2.1.

Animal Model

The animal study was completed using the hamster cheek pouch carcinogenesis model. For the animal study, all experimental protocols were approved by the Institutional Animal Care and Use Committees of Duke University and North Carolina Central University and in accordance with the National Institutes of Health (NIH). Male Syrian golden hamsters, six weeks of age, were obtained from Harlan Laboratories (Indianapolis, Indiana) and housed at North Carolina Central University. The animals were housed four per cage in a room with controlled temperature and humidity and in a $12\text{-}\mathrm{h}$ light/dark cycle. Regular cage changes ensured maintenance of hygienic conditions. All animals were given the AIN-93M diet (Research Diets, New Brunswick, New Jersey). The diet consisted of 14% casein, 0.18% l-cystine, 49.5% corn starch, 12.5% maltodextrim 10, 10% sucrose, 5% cellulose, 4% soybean oil, 0.0008% t-Butylhydroquinone, 3.5% mineral mix, 1% vitamin mix, and 0.25% choline bitartrate. Tap water was available ad libitum. After an acclimatization period of one week, the left cheek pouch of each animal was topically treated with $100\mu \mathrm{l}$ of 0.5% 7,12- $\text{dimethylbenz}\left[a\right]\text{anthracene}$ (DMBA; Sigma Chemical Company, St. Louis, Missouri) in mineral oil with a paintbrush three times per week for six weeks. The right cheek pouch was left untreated and served as the control group.

2.2.

Experimental Protocol

At $24\text{weeks}$ after the initial treatment of DMBA, the hamsters were shipped to Duke University for optical spectroscopic analysis. The hamsters were euthanasia by $\mathrm{C}{\mathrm{O}}_2$ asphyxiation before being subjected to gross necropsy. The entire left and right cheek pouches were excised and cut into two pieces. The samples were laid flat between two coverglasses, moistened with phosphate buffered saline (PBS), and immediately scanned by the parallel frequency domain optical coherence tomography (pfdOCT) system. Following the optical measurements, scanned areas were marked with India ink and the tissue samples were fixed in 10% PBS buffered formalin. The fixed samples were later embedded in paraffin, sectioned, and stained with hematoxylin and eosin (H&E) for histopathological analysis.

The complete animal trial analyzed tissue samples from 21 hamsters. Although one treated and one untreated sample were extracted from each animal and scanned by the fLCI system, only 16 of 21 untreated samples were used in the study. The signal-to-noise ratio of the scans from the remaining five untreated samples was insufficient to provide useful data. Therefore, these scans were not included in the spectroscopic analysis.

2.3.

Parallel Frequency Domain Optical Coherence Tomography

Ex vivo tissue samples were examined using the pfdOCT system first described by Graf ¹¹ The pfdOCT system, shown in Fig. 1, is based on a modified Michelson interferometer geometry and utilizes a 4f interferometer first demonstrated by Wax ¹² The system utilizes a Xenon arc-lamp source ( $150\mathrm{W}$ , Newport Oriel, Stratford, Connecticut) for illumination, and the 4f interferometer uses two 4f imaging systems to spatially resolve light from the source to the detector. The detection plane of the imaging system coincides with the entrance slit of an imaging spectrometer (Shamrock 303i, Andor Technology, South Windsor, Connecticut), which spatially resolves 255 detection channels, each $25\mu \mathrm{m}$ in width. The imaging spectrometer optics, along with the combination of the $600\text{lines}∕\mathrm{mm}$ grating and the $1024\text{-pixel}$ CCD array, limits the detected spectrum to the $500\text{-}\text{to}625\text{-}\mathrm{nm}$ range. Data from the spectrometer is downloaded in real time to a laptop PC via the USB 2.0 interface, and spectrometer control and data acquisition is achieved using custom LabVIEW (National Instruments, Austin, Texas) software.

Fig. 1

Schematic of the pfdOCT system. Light source— $250\text{-}\mathrm{W}$ Xe arc-lamp. L1 through L5—lenses. BS—beamsplitter. M—reference mirror. Inset: Incoming light incident on the spectrometer slit. Slit allows only a small slice of incoming light to enter the imaging spectrometer. Adapted from Ref. 11.

The fLCI method seeks to recover structural information about scatterers by examining the wavelength dependence of the intensity of elastically scattered light. The technique determines scatterer sizes by analyzing the Fourier transform of the spectra originating from specific subsurface layers of a sample. Depth resolution is obtained by employing the coherence gating methods commonly used in frequency domain OCT. By exploiting the low temporal coherence length of a broadband light source in an interferometry scheme, fLCI can selectively analyze spectral information from the most diagnostically relevant layers in probed samples.

In order to perform depth-resolved spectroscopy, fLCI data must be processed to simultaneously obtain depth resolution and spectral resolution, from data acquired in a single domain. To implement this processing, fLCI and spectroscopic OCT (SOCT) have typically employed a short-time Fourier transform (STFT) in which a Gaussian window is applied to the interference signal before taking a Fourier transform, yielding a depth scan centered about a particular center wave number. By shifting the center of the Gaussian window and repeating the process, a data set with both depth and spectral resolution can be generated. It should be noted, however, that with this approach, any attempt to increase spectral resolution results in degradation of depth resolution and vice versa. Most recently, Robles introduced the dual window (DW) method for processing SOCT signals, which can be incorporated into the fLCI analysis.¹³ The DW method is based on performing two separate STFTs and combining the results to achieve simultaneously high depth and spectral resolution.

From the depth-resolved spectroscopic information, fLCI seeks to determine structural information by analyzing oscillations in the spectrum of light returned from a specific depth of interest. More specifically, fLCI seeks to distinguish between normal and dysplastic epithelial tissue by detecting the nuclear enlargement that occurs at the earliest stages of precancerous development. Figure 2a shows an illustration representing two nuclei as well as the scattering events that take place at both the front and back surfaces of each nucleus where an index of refraction change is present. Depending on the coherence of the field induced by the sample,¹⁴ the reflections from the front and back surfaces of the nuclei will interfere with one another, producing constructive or destructive interference, as shown in Fig. 2b. The frequency of this oscillation is directly dependent on the diameter and refractive index of the scatterer, with larger particles resulting in a higher frequency of oscillation and smaller particles resulting in a lower frequency of oscillation. The fLCI method seeks to detect and analyze these spectral oscillations to measure nuclear diameter.

Fig. 2

(a) Cell nuclei with incident and scattered fields indicated. (b) Interference spectra with wave number–dependent oscillations caused by interference between front and back surface reflections.

2.4.

Data Processing

The raw data acquired by the pfdOCT system consisted of 120 spectra, each of which originates from adjacent $25\text{-}\mu \mathrm{m}$ -diam spatial points on the experimental sample. The raw data, along with the plots of three such spectra, are shown in Fig. 3a . The diameter of the signal beam was shaped to illuminate only 120 of the 255 spectral channels of the imaging spectrometer to preserve the signal-to-noise ratio of the measurements.

Fig. 3

(a) Raw data from the complete animal trial with spectra from three spectrometer channels shown. (b) Three typical depth-resolved spectroscopic plots produced by DW processing the spectra in (a). Summing the plots from all 120 channels produces the final TFD, as shown.

To analyze spectra from specific tissue layers, the spectrum detected by each channel of the imaging spectrometer was processed using the DW processing method.¹³ Briefly, the DW method uses the product of two STFTs to reconstruct the time-frequency distribution (TFD) of the interferometric signal: one STFT with a narrow window for high spectral resolution and another with a wide window for high spatial resolution. Equation 1 gives a mathematical description of the distribution obtained with the DW method from a single spatial line:

Robles have shown that the distribution obtained from the DW method can be related to Cohen’s class of bilinear functions,¹³ even though it is constructed using two linear operations. In one limit, where $a^2∕b^2⪡1$ , the DW distribution gives a measurement of the Wigner time-frequency distribution (TFD) with spectral and depth resolution set independently by the width of the two orthogonal windows, $a$ and $b$ . Significantly, the use of the two orthogonal windows eliminates many common artifacts in other TFDs, such as the cross-term artifacts from the Wigner TFD and the reflections in time artifacts from the Margenau and Hill TFD.¹³ Further, the DW contains local oscillations in the spectral dimension that reveal morphological information about the sample—specifically, the distance between scattering surfaces in the vicinity of the point of analysis.

The DW method was implemented using a custom MATLAB program to process the data with both a narrow spectral window of $0.0405\mu {\mathrm{m}}^{-1}$ FWHM and a wide spectral window of $0.665\mu {\mathrm{m}}^{-1}$ FWHM. The depth-resolved spectra generated by each window were multiplied together to produce a plot with simultaneously high spectral and depth resolution. The resulting 120 depth-resolved spectroscopic plots were summed together to improve the signal-to-noise ratio, producing a single depth-resolved spectroscopic plot for each tissue sample, as shown in Fig. 3b.

In neoplastic transformation, nuclear morphology changes are first observed in the basal layer of the epithelial tissue. In hamster buccal pouch tissue, the basal layer lies approximately $30\text{to}50\mu \mathrm{m}$ beneath the surface for normal tissue, and approximately $50\text{to}150\mu \mathrm{m}$ beneath the surface for dysplastic tissue. Because examination of the basal layer offers the earliest opportunity for detecting developing dysplasia, it is the target tissue layer for the fLCI technique and for this study.

In order to target the basal layer of the epithelium, the raw experimental data were first processed to yield a pfdOCT image by a line-by-line Fourier transform. These B-mode images were summed across the transverse axis to generate single depth plots (A-scan) like those presented in Fig. 4 . Several important histological features can be identified in the depth scans and co-registered with the corresponding histopathology images. Figure 4 indicates the location of a keratinized layer (green arrow), the basal layer of the epithelium (red arrow), and the underlying lamina propria (blue arrow) in the micrographs of fixed and stained histological sections from untreated and treated tissue samples. Scattering peaks corresponding to the same tissue layers were identified in each depth scan. To correlate the distances in the histology images with distances in the depth scans, the index of refraction of the tissue was taken into account. An average refractive index for the tissue of $n=1.38$ was used to convert depth scan distances to optical path lengths.^{15, 16} Variation of the refractive index within the tissue is a potential limitation of the current method and is discussed further in the following.

Fig. 4

Histopathology image and corresponding depth plot for (a) untreated and (b) treated epithelium. Arrows indicate keratinized layer (green), basal layer (red), and lamina propria (blue). (Color online only.)

For each sample, a $15\text{-}\mu \mathrm{m}$ -depth segment corresponding to the location of the basal layer was selected from the depth scan and used to guide analysis of the depth-resolved spectroscopic plot, as shown in Fig. 5a . The spectra from the depth identified with the basal layer in each A-scan were averaged to generate a single spectrum for light scattered by the basal layer. As shown in Fig. 5b, a power law curve of the form $y=b\cdot x^{\alpha }$ was initially fit to each spectrum, modeling the spectral dependence resulting from the fractal structure of cellular organelles,^{17, 18, 19} including heterogeneity of the substructure of the nucleus. The residual of each spectrum was calculated by subtracting the power-law curve from the experimental spectrum to produce a normalized spectrum that isolates the oscillatory features, as shown in Fig. 5c.

Fig. 5

(a) Depth-resolved spectroscopic plot with basal layer indicated by dashed red box. (b) Spectrum from basal tissue layer along with power-law fit. (c) Residual spectrum from basal tissue layer. (d) Correlation plot generated by Fourier transforming spectrum in (c). Peak correlation distance can be related directly to scatterer size. (Color online only.)

The normalized spectra showed clear oscillations resulting from interference produced by scattering from the front and back surfaces of basal cell nuclei. Each normalized spectrum was Fourier transformed to generate a correlation plot similar to that shown in Fig. 5d, which shows a clear peak corresponding to the dominant frequency in the normalized spectrum. Peak detection was carried out by an automated, custom MATLAB program (Mathworks, Natick, Massachusetts). The script first high-pass filtered the spectrum with a cutoff of four cycles in order to remove any low-frequency content not removed by the power-law fit. The location of the peak in the correlation plot was then automatically detected by the MATLAB script and related to scatterer diameter with the simple equation $d=\text{correlation}\text{distance}∕(2\cdot n)$ , where $n$ is the refractive index, and $d$ is the diameter of the cell nuclei. An nuclear index of refraction of $n=1.395$ was assumed.⁹