2.1.

Instrumentation

The experimental setup was similar to that described previously.²⁵ We use a tunable diode laser (Sacher, Littmann configuration, 785 nm, maximum power 1 W, total tuning range 200 GHz). The laser beam was coupled to the excitation part of the Raman probe (Emvision LLC) after passing through a line filter. The collection part of the probe was then coupled to the spectrometer (Shamrock SR-303i, Andor Technology) using an F-number matcher. The F-number of the spectrometer was 4, and that of the collection fiber was 2.27. Hence a pair of lenses with focal lengths of 16 and 10 mm, respectively were used in the F-number matcher, between which an edge filter (SEMROCK) was inserted to filter out 785 nm photons. The spectrometer was equipped with a $400\text{lines}/\mathrm{mm}$ grating with a deep depletion, back illuminated and thermo-electrically cooled CCD camera (Newton, Andor Technology).

The fiber probe used for this study had one optical fiber with a 200-μm diameter to deliver the excitation beam and seven optical fibers each of 200-μm diameter to collect the Raman signal. The excitation part in the probe head included a line filter to filter out background fluorescence from the fiber, and the collection part included a high-pass filter which rejects the 785-nm photons. A GRIN lens at the tip of the probe head helped to increase the collection efficiency. The probe had a 1-m-long fiber pigtail, and the probe head had a diameter of 4.2 mm. With a working distance of $\sim 1\mathrm{mm}$, the output excitation beam from the probe at 1 mm had a diameter of $\sim 500\mu \text{m}$. In order to avoid signal fluctuations due to variation in the axial height of the probe head from the tissue surface, the probe was converted to a contact probe by introducing a sapphire window at the tip of the probe head. This was achieved by inserting the probe head into a flexible sleeve (Tygon) of inner diameter 4.8 mm, outer diameter 6.8 and length 40 mm, to the end of which a 1 mm thick sapphire window (Comar optics) was fixed. This contact probe head was pressed against the tissue surface while acquiring Raman spectra. The average excitation power at the sample was $\sim 200\mathrm{mW}$, which corresponds to a power density of $\sim 25\mathrm{W}/{\mathrm{cm}}^2$. In order to compare the probe performances, another custom-designed fiber probe (Emvision LLC) with a 5-m fiber pigtail length was also used to acquire Raman spectra from bone tissue. The optical design of the second fiber probe was similar to the one described previously, except for the longer fiber pigtail length, which would contribute to higher background from the fiber. Although the probe head design was similar it was experimentally observed that the longer probe had higher collection efficiency while the fluorescence background was higher due to the longer fiber pigtail.

All the data presented in this paper were acquired at a single grating position of the spectrometer. The wave-number axis was calibrated using a quadratic interpolation for four standard Raman peaks of ethanol in the fingerprint region.

2.2.

Sample

The samples used for the investigation were bovine bone tissues and adipose tissue derived from chicken, pork, beef, and lamb. The bovine bone tissue was dissected from the rib of beef; the adipose tissue for the experiment was dissected adjacent to the rib of the beef, pork, and lamb. In the case of chicken, the adipose tissue was dissected from above the chicken breast, all within 48 h of slaughter.²⁶ Care was taken to ensure that for adipose tissue only the part next to the skin was chosen for the experiment since the fatty acid composition changes across the cross-section of the tissue.^26,27 The Raman signals from all these samples were collected at the ambient temperature of 19 to 21°C. For each sample, 20 different spectra were collected to ensure we could obtain reliable statistics.

2.3.

Raman Acquisition Method

The principle of WMRS depends on the fact that, in contrast to the fluorescence background, the Raman peaks synchronously modulate along with the modulation of the excitation wavelength. This differential signal is then extracted from the constant background by using PCA, where a singular value decomposition (SVD) algorithm was used to estimate the principal components.¹³ The experiment consists of acquiring a series of Raman spectra each with 5-s single exposure time; each spectra corresponds to a slightly differing excitation wavelength in discrete steps with equal intervals. Each Raman spectrum was acquired by full vertical binning (FVB) the CCD. The wavelength of the excitation beam was varied in discrete steps in a symmetric trapezoidal pattern with time period 30 s. The amplitude of the modulation was $\mathrm{\Delta }\nu =160\mathrm{GHz}$, which corresponds to a wavelength shift of $\mathrm{\Delta }\lambda =0.32\mathrm{nm}$. For the tissue analysis, the total acquisition time was 30 s corresponding to 6 spectra. The SNR of the modulated signal depends on the modulating amplitude, frequency and acquisition time, and in turn these parameters depend on the fluorescence characteristics and Raman cross-section of the sample, the shape of the Raman peak, and the detection geometry itself.²⁵

From a measured dataset, a standard Raman spectrum was obtained by accumulating all modulated spectra belonging to a set.

2.4.

Data Treatment

Each of the acquired spectra was normalized (as explained in the section below), and PCA was then applied to obtain the modulated Raman spectra from the constant fluorescence background. The spectrum of the first principal component corresponds with the maximum variation in the data set originating from the continuous shift of the Raman peaks. This spectrum gives a derivative-like Raman signal with most of the fluorescence background suppressed.

The standard Raman spectrum was smoothed with a Savitzky Golay filter of smoothing width 9 and order 3.

The SNR measured from the differential spectrum was compared with that from the standard Raman spectrum obtained for the same peak. The SNR of a particular Raman peak was estimated as the ratio of the peak intensity value (peak-to-peak value for modulated Raman spectra) to the standard deviation of the signal in a Raman free spectral region (here we chose 2000 to $2060{\mathrm{cm}}^{-1}$ as the noise region).¹³ For a fair comparison between the SNR of standard and modulated Raman, SNR from the first differential of the standard Raman spectrum was also estimated. Signal to background (S/B) of a Raman peak was estimated as the ratio of the height of the Raman peak from the fluorescence background to the height of the fluorescence background,²⁸ after subtracting the dark current of the CCD from the recorded spectrum.

2.4.1.

Correction for intensity fluctuations and photo-bleaching

The basic assumption of the WMRS technique is that only Raman peaks get shifted between different short acquisitions while the fluorescence background remains constant. In practice, this is not completely true. The power of the laser source fluctuates [$\sim 5\%$ as shown in Fig. 1(b)] while the wavelength of the source is modulated.⁷ Another source of fluctuation contained within the fluorescence background (especially when biological samples are probed) is the photo-bleaching of the sample. In order for the PCA procedure to pick out the Raman peak, both of these two fluctuations have to be corrected simultaneously.

Fig. 1

Illustration of the procedure to correct the signal for power fluctuation and photo-bleaching, using a dataset of 36 Raman spectra. Each point in the plots [(a)] to [(h)] corresponds to the pixel intensity of a single Raman spectrum. (a) Mean value of adjacent 10 pixels from $2000{\mathrm{cm}}^{-1}$ in the noise region. (b) Normalization factors obtained from the data shown in [(a)] after correcting this data for photo-bleaching by subtracting a second order polynomial fit. (c) Variation of signal intensity of a pixel at a Raman band region. (d) Signal shown in [(c)] after normalization using the normalization factors shown in [(b)]. (e) Signal shown in [(d)] after correcting for photo-bleaching by subtracting second order polynomial fit. It can be seen that the signal periodically fluctuates which corresponds to the modulation of the wavelength of excitation. (f) Variation of signal intensity of a pixel at a noise region. (g) Signal shown in [(f)] after normalization using the normalization factors shown in [(b)]. (h) Signal shown in [(g)] after correcting for photo-bleaching subtracting a second order polynomial fit. (Std.) Standard deviation of signals. (i) flow chart showing the correction procedure.

Although photo-bleaching results in an exponentially decaying fluorescence signal, within a relatively short acquisition time (10 s), this can be corrected by subtracting the signal with a linear or second order polynomial fit. The power fluctuation can be normalized by estimating the normalization factors from the Raman free spectral region (background region).

In order to demonstrate the process of correcting for power fluctuation and photo-bleaching, a set of 36 wavelength-modulated Raman spectra, corresponding to six modulation cycles is chosen. Each spectrum is acquired with 5-s acquisition time and the dataset is represented as a $1024\times 36$ matrix. Each column in this dataset corresponds to a single Raman spectrum and each row corresponds to variation of intensity of the spectra at a particular pixel. Without these fluctuations, a row in a Raman band region should be periodically fluctuating while a row in a background region should ideally be non-fluctuating.

Figure 1(a) to 1(h) details the correction procedure, where in each plot, one data point corresponds to the pixel intensity of a single Raman spectrum. The first step is to find the factors to correct for power fluctuation. The mean of 10 rows in the background region (10 pixels from $2000{\mathrm{cm}}^{-1}$ in this case) was estimated for this [Fig. 1(a)]. To separate the power fluctuation information from variation due to photo-bleaching, a second order polynomial fit was performed on this mean data. Subsequently this fit was subtracted from this mean data. Each of the elements in this fit-subtracted data was subsequently divided with the intensity value of the 14th acquisition (chosen randomly) to obtain factors to correct for power fluctuation [Fig. 1(b)].

After obtaining the correction factors, each row in the dataset was corrected for photo-bleaching by subtracting the data with a second order polynomial fit, followed by dividing with the correction factors [Fig. 1(b)] in order to compensate for the power fluctuations.

The effect of this correction is illustrated in Fig. 1(c) to 1(h). Figure 1(c) shows a row in the region of a Raman peak. After correcting for power fluctuation the resultant data is shown in Fig. 1(d), where it can be seen that, although there is a periodicity in the signal, the signal is decaying due to the photo-bleaching effect. After correcting for photo-bleaching, the data as seen in Fig. 1(e) is periodically fluctuating. This corresponds to the wavelength modulation of the Raman peak. Figure 1(f) shows a row in a fluorescence background region. Figure 1(g) shows the data after power fluctuation corrections and Fig. 1(h) shows the data after correction for photo-bleaching. It can be seen that the ratio of standard deviation of the array in Fig. 1(e) and 1(h) is one order of magnitude higher than the ratio of standard deviation between the arrays in Fig. 1(c) and 1(f). This naturally translates into a better distinction of Raman peak and fluorescence background when PCA is applied to this series of modulated Raman spectra. The Matlab code for the correction procedure is given as an Appendix to this article.

3.1.

Bovine Bone Tissue Fluorescence Suppression

Raman spectra of bovine bone tissue were acquired using fiber Raman probes. A comparison of the standard and modulated signals is shown in Fig. 2. It can be seen that the modulation technique has suppressed the fluorescence background and enhanced the SNR of Raman peaks. We performed this experiment with two different Raman probes of differing pig tail lengths: 1 m and 5 m. It can be seen that the signal to background ratio for the phosphate $v_1$ peak (marked in the figure by a star symbol) at $960{\mathrm{cm}}^{-1}$ is higher for the probe with the shorter fiber pigtail due to the higher fluorescence contribution from the material of the fiber. However it was observed that there was a significant enhancement ($\sim 8\text{times}$) in the phosphate ${\nu }_1$ Raman peak in the modulated signal when compared with the standard for both of the fiber Raman probes. When compared with the SNR of the first order differential of the standard Raman spectrum, the phosphate ${\nu }_1$ Raman peak improved the SNR by a factor of approximately two. However, the first order differential contained some artifacts in the form of spurious peaks. The higher SNR for the fiber probe with the longer fiber pigtail is due to the higher collection efficiency of that probe. With this technique even the other weak peaks at $1070{\mathrm{cm}}^{-1}$ corresponding to the carbonate ${\nu }_1$, the amide III peak at $1245{\mathrm{cm}}^{-1}$, ${\mathrm{CH}}_2$ scissoring at $1451{\mathrm{cm}}^{-1}$, and the amide I peak at $1672{\mathrm{cm}}^{-1}$ were amplified as shown in the Fig. 2.³ The total acquisition time required to collect this signal was 30 s.

Fig. 2

Comparison between the standard and modulated Raman signal of bovine bone tissue recorded using two fiber probes with different fiber pigtail length. The SNR has been estimated from 20 Raman spectra acquired with a total acquisition time of 30 s each. The modulated Raman spectrum has been obtained by applying PCA to six Raman spectra acquired with single acquisition time of 5 s each. Standard Raman spectrum has been obtained by accumulating six Raman spectra acquired with single acquisition time of 5 s each. (Peak assignment: phosphate ${\nu }_1$: $960{\mathrm{cm}}^{-1}$, carbonate ${\nu }_1$: $1070{\mathrm{cm}}^{-1}$, amide III: $1245{\mathrm{cm}}^{-1}$, CH2 scissoring: $1451{\mathrm{cm}}^{-1}$, amide I: $1672{\mathrm{cm}}^{-1}$).

3.2.

Fluorescence Suppression in Adipose Tissue

The WMRS technique may also be used to analyze samples with relatively high Raman cross-section, since WMRS can enhance the SNR of the Raman peaks. In order to demonstrate this, Raman spectra of adipose tissues from four different species—namely pig, lamb, chicken, and beef—were analyzed. The Raman spectra of the adipose tissues may be an indicator towards the quality of different animal fats.^26,27

Figure 3 shows the standard and modulated Raman spectra acquired from adipose tissues of four species. It can be seen that the SNR has been enhanced by a factor of approximately two when WMRS is implemented.

Fig. 3

Comparison between the standard and modulated Raman mean spectra of the adipose tissue from pork, lamb, chicken, and beef. The SNR has been estimated from 20 Raman spectra acquired with a total acquisition time of 30 s each. Modulated Raman spectra have been obtained by applying PCA to six Raman spectra acquired with single acquisition times of 5 s each. Standard Raman spectra were obtained by accumulating six Raman spectra acquired with single acquisition times of 5 s each. (Peak assignment: out-of-phase aliphatic $\mathrm{C}\text{-}\mathrm{C}$ stretch all-trans: $1060{\mathrm{cm}}^{-1}$, in-phase aliphatic $\mathrm{C}\text{-}\mathrm{C}$ stretch all trans: $\text{1125\hspace{0.17em}\hspace{0.17em}}{\mathrm{cm}}^{-1}$, methylene twisting deformations: $1295{\mathrm{cm}}^{-1}$, methylene scissor deformations: $1436{\mathrm{cm}}^{-1}$, olefinic stretch: $1650{\mathrm{cm}}^{-1}$, carbonyl stretch $1730{\mathrm{cm}}^{-1}$).