter to the unshifted pulse, yielding a 20-nJ pulse with a 73-fs transform-limited duration. This is done to improve the spectral separation between the pulses for imaging experiments.  While the filtered, unshifted pulse is compressible with minimal structure to 90 fs ( Fig. 3(a)), unfortunately the Stokes pulse is not highly compressible with group delay dispersion (GDD) compensation only, as can be seen in Fig. 3(d). When optimizing for the peak power of the dechirped pulse, we obtain a pulse with a ∼100-fs (full-width half-maximum) peak and a 2-ps pedestal that contains a significant fraction of the pulse energy. Insight into the compressibility of these two pulses can be gained from the spectrograms in Figs. 3(c) and 3(f). While the frequency sweep of the unshifted pulse is quite linear, that of the Stokes pulse shows significant contributions from higher-order spectral phase, especially on the blue side of the spectrum. Despite the existence of the pedestal in the compressed Stokes pulse, the peak power that is predicted from these simulations (∼70 kW) should be sufficient for 2PM.

Laser experimental results
The experimental realization of the system simulated in Section 2 consists of two distinct components. A normal-dispersion fiber laser generates 3.5-nJ pulses at 1030 nm and 7.5-MHz repetition rate that are compressed, partially filtered, and attenuated to produce 0.6-nJ pulses that seed the gain-managed regime. In this system a transform-limited (or nearly transform-limited) seed pulse roughly 500 fs to 1 ps in duration with roughly 0.5 -1 nJ energy will result in nearly identical amplifier behavior, as the gain-managed evolution does not depend sensitively on the properties of the seed [33]. These processed pulses enter the cavity described in Section 2. Nominally, the only difference between the experimental and simulated cavities is the presence of a high-power fiber pump combiner (commercial, 10 µm signal input and output fibers, 976 nm pump port) and fiber pigtails fused to collimating optics. The gain and feedback fibers are 4 m of Nufern PLMA-YDF-10/125-VIII and 22 m of Nufern PM980-XP, respectively (as modelled the simulations). The pulses from the two cavity outputs (as shown in Fig. 1) are then compressed with transmission-grating pairs and the unshifted pulse is further filtered to increase the spectral separation between the pulses. Figure 4 shows the measured and simulated output spectra. The energies of the experimental Stokes and unshifted pulses in Fig. 4 are 17 nJ and 68 nJ, respectively. The energies of the numerical Stokes and unshifted pulses are 26 nJ and 73 nJ, respectively. These similar pulse energies coupled with strong qualitative agreement between the spectra in simulations and experiments suggest that we are replicating the real behavior of the system well in our simulations. While the agreement between simulation and experiments is promising, temporal characterization is critical to ascertaining the suitability of these pulses for intensity-dependent processes like 2PM. To do this, we compress the Stokes and unshifted output pulses with separate transmission grating pairs and measure the intensity autocorrelations (Fig. 5). The width of the autocorrelation of the unshifted pulse is ∼150 fs, which corresponds to a pulse width of 110 fs assuming a gaussian deconvolution factor (the transform-limited pulse duration is ∼70 fs). The full-width half-maximum of the Stokes pulse autocorrelation is 380 fs, but there is a significant pedestal of approximately 2.2 ps duration. Both observations are consistent with the simulations in Section 2. While the autocorrelation is not a direct measurement of peak power, these measurements do demonstrate the generation of short pulses at widely separated wavelengths that should be useful for 2PM. Lastly, noise is generally a concern with pulses generated by nonlinear frequency conversion. 2PM can enhance intensity fluctuations because it is an intensity-squared dependent process. For the nondegenerate excitation, both the amplitude fluctuations and the relative timing jitter of the pulses contribute to noise in the signal. In the extreme case, if the relative timing jitter was much larger than the pulse duration it would be impossible to use the nondegenerate excitation channel. In order to test whether we should be able to use all excitation channels, we measured the sum-frequency generation (SFG) and second-harmonic generation (SHG) signals as proxies for the two-photon excited fluorescence (2PEF) signals in imaging experiments. We directly measured the pulse trains generated by 1) SHG of the gain-managed pulse, 2) SHG of the Stokes pulse, and 3) SFG of the gain-managed and Stokes pulses. The results are summarized in Table 1. Since the response time of the photodetector is much longer than the pulse duration, the signal voltage is proportional to the pulse energy incident on the detector. We find that the unshifted SHG fluctuations are ∼2.3% of the pulse energy, whereas the Stokes SHG fluctuation is much larger at ∼11%. The SFG fluctuations are ∼18% which is not surprising given the Stokes SHG fluctuations. Using the same method, we measured the pulse energy fluctuations of the unshifted and Stokes pulses to be 2.2% and 2.7%, respectively. The low Stokes energy fluctuation relative to the SHG fluctuation suggests significant phase noise. We currently believe that this is due to unsaturated Raman gain at the trailing edge of the Stokes pulse. While the SFG fluctuation is more significant than the Stokes SHG fluctuation, it demonstrates that the relative timing jitter is not much larger than the compressed pulse durations. While these fluctuations are certainly larger than desirable, we show below that the source is quite adequate for 2PM. To investigate why the fluctuations do not degrade imaging performance excessively, we measured the fluctuation in the 2PEF signals from a pool of dye excited with a commercial titanium-sapphire laser. We observed fluctuations of ∼30% even with the quiet source, so we attribute them to the detection process. With such large fluctuations in the detection process, it is not surprising that even 18% fluctuations are not the limiting factor in the overall signal-to-noise ratio.

