We demonstrate near shot-noise limited hyperspectral stimulated Raman scattering (SRS) spectroscopy using oscillator-only excitation conditions. Using a fast CMOS camera synchronized to an acousto-optic modulator and subtracting subsequent frames acquired at up to 1 MHz frame rates, we demonstrate demodulation and recovery of the SRS spectrum. Surprisingly, we observe that the signal-to-noise of SRS spectra is invariant at modulation frequencies down to 2.5 kHz. Our approach allows for a direct comparison of SRS with coherent anti-Stokes Raman scattering (CARS) spectroscopy under identical experimental conditions. Our findings suggest that hyperspectral SRS imaging with shot-noise limited performance at biologically compatible excitation energies is feasible after minor modifications to fast frame-rate CMOS array technology.
©2013 Optical Society of America
Coherent Raman microscopy has emerged as a fast and valuable technique for label-free imaging with coherent anti-Stokes Raman scattering (CARS) microscopy leading the way . Recently, stimulated Raman scattering (SRS) microscopy, an alternative to CARS, has gained popularity [2, 3]. Both techniques produce signal when a sample is excited by two coherent sources with an energy difference equal to a Raman transition. In CARS, a non-resonant background (NRB) interferes coherently with the resonant signal, which distorts the Raman lineshape and prevents direct quantitative analysis of CARS spectra. Several experimental and computational approaches have been developed to enable resonant signal extraction and quantitative analysis of CARS spectra (reviewed in ); however, these methods add complexity to the data collection or analysis, respectively.
In contrast, SRS only occurs when the energy difference between the excitation sources is resonant with a Raman transition; therefore SRS does not produce an NRB via molecular transitions that distorts the Raman lineshape, though a relatively weak baseline may still appear. Because of the method used to detect the SRS signal, it is linearly proportional to the intensity of both excitation sources and the number of excited molecules in the focal volume.
Unlike CARS, the SRS signal exists at the same wavelengths as the excitation pulses, making it impossible to spectrally separate the relatively weak SRS signal from the excitation beams. The solution to this problem is to modulate one beam and use lock-in detection. Modulation rates are typically 1-20 MHz for biological imaging [2, 3, 5, 6]. Current array technology cannot demodulate at these rates, so SRS of biological samples is typically limited to single photodiodes. This practical limitation makes hyperspectral biological SRS imaging difficult despite the rapid progression of single-frequency (channel) SRS imaging. However, spectral detection of CARS is well established, making it the current method of choice for coherent hyperspectral imaging [7, 8].
Parallel, multichannel detection of several discrete frequencies (multiplex SRS) has recently been demonstrated [9, 10]. Sequential, frequency-scanning solutions have also been successfully implemented for hyperspectral SRS recently [11–14]; images including complete spectra are desirable given the increased information content of the hyperspectral images such an approach provides and the synchronized acquisition of all Raman modes of interest. Hyperspectral SRS has been reported for low repetition-rate, high-energy laser systems [15–17], but the peak laser powers in those experiments are incompatible with living biological samples. An array detector capable of high-speed signal demodulation could enable the facile implementation of hyperspectral SRS at low peak powers similar to that enjoyed for hyperspectral CARS.
In this work, we demonstrate hyperspectral SRS detection using a CMOS imaging array with a maximum frame rate of 1 MHz (Fastcam SA5, Photron) and unamplified lasers to detect near shot-noise limited CARS and SRS spectra. This CMOS array uses an on-board memory buffer to store acquired frames in real-time before batch transferring the frames to the control computer via gigabit ethernet. We perform simplified lock-in spectral detection of a dispersed (broadband) Stokes beam at up to 500 kHz by subtracting subsequent frames. Subtracting subsequent frames essentially forms a high-pass filter whereas averaging over many differences creates a low-pass filter. Together, these functions approximate a band-pass filter, which is the core of lock-in detection. Because of the versatility of the CMOS array and its simple integration into our system for SRS, we could quickly convert between SRS and CARS spectroscopy to acquire both spectra under identical experimental conditions.
Figure 1 is a simplified schematic of our apparatus and shows the necessary components for hyperspectral SRS in green, which are the modulator, pump filter, and SA5 CMOS array; only three minor alterations are necessary to convert the SRS setup to a CARS setup. By switching the SRS pump filter to a CARS short-pass filter, moving the grating in the monochromator, and permanently switching the modulator into the “on” state, the system is ready for CARS.
