Design of the pulse-picking nonlinear optical microscope
Up to date, the most widely used light sources for nonlinear optical microscopy are high-repetition-rate femtosecond or picosecond lasers. Assuming an 80 MHz laser repetition rate and a 10 µs pixel dwell time, there are 800 pulses on each image pixel. Since nonlinear optical signals are usually proportional to the square or cubic of laser peak power, reducing the number of pulses at each pixel with higher pulse energy can largely improve the sensitivity while maintaining the same average input power. We develop a pulse-picking technology to reduce the duty cycle of high repetition rate lasers for high sensitivity nonlinear optical microscopy. The design of our microscope is illustrated in Fig. 1a. We use a femtosecond laser with synchronized dual outputs: one as the Stokes beam with a fixed wavelength at 1045 nm and the other as the pump beam with a tunable wavelength from 690–1300 nm. The two beams are combined and chirped by multiple SF-57 glass rods for spectral focusing. Before the microscope, the Stokes pulse is chirped to 1.8 ps while the pump pulse is chirped to 3.4 ps. The ratio of pulse durations matches the ratio of spectral widths to ensure the best spectral resolution using spectral focusing [29]. The combined beams are sent to an acousto-optic modulator (AOM) that is controlled by a function generator. Square waves with tunable duty cycles from 1.4–97% at various modulation frequencies are sent to the AOM for pulse picking. We direct the 1st order AOM output to a lab-designed upright laser-scanning microscope with two photomultiplier tubes (PMTs) in the epi-direction and one in the forward direction. As shown in Fig. 1a, we use PMT1 for forward CARS (FCARS) detection, PMT2 for the acquisition of TPEF signals at 450 nm, and PMT3 for collection of either TPEF signals at 570 nm, or SHG signals, or epi CARS (ECARS) signals.
We choose the 1st diffraction order from the AOM over the 0th order for imaging because it will completely shut off the laser beam at the ‘time-off’ periods. This maximizes the nonlinear optical signal generation at a fixed input average power. However, the AOM Bragg angles for the pump and Stokes wavelengths are different. To ensure beam overlap along the 1st order of diffraction, one of the beams needs to be slightly misaligned from the perfect Bragg angle. The AOM beam separation angle between the 0th and 1st orders, regardless of the incidence angle, is
$${\theta _s}=\frac{{\lambda f}}{V}$$
1
where λ is the beam wavelength, f is the acoustic frequency, and V is the acoustic velocity. The Bragg angle is half of the separation angle
$${\theta _B}=\frac{{\lambda f}}{{2V}}$$
2
The laser beam geometry at the AOM for spatially overlapping pump and Stokes beams along the 1st order diffraction is illustrated in Fig. 1b. We first let the pump beam at 800 nm enter the AOM at the Bragg angle \({\theta _B}\). Both the 0th and the 1st diffraction orders of this wavelength have an angle of \({\theta _B}\) to the crystal surface normal. Assuming the Stokes beam at 1045 nm has a Bragg angle \(\theta _{B}^{'}={\theta _B}+\theta ^{\prime}\), the incidence angle of this beam needs to be slightly detuned from \(\theta _{B}^{'}\) to generate the 1st order diffraction at the same direction of the 1st order pump beam. If \(\delta\)is the angle between the incidence angle and the Bragg angle of the Stokes beam, it satisfies
$${\theta _{s{\text{,Stokes}}}}=2\left( {{\theta _B}+\theta ^{\prime}} \right)=2{\theta _B}+\theta ^{\prime}+\delta$$
3
This gives \(\delta =\theta ^{\prime}\),
and indicates that when the pump and Stokes beams are collinear at the 1st order of pump, the angle between the incidence and the Bragg angle of the Stokes beam equals the angle difference between the pump and Stokes Bragg angles. In our optical configuration, 𝛿=0.46°. Using two mirrors in the Stokes-only beam path, we can fine-tune the incidence angle of the Stokes beam at the AOM to satisfy this condition. Using this method, we can reach a 60% efficiency for the pump and 42% efficiency for the Stokes beam using a 90% duty cycle. The loss of efficiency is due to the suboptimal crystal anti-reflective coating and Bragg angle errors.
Pulse-picking for sensitivity improvement
Laser pulses from both the pump and Stokes beams picked by the AOM at different duty cycle values are displayed in Fig. 1c. We estimate the rise time of the AOM in our experiment is ~ 23 ns for the Stokes and ~ 17 ns for the pump (Supplementary Information 1), slightly longer than the time interval between adjacent pulses from the laser, which can also be inferred from Fig. 1c. At high duty cycles, the number of pulses picked by the AOM is proportional to the duty cycle. At 5% duty cycle, we were able to pick 4 major pulses for both pump and Stokes beams at 1.1 MHz modulation frequency. When the duty cycle is reduced to 2% or below, as few as one major pulse can be picked.
