2.1.

Nonlinear Optical SRS Process

SRS occurs when there is an energy transfer from a high-energy pump photon to a low-energy Stokes photon via vibrational interactions with a molecule, as shown in Fig. 1(a). In this case, the pump beam excites a biomolecule from its ground state to a virtual state. Then when the Stokes beam (instantaneously) stimulates the molecule back down to a vibrational state, a coherent Stokes photon is generated. The frequency difference between the pump and Stokes beams is denoted as $\mathrm{\Omega }={\omega }_{\text{pump}}-\omega$. The nonlinear change experienced by the pump field is expressed as

Fig. 1

Schematic of SRS interaction and experimental SRS-SOCT setup. (a) SRS process energy diagram. (b) An ultrafast laser system designed to deliver pump and Stokes beams for SRS-SOCT. (c) The imaging setup, with a spectroscopic detection using a Michelson interferometer. DM, dichroic mirror; BS, beam splitter; OPO, optical parametric oscillator; OPA, optical parametric amplifier; AOM, acoustic optical modulator; and PPLN, periodically poled lithium niobate. In the inset of (b), we show measured signals in the frequency domain, with the SRS-response peak at 15 kHz, given by the AOM modulation frequency. The peak only appears when Stokes beam is on, indicating the nonlinear optical process.

2.2.

Ultrafast Laser System for SRS-SOCT

We recently introduced an ultrafast laser system specially designed for SRS-SOCT.²⁶ The schematic of the system is shown in Fig. 1(b). Overall, this system operates at a 40-MHz repetition rate with two outputs: one for the Stokes beam and the other for the pump beam, which also serves at the SOCT light source, as illustrated in Fig. 1.

For the SRS pump pulse, first, a high-power ($\sim 8\mathrm{W}$) Yb:KGW oscillator ($\sim 450\mathrm{fs}$, 1040 nm) is used as the pump source for the laser system. The output is then sent to an optical parametric oscillator (OPO) and an optical parametric amplifier (OPA) to produce a tunable output from 1.4 to $2\mu \mathrm{m}$. Then the pulse is frequency doubled using a periodically poled lithium niobate (PPLN) crystal, which produces a narrow band pulse at 790 nm. Next, the pulse is spectrally broadened via self-phase modulation using a tapered SMF28 fiber (78-mm length). This step is critical to generating a low noise, compressible broadband pulse that is well suited for SRS and SOCT. Finally, a prism pulse compressor is used to compensate for chromatic dispersion and shorten the duration of the output pulse ($\sim 50\mathrm{fs}$).

For the SRS Stokes pulse, we directly split a portion of the Yb:KGW oscillator’s output. This pulse is also actively modulated at 15 kHz by an acoustic optical modulator (AOM, Gooch & Housego, AOMO 3200-124) for lock-in detection.

Overall, our laser system outputs a pump pulse centered at 790 nm (but tunable from 740 to 975 nm) with a bandwidth of $\sim 40\mathrm{nm}$ and compressible duration down to $<50\mathrm{fs}$. The average pump output power from the laser is set to $\sim 115\mathrm{mW}$, of which $<50\mathrm{mW}$ is incident on the sample. The Stokes pulse is fixed at 1040 nm, with 450 fs duration and an average output power of $\sim 200\mathrm{mW}$ ($\sim 95\mathrm{mW}$ is incident on the sample). As with other studies,^13,17,18 no noticeable tissue damage was produced at these power levels. With this configuration, we achieve an axial resolution with OCT of $\sim 10\mu \mathrm{m}$ and cover a Raman spectral region of $\sim 2700$ to $3200{\mathrm{cm}}^{-1}$. The SRS signal shares the same axial resolution as OCT, but as with any SOCT method, the molecular signatures accumulate along depth.^25,26 The SRS-SOCT signal scaling with pump and Stokes power has been rigorously characterized in our prior work using well controlled synthetic samples and fresh tissues.^25,26

2.3.

