A Time-Encoded Technique for fibre-based hyperspectral broadband stimulated Raman microscopy

Raman sensing and microscopy are among the most specific optical technologies to identify the chemical compounds of unknown samples, and to enable label-free biomedical imaging. Here we present a method for stimulated Raman scattering spectroscopy and imaging with a time-encoded (TICO) Raman concept. We use continuous wave, rapidly wavelength-swept probe lasers and combine them with a short-duty-cycle actively modulated pump laser. Hence, we achieve high stimulated Raman gain signal levels, while still benefitting from the narrow linewidth and low noise of continuous wave operation. Our all-fibre TICO-Raman setup uses a Fourier domain mode-locked laser source to achieve a unique combination of high speed, broad spectral coverage (750–3,150 cm−1) and high resolution (0.5 cm−1). The Raman information is directly encoded and acquired in time. We demonstrate quantitative chemical analysis of a solvent mixture and hyperspectral Raman microscopy with molecular contrast of plant cells.


TICO-Raman spectrum compared with literature spectrum
The high quality of the TICO-Raman spectra is best seen when compared with a published reference spectrum given from 1 . The data from 1 was extracted manually and adjusted to a linear x-scale with uniform increments. For the comparison (Supplementary Figure 7), the height of the TICO-Raman spectrum of cyclohexane was scaled to the peak at 1267 cm -1 of the reference spectrum, since this peak is least resolution-critical. The spectral positions of all peaks match exactly showing the quality of the time-to-wavenumber mapping of TICO-Raman. Furthermore, the intensities match very well. However, the intensity of the sharp, narrowband peak at 802 cm -1 is higher in our spectrum than in the literature spectrum. This is probably because we can better resolve the narrowband peak. The resolution of our system in this region is 1.5 cm -1 , limited by sampling density of the broadband sweep. Nevertheless, this resolution is sufficient, as can be seen for a comparable sharp line of benzene at 992 cm -1 measured with lower and higher resolution (Fig. 3A).

Fast non-averaged spectrum
The good raw signal-to-noise ratio (SNR) enables acquiring spectra in a fast, non-averaged measurement. This is shown with a 512 point spectrum of benzene at 1000 cm -1 (Supplementary Figure 8). The measurement time for the entire spectrum was 9.2 ms, the time per spectral point was 18 µs (acquired with the 55-kHz probe laser). The SNR of 34 in this non averaged acquisition was achieved with an average power of 2.6 mW for the probe and 175 mW for the pump laser. With the 415 kHz System, the measurement time can be reduced down to 2.5 µs per spectral point.
By employing faster FDML lasers and higher pulse repetition rates, we recorded a 64-point spectrum of toluene in 157 µs (Supplementary Figure 9). The 415 kHz FDML allows for these spectra with 6.4 kHz acquisition rate. The SNR of 11 was achieved with an average power of 2.6 mW for the probe and 500 mW for the pump laser (higher than in Supplementary Figure 8 due to the higher repetition rate).

Imaging of polystyrene (PS) and poly methyl methacrylate (PMMA) beads
Supplementary Figure 10 shows a faster hyperspectral TICO-Raman image of a mixture of PS and PMMA beads dispersed in ultrasound gel. The beads are 20 µm in diameter, pixel size is 1 µm 2 and the image size is 480 x 320 pixels. 2 mW of probe power and 400 W of instantaneous pump power (500 mW average power) were used. At each pixel a 32 point spectrum was acquired and averaged 8 times, resulting in a pixel dwell time of 636 µs, or 19 µs/spectral point with the 415 kHz system. The spectra were then Savitzky-Golay filtered using three side points and second order polynomials. Supplementary Figure 10B shows two spectra of PMMA and PS, where 5 x 5 pixels were averaged at the points P1 and P2, respectively. The coloured areas indicate the spectral points over which the coloured images were averaged.

