2.1.

Tissue Specimens

All studies were performed in conjunction with University of South Alabama Institutional Review Board protocol: IRB # 445452-3. Colorectal tissue samples were obtained by the University of South Alabama Departments of Surgery and Pathology from colorectal surgical resection specimens. The samples were collected as pairs of adenocarcinoma and normal mucosa and were $\sim 0.5$ to 1.0 mm in thickness. The tissues were flash frozen in liquid nitrogen. Confirmation of adenocarcinoma was determined by histologic evaluation of hematoxylin and eosin (H&E) permanent sections. Samples were maintained at $-80°\mathrm{C}$ until prior to hyperspectral imaging, typically within 1 to 2 days. For imaging, tissues were reconstituted in cold phosphate-buffered saline (PBS) and placed onto a 25-mm round coverslip.

A total of eight patients have been enrolled thus far in the study. However, data from the first three patients are not presented here, due to variations in specimen processing and imaging protocol. In addition, one patient was not considered, as surgical exam revealed extreme fibrosis, possibly due to perforation of the bowel, but no evidence of adenocarcinoma.

Fresh rat tissues were obtained postmortem from an ongoing, unrelated animal study. All procedures in the unrelated animal study were conducted in accordance with approved University of South Alabmama Institional Animal Care and Use Committee (IACUC) protocol #623125.

2.2.

Hyperspectral Microscope

An inverted fluorescence microscope (TE2000-U, Nikon Instruments) with a $10\times$ objective ($10\times /.25$ Ph1 ADL $\infty /1.2$ WD 6.2, Nikon Instruments) was used as the imaging platform. Illumination light was provided by a 300-W Xe arc lamp (TITAN 300ST-K, Sunoptics Technologies). An evaluation prototype thin-film tunable excitation filter system (Semrock, Inc.) containing six separate tunable filters (Versachrome, Semrock, Inc.) was used to filter the excitation light illuminating the sample. Our group has previously described this tunable filter system in detail.^55,61,62 Output from the tunable filter was coupled to the microscope through liquid light guide and supplied in either an epifluorescence configuration or a transmitted light configuration. An electron-multiplied charge-coupled device camera ($\text{Rolera}\text{-}\mathrm{em}\text{-}{\mathrm{c}}^2$, Q-Imaging) was used to acquire images of the illuminated tissues. For hyperspectral imaging, fluorescence excitation was scanned using four different wavelength ranges with corresponding long-pass dichroic beamsplitters at 458, 495, 555, and 594 nm cut-off wavelengths (Table 1). The optical path for the microscope in both fluorescence and reflectance configurations is shown in Fig. 1.

Table 1

Equipment settings for each type of hyperspectral imaging scan performed. Different dichroic beamsplitters were used to allow a range of fluorescence excitation scanning, with cut-off wavelengths at 458, 495, 555, and 594 nm cut-off wavelengths (part numbers: BLP01-458R-25, FF01-495/LP-25, BLP02-561R-25, and BLP01-594R-25, respectively, Semrock, Inc.). For transmission and absorbance scanning, no dichroic beamsplitter was used.

Fig. 1

Lightpaths for optical microscope setup: (a) fluorescence excitation-scanning mode and (b) transmission scanning mode. Note that when operating in fluorescence excitation-scanning mode, all fluorescence emission was detected using a long-pass emission filter, for each fluorescence excitation band used.

2.3.

Image Acquisition

Fields of view within tissue specimens were identified for imaging using fluorescence with 480-nm excitation and 495-nm long-pass emission and transmission at 515 nm. Multiple fields of view were acquired to sample the range of structures in each specimen. An additional field of view was acquired that included the tissue edge and surrounding background region (with no tissue). The background spectrum was used during spectral correction (described below). All fluorescence hyperspectral images were acquired with the following detector settings: electron multiplying (EM) gain of 3800, 14-bit dynamic range, and binning of $2\times 2$. Transmission images were acquired using an EM gain of 1. Image acquisition for each type of fluorescence and transmission spectral image stack was performed sequentially.

2.4.

