2.1.

Endoscopic Spectral Domain OCT

The imaging system is an endoscopic, spectral domain OCT system built for in vivo mouse colon imaging (shown in Fig. 1 ). The light source is a superluminescent diode (Superlum Broadlighter, Moscow, Russia) centered at $890\mathrm{nm}$ with a $150\text{-}\mathrm{nm}$ full-width-at-half-maximum (FWHM) bandwidth. On the basis of this bandwidth, the theoretical point spread function in air is $\sim 2.3\mu \mathrm{m}$ and in tissue $1.7\mu \mathrm{m}$ assuming a refractive index of 1.4. The source optical power is split by a 50:50 coupler of Corning HI780 fiber (AC Photonics, Santa Clara California). The reference arm is attenuated using an neutral density filter with 0.6 optical density, and SF2 right-angle prisms (Thorlabs, Newton, New Jersey) are used for dispersion compensation.

Fig. 1

Imaging system is an endoscopic, spectral domain OCT system. The light source is a superluminescent diode ( $890\mathrm{nm}$ , $150\mathrm{nm}$ FWHM) and is coupled into a fiber-based Michelson interferometer. The detector is a custom built spectrometer that images light from $800\text{to}1000\mathrm{nm}$ .

The sample arm consists of a $2\text{-}\mathrm{mm}$ -diam endoscope. The endoscope fiber length is $\sim 2\mathrm{m}$ with focusing optics at the tip. The optics consist of a $1\text{-}\mathrm{mm}$ -diam, $1.435\text{-}\mathrm{mm}$ -long glass spacer, angled at $8\mathrm{deg}$ to limit backreflections, followed by a $1\text{-}\mathrm{mm}$ -diam, $1.906\text{-}\mathrm{mm}$ -long GRIN lens (GRINTECH GmbH, Jena, Germany). The focusing optics produce a numerical aperture of 0.1 ( $1∕{\mathrm{e}}^2$ point), resulting in a FWHM spot diameter of $5\mu \mathrm{m}$ at focus. A side-firing endoscope is accomplished using a prism with aluminized face oriented at $48\mathrm{deg}$ relative to the incident light. The tip optics are enclosed in a $2\text{-}\mathrm{mm}$ -diam glass tube. Light exits the tube $3\mathrm{deg}$ from the normal to limit light coupling from the tube interfaces.

The detector consists of a custom-built CCD-based spectrometer. Light exits the detection fiber and is collimated by an off-axis parabolic mirror (Edmund Optics, Barrington, New Jersey). Collimated light is dispersed by a volume phase holographic grating with 1200 line pairs per millimeter, $830\mathrm{nm}$ blaze (Wasatch Photonics, Logan, Utah). Light is incident at $31\mathrm{deg}$ and exits at $26–43\mathrm{deg}$ ( $800–1000\text{-}\mathrm{nm}$ wavelength). A custom focusing optic, Cooke triplet design using off-the-shelf lenses, focuses the $17\mathrm{deg}$ field of view onto a $2048\text{-pixel}$ silicon CCD array (Atmel AT71SM2CL2014-BAO, San Jose, California). Data are collected using a frame grabber, Meteor2-CL/32 from Matrox Imaging (Dorval, Quebec). Roll-off inherent to Fourier domain systems is affected by pixel size as well as optical diffraction and aberrations introduced by the grating and imaging optics. In practice with this detector, the signal drops by $\sim 6\mathrm{dB}$ at $0.85\text{-}\mathrm{mm}$ imaging depth in air.

The system axial point-spread function is affected by the source bandwidth imaged at the detector as well as system and media dispersion. The detector images over $200\mathrm{nm}$ of the source bandwidth, retaining the theoretical $2.3\mu \mathrm{m}$ . System dispersion is corrected in system calibration, similar to Ref. 30, and is summarized as follows. Images are collected from a single scatterer at two different depths. A minimum of 1000 A-scans are collected from each depth and averaged to approximate the background signal. Because of phase instabilities, the average signal does not retain phase information about the scatterers, only the background signal. The background is subtracted from each signal and then the Hilbert transform is performed to calculate the phase. In the absence of system dispersion, the phase of each scatterer should be linear with respect to wavenumber; however, system dispersion introduces a nonlinear phase component at each wavenumber so that the resultant phase is nonlinear. Because the system dispersion phase component is common to both scatterer signals, the subtraction of these two phases eliminates the system dispersion phase component. This resultant phase should now be linear with respect to wavenumber. Linear interpolation is performed to approximate even sampling with respect to wavenumber. After resampling the phase, the system dispersion phase component can be measured by subtracting the linear portion of the phase. The known system dispersion phase component can now be removed for all subsequent measurements by multiplying the measured signal by $\exp (-{\mathrm{i}}^{*}{\text{phase}}_{\text{system}})$ at each pixel. The resultant axial point spread function in practice is under $3\mu \mathrm{m}$ in air.

Sensitivity and dynamic range were measured using the method by Tumlinson ³¹ for endoscopic systems. The sensitivity is $88\mathrm{dB}$ . The dynamic range was measured for multiple OCT images of mouse colon and found to vary between 44 and $68\mathrm{dB}$ .

2.2.

Gold Nanoshell Fabrication and Functionalization

Highly scattering gold nanoshells were produced as described previously.²⁸ The nanoshells were $\sim 120\mathrm{nm}$ diam with a resonant peak near $800\mathrm{nm}$ . Nanoshells were coated in polyethyleneglycol and functionalized with epidermal growth factor receptor antibody. Both functionalized (EGFR-targeted nanoshells) and unfunctionalized (pegylated nanoshells) were imaged in mouse colon.

2.3.

Imaging Protocol

A total of 11 A/J mice were imaged in this study. Of the 11, nine were treated with AOM to induce colorectal cancer. AOM $(10\mathrm{mg}∕\mathrm{kg})$ was administered once a week for five weeks, starting at age seven weeks. The other two mice were treated with saline of equivalent volume with the same schedule (control).

Twelve to $24\mathrm{h}$ prior to imaging, mice were separated without food and given Pedialyte in place of water. Immediately prior to imaging, mice were anesthetized with 2.5% Avertin delivered intraperitoneally. The distal colon was gently flushed with $3–10\mathrm{mL}$ of warm phosphate buffered saline (PBS). The endoscope was lubricated with a water-based lubricant and inserted $\sim 35\mathrm{mm}$ inside the colon. For all mice, OCT images were collected in eight positions, $30\text{-}\mathrm{mm}$ long and spaced $45\mathrm{deg}$ apart, without any contrast agent. Subsequently, highly scattering gold nanoshells targeted to EGFR fabricated at Rice University, $0.2\mathrm{mL}$ with $1\times {10}^{10}\text{nanoshells}∕\mathrm{mL}$ ( $2\times {10}^9$ nanoshells), were intraluminally applied in the colons of five AOM-treated mice and one control mouse. After $20\min$ , the colon was flushed with $10\mathrm{mL}$ PBS. Untargeted, pegylated gold nanoshells (Rice University) were similarly applied in the colons of three AOM-treated mice and one control mouse. Protocols were approved by the University of Arizona Institutional Animal Care and Use Committee.

Eight AOM mice (AOM 1–8) and two control mice (control 1–2) were used to test the quantitative metrics for distinguishing tissue types and detecting gold nanoshells. An additional AOM mouse (AOM 9) was used to demonstrate the ability of the algorithm to detect disease in a three-dimensional data set. The data set was acquired by taking 120 OCT images, $20\text{-}\mathrm{mm}$ long, spaced $3\mathrm{deg}$ apart in the colon. Imaging procedures per each animal are outlined in Table 1 .

Table 1

Animal information.

2.4.

Histology and Immunohistochemistry

After imaging, mice were euthanized and $30\mathrm{mm}$ of colon was excised. Colons were fixed with Poly/LEM Fixative (Polysciences) for $12\mathrm{h}$ , processed, embedded routinely in paraffin, and sectioned $6\text{-}\mu \mathrm{m}$ thick. Sectioned tissue was stained with hematoxylin and eosin (H&E) or EGFR immunostain as described below.

After antigen retrieval, sections were washed, blocked with goat serum, and incubated overnight with primary rabbit polyclonal anti-EGFR (1005):sc-03 (Santa Cruz Biotechnology). Endogeneous peroxidase activity was blocked by immersion in 3% hydrogen peroxide in water followed by a $1\text{-}\mathrm{h}$ incubation at room temperature with Biotinylated Anti-Rabbit IgG $(\mathrm{H}+\mathrm{L})$ . Sections were incubated for $10\min$ at room temperature with streptavidin-HRP (Dako). Complexes were visualized (brown) with $3,3^{\prime }$ -diaminobenzidine (Dako), and the sections were counterstained with 0.5% methyl green.

Disease was positively identified using H&E slides and correlated with disease features in OCT as described by Hariri ¹⁴ Diseased areas identified in the H&E slides were adenomas and GIN; all other areas were classified as normal. EGFR immunostained slides were used to quantify EGFR expression in adenomas, GINs, and selected normal regions. Each region was given a grade between 0 and 4, judged by eye, representing the average color hue and saturation. The darker the brown was in the slides, the higher the EGFR grade was, as shown in Fig. 2 .

