1.1.

Collagen and Second Harmonic Generation

Tumor cells grow in an uncontrolled manner, in part because of biochemical and mechanical changes in their microenvironment, the extracellular matrix. Collagen, a major structural component in the extracellular matrix, has been shown to have a pivotal role in cancer development.⁵ Studies have shown that structural changes happen in the collagen surrounding a tumor that can be difficult to detect with traditional histology.⁶ Additional methods for studying the structural changes in collagen associated with cancer development can provide insight into cancer diagnosis and progression status.

Second harmonic generation (SHG) can be used to quantitatively study collagen’s structural changes. SHG is a multiphoton optical phenomenon that certain nonlinear materials, including collagen, can produce through a scattering mechanism.^7,8 When two photons of the same frequency simultaneously interact within these materials, one photon with double the energy and half the wavelength of the original photons may be produced. The intensity of the SHG signal that collagen can produce is dependent on the nonlinear optical susceptibility of the collagen.⁹ Nonlinear optical susceptibility depends on the thickness of fiber thickness, fiber density, and fibril “packing,” which is the three-dimensional structure of the fibrils that make up collagen fibers. Many studies have used backscattered SHG and other backscattering techniques as quantitative indicators of extracellular matrix alterations in various cancer types, including CRC.^10–14 Researchers have also shown that an SHG directionality comparison between the forward- and backward-scattered signal can be utilized ex vivo as an independent prognostic indicator for metastasis.^15,16

There are two main advantages to studying collagen with SHG. The primary advantage is that it can serve as a label-free imaging technique that visually isolates collagen from its surroundings because collagen is the dominant material in tissue that produces a significant SHG signal.^7,8 The secondary advantage is that to produce a strong SHG signal, it is necessary to confine the excitation light to a small focal volume to increase the probability of a two-photon effect taking place. Given this selective excitation only within this small focal volume, three-dimensional information about the collagen structure can be easily obtained by gathering intensity data from each focal volume and combining these individual voxels (volumetric pixels) to construct a three-dimensional image.^7,8

1.2.

Multiphoton Microscopy

The data for an SHG image is traditionally obtained ex vivo with a multiphoton microscope.¹⁷ The key components of a multiphoton microscope are a femtosecond laser, a high numerical aperture (NA) focusing lens or objective, and scanning components that move the laser beam, and thus, the focal volume, across the field of view. Lateral scanning usually is accomplished with galvanometer-based mirrors that rotate to move the beam across the field of view.¹⁸ Alternative scanning methods may maximize acquisition speed such as resonant galvanometer systems,¹⁹ hexagonal mirror scanning,²⁰ and microelectromechanical systems (MEMS) mirrors.²¹

1.3.

Multiphoton Endoscopy

Our group and others have previously obtained multiphoton images through an endoscopic architecture.^22,23 Because scanning mirrors and MEMS devices are challenging to fit in the distal assembly of very small diameter endoscopes, we used a tube piezoelectric device with a cantilevered fiber. However, miniature two-dimensional scanning mechanisms are generally expensive and complex and may create a scan pattern that deviates from the standard rectilinear grid. There is therefore a need for a simpler way to obtain SHG intensity data endoscopically. This study investigates to what extent scanning is necessary to obtain meaningful SHG intensity data for the purpose of colorectal cancer diagnosis. We simulate random one-dimensional line-scanning and random point-sampling techniques (Table 1). As opposed to traditional two-dimensional synchronized scanning, these techniques can be implemented with one-dimensional synchronized scanning (e.g., spectro-temporal encoded methods²⁴) or non-synchronized manual point sampling (e.g., endoscopist’s random hand movements), respectively. We hypothesize that non-imaging, randomly sampled SHG intensity measurements are sufficient for differentiating between tumor and normal tissue.

Table 1

Simulated image types, corresponding data localization type, and potential scanning hardware.

2.1.

Sample Preparation

Surgically resected colon tissue samples from 10 colon cancer subjects were obtained under a University of Arizona Institutional Review Board-approved protocol. Patients gave informed consent. Three different samples were obtained from each resection: tumor, tumor-adjacent, and normal. Tumor samples were obtained from the bulk of the tumor. Tumor-adjacent samples were taken from the first normal-appearing (to the physician) tissue adjacent to the tumor. Finally, normal tissue was obtained from the edge of the resection where the physician assumed a clear margin, typically 1 cm or more away from the tumor. The samples were fixed in 10% formalin, embedded in paraffin, cut into $6\text{-}\mu \mathrm{m}$ thick sections using a microtome, and placed on glass slides. Serial dilutions of ethanol were used for deparaffinization, and the slides were stored in water until imaging.

Additional adjacent sections were stained with hematoxylin and eosin, and a diagnosis was provided by a pathologist. Pathology diagnosis of the actual imaged tissue sample could differ from the overall patient diagnosis.

2.2.

Imaging

Images were obtained with a Zeiss LSM 880 multiphoton microscope with linearly polarized 850 nm excitation from a tunable Spectra-Physics Mai Tai DeepSee laser (tunable range 690 to 1040 nm). The objective was a 30X/0.8NA Plan-Apochromat (Zeiss #440640-9903-000), located in a backward (epireflectance) orientation. Emission from 400 to 430 nm was collected using a bandpass filter embedded in the microscope. The image size was $1024\times 1024\text{pixels}$ over $425\times 425\mu \mathrm{m}$. Power was kept constant. Because slides were only $6\text{-}\mu \mathrm{m}$ thick and axial resolution is limited to about 1 to $2\mu \mathrm{m}$, only one image plane was obtained, resulting in two-dimensional data.

One region per slide was randomly selected by the microscope user with the selection criteria being (1) the image contained mucosa only and (2) no image artifacts were present from sectioning. The mucosa is the most luminal layer of the colon and is thus considered to be comparable with the tissue that would be imaged by an endoscope in vivo.

2.3.

Data Analysis

Once images were obtained, the data from each image were processed in MATLAB. As shown in Fig. 1, low and high thresholds were applied that discarded pixel values that were $<4$ (background noise and autofluorescence leakage into the SHG channel) and $>254$ (saturated pixels). The low (background/autofluorescence) threshold was determined by measuring the average pixel value of a section of the image that contained a faint cytoplasmic signal (assumed to be autofluorescence) but no fibrillar structure (assumed to be SHG from collagen). The low threshold was relatively consistent for each image, with a mean of $3.82\pm 0.25$ (SEM), a minimum of 1, and a maximum of 6. The overall average threshold (4) was applied to every image. Saturated pixels occurred occasionally generally due to contamination/dust on the slide, but this incidence was rare. After applying the thresholds, the mean SHG intensity of each image was calculated.

Fig. 1

Representation of colon mucosa analysis algorithm with enhanced brightness and contrast. Starting with the raw image (a), the image was thresholded to exclude low- and high-value pixels (b), point-sampled (c), and line-sampled (d). Panel (c), 200,000 pixels, and panel (d), 400 lines, were chosen for easy visualization of the algorithm. Different sample sizes (10, 100, 1000, and 10,000) were implemented in actual data processing. Image size $425\times 425\mu \mathrm{m}$.

To simulate line-scanning, an algorithm was applied that selected random columns from the two-dimensional images. To simulate point scanning, a random pixel sampling algorithm was applied.

For the lowest sample size of 10, an array of 10 random numbers with values between 1 and the number of columns in the image was created with the “randi” function in MATLAB (line scanning), or 10 random numbers with values between 1 and the number of suprathreshold pixels in the image was generated (point scanning). These arrays contained the indices of the selected lines or pixels for analysis. The mean SHG intensity of the randomly selected lines or pixels in each image was then calculated. This process was repeated logarithmically increasing sample sizes, including 10, 100, 1000, and 10,000 lines or pixels. This process was repeated ten times to estimate variation in sampling. Figure 1 shows an example, with the original image, the thresholded image, and simulated point and line sampling. The images are false-color mapped (rainbow, black, and blue representing low grayscale intensity, red representing high).