Degenerate and nondegenerate 2PM imaging
We used the Raman laser for 2PM experiments with the gain-managed and Stokes pulses individually as well as the nondegenerate signal from the spatiotemporal overlap of the two. Although our goal was to image in vivo mice labelled with various fluorophores, we first tested the source by performing simultaneous SHG and SFG microscopy on collagen from the tail of a mouse. To obtain these images, we used a 25X 1.05 NA water immersion objective (Olympus XLPLN25XWMP2) with a working distance of 2 mm. The signal was spectrally resolved with long-pass dichroic mirrors with cut-on wavelengths at 532 nm and 562 nm and detected with three GaAsP photomultiplier tubes (PMTs) (H10770PB-40SEL, Hamamatsu). For each data stack frames of 512 by 512 pixels were acquired at intervals of 1-µm between slices. The pixel dwell time was 0.8 µsec, with averaging of four digital samples of the PMT signal for each pixel. The system is conventional for a two photon microscope -an expanded system diagram may be found in the Supplement 1 ( Figure S3). The results of this imaging experiment are shown in Fig. 6. As SHG in collagen is not very wavelength sensitive, the images in each tile should be close to identical if the excitation condition produces a signal within that channel's band. When we illuminate the sample with only the gain-managed pulse, we see a signal only on channel 1 (corresponding to detection wavelengths λ s shorter than 532 nm). When we illuminate the sample with only the Stokes pulse, we see a signal only on channel 3 (corresponding to λ s longer than 562 nm). However, when we illuminate the sample with both the gain-managed and Stokes pulses we see an additional signal on channel 2 (corresponding to λ s between 532 nm and 562 nm). Additionally, this signal is only visible when the relative temporal delay of the pulses is tuned. This signal is necessarily due to SFG between the gain-managed and Stokes pulses.
Encouraged by these results, we next used this system for in vivo 2PEF imaging in the cortex of a live, anesthetized mouse labeled with three different color fluorescent labels. All animal procedures were approved by the Cornell Institutional Animal Care and Use Committee (protocol #: 2015-0029). Briefly, Thy1-YFPH (B6.Cg-Tg(Thy1-YFP)HJrs/J, Jackson Labs) mice were anesthetized using isoflurane and a 5-mm diameter thinned skull cortical window surgery was performed using a dental drill over the parietal cortex. These mice express yellow fluorescent protein (YFP) in excitatory neurons in the brain. We prepared CellTracker Red Fig. 6. A tiling of SHG images across different excitation conditions (columns, λ p is the incident pulse wavelength) and detection spectral bands (rows, λ s is the signal wavelength). Notably, the signal in the intermediate band is present only when the sample is illuminated by both beams -this signal is necessarily from sum harmonic generation. Scale bar: 50 µm.
CMTPX dye (ThermoFisher), a cell membrane-permeable dye that becomes cell impermeant after an intracellular reaction, according to manufacturer instructions and microinjected it into the cortex using a Nanoject (Drummond) through a pulled glass pipette lowered to depths of 50 µm and 100 µm beneath the cortical surface. The cranial window was then sealed by gluing a glass coverslip to the skull. The mouse was intraperitoneally injected with dexamethasone (0.025 mg per 100 g mouse weight) and ketoprofen (0.5 mg per 100 g mouse weight) to reduce post-surgical inflammation and pain, respectively. The mouse was then removed from anesthesia for ∼4 hours, as we waited for the CellTracker dye to accumulate in cells. We then re-anesthetized the mouse and retro-orbitally injected 50 µL of 1% (weight/volume) Alexa 647 dye conjugated to 10 kDa dextran in saline to fluorescently label the blood plasma.