Sparse-Sampling: Specifically tailoring the number of spectral points
For most Raman microscopy applications, more than one spectral point per pixel is desired but less than 100 are required. For most cases 4-20 points will be sufficient to provide enough information for a molecular contrast and at the same time are few enough to still enable high speed imaging. In our TICO-Raman system, the number of spectral points can be adjusted and

Noise suppression
The FDML probe laser already provides low intensity noise, only about 20 dB above the shot noise for 1 mW power on the photodiode and a detection bandwidth of 100 MHz. Low intensity noise is a prerequisite for measuring small signal changes on top of a large offset. A further, very unique feature of FDML is the good correlation between consecutive sweeps.
Because an FDML laser optically stores the light field of the wavelengths sweep, every sweep is an optical copy of the last one and the remaining noise is strongly correlated. This enables extremely efficient digital referencing algorithms.
A second mechanism to reduce noise is the use of dual balanced, differential detection with a pair of matched photodiodes. For this purpose the probe laser is split into a reference and a sample arm in the beam delivery unit (Fig. 2C). Both signals are detected with a photoreceiver outputting only the difference in the signal. This allows electronic common mode rejection of the probe laser light of up to 40 dB to be achieved. Additionally, the balancing subtracts the probe offset (Fig. 2E) so the depth resolution of the analogue-to-digital converter (ADC) is optimally utilized for signal digitization.
A third mechanism to reduce noise and, more importantly, to reduce artefacts is digitally subtracting signals from two consecutive sweeps of the probe laser, one with and one without pump light. This step eliminates chromatic differences in the splitting ratio between the two optical pathways and interference effects causing spectral ripple and it reduces acoustic and thermo optical crosstalk. This concept also functions as an additional noise filter for the probe laser to further suppress the 1/f-noise.
With the combination of these measures, our system achieves shot-noise limited relative sensitivity of 1.8 10 √Hz at 2 mW probe light. With a 1.8 ns gate time, a relative transmission change of 3.6·10 -4 can be measured without averaging. This value was confirmed experimentally.

TICO Modulation pattern for successive spectral sampling
The spectra are generated by scanning the energy difference between pump and probe laser (Supplementary Figure 3). This is achieved by using arbitrary waveforms to drive the whole laser system. The  Figure 3A). Furthermore, the SRG on the probe laser is measured, instead of the SRL on the pump, as the signal scales as:

∝
The analog holds for the SRL on the pump, but as the pump power is about 5 orders of magnitude higher than the probe power, the relative change on the probe is on the order of The ADC card is triggered once for every spectrum, so every 513th FDML period. Although after 512 periods the whole spectrum is already scanned by the pulse, another sweep without pulse is added to the end of one acquisition to serve as background reference. This is due to the pulse height calculation, where we subtract consecutive sweeps for digital balancing. The 513th blank sweep at the end of each waveform provides the subtrahend to the last pump pulse position.

Time-to-wavenumber encoding
The sinusoidally-driven FDML probe laser provides inherent time encoded wavelength information. This is due to the phase-locked, electronically driven filter element. The wavelength sweep can be expressed as: where , the inverse of the drive frequency, is the sweep period, and are the center wavelength and the sweep range, respectively. The phase Φ depends on the initial position of the filter and is set to zero on the AWG. One sweep direction is used for spectral acquisition and sampled on the ADC with 512 spectral points. Therefore, the wavelength of a sample # is: The Raman transition wavenumber corresponding to the energy difference between pump and probe light for a given sample is given by: The centre wavelength and the sweep span are recorded with an optical spectrum analyzer (Yokogawa, AQ6370). The sample point array was then computed using a LabVIEW measurement program.
Since the FDML sweeps a cosine function in wavelength, the wavelength difference between consecutive samples varies. At the edges of the sweep range, successive samples have less difference in wavelength than near the centre wavelength. For the resolution of the Raman spectra, the maximum wavenumber step between two adjacent spectral points was taken. This holds as long as the linewidths of the applied lasers do not dominate the resolution. Currently the FDML lasers can provide linewidths better than 50 pm and the pump laser better than 60 pm. This results in a minimal spectral resolution of below 0.5 cm -1 for our current setup.