Implementation and programming of the Fastcam SA5 array is the salient feature of this work that allows low-energy SRS detection. The SA5 CMOS array is a 12-bit camera with 20 µm x 20 µm pixels that operates in global shutter mode at frame rates up to 1 MHz. It has the unique feature of allowing independent control of the frame rate and exposure time. It also provides multiple programmable outputs that are synchronized to the camera exposure, with each output having nanosecond accuracy. Timing synchronization of the array exposure to beam modulation is critical for SRS and by making the SA5 the master clock, this ensures that the pump beam modulation is synchronous with frame acquisition; however, there is no timing synchronization between the laser and camera. A programmable output from the camera is configured to produce a trigger signal at half the frame rate that drives the acousto-optic modulator (AOM). The AOM (1205C-2-804B, Isomet) chops the pump beam with 100% modulation depth. Taking the difference between adjacent frames isolates the modulated signal and produces an SRS spectrum.
Two separate Ti:sapphire oscillators (Tsunami, Spectra Physics) produce the narrowband pump and broadband Stokes (probe) light. The 10-picosecond (ps) pump laser is centered at 710 nm while the Stokes laser is tuned for different experiments. A “Lok-to-Clock” system (Spectra Physics) electronically synchronizes the pulse trains from the two laser sources (at ~80 MHz). Further oscillator synchronization is provided by a homebuilt cross-correlator to reduce the timing jitter to less than 1 ps. The beams were combined, fed into a Nikon Ti Eclipse microscope, and focused with a 40X, 1.3 NA (Zeiss) or 60X, 1.49 NA objective (Nikon). The power of the pump and Stokes beams at the sample plane were 75 mW and 5 mW, respectively, unless otherwise stated. The Stokes power was limited by the saturation of the camera, and the pump power was limited to 75 mW to maintain excitation conditions under the damage threshold for SRS imaging of most biological samples [2, 6, 7, 13]. A 40X, 0.6 NA objective (Zeiss) or 40X, 1.3 NA (Zeiss) is used to collect the signal. Short-pass 694 nm (Semrock), for CARS, or band-pass 809 nm (Semrock), for SRS, filters reject unwanted radiation, and a Shamrock 303i spectrometer (Andor) disperses the signal on the SA5 [Fig. 1]. The spectra from the CMOS array contain ~85 cm−1 of bandwidth over 64 detector pixels; 64 pixels is the maximum for 1 MHz frame rates. Spontaneous Raman spectra were collected with a commercial confocal Raman microscope using the 488 nm laser illumination (Bruker).
3. SRS signal characterization
In this initial demonstration, neat benzonitrile (BN) was used in all experiments. Figures 2(a) and 2(b) show SRS spectra of the 1600 cm−1 mode of BN at 500 kHz modulation. Control experiments verified the identity of the signal as SRS in terms of laser power dependence and concentration dependence. As shown in Fig. 2(a), a small and broad baseline, potentially due to cross-phase modulation  and/or two-color two-photon absorption , accompanies the SRS peak at 1600 cm−1. Cross-phase modulation is likely the dominating effect, because the offset is much smaller when using matched NA objectives. Therefore, each raw spectrum is fit to a second-order polynomial while masking the sample resonance [Fig. 2(a)]. The fit is subtracted from the spectrum, and the resulting baseline-corrected spectrum is fit to a Lorentzian [Fig. 2(b)]. The maximum of the Lorentzian fit defined the ‘signal’ for signal-to-noise ratio (SNR) measurements, whereas the standard deviation of the baseline off-resonance was the noise.
To further explore the SRS signal, we characterized the signal (S), noise (N), and the SNR as a function of modulation frequency and exposure time. Figure 3 (left) shows that the SRS signal, noise, and SNR are constant with respect to the modulation frequency between 2.5 and 500 kHz (5 kHz – 1 MHz frame rates). The noise increases at the lowest frequency (0.5 kHz). We could not measure performance at lower frequencies because the camera cannot create a synchronous trigger signal at slower frame rates. Figure 3 (right) shows the expected linear signal increase with exposure time, due to integration of more photons, while the noise and SNR both show a square-root dependence [Fig. 3, right] at 50 kHz modulation. All spectra compared in Fig. 3 are the average of 105 frames and were taken sequentially to minimize the effect of long-term system drift. This combination of metrics suggests that the SRS spectra obtained in our apparatus are shot-noise limited, perhaps even at modulation frequencies as low as 2.5 kHz as the N values shown in Fig. 3 (left) do not significantly change until 1 kHz modulation.