Interferometric Spectroscopic Detection in SRS-SOCT

The imaging portion of the SRS-SOCT system is shown in Fig. 1(c). First, the SRS pump beam output from the laser is sent to a Michelson interferometer where the light is split into two identical beams: one used as the reference beam and the other used as the sample beam. The Stokes beam, modulated at 15 kHz by an AOM as described above, is combined with the pump beam through a dichroic mirror. Both the pump and Stokes beams are focused onto the biological sample (rat brain, details below in Sec. 2.4) using a $5\times$ objective (Zeiss, A-Plan $5\times$, NA = 0.12). With a beam diameter of the laser of 2 mm, this yields a lateral resolution of $5.2\mu \mathrm{m}$, which applies to both the OCT and SRS signals. The generated SRS signal, encoded in the sample beam, is collected after light is backscattered and detected interferometrically as in conventional Fourier-domain SOCT. Here we use a fast spectrometer (BaySpec, OCTS-0780-0840-0900-SLIT) to record the interferometric signal. Spectra are recorded at 50 kHz (camera line scan rate), and for each A-scan, 2048 spectra are recorded. The 15-kHz Stokes beam modulation frequency is selected to be well above the thermal bandwidth of the sample to avoid any thermal contribution to the SRS signal^29–31 {Choi, 2001 #55}. We also use a 2D Galvo-System (Thorlabs, GVS012) to steer the beams and perform lateral scans.

The recorded spectral interference between the back reflected sample field and the reference field $E_0(\omega )$ (same as the initial pump field) is expressed as $I_{\mathrm{spec}}(\omega )={|E_0(\omega )+rE(\omega )\exp (-i\frac{\omega ·2z}{c})|}^2$, where $E(\omega )$ is given by Eq. (1), $r$ is reflectivity, and $z$ is the pathlength difference between the two beams (i.e., depth). From $I_{\mathrm{spec}}(\omega )$, the term of interest is the oscillating term that contains the nonlinear refractive index ${\tilde{n}}_{\mathrm{NL}}$, embedded in $E(\omega )$. Considering Eqs. (1) and (2), we express the oscillating term as

These expressions assume a weak nonlinear perturbation, which is a reasonable approximation for SRS-SOCT.^25,32 It is worth noting that the SRS intensity (the conventional SRS measurement) shows a dependence on the initial intensity of the pump field $I_0$ and sample reflectivity $r$, whereas the SRS phase does not. This results in the phase signal being more robust to laser power fluctuations and independent of the samples’ reflectivity,³² thus providing a better signal-to-noise ratio.

2.4.

Biological Sample Preparation

To test the ability of SRS-SOCT to detect brain tumors, we use the 9L gliosarcoma rat tumor model, which mimics human high-grade gliomas, including high proliferative capability, vascularization, and infiltrative pattern.²⁸ All animal experimental protocols were approved by the Institutional Animal Care and Use Committee of Georgia Institute of Technology and Emory University. As shown in Fig. 2(a), cloned 9L glioblastoma cancer cells were implanted in $N=6$ Fischer rat brains. Tumor was allowed to develop for 9 to 12 days before the rats were sacrificed. To gain access to the tumor cross section, the surgically removed brains were coronally cut in two halves across the approximate tumor implantation point. SRS-SOCT imaging was performed on the tumor (identified via visual inspection) as well as the counter-lateral, healthy brain region, which was used as control. A total of 54 regions were imaged (34 from tumor regions and 20 from healthy regions). We acquired 8 to 10 scans from each brain ($\sim 4$ to 5 points from tumor regions and $\sim 4$ to 5 points from healthy regions). Figs. 2(b) and 2(c) show photographs of a representative harvested and cut rat brain.

Fig. 2

(a) Schematic of rat brain implantation. (b) Top view of the harvested brain. The implant point is indicated by the black arrow. (c) Cross section of the brain with tumor region (marked in dashed circle).

2.5.

Measuring SRS-SOCT Signal

Using the SRS-SOCT system, we measured interferometric signals that contain the SRS response of rat brains. For each of the freshly cut brains ($n=6$), we took 8 to 10 B-scans (total of 54 B-scans), each composed of 200 A-scans (synthesized from 2048 sequential recorded spectra) along a $\sim 480\text{-}\mu \mathrm{m}$ line shift. In the axial direction, the images cover a range of $\sim 2\mathrm{mm}$, but the penetration into the brain tissue is only $\sim 0.7$ to 1 mm. The scanned points were selected at equidistant positions to cover both tumor in one side and healthy regions in the counter lateral side. No staining or other sample preparations were performed for SRS-SOCT imaging. For this feasibility study (and given the use of an established animal model), all measurements were taken as independent even if they were from the same animal. Tumor regions from adjacent regions were analyzed with histopathology (using H&E stained tissue sections), and the presence of tumors was confirmed in all imaged animals. Healthy regions were also confirmed.

3.1.

Processing SRS-SOCT Dataset

To extract the complex nonlinear SRS signals, the interferometric spectra first are interpolated from wavelength to a linear wavenumber array ($k=2\pi /\lambda$) and then Fourier transformed from the spectral to spatial domain to get depth information (A-scan). For each lateral position in the B-sacn, we have 2048 scans that were recorded at 50 kHz. Next, to extract the SRS response of the sample, we perform another Fourier transform along the temporal dimension to get the frequency response of the A-scans. We select only the signal corresponding to the 15-kHz modulation (i.e., lock-in detection), as shown in the inset of Fig. 1(c). This way, we obtain both the SRS amplitude and phase terms, which are then normalized by their respective background noise, as was done in Refs. 25 and 32. Finally, 200 acquisitions are synthesized to provide cross-sectional images (B-scan).

An example of the SRS-SOCT B-mode image ($x-z$ dimension) is shown in Fig. 3. Figure 3(a) shows the conventional OCT image, and Figs. 3(b) and 3(c) show SRS amplitude and phase images (average response across the high-wavenumber region). Again, the OCT image exhibits spatially resolved information from linear photon scattering, whereas the SRS-SOCT images contain spatially resolved information regarding the third-order nonlinear response of the sample.

Fig. 3

SRS-SOCT spatially resolved images (B-scans) of rat brain: (a) convectional OCT image, (b) SRS amplitude image, and (c) SRS phase image of a tumor region in rat brain. The B-scan images are synthesized from 200 A-scans ($x$ dimension), which covers $480\mu \mathrm{m}$.

The SRS-SOCT dataset contains both spatial and spectral information. In the following sections, we first analyze the SRS-SOCT data spectrally to give molecular specific information, and then we analyze the attenuation behavior as a function of depth from the OCT and SRS signals to show additional biomarkers that differentiate between tumor and healthy tissues.

3.2.

Spectral Analysis of SRS-SOCT Data

As previously reported,^18,33 the lipid and protein content of the brain tends to be lipid-rich for white matter, protein-rich for tumor regions, and lipid-and-protein-rich for cortex. These subtle molecular differences result in slight differences in the Raman spectral response, with most spectral differences arising in two particular regions in the high-wavenumber region.^18,33 The first region is attributed to spectral signatures of lipids and roughly covers the range between 2800 and $2900{\mathrm{cm}}^{-1}$. The second window signifies spectral signatures of proteins and covers the regions between 2900 and $3000{\mathrm{cm}}^{-1}$. Thus to provide molecular specific information that yields contrast for tumors, we considered two spectral regions: a high-frequency Raman band (centered at $2947{\mathrm{cm}}^{-1}$) to detect proteins and a low-frequency Raman band (centered at $2861{\mathrm{cm}}^{-1}$) to detect lipids. These two windows provide an average cumulative SRS response of the lipid and protein contents of the sample. In this analysis, the interferometric SOCT data are filtered using a Butterworth filter with an $80\text{-}{\mathrm{cm}}^{-1}$ ($\sim 5\mathrm{nm}$) bandwidth at the corresponding spectral regions, and then the nonlinear amplitude and phase changes are calculated following the procedure described above in Sec. 3.1. We also note that using two windows, instead of a full time-frequency-analysis with higher spectral resolution, we achieve higher signal-to-noise ratio data, improved spatial resolution, and faster, less computationally intensive processing.

To differentiate between tumor and healthy regions, we define the following metrics: For the SRS amplitude, we chose the ratio of the high-frequency band value to the low-freuqency band value. For the SRS phase, we simply take the difference of phase values from those two bands. Thus for both of these metrics, we expect protein-rich tumor regions to have a higher amplitude ratio and higher phase difference than the normal regions. Moreover, because the SRS-SOCT signal is cummulative along depth,²⁵ we extract these metrics from a depth range 100 to $150\mu \mathrm{m}$ from the brain surface, which allows for sufficient signal accumulation and avoids noisy, low-amplitude signals at deeper depths. Signals are averaged across each B-scan (54 regions in total) for all 6 rat brains.

Results from this analysis are presented in Fig. 4 in the form of standard box plots with interquartile ranges inside the box and the whiskers indicating the nonoutlier maximum and minimum ranges. Figure 4(a) shows the SRS amplitude ratio, which shows a slightly higher amplitude ratio for tumor than healthy regions, but the difference is not statisitically significant. Figure 4(b) also shows a higher SRS phase difference for tumor regions with a $p\text{value}<0.05$ using a two-sided $t$ test. As previously discussed, these results agree with our expectation that tumor regions, with a relatively rich protein composition, will have a higher phase difference, but indeed the differences are weak. Higher amplitude ratios were also expected, and while slight differences are observed, which are in agreement with these expectations, these are not statistically significant.

Fig. 4

(a) Ratio of SRS amplitude of the high-frequency band value to the low-frequency band value. Differences are not statistically significant (n.s.) with a $p$ value of 0.0973 (two-tailed $t$ test). (b) Difference of SRS phases between the same frequency bands. $p$ value is 0.0243 (two-tailed $t$ test), which is statistically significant ($p\text{value}<0.05$), denoted by the star (*).

The spectral analysis provides insights on the local biomolecular contents, but we note that the separation between tumor and healthy regions is still small and only statistically significant for the phase signal. Further, the ability to separate between these two groups based on spectral signatures alone is very poor. These results are consistent with SRS strategies to detect tumor,^18,34 where the spectral signatures have to be used in combination with structural parameters to achieve robust tumor detection. Thus to improve tumor identification, we incorporate structural signals from SRS-SOCT.

3.3.

Spatial Analysis of SRS-SOCT Data

One important property of the SRS-SOCT data is the rate at which the signals attenuate as a function of imaging depth. This parameter is affected by properties of brain tissue that affect scattering and the Raman signals such as cell density, protein concentration, nuclear-to-cytoplasmic (NC) ratios, changes in myelin content, existence of foci of necrosis, and tissue heterogeneity.^12,35–39 Further, because the SRS phase signal is contained in the argument of the interferometric signal, it is less prone to intensity variation of the pump beam and provides a more robust measurement than the nonlinear amplitude signal, and therefore it provides a higher SNR by a factor of 4 or more.^25,26 Therefore, we only use SRS phase signal in this analysis and compare it with OCT.

To calculate the decay rate at a point of interest, we perform an exponential fit of the OCT and SRS phase signals and calculate the average decay coefficients for each region (B-scan). To avoid artifacts generated by surface inhomogeneities, we set the starting point of the exponential fit at $\sim 50\mu \mathrm{m}$ below the sample surface. Figures 5(a) and 5(b) show plots of the decay coefficients of OCT and SRS-phase, respectively. The decay rate of healthy regions is significantly greater than tumor regions for both OCT and SRS-phase. We attribute the lower decay rate in tumor regions to changes in NC ratios as well as degradation of myelin content.¹² Specifically, those rates associated with tumor and healthy regions in OCT signal have an average value of 3.73 and $8.17{\mathrm{mm}}^{-1}$, respectively, with a discriminant threshold value of $\sim 6{\mathrm{mm}}^{-1}$. Similarly, in the SRS phase signal, tumor and healthy regions exhibit average values of 3.14 and $7.83{\mathrm{mm}}^{-1}$, respectively, with a discriminant value of $\sim 6{\mathrm{mm}}^{-1}$. These results are in excellent agreement with the previous OCT studies that report attenuation values of 2.7 to $3.5{\mathrm{mm}}^{-1}$ for cancer and $7.0{\mathrm{mm}}^{-1}$ for healthy regions with a discriminant threshold of $\sim 5.5{\mathrm{mm}}^{-1}$.^12,40 Our slightly higher attenuation cutoff may be attributed to the shorter center wavelength in our OCT systems (790 versus 1310 nm). The discrimination in decay rate of tumor and healthy regions suggests that OCT and SRS can both provide excellent separation between tumor and healthy tissues, with $p\text{values}\ll {10}^{-4}$.

Fig. 5

(a) Decay rate for OCT signal for rat brains. $p$ value for tumor versus healthy is $1.74\times {10}^{-16}$ (two-tailed $t$ test). (b) Decay rate for SRS phase signal for the same brains. $p$ value for tumor versus healthy is $3.13\times {10}^{-10}$ (two-tailed $t$ test) ($n=54$).

Finally, we note that both the OCT and SRS-phase A-scan signals in our data show a weaker backscattering response in tumor regions, despite having smaller attenuation coefficients. In other words, the A-scans of tumor regions contain weak but long-tailed signals, as opposed to the healthy regions with strong but short-tailed signals [Figs. 6(a) and 6(b)]. To show the differences in A-scans of the tumor versus healthy regions, we calculate the ratio of the area under curve for the first $100\mu \mathrm{m}$ from the surface, relative to the total area under the curve. As shown in Figs. 6(c) and 6(d), it is clear that the tumor regions exhibit a weaker scattering and SRS-phase response compared with healthy regions. Decreased scattering coefficient of brain tumor is attributed to several key factors such as cell density, NC ratio, and alterations in myelin content.¹² This results in a lower decay rate in brain tumor regions.

Fig. 6

(a), (b) A-scan profiles from tumor (red) and healthy (blue) regions from (a) OCT and (b) SRS phase. (c), (d) Ratio of the area under the curve for first $100\mu \mathrm{m}$ from the surface to the total area under the curve, averaged across each B-scan, for (a) OCT and (b) SRS phase. $p$ value for healthy versus tumor is $2.71\times {10}^{-5}$ for OCT and (d) $1.63\times {10}^{-14}$ for SRS phase. The A-scan from tumor regions show a smaller decay rate and weaker backscattering whereas healthy regions exhibit greater decay rate as well as stronger backscattering.

Although indeed the OCT and SRS-SOCT parameters show similar discrimination levels between healthy and tumor regions alone, the combination of these metrics provides complimentary information that enhances the combined method’s ability to identify tumors. To illustrate this point, Fig. 7(a) plots the attenuation coefficient of OCT versus the SRS-SOCT phase-signal area ratio, where this combined 2D parameter space provides better differentiation between tumors and healthy tissues than either of the two parameters alone. Next, we compute the coordinates of each point along the dashed line in Fig. 7(a) (rescaled from 0 to 1), which represents the direction along which the two groups are most distinguishable. Of note, the line is a diagonal, meaning that the two parameters together yield the best separation. These coordinates are then used to calculate statistical significance and generate a boxplot as shown in Fig. 7(b). Results yield a $p$ value of $6.63\times {10}^{-23}$, which is highly statistically significant.

Fig. 7

(a) Scatter plot of attenuation coefficient of OCT signal versus SRS-SOCT phase-signal area ratio. (b) The box plot of the projected distance of the scattered points from a reference point along the projection line in (a).