Reducing the laser duty cycle would enhance the sensitivity of CARS microscopy at the same input average power. The intensity of the CARS signal can be expressed as
$${I_{CARS}} \propto {\left| {{\chi ^{\left( 3 \right)}}} \right|^2}I_{p}^{2}{I_s}$$
4
where \({I_{CARS}}\),\({I_p}\), and \({I_s}\) are, respectively, the intensities of the CARS, pump, and Stokes beams.\({\chi ^{\left( 3 \right)}}\) is the third-order nonlinear optical susceptibility. The intensity of a laser pulse can be expressed as
$$I=\frac{\mathcal{E}}{{\tau A}}=\frac{P}{{f\tau A}}$$
5
Here, \(\mathcal{E}\), \(\tau\), and A are pulse energy, pulse width, and laser focus area of the laser beam, while P and are the laser average power and repetition rate, respectively. By modulating combined laser beams at a lower frequency and applying a duty cycle of D, we have
$${P_{CARS}}=fD{\tau _{CARS}}A\cdot{I_{CARS}}=fD{\tau _{CARS}}A\cdot{\left| {{\chi ^{\left( 3 \right)}}} \right|^2}{\left( {\frac{{{P_p}}}{{fD{\tau _p}A}}} \right)^2}\frac{{{P_s}}}{{fD{\tau _s}A}}={\left| {{\chi ^{\left( 3 \right)}}} \right|^2}\frac{{P_{p}^{2}{P_s}}}{{{f^2}{D^2}{A^2}}}\frac{{{\tau _{CARS}}}}{{\tau _{p}^{2}{\tau _s}}} \propto \frac{1}{{{D^2}}}$$
6
This indicates the pulse picking CARS (PPCARS) average signal is reciprocal to the square of the duty cycle. Similarly, we can derive that for the TPEF and SHG processes, the average signal is proportional to the reciprocal of D.
$${P_{{\text{TPEF}}}},{P_{{\text{SHG}}}} \propto \frac{1}{D}$$
7
Chemical imaging by the pulse-picking nonlinear optical microscope
Figure 1d shows the relationship between the sensitivity enhancement and duty cycle for FCARS signals at different modulation frequencies. We used the dimethyl sulfoxide (DMSO) CH3 symmetric stretching peak at 2915 cm− 1 for the signal-to-noise ratio (SNR) analysis. A boundary of a DMSO drop sandwiched between two glass coverslips was imaged for SNR calculation. We measured the SNR of the DMSO by dividing the average value of the DMSO signal by the standard deviation of the empty area. For each duty cycle measured in Fig. 1d, the SNR was divided by the SNR of the 97% duty cycle to calculate the sensitivity increase. A 1/D2 curve is plotted as a reference. We find that the experimental data matches the theoretical curve very well at high duty cycles but starts to deviate from the theory at very low duty cycles. The maximum sensitivity enhancement we were able to obtain is 1078 at 1.4% duty cycle, 700 kHz. The major cause of sensitivity drop at very low duty cycles might be the unlocked phase between the function generator modulation and the laser pulse train. Frequency drifts between the two are less significant when the duty cycle is high since almost the same number of pulses can always be picked at any phase difference. However, when the duty cycle becomes low, especially below 4%, the phase drifts can greatly impact the number of pulses picked by the AOM. Figure 1e displays correlations between modulation frequency and SNR increase at different duty cycles. These results show that at high duty cycles (> 20%), the SNR increase is very similar at different modulation frequencies, while at low duty cycles (< 20%), lower modulation frequency gives stronger SNR. The SNR decrease at 300 kHz is likely due to the drift between modulation and acquisition of image pixels.
We used fluorescent polystyrene beads and measured the fluorescence signal at 450 nm excited by 800 nm laser pulses (Fig. 1f) to evaluate the sensitivity enhancement of TPEF, which shows a near 1/D relation at high duty cycles and starts to deviate from the theoretical curve at lower duty cycles. We can obtain a 16.3 sensitivity increase at a 2% duty cycle, 700 kHz. SHG signal improvement, which shows a similar dependence as the TPEF, was measured using a mouse tail tendon specimen and 1045/522 nm excitation/detection (Fig. 1g). A sensitivity increase of 14.5 was achieved at a 2% duty cycle, 1.1 MHz. Plots of TPEF/SHG SNR improvement versus duty cycle at other modulation frequencies are plotted in Figure S2, showing the maximum sensitivity enhancement of ~ 20 folds for both TPEF and SHG.
To measure the absolute sensitivity of our microscope, we performed hyperspectral CARS imaging. A spectral phase retrieval method based on Kramers–Kronig relations was used to derive Raman spectra from chemical compounds using FCARS spectra [30–32]. Figure 2a shows CARS spectra of DMSO and methanol in the C-H stretching region. We measured the sensitivity of the PPCARS system using the 2915 cm− 1 peak of DMSO diluted in D2O. The retrieved CARS signal intensity versus DMSO concentration is shown in Fig. 2b, from which a quadratic relationship, which agrees with Eq. 4, can be identified. Retrieved Raman spectra of DMSO below 1% concentration are shown in Fig. 2c. SNR calculations (Supplementary Information 4) indicate that the lowest concentration detectable using 1.1 MHz modulation is 0.1%, corresponding to 14.2 mM DMSO. To detect such a concentration, only 5.2 mW pump and 6.2 mW Stokes were used at the sample. Using 700 kHz modulation, we can detect 0.05% DMSO, equaling to 7.1 mM concentration using only 2.0 mW pump and 3.7 mW Stokes beams at the sample (Figure S3). By fitting the 2915 cm− 1 DMSO peak using a Lorentzian function, as shown in Fig. 2e, we measured the spectral resolution of our system to be 16.1 cm− 1 in the C-H region.