We imaged this mouse with the same three excitation conditions used with the mouse collagen: gain-managed pulses only, Stokes pulses only, or both pulses present. The fluorescence was separated into three channels with two long-pass dichroic mirrors with cut-on wavelengths at 562 nm and 660 nm. Before taking image stacks, we observe that the signal increases to a maximum as a function of the delay between the pulses. We attribute this effect to non-degenerate two-photon absorption and set the delay to this value for the rest of the experiment. Next, we took 1-µm spaced image stacks of 1024 by 1024 pixels through a 30 µm thick volume at the cortical surface (Supplement 1). We used the same 25X 1.05 NA water-immersion objective (Olympus, XLPLN25XWMP2) and PMTs (H10770PB-40SEL, Hamamatsu) as for the SFG experiment. The YFP signal was largely visible in channel 1 and was only excited well by light from the gain-managed laser (1050 nm). The Alexa 647 and CellTracker signals were both visible in both channels 2 and 3. The Alexa 647 was well excited by the Stokes pulse (1200 nm) but not with the gain-managed pulse. There was some signal increase evident when both the gain-managed and Stokes pulses were used with Alexa 647, indicating non-degenerate 2PEF. The CellTracker signal was visible with excitation using the gain-managed pulses, but was much brighter when both pulses were present, again suggesting non-degenerate two photon excitation. We manually identified bright regions of interest (ROIs) that contained readily-identifiable tissue or cell features that were labeled with a single fluorescent marker (e.g. a dendrite for YFP, a blood vessel for Alexa 647, and a bright cell-shaped object for CellTracker). We used the emission values in each channel from these ROIs across each excitation condition to extract basis spectra (Supplement 1) that we used to linearly unmix the image data to obtain images of each fluorophore ( Fig. 7(a)). In these images, the YFP-labeled dendritic arbor of deep-lying cortical neurons, some larger surface blood vessels, as well as cortical capillaries labeled with Alexa 647, Fig. 7. (a) In vivo degenerate and non-degenerate 2PEF imaging in the brain of a live, anesthetized transgenic mouse with YFP-labeled dendrites (blue), Alexa 647 dextran-labeled blood vessels (red), and some cortical cells labeled by intracortical injection of CellTracker Red CMTPX (green). The image is a maximum projection of a linearly-unmixed image stack (30 µm thick) based on nine separate image stacks acquired using three emission channels and under two degenerate (e.g. gain-managed and Raman laser excitation, individually) and one non-degenerate (e.g. both pulses present) excitation condition. (b) Unmixed CellTracker images using (i) all excitation conditions or (ii) only the two degenerate excitation conditions. The arrows in panel (i) indicate labeled microvessels. Note that the CellTracker images in panel (b) are shown on a different contrast scale, with brighter features saturated, than these same images in Supplement 1. Scale bars: 50 µm. and some CellTracker labeled cells located primarily in vessel-adjacent positions, were all clearly visible and distinguishable. To assess the value of incorporating the non-degenerate excitation, we also performed this unmixing excluding the non-degenerate excitation condition (Supplement 1). Comparing the CellTracker image with degenerate and non-degenerate excitation conditions ( Fig. 7(b)(i)) to that with only the degenerate excitation conditions (Fig. 7(b)(ii)), we found subtle labeling of small blood vessels, likely due to CellTracker dye being taken up by endothelial cells, that could only be seen when all excitation conditions were used in the unmixing.

Discussion and conclusion
For intensity-dependent processes (such as 2PEF) it is critical to generate pulses with high peak power to maximize the signal-to-background ratio, especially for imaging deep into scattering media. In our system, pulse compressibility is limited by nonlinear phase accumulation. We may then either directly compensate the spectral phase generated by the Kerr nonlinearity or modify the system to reduce this effect to increase the peak power of the pulses. Techniques for compensation of high-order spectral phase are known, but they increase complexity of the system.
In fiber-based systems, mitigating nonlinearity is typically done by increasing the mode area of the fiber. Recently-published work shows that the pulse energy of a gain-managed nonlinear amplifier constructed with 30 µm gain fiber can reach 1.2 µJ [35]. If we assume the Stokes pulse can be generated with same efficiency as the in the system described above, we would estimate a roughly tenfold increase in the Stokes pulse energy, to 170 nJ. Thus, scaling by fiber mode area is a promising route to higher peak powers.
Cavity optimization may also reduce the effect of Kerr nonlinearity on the spectral phase of the pulse. Since gain-managed nonlinear amplification is highly nonlinear, the seed pulse experiences significant amounts of cross-phase modulation as it copropagates with the gain-managed pulse. The simplest way to solve this is probably to take the ytterbium amplifier out of the cavity. In this design, a gain-managed nonlinear amplifier would first be filtered to generate a pulse around 1150 nm, which would then be injected into a Raman cavity composed of only passive fiber. This would minimize the nonlinear phase accumulated by the Stokes pulse and would also reduce the issue with unsaturated Raman gain, which we believe is the most likely reason our system exhibits significant phase noise.
In summary, we have demonstrated a fiber system capable of generating synchronized pulses at 1050 nm and 1200 nm and demonstrated their utility for simultaneous degenerate and nondegenerate 2PM. This system can form the basis of a compact and inexpensive source for hyperspectral two-photon microscopy. We believe there is room for improvement in pulse peak power by using larger mode-area fiber and further optimizations of the cavity, which would enable deeper imaging.