2.1.

Instrumentation

The experiments were conducted with a clinically compatible endoscopic a/LCI system (Fig. 1) that has been previously introduced.²² Fiber-optic coupled light generated by an 830-nm superluminescent diode (Superlum Diodes, Moscow, Russia, Δλ_FWHM = 19 nm) passes through an optical isolator (AC Photonics, California) and is split into reference and sample arms at a 5:95 ratio by a fiber splitter (AC Photonics, California). A polarization controller (Thorlabs, New Jersey) is used in the sample arm in order to maximize the power coupled by a fiber polarizer (Chiral Photonics, Inc., New Jersey) into a polarization-maintaining delivery fiber (Corning, Inc., New York); the use of polarized light in the sample arm of a/LCI systems has been previously shown to improve their ability to size particles accurately.²³ The polarized light is transmitted through the fiber probe assembly to the probe tip [Fig. 1b], where it is delivered to the sample as a collimated light beam (0.5-mm diameter) at an oblique angle by a drum lens (diameter: 1.5 mm, length: 2.5 mm; Edmund Optics, New Jersey). The delivered light passes through a protective window that is angled 8 deg out-of-plane in order to prevent reflected light from being collected. The angular distribution of the scattered light is imaged across the face of a coherent imaging fiber bundle (Schott NA, New York) placed in the Fourier plane at the back focal plane of the drum lens. By using a 4f system comprised of lenses L1 and L2 to image the proximal face of the fiber bundle onto the input slit of an imaging spectrometer (Spectrapro 2150i; Acton Research Corporation, Massachusetts), the spectral profile of the angular distribution of scattered light collected across a strip through the center of the fiber bundle is collected by a CCD (Pixis:100; Princeton Instruments, New Jersey). Reference light that has passed through a pathlength-matched single-mode optical fiber and been collimated by L3 is mixed with the sample beam at a beam splitter in order to produce spectral interference fringes whose frequency is a function of pathlength mismatch between the two arms. These interference fringes are processed in a manner similar to Fourier-domain OCT to produce individual depth scans for each scattered angle in a procedure previously described by Brown ²¹

Fig. 1

a/LCI system diagram. (a) System illustration, (b) probe tip assembly illustration showing delivered (dark gray) and scattered (light gray) light, and (c) photograph of the endoscopic probe pictured next to a U.S. dime for scale. Illustration adapted from Zhu (Ref. 22).

2.2.

Study Design

Twenty-seven subjects undergoing partial colonic resection surgery at Duke University Medical Center were enrolled in this study. For each subject, the tissue sample was opened longitudinally and multiple (typically 5 to 6) epithelial locations were evaluated using the a/LCI probe within 2 h of resection for a total of 138 paired biopsy locations. Scan locations were selected on the margins of the gross lesions in an effort to find dysplastic tissue sites. An additional single scan was taken from each tissue sample at a remote location, believed to contain healthy tissue upon gross examination. At each measurement point, the probe was brought into contact with the tissue and 15 to 25 acquisitions of 25 ms each were taken. All scans for each single optical biopsy were collected in under 30 s. Across all patients, a total of 2734 individual scattering spectra were acquired. Following imaging using the a/LCI system, India ink was used to mark the sampled points and physical biopsies were collected for histopathological analysis. The samples were analyzed to determine the disease state by a trained pathologist. This study was conducted with the oversight of the Institutional Review Board at the Duke University Medical Center in Durham, North Carolina.

2.3.

Data Processing

The a/LCI data were processed in a manner described previously in order to determine depth-resolved nuclear morphology characteristics of the epithelial tissue.^{19, 21, 22} Briefly, the data were processed to isolate the scattering components from cell nuclei in three depth segments of the tissue: the segment from the tissue surface to 100 μm in depth, the segment between 100 and 200 μm in depth, and the segment from 200 to 300 μm in depth, identified as the basal layer of the epithelium. Periodic oscillations in the angular spectra, which are characteristic of the underlying nuclear morphology, were compared to theoretical predictions to determine the average nuclear diameter and nuclear density of the tissue in question. In this analysis, nuclear density is measured as the relative ratio of the average refractive index of the cell nuclei and the surrounding cytoplasm. Each scan was examined to ensure a scattering signature of sufficient strength, and those that did not fulfill this requirement were rejected. Morphological characteristics, as measured using a/LCI, were compared to pathologic classification of co-registered tissue samples as determined by a trained pathologist.

Statistical analyses were used to assess the association of dysplasia with the a/LCI measurements of nuclear morphology. For each model, the pathological diagnosis was dichotomized as dysplasia versus no dysplasia. Statistical significance was determined between average nuclear morphological characteristics as measured by a/LCI for dysplastic and nondysplastic biopsies, which are given below as the mean value with the standard deviation noted in parentheses (SD). Normality of these data was verified using the Shapiro–Wilk W-test. For normally-distributed data, the Student's t-test was used to assess statistical significance. For data that were not normally distributed, the Wilcoxon Rank Sum test was used. P values associated with type I error rates of less than 0.05 were considered significant in this study. All analyses were performed using JMP version 9.0 (SAS Institute, Inc., Cary, North Carolina).

The relationship between sensitivity and specificity was determined through the development of a receiver operating characteristic (ROC) for each of the three epithelial depth layers. For these analyses, the presence of dysplasia was considered a binary classifier and nuclear diameter and density were used as discriminants. To evaluate the predictive ability of these measures, an area under the ROC curve (AUC) was calculated as a fraction of the maximum possible AUC.

3.1.

Sample Data

Typical a/LCI data are shown in Fig. 2. Figure 2a shows an angle- and depth-resolved scattering map, which contains the angular distribution of light scattered by the sample as resolved by depth. Light shades of gray indicate higher intensities, which correspond to increased amounts of scattered light. The horizontal axis indicates depth in the sample, with 0 μm corresponding to the interface between the probe and the tissue. A depth-resolved reflection profile is created by summing across the angular range of the scattering map [Fig. 2b]. During signal processing, the angular scattering pattern is analyzed in 50-μm segments, indicated by gray lines. Fig. 2c shows an example angular profile recovered from a single 50-μm depth segment between 200 and 250 μm beneath the epithelial surface of healthy tissue. This angular profile (solid line) is overlaid with the best theoretical fit (dashed line). The corresponding average nuclear size for this depth segment of tissue is indicated.

Fig. 2

Typical a/LCI data. (a) Angle-resolved depth scan of light scattered from tissue. Lighter shades of gray indicate an increased amount of scattered light. (b) A-scan indicating depth increments used for processing. Fifty micrometer segments of the tissue used for processing are indicated. (c) Example angular scan pictured (solid line) with best-fit Mie theory solution (dashed line) and size indicated.

3.2.

Biopsy Results

Of the 27 tissue samples examined with the a/LCI system, histological evaluation showed that 8 were pathologically normal, 7 presented for low- and high-grade dysplasia, 2 showed both dysplasia and adenocarcinoma, 5 displayed only invasive adenocarcinoma, and 5 were characterized as ulcerative diseases. In the presented analysis, only biopsies that were pathologically normal or that had a diagnosis of low- (LGD) or high-grade (HGD) dysplasia were considered. This restriction resulted in a data set that consisted of 14 dysplastic biopsies from adenomatous polyps (both LGD and HGD) and 67 biopsies from healthy tissue sites, with 26.4% of acquired scans omitted due to low signal strength.

3.3.

Statistical Analyses

Nuclear diameter and density were measured at each of three epithelial depth segments. These results are given in Table 1. All data were normally-distributed, with the exception of the nuclear density of pathologically-normal biopsies from 0 to 100 μm in depth and the nuclear diameter of dysplastic biopsies from 100 to 200 μm in depth. Measurements of nuclear diameter and density between the epithelial surface and a depth of 100 μm did not differ in a statistically-significant manner, nor did nuclear density between 100 and 200 μm. A decrease in nuclear diameter between 100 and 200 μm was statistically associated with the presence of dysplasia at the p < 0.05 level. For the depth layer between 200 and 300 μm, both a reduced nuclear density and an increased nuclear diameter were statistically associated with the presence of dysplasia at the p < 0.0001 level.

Table 1

Average nuclear morphology measurements at each of the 100-μm epithelial depth segments for nondysplastic and dysplastic biopsies as measured by a/LCI. Data are given as mean (SD) and statistical significance as indicated with associated p-value. NS indicates that the specified characteristic was unable to differentiate dysplasia at a statistically-significant level (P < 0.05). Single and triple asterisks indicate statistical significance at the P < 0.05 and P < 0.001 level, respectively.

To determine the relationship between sensitivity and specificity, ROC curves were created using each of the six nuclear morphological measurements as discriminants. For these analyses, the presence of dysplasia was considered as a binary classifier. Nuclear diameter and density of tissue from the epithelial depth of 200 to 300 μm demonstrated a good ability to predict the presence of dysplasia, and were individually associated with AUC values of 0.87 and 0.79, respectively. When these measures were linearly combined to form a single discriminant, a ROC with an AUC of 0.91was obtained. The three aforementioned ROC curves for the 200 to 300 μm depth segment appear in Fig. 3. To a lesser extent, average nuclear diameter of tissue located 100 to 200 μm beneath the mucosal surface was effective at discriminating between dysplastic and nondysplastic populations, with an AUC value of 0.68. Neither the average nuclear diameter of tissue between 0 and 100 μm in depth (AUC = 0.54), the average nuclear density of tissue between 0 and 100 μm (AUC = 0.53), nor the average nuclear density of tissue between 100 and 200 μm (AUC = 0.56) were able to effectively detect dysplasia.

Fig. 3

ROC curves corresponding to various discriminators for the third 100-μm segment of the epithelium. AUC (gray area) ratios are indicated. Circles indicate optimal value for each discriminant. (a) Nuclear diameter, AUC = 0.87; (b) nuclear density, AUC = 0.79; (c) nuclear diameter and nuclear density, AUC = 0.91.

A scatter plot of average nuclear diameter and nuclear density for the epithelium from 200 to 300 μm in depth is shown in Fig. 4. Because both nuclear diameter and nuclear density are statistically-significant predictors of the presence of dysplasia, it is possible to draw an optimal decision line (dashed line in Fig. 4) as determined by ROC analysis that efficiently and accurately distinguishes the dysplastic biopsies, which have increased nuclear diameter and decreased nuclear density, from those that are pathologically normal. This optimal discriminator corresponds to the point on the ROC curve that is geometrically closest to the upper left, which represents perfect sensitivity and specificity. By using this decision line, which is nearly identical to that previously determined by Pyhtila,²⁰ the biopsies can be classified with a sensitivity of 92.9% (13/14), a specificity of 83.6% (56/67), and an overall accuracy of 85.2% (69/81). Furthermore, this decision line yields a positive-predictive value of 54.2% (13/24) and a negative-predictive value of 98.2% (56/57).

Fig. 4

Scatter plot showing nuclear size (micrometer) versus nuclear density for the epithelial depth segment 200 to 300 μm beneath the mucosal surface. Each point represents a single optical biopsy, and points are color-coded with respect to their pathological assessment. The dashed black line indicates the ideal decision line for these data for the prediction of dysplasia.