Fig. 2

Normal colon sections with EGFR immunostaining are shown, ranging from less (grade 0) to more (grade 4) EGFR. Note the increasing brown color saturation from grades 0 to 4. (In gray-scale print, increasing brown saturation appears as a darkening around the crypts and apical cell layer.) (Color online only.)

2.5.

Image Processing and Quantitative Metrics

OCT image properties associated with normal, GIN, and adenoma were previously determined in a study of 18 AOM-treated mice.¹⁴ These imaging features are listed in Table 2 as well as illustrated in Fig. 3 .

Fig. 3

OCT imaging of mouse colon clearly distinguished normal from diseased tissue. (a) A $22\times 1.5\mathrm{mm}$ logged OCT image of AOM-treated mouse colon. (b) The same OCT image after median filtering with a large ( $21\times 21\text{pixel}$ , $125\times 25\mu \mathrm{m}$ ) kernel. (c) Plots corresponding to the logged backscattered intensity as a function of depth. The position and length of the plots are indicated by the white lines in the OCT images labeled “Normal,” “Adenoma,” and “GIN.” Anatomical features are indicated on the Normal and GIN plots. In both plots, the mucosal (M) surface and mucosa (M)/submucosa (SM) boundary are visible. The Normal plot also clearly shows the boundaries between the submucosa (SM), muscularis propria (MP), and adventitia. Because of mucosal thickening and heightened attenuation with depth, these boundaries are not visible in the GIN. The adenoma displays no clearly discernable boundaries. Note also the marked mucosal thickening between the normal and GIN tissues.

Table 2

Imaging properties of normal, GIN, and adenoma.

To differentiate tissue based on the image features described in Table 2, specifically boundary visibility, attenuation in the mucosa, and mucosal thickening, three image properties were measured: (i) Average gray-scale value in the mucosa after logarithmic transformation, (ii) slope of a line fitted to log-transformed gray-scale values in the mucosa, and (iii) slope standard deviation from the tissue surface to $\sim 360\text{-}\mu \mathrm{m}$ deep, which encompasses the average depth of normal mouse colon to the muscularis propria/adventitia boundary. These three metrics can be described as zero-, first-, and second-order properties, respectively. The gray-scale value and slope in the mucosa quantify the amount of attenuation in the mucosa, while the slope standard deviation is affected by boundary visibility. Mucosal thickening decreases boundary visibility in at least two ways; first, it pushes the boundaries further from the system focus and, second, it increases the number of scatterers ahead of the boundaries. In effect, slope standard deviation is affected by boundary visibility, which is affected by mucosal thickening.

Sections 2.5.1, 2.5.2, 2.5.3 describe the specific steps in the automated disease detection algorithm, including image preprocessing, calculating the zero-, first-, and second-order image metrics, and assessing the efficacy of each metric for identifying disease.

3.1.

Distinguishing Adenoma, GIN, and Normal Regions

Correlating OCT and histology revealed 140 adenomas, 32 GINs, and 269 normal segments. One saline-treated mouse, control 2, was not classified as normal due to being diagnosed with diffuse mucosal edema. Using the surface depth and segment size criteria, 5 adenomas, 4 GINs, and 42 normal segments were excluded. The included tissue segments totaled 135 adenomas, 28 GINs, and 227 normal segments.

The best metric for distinguishing adenomas, GINs, and normals was the slope standard deviation metric. As predicted, this metric was highest for normals and decreases for GINs and adenomas, and it was statistically significantly different between the three tissue types with $p<0.0002$ in all cases. At one threshold value, this metric was 89% sensitive and 70% specific for identifying adenomas and GINs from normals. At a lower threshold value, this metric distinguished adenomas from GINs with 70% sensitivity and 76% specificity.

Average gray scale and slope were less predictive. Average gray scale decreased as predicted from normals to GINs to adenomas. This difference was statistically significant for differentiating adenomas from normals and adenomas from GINs but not for differentiating normals from GINs. The difference in slope was statistically significant between normals and GINs, with GINs have steeper slopes than normals. The slope for adenomas was steeper than the slope for normals, as predicted, but less steep than the slope for GINs, contrary to prediction. The metrics are shown graphically for each tissue type in Fig. 4 .

Fig. 4

Metric values (on an arbitrary scale) for three different tissue types: Normals, GINs, and Adenomas. From Normal to GIN to Adenoma, the gray-scale and slope standard deviation decrease, as anticipated. From Normal to GIN and from Normal to Adenoma, the slope becomes steeper, as anticipated, but from GIN to Adenoma, the slope becomes less negative.

3.2.

Utility for 3D Imaging

The slope standard deviation metric was calculated for a three-dimensional data set from one AOM-treated mouse, shown in Fig. 5 . This metric was mapped to a gray-scale value, where dark gray (blue online) indicated the highest slope standard deviation or most “normal” and white (pink online) indicated the lowest slope standard deviation or most “diseased.” A-scans with the colon surface deeper than $0.2\mathrm{mm}$ , which could not be classified, were mapped to black. The image is overlaid with histologically verified features, specifically adenoma (Ad), GIN (G), and lymphoid aggregate (LA). LAs are normal immune tissue in the colon. As the image suggests, LAs can be confounders for adenomas and GINs.

Fig. 5

(a) Two-dimensional representation (“disease map”) of the mouse colon, $20\mathrm{mm}$ long in which left is rostral and right is caudal, and in which the top and bottom wrap around to complete the cylindrically shaped colon $\sim 6\mathrm{mm}$ in circumference. The metric plotted is the slope standard deviation, ranging from 0 (diseased) to 0.1 (normal) arbitrary units. Locations of adenomas (Ad), GINs (G), and LA, verified both in OCT and histology, are labeled. Note that the white (pink online) areas, which indicate a low slope standard deviation or “disease,” correspond with the locations of adenomas, GINs, and lymphoid aggregates. Although lymphoid aggregates are considered confounders for disease in OCT, note the colocalization of lymphoid aggregates and adenomas, which is typical in the AOM-treated mouse model. Dark gray (blue online) areas indicate high slope standard deviation or “normal.” Black areas indicate that the colon surface fell below the minimum depth, $0.2\mathrm{mm}$ , from the endoscope window and could not be classified. The logged OCT images ( $20\mathrm{mm}$ long by $0.3\mathrm{mm}$ deep) at the bottom were taken at locations indicated by the dotted lines in the “disease map” at the top. (Color online only.)

3.3.

Identifying Nanoshells

Data from EGFR-targeted nanoshells and untargeted pegylated nanoshells were first combined in order to show sensitivity to nanoshells overall regardless of specificity to EGFR. For all tissue types, average gray scale increases as predicted with $p<0.05$ between pre- and postnanoshell images. Slope and slope standard deviation, however, did not have a consistent response to nanoshells for all three tissue types.

3.4.

Visualizing EGFR with Targeted Nanoshells

The data were graded using EGFR immunostained slides and each tissue type was subdivided into low, medium, and high grade. The number of GINs after subdividing into grades was too small to analyze; therefore, only normals and adenomas were considered for this analysis. To grade normals, normal sections were correlated with histology based on surrounding features, including adenomas, GINs, and lymphoid aggregates. Some normal sections were excluded from analysis for lack of an identifiable landmark.

Figure 6 shows the percent change in the metrics after adding nanoshells for each EGFR grade, where the percent change was calculated as the value postnanoshells $\left(x_{\mathrm{post}}\right)$ minus the value prenanoshells $\left(x_{\mathrm{pre}}\right)$ divided by the value prenanoshells times 100%, $(x_{\mathrm{post}}-x_{\mathrm{pre}})∕x_{\mathrm{pre}}{}^{*}100\mathrm{\%}$ . Figure 7 shows representative OCT images before and after adding nanoshells. For both normals and adenomas, the average gray scale increases postnanoshells and the percent change in gray scale increases with EGFR grade, suggesting higher nanoshell density with higher EGFR expression. The percent change in slope and slope standard deviation postnanoshells decreases with EGFR grade in normals, which suggests that other image effects may be present besides increased attenuation. The percent change in slope and slope standard deviation for adenomas showed no trend with EGFR grade.

Fig. 6

Charts show the percent change in image metrics post-EGFR-targeted nanoshells. The average gray-scale metric is plotted using the left axis, with values indicated. The slope and slope standard deviation metrics are plotted using the right axis. For the normal segments, gray scale increases and slope becomes less negative with EGFR-grade. The slope standard deviation decreases with nanoshells, as anticipated, but the percent change decreases with EGFR-grade. For adenomas, grayscale increases with EGFR-grade, while slope and slope standard deviation do not correlate with EGFR-grade. Note that for both normal segments and adenomas, the percent change in average gray scale increases with EGFR-grade postnanoshells.

Fig. 7

Logged OCT images ( $30\mathrm{mm}$ long by $0.3\mathrm{mm}$ deep) are in the same colon location in the same mouse (a) pre- and (b) post nanoshells. Note the increase in surface signal in the postnanoshell image at the positions indicated by the arrows.

Image metrics for untargeted pegylated nanoshells do not show a consistent response for all three tissue types nor show a trend with EGFR grade for any of the metrics, suggesting nonspecific adherence to tissue.