A theoretical estimate of the shot-noise contribution to the SNR in these experiments indicates that, indeed, shot-noise is likely the dominant noise source. Shot-noise is the variation in photon counting intrinsic to photon statistics, which obey a Poisson distribution. Therefore, if shot-noise limited detection is achieved, the noise would be the square-root of the number of photons that hit each pixel in the CMOS array per frame. Using the pixel well depth (45000 e-), the bit depth of the camera (12-bit), and the quantum efficiency (~0.33 e-/photon at 800 nm), we can convert digital counts to the number of photons that reached the array (~33 photons/digital count).
The signal intensity is a measure of the number of SRS signal photons that illuminate the detector per frame pair (signal is only present on every other frame due modulation). Figure 4(a) shows the SRS signal, which is the sum of 27100 differences (frame pairs) divided by the number of frame pairs. The peak height is 4.34 digital counts or about 143 photons. The noise due to shot noise is the square-root of the number of photons that hit the detector each frame. The total number of photons per frame [Fig. 4(b)] yields an average of 6885 counts per frame or about 227300 photons. The resulting SNR must be divided by √2 to account for the noise propagation from differencing two sequential frames. Therefore, the estimated shot-noise SNR from each difference is 0.21. Averaging over many differences improves the SNR by √N where N is the number of differences (27100). Using these numbers, we obtain a theoretical shot-noise limited SNR of ~35, which is within a factor of 1.75 of the SNR extracted from our measurements (~21 for Fig. 4(a)). This correspondence strongly suggests near shot-noise limited performance with our apparatus under low energy, biologically compatible excitations.
4. Hyperspectral CARS and SRS
The system presented here can be quickly converted between CARS and SRS. Given this ability, we sought to directly compare CARS (non-resonant plus resonant) and SRS spectra under identical experimental conditions. For these comparison experiments, a matched pair of 40X 1.3 NA oil immersion objectives was used. The phase-matching requirement in CARS narrows the emission NA, so a lower NA collection objective still gathers almost all CARS light. However, there are no phase-matching requirements in SRS, so the size of the emission cone is equivalent to the size of the excitation cone. Matching the emission and collection NA ensures collection of all the SRS signal (in the absence of cross-phase modulation).
Figure 4 shows SRS (a) and CARS (c) spectra acquired under equivalent exposure conditions. The SRS data was collected with a 150 kHz frame rate (75 kHz modulation frequency). The CARS spectrum was acquired with a single 20 ms exposure. The SRS spectrum was acquired with 54200 frames (27100 differences) each exposed for 369 ns (for a total exposure time of 20 ms in the SRS measurement). This allows a direct comparison of the lineshapes and SNR of these two coherent Raman spectroscopies. Figure 5 shows an overlay of CARS, SRS, and Raman spectra of BN centered at 1185 cm−1 where the BN spectrum shows multiple peaks. This shows the ability of hyperspectral SRS to simultaneously resolve multiple resonances and emphasizes the distortion of the CARS spectrum by the NRB. It should be noted that BN has an extremely large Raman cross-section, and the lineshape distortion seen in CARS would be even larger if the ratio of resonant to non-resonant signal is smaller e.g. in a biological sample.
It is clear from Fig. 5 that the SRS spectrum matches the lineshape of the spontaneous Raman response, as expected, whereas the CARS peaks are distorted. We also observe that the CARS spectrum has a higher SNR than the SRS spectrum. The difference in SNR between the CARS and SRS spectra is interesting in light of previous theoretical work , which suggests that the SNR of CARS and SRS to be roughly equal. We do not know the source of this discrepancy at this time, and it will be further explored in a future study.
While our setup is capable of detecting small relative intensity changes, ΔIprobe/Iprobe ~10−4, the SNR of our measurements is lower than previous single channel SRS implementations [2, 5]. This can be rationalized by our use of a 5 mW, broadband probe pulse and array detection, which distributes the power of the probe beam over roughly 300 cm−1 and 225 detector elements. This is in contrast to previous SRS measurements with unamplified systems that used ~10 – 40 mW of power for each beam and a single photodiode detector. Increasing the probe power will increase the SNR similarly to that shown for increasing exposure time [Fig. 3, right]. Unfortunately, the CMOS array used here has a limited dynamic range and prevented us from increasing the probe power beyond 5 mW, so were unable to decrease the minimum detectable modulation.
A more important limitation regarding the tolerable level of incident laser power is the threshold of physical or functional damage to the sample, which is especially important for biological tissue and cell samples. This empirical limit depends on the incident pulse duration and intensity, and is unique for each biological sample. In this work the excitation intensity was compatible with most biological systems imaged on apparatuses with similar excitation sources. According to previous reports , longer pulse durations would allow for a greater excitation intensity without sample damage; however, in this demonstration, the Stokes intensity is limited by the saturation of the camera pixels, which is independent of pulse duration. For a given total excitation power (pump plus probe power) with each laser having equal pulse spectral width, a 50:50 split in pump and probe power results in the largest SRS signal. In our case of unequal spectral widths, the optimal configuration is obtained by maximizing the instantaneous power of the Stokes beam via attenuation (to operate near full well capacity of the camera) and dispersion compensation (to optimize nonlinear interaction with the pump) without burning the sample. The power of the pump beam is then increased as necessary to promote more efficient SRS generation, assuming sample damage is dominated by Stokes irradiation and that pump irradiation is reasonably benign.
Beyond SNR and sample damage, the most relevant comparison metric between CARS and SRS is actual measurement time. In this paper, we normalize to the actual exposure time. However, a substantial bottleneck in the system presented here is data transfer from the CMOS array to the computer. After collecting all the frames in an SRS spectrum (105 frames for most spectra in this manuscript), the data must be transferred to a computer for differencing and averaging. This transfer process takes about 5 minutes for 105 frames, which is prohibitively long for imaging experiments – a 100x100 pixel image would take about 14 days to transfer, even though the actual imaging (total exposure) time is just over 6 minutes. Recently, sequential frequency scanning solutions demonstrated high-speed image acquisition; although this SRS imaging implementation is optically more complex than the apparatus presented here, it is currently a more practical technology. Onboard mathematics would allow the transfer of a single frame and make SRS imaging with a high-speed array more practical for hyperspectral SRS imaging.
In summary, we demonstrate simple, near shot-noise limited hyperspectral SRS detection under low energy excitation conditions, which would be compatible with biological imaging, at modulation rates between 2.5 – 500 kHz using a fast CMOS camera. We use this setup to directly acquire hyperspectral CARS and SRS under identical experimental conditions. Though shot-noise limited performance with the SA5 detector is achievable at low modulation rates, faster modulation rates are still preferred because they allow higher probe beam intensity and faster signal averaging, resulting in faster acquisition. Further improvements in CMOS array technology, including faster modulation, onboard mathematics, and a larger dynamic range will enable hyperspectral SRS imaging with similar accessibility as hyperspectral CARS imaging.
We thank Andreas Mangold from VKT for generous use of the SA5 camera and Hansjörg Menges for technical assistance with the spontaneous Raman spectroscopy. This work was supported by Marie Curie Foundation grant #CIG322284 to SHP.
References and links
2. C. W. Freudiger, W. Min, B. G. Saar, S. Lu, G. R. Holtom, C. He, J. C. Tsai, J. X. Kang, and X. S. Xie, “Label-free biomedical imaging with high sensitivity by stimulated Raman scattering microscopy,” Science 322(5909), 1857–1861 (2008). [CrossRef] [PubMed]
3. P. Nandakumar, A. Kovalev, and A. Volkmer, “Vibrational imaging based on stimulated Raman scattering microscopy,” New J. Phys. 11(3), 033026 (2009). [CrossRef]
4. J. P. Day, K. F. Domke, G. Rago, H. Kano, H. O. Hamaguchi, E. M. Vartiainen, and M. Bonn, “Quantitative coherent anti-Stokes Raman scattering (CARS) microscopy,” J. Phys. Chem. B 115(24), 7713–7725 (2011). [CrossRef] [PubMed]
5. B. G. Saar, C. W. Freudiger, J. Reichman, C. M. Stanley, G. R. Holtom, and X. S. Xie, “Video-rate molecular imaging in vivo with stimulated Raman scattering,” Science 330(6009), 1368–1370 (2010). [CrossRef] [PubMed]
7. K. Bito, M. Okuno, H. Kano, S. Tokuhara, S. Naito, Y. Masukawa, P. Leproux, V. Couderc, and H. O. Hamaguchi, “Protein secondary structure imaging with ultrabroadband multiplex coherent anti-Stokes Raman scattering (CARS) microspectroscopy,” J. Phys. Chem. B 116(4), 1452–1457 (2012). [CrossRef] [PubMed]
8. Y. J. Lee, D. Moon, K. B. Migler, and M. T. Cicerone, “Quantitative Image Analysis of Broadband CARS Hyperspectral Images of Polymer Blends,” Anal. Chem. 83(7), 2733–2739 (2011). [CrossRef] [PubMed]
10. D. Fu, F.-K. Lu, X. Zhang, C. Freudiger, D. R. Pernik, G. Holtom, and X. S. Xie, “Quantitative Chemical Imaging with Multiplex Stimulated Raman Scattering Microscopy,” J. Am. Chem. Soc. 134(8), 3623–3626 (2012). [CrossRef] [PubMed]
11. Y. Ozeki, W. Umemura, K. Sumimura, N. Nishizawa, K. Fukui, and K. Itoh, “Stimulated Raman hyperspectral imaging based on spectral filtering of broadband fiber laser pulses,” Opt. Lett. 37(3), 431–433 (2012). [CrossRef] [PubMed]
12. D. Zhang, P. Wang, M. N. Slipchenko, D. Ben-Amotz, A. M. Weiner, and J.-X. Cheng, “Quantitative Vibrational Imaging by Hyperspectral Stimulated Raman Scattering Microscopy and Multivariate Curve Resolution Analysis,” Anal. Chem. 85(1), 98–106 (2013). [CrossRef] [PubMed]
13. Y. Ozeki, W. Umemura, Y. Otsuka, S. Satoh, H. Hashimoto, K. Sumimura, N. Nishizawa, K. Fukui, and K. Itoh, “High-speed molecular spectral imaging of tissue with stimulated Raman scattering,” Nat. Photonics 6(12), 845–850 (2012). [CrossRef]
14. D. Fu, G. Holtom, C. Freudiger, X. Zhang, and X. S. Xie, “Hyperspectral imaging with stimulated Raman scattering by chirped femtosecond lasers,” J. Phys. Chem. B 117(16), 4634–4640 (2013). [PubMed]
15. E. Ploetz, S. Laimgruber, S. Berner, W. Zinth, and P. Gilch, “Femtosecond stimulated Raman microscopy,” Appl. Phys. B 87(3), 389–393 (2007). [CrossRef]
17. D. W. McCamant, P. Kukura, S. Yoon, and R. A. Mathies, “Femtosecond broadband stimulated Raman spectroscopy: Apparatus and methods,” Rev. Sci. Instrum. 75(11), 4971–4980 (2004). [CrossRef] [PubMed]
18. K. Ekvall, P. van der Meulen, C. Dhollande, L. E. Berg, S. Pommeret, R. Naskrecki, and J. C. Mialocq, “Cross phase modulation artifact in liquid phase transient absorption spectroscopy,” J. Appl. Phys. 87(5), 2340–2352 (2000). [CrossRef]
19. D. Fu, T. Ye, T. E. Matthews, B. J. Chen, G. Yurtserver, and W. S. Warren, “High-resolution in vivo imaging of blood vessels without labeling,” Opt. Lett. 32(18), 2641–2643 (2007). [CrossRef] [PubMed]
20. Y. Ozeki, F. Dake, S. Kajiyama, K. Fukui, and K. Itoh, “Analysis and experimental assessment of the sensitivity of stimulated Raman scattering microscopy,” Opt. Express 17(5), 3651–3658 (2009). [CrossRef] [PubMed]
21. K. König, T. W. Becker, P. Fischer, I. Riemann, and K. J. Halbhuber, “Pulse-length dependence of cellular response to intense near-infrared laser pulses in multiphoton microscopes,” Opt. Lett. 24(2), 113–115 (1999). [CrossRef] [PubMed]