Beam delivery unit
The pump and probe light is delivered to the spectroscopy and microscopy setup in optical

Dynamical spectral zooming
Spectral zooming is achieved by reducing the span of the probe FDML laser while keeping the number of spectral samples constant. Spectral zooming allows a narrow region of interest to be sampled more densely, resulting in higher spectral resolution. The region can be chosen freely within the bandwidth of the probe laser. The spectra in Supplementary Figure 4 were acquired with the same FDML laser. For the overview record in blue, the wavelength span was set to 106 nm around a centre wavelength of 1276 nm. With the 1122 nm pump, this results in a spectral coverage from 736 cm -1 to 1388 cm -1 with a spectral resolution of < 3 cm -1 . The spectrum with higher resolution in red was acquired with a 15 nm span around 1264 nm, thus covering the range from 960 cm -1 to 1055 cm -1 with a resolution < 0.5 cm -1 . In this example, the spectral zooming allows for a more accurate determination of the peak height and the neighbouring peaks of toluene (1005 cm -1 ) and benzene (992 cm -1 ) are more readily resolved.

Sample preparation for spectroscopy
For the chemical samples, we used thin strips of 30 µm thick PTFE sheets placed on microscope slides and covered with a No.1 coverslip (~ 160 µm in thickness). The liquid samples could then be deposited on the edge of the coverslip and entered the compartment by capillary forces. To avoid evaporation, the compartments were then sealed with conventional glue which did not enter the sample volume due to its higher viscosity. We used pure liquid chemicals for spectroscopy (SigmaAldrich). The beams enter through the coverslip side to minimize chromatic aberrations introduced by the glass, as the coverslip is thinner than the microscope slide.  (Fig. 4), we cut about 60 µm thick slices out of the stem with a microtome (Euromex, MT.5503) and placed it on a microscope slide. A droplet of conventional olive oil was added.

Spectra data processing
The spectra of the chemicals were recorded with a 55 kHz sweep rate. Two FDML lasers centred at 1300 nm and 1550 nm can be combined with the 1064 nm or the 1122 nm pump, resulting in a total coverage from 750 cm -1 to 3150 cm -1 . The acquisition time of a single raw 512 point spectrum was 9.2 ms (time = [FDML laser frequency / (no of spectral points+1)] -1 ).
The spectra were averaged 1,000 times for a high signal to noise ratio. Since the system operates at the shot noise limit, more averaging improves the quality of the spectra even further.
A small remaining offset was compensated by an automated baseline subtraction in Origin Software (Supplementary Figure 5). No subtraction was performed for the linearity measurement of the chemical mixture (Fig. 2B,C) and the non-averaged spectra (Supplementary Figure 8,9). For the broadband spectrum (Fig. 3), the four spectra were merged to 1565 spectral points, since redundant points in the overlap regions were omitted.

Image data processing-molecular contrasting
The TICO-Raman microscopy images were acquired by raster scanning the sample. The resolution was set to 600 x 400 pixels with 1 µm x 1 µm pixel spacing. At each point a 64point spectrum was acquired. The 100-times averaged spectra were Savitzky-Golay filtered in 2 nd -order with 3 neighbouring points. A remaining offset was corrected with a 2 nd -order polynomial fit (cp. fig S6). In post-processing the optimal signal-to-noise ratio was achieved by averaging over 11 neighbouring spectral points of the broad Raman transitions of olive oil and 17 spectral points for lignin (coloured areas in Supplementary Figure 6). The obtained intensity distributions were then exported to an 8-bit grey scale image. After applying appropriate intensity cut levels and rescaling to 8-bit, the images were coloured in an image processing program (Gimp 2) and overlaid by adding the single coloured layers to an RGB image.
Since the sample was removed between the measurements of the Raman contrast and the transmission microscopy image, the images had to be aligned and slightly resized before image fusion (Fig. 4E). There are some partially black lines visible in the image, the origin of these small artefacts is not known.