Figure 2e plots retrieved Raman spectra of polystyrene (PS) and polymethyl methacrylate (PMMA) in the 1570–1750 cm− 1 Raman fingerprint region acquired using 1 µm PMMA and PS mixed particles. The PS peaks at 1583 (C = C stretching) and 1602 cm− 1 (ring-skeletal stretching) can be resolved, while the PMMA peak at 1736 cm− 1 is also detected. Using the strong peak at 1602 cm− 1, we measured a 9.1 cm− 1 CARS spectral resolution of our microscope in this region. Figure 2g compares CARS images of mixed PMMA and PS beads at ~ 1602 cm− 1 using 97% and 4% duty cycles. A clear SNR and contrast improvement can be seen at the reduced duty cycle. By spectral phasor analysis of hyperspectral CARS images in the fingerprint region, we can separate PMMA and PS microparticles, as shown in Fig. 2h. For all the fingerprint imaging, 13.7 mW pump and 6.2 mW Stokes beams were used at the sample with a 10 µs pixel dwell time.
Cell imaging by the pulse-picking nonlinear optical microscope
Next, we applied PPCARS for cell imaging. Fig. 3a compares single-colour FCARS (top) and ECARS (bottom) images from Mia PaCa-2 cells at different duty cycles. We used 700 kHz modulation frequency and observed a continuous increase in signal and sensitivity for both FCARS and ECARS as the duty cycle decreased from 97% to 1.4%. To better compare the sensitivity improvement, we plot the intensity profiles along the lines in Fig. 3a (see Fig. 3b), which show a signal improvement of ~250x at the 1.4% duty cycle. We note that the sensitivity enhancement for small lipid droplets in the cells is less than the pure samples shown in Fig. 1. This is due to the higher ratio of nonresonant contribution at laser focus. Power at the sample for imaging is 10.8 mW for the pump and 5.0 mW for the Stokes. Cell imaging results using different duty cycles at 1.1 MHz modulation frequency are shown in Figure S4. We have also performed live-cell imaging of lipid droplets and mitochondria using CARS and TPEF signals from a mitochondria marker. Images comparing 97%, 10%, and 4% duty cycles are shown in Figure S5.
Hyperspectral CARS images of cells were performed using 2.0 mW pump and 3.7 mW Stokes excitation power at 10 µs pixel dwell time. By spectral focusing and spectral phasor unmixing, we can separate major cellular compositions including cytosol, endoplasmic reticulum, nuclei, and lipid droplets in cells using both FCARS (Fig. 3c) and ECARS (Figure S6) images. The composited chemical map of cells and retrieved Raman spectra of four major components using FCARS are shown in Figs. 3d and 3e, respectively. The separation capability of our hyperspectral CARS microscopy is comparable to spectral-focusing-based hyperspectral SRS [29, 33]. We also performed 3D imaging of a Mia PaCa-2 cell as shown in Fig S7 and Supplementary video 1, demonstrating the 3D chemical imaging capability of the PPCARS microscope.
Tissue imaging by the pulse-picking nonlinear optical microscope
To evaluate sensitivity enhancement of our multimodal microscope for tissue imaging, we compared FCARS, ECARS, TPEF at 450 nm, and TPEF at 570 nm images of mouse liver tissue at 97% and 4% duty cycles under 1.1 MHz modulation (Figs. 4a-d). CARS excitation wavelengths are tuned to the CH2 stretching at 2855 cm-1. Signals in the TPEF 450 nm channel are majorly contributed by the autofluorescence from nicotinamide adenine dinucleotide (NADH) while in the TPEF 570 nm channel are contributed by the autofluorescence from flavin adenine dinucleotide (FAD). To better compare the contrast enhancement, we combine two duty cycle images into one and display half of each. We also select a smaller field of view and show a magnified image on the right of each large area image. Intensity profiles along the lines in Figs. 4a-d are plotted in Fig. 4e. These results show strong SNR enhancement for all modalities at 4% duty cycle. We also display ECARS mouse kidney tissue images at 97% and 10% duty cycle to demonstrate sensitivity increase at a moderate duty cycle (Fig. 4f). As well, we show sensitivity enhancement of SHG imaging using mouse tail tendon in Fig. 4g. Images from other tissue samples and at other modulation frequency/duty cycles can be found in the Supplementary Information (Figure S8-12). These results highlight the potential of pulse-picking technology for the chemical analysis of intact biopsy samples for diagnostics.