3.1.

Average Second Harmonic Generation Signal Shows Significant Differences Between Tissue Types

Figure 2 shows that the overall mean grayscale value of SHG images from each tissue type (normal, tumor-adjacent, and tumor) was significantly different from each other tissue type. The significance levels shown in Fig. 2 are based on a paired $t$-test; paired and unpaired $t$-test results are summarized in Table 2.

Fig. 2

Mean SHG pixel value in images of normal, tumor-adjacent, and tumor tissue. Using a paired $t$-test, $p\text{-}\text{value}<0.05=*$, $p\text{-}\text{value}<0.01=**$.

Table 2

Summary of paired and unpaired significance, with p<0.05=* and p<0.01=**.

Images of normal tissue had the highest mean SHG image grayscale value, followed by tumor-adjacent and tumor tissue.

The significance levels shown in Table 2 suggest that the average grayscale value of SHG images is different in mucosal collagen images obtained in the three different tissue types, with an emphasis on the high level of significance between normal and tumor tissue ($p<0.01$ for both paired and unpaired $t$-tests).

3.2.

Qualitative Image Comparison Matches Quantitative Trends of Second Harmonic Generation Intensity

Figure 3 is an example of normal, tumor-adjacent, and tumor tissue. These images show the overall decreasing grayscale value trend of the quantitative data shown in Fig. 2. Note that, in addition to the lower density of the collagen structures in the tumor, there is also an average decreasing intensity of the signal from collagen fibers progressing from the normal tissue to tumor tissue.

Fig. 3

SHG images of normal (a), tumor-adjacent (b), and tumor (c) mucosal structures in the same subject. These images have been thresholded and have the same brightness and contrast settings to allow for a true intensity comparison.

3.3.

Random Point and Random Line Sampling Shows Significance for Differentiating Tissue Types

Analysis of the randomly sampled images reveals that the paired $p$-value for comparison of grayscale value among all three tissue types shows significance and generally resembles the entire image, after sampling only 1000 pixels [Fig. 4(a)] or 100 lines [Fig. 4(b)]. As expected, line sampling performs better than point sampling for 10 and 100 lines/points, because there were tens to hundreds of analyzable points per line. These results suggest that non-imaging, randomly sampled SHG intensity measurements are likely to be sufficient for differentiating between normal, tumor, and tumor-adjacent tissue in the colon.

Fig. 4

$p$-value improves as more (a) points and (b) lines are sampled in the thresholded image for comparisons between adjacent/tumor (A/T), normal/adjacent (N/A), and normal/tumor (N/T) groups, for both paired (solid line) and unpaired (dashed line) $t$-tests. (a) Significance levels of $p<0.05$ are achieved for groups N/A (paired and unpaired) and A/T (paired) for sampling levels above 1000 points. Significance levels of $p<0.01$ are achieved for group N/T (paired and unpaired) for sampling levels above 1000 points. (b) Significance levels of $p<0.05$ are achieved for groups N/A (paired and unpaired) and A/T (paired) for sampling levels above 100 lines. The significance level of $p<0.01$ is achieved for group N/T (paired and unpaired) above 100 lines.

4.1.

Future Directions

Overall, our study suggests that two-dimensional scanning to form an SHG image of collagen may not be necessary for diagnosing colorectal cancer. If these results are also found in vivo in intact tissue, there may be potential for the development of a simplified non-imaging SHG system. Removing the need for scanning with an endoscope greatly simplifies design and reduces cost. This study is the first step toward building a non-imaging endoscope that may provide additional information and help improve diagnostic accuracy during or after a colonoscopy.