Image Processing and Analysis

Spectral images were postprocessed using MATLAB software (MathWorks). Absorbance data were calculated from transmission values, using the transmission spectrum from a background region as reference. ENVI software (Exelis Visual Information Solutions, Boulder, Colorado) was used to identify regions of interest (ROIs) and extract spectral data. Fluorescence images were corrected to flat spectral response using background subtraction and multiplication by a correction factor. Correction coefficients were determined using a NIST-traceable lamp (LS-1-CAL-INT, Ocean Optics, Inc.) and fiber-coupled spectrometer (QE65000, Ocean Optics, Inc.), as described previously.^41,55,61

Multiple ROIs were identified according to the structural features of each specimen. In addition, a region encompassing the entire field of view was selected. The average spectrum from each region was extracted. Extracted spectra were plotted using Excel (Office 2010, Microsoft Corporation).

2.5.

Results and Discussion

Colonic adenocarcinoma and surrounding normal tissue were imaged using a hyperspectral imaging fluorescence excitation (HIFEX)-scanning microscope configuration. An example image of normal mucosa is shown in Fig. 2. For each pixel in the image, fluorescence excitation, transmission, and absorbance spectra were acquired [Figs. 2(a)–2(c)]. These images can be visualized as three-dimensional spectral image stacks, in which the first two dimensions represent spatial data and the third dimension represents spectral data. ROIs were selected based on structural/anatomical features in the image. For each ROI, the pixel-averaged spectrum was calculated for each hyperspectral imaging mode [Figs. 2(d)–2(h)]. In cases in which a desired structure of the specimen (e.g., mucosa) spanned several fields of view, the spectra from multiple ROIs were averaged to obtain a representative spectral signature for the structure.

Fig. 2

Example image of normal colonic mucosa. For each pixel in the image, fluorescence excitation, transmission, and absorbance spectra were acquired. These images can be visualized as three-dimensional spectral image stacks, in which the first two dimensions represent spatial data and the third dimension represents spectral data. (a) HIFEX scan acquired from 390 to 480 nm. Three wavelength bands are shown for visualization ($\text{red}=465\mathrm{nm}$, $\text{green}=430\mathrm{nm}$, and $\text{blue}=390\mathrm{nm}$). Spectral data are shown in the $z$-axis (going into the page). Arrows point to colonic crypts. (b) Transmission scan and (c) absorbance scan from 390 to 700 nm ($\text{red}=650\mathrm{nm}$, $\text{green}=525\mathrm{nm}$, and $\text{blue}=410\mathrm{nm}$). (d–h) Pixel-averaged spectral data for a representative ROI: (d) fluorescence excitation scan from 390 to 450 nm, (e) fluorescence excitation scan from 390 to 480 nm (corresponding to the image in panel a), (f) fluorescence excitation scan from 390 to 550 nm, (g) transmission scan from 390 to 700 nm (corresponding to the image in panel b), and (h) absorbance scan from 390 to 700 nm (corresponding to the image in panel c).

To assess tissue heterogeneity, vascular, adipose, and mucosal structures of noncancerous colon were imaged (Fig. 3). Different tissue structures had similar fluorescence spectral shapes with large differences in magnitude [Figs. 3(d)–3(e)], suggesting that the bulk fluorescence excitation spectrum of a healthy colon is homogeneous. However, normalizing fluorescence data to a peak value of unity [Figs. 3(g)–3(h)] revealed subtle differences, likely indicating that many autofluorescent molecules contribute to the overall spectrum. A possible explanation is that less stable autofluorescent molecules—such as NADH or FAD, whose fluorescence properties depend on the oxidative state of the molecule—may have been degraded or compromised during the flash freezing and sample preparation process.

Fig. 3

Corrected images for three fields of view in normal tissue, showing (a) blood vessel, (b) adipose tissue, and (c) mucosa. Spectral data from the representative tissue types: (d) fluorescence excitation scan from 390 to 450 nm, (e) fluorescence excitation scan from 390 to 550 nm, (f) absorbance scan from 390 to 700 nm. Normalizing the fluorescence data to a peak value of unity (g and h) revealed that there may be subtle differences in the concentrations of two autofluorescent molecules, one with a peak excitation wavelength of 400 nm and another with a peak excitation wavelength of 440 nm. The absorbance spectrum of mucosa was smaller, and of a different shape, than that of vascular and adipose tissues (compare panels f and i), indicating the composition of absorbing molecules in the mucosa is different from that of vascular and adipose tissues. The heterogeneity in fluorescence and absorbance spectra is likely due to one or two autofluorescent molecules that are present at high concentrations, such as collagen or elastin, and additional autofluorescent molecules present at lower concentrations.

Unfortunately, there are relatively few spectroscopic studies of tissue autofluorescence that report fluorescence excitation spectral data. Richards-Kortum and Sevick-Muraca³⁵ have reported flavomononucleotides, with a peak fluorescence excitation wavelength of 436 nm. However, Wagnieres et al.³⁷ (in a comprehensive review) have also reported that flavins have a peak excitation wavelength of 455 nm and that porphyrins have a peak excitation wavelength of 400 nm. A definitive answer as to which molecules contribute to colon autofluorescence is not possible until a more detailed study of the components of fluorescence excitation spectra is performed. Further work is also warranted to develop specialized spectral analysis approaches that can identify weak signatures in the presence of much stronger signatures, such as matched filtering and energy minimization algorithms.^63–65

Interestingly, the absorbance data of colonic mucosa were much smaller, and of a different spectral shape, than that of vascular and adipose tissues [Figs. 3(f) and 3(i)]. This indicates that the composition of absorbing (but not fluorescing) molecules in the mucosa is different from that of vascular and adipose tissues. In addition, mucosa may be more transparent than either of the other tissue types and may allow a higher optical penetration depth, as indicated by the lower magnitude of absorbance. This observation indicates that absorbance or reflectance spectral image data may provide complimentary information to fluorescence spectral image data in differentiating among tissue types.

In this pilot study, four pairs of adenocarcinoma and surrounding normal mucosa were analyzed. While this is not a large enough sample size to calculate sensitivity and specificity metrics, it does provide a proof-of-principle or early feasibility demonstration for visualizing differences in spectral signatures of cancerous and normal tissues. At short excitation wavelengths, the fluorescence total intensity of adenocarcinomas was lower than normal tissue [Fig. 4(a), data shown for two specimen pairs]. This is consistent with prior literature, which indicates that fluorescence emission decreases in colonic adenoma and carcinoma,^29,53 as well as cancer progression in bronchial and oral mucosa.^27,30 However, fluorescence resulting from excitation at higher wavelengths was increased, and in the S4 sample was higher than normal tissue [Figs. 4(b) and 4(c)]. Transmission and absorbance spectral data indicate that adenocarcinoma displays increased optical absorbance, as compared to surrounding normal tissue [Figs. 4(d) and 4(e)], with additional spectral differences that could be exploited to increase sensitivity and specificity for tumor detection. These preliminary data demonstrate that there may be differences in spectral signature (also called spectral fingerprint) between cancerous and normal tissue and indicate that the excitation-scanning approach could provide high specificity (and likely high sensitivity) for detecting adenocarcinomas. It is likely that hyperspectral imaging can detect subtle changes in spectral signatures coincident with cancer stage that may not be detectable with NBI or AFI. In addition, due to the high sensitivity of fluorescence excitation-scanning hyperspectral imaging,⁵⁵ this technology could likely be translated to an endoscope or similar clinical imaging platform, enabling real-time hyperspectral imaging for endoscopy procedures. Finally, these data indicate that lesions are molecularly complex, likely heterogeneous, and should be investigated further to understand the relationship between cancer progression and spectral signature.

Fig. 4

Spectral differences between normal and cancerous tissue for two specimen pairs: S4 and S5. Preliminary data demonstrate the differences in spectral signature (fingerprint) between cancerous and normal tissue. Transmission and absorbance spectral data indicate that adenocarcinoma displays increased optical absorbance, as compared to surrounding normal tissue, with additional spectral differences that could be exploited to increase sensitivity and specificity for tumor detection. Spectral scan types are as follows: (a) fluorescence excitation scan from 390 to 450 nm, (b) fluorescence excitation scan from 390 to 450 nm, (c) fluorescence excitation scan from 390 to 450 nm, (d) transmission scan from 390 to 700 nm, and (e) absorbance scan from 390 to 700 nm.

The wavelength range of 390 to 450 nm was selected to visualize key spectral differences for all specimen pairs (Fig. 5). For all tissue pairs, the fluorescence intensity of normal tissues was higher than cancerous tissues. This is consistent with prior literature that indicates that fluorescence emission decreases with adenocarcinoma,^29,53 as well as bronchial and oral cancer.^27,30 In addition, cancerous tissues presented several common spectral features, including a local peak excitation wavelength of 400 nm and a local minimum at 430 nm [Figs. 5(e) and 5(h)]. This indicates that adenocarcinomas have common aspects of molecular composition.

Fig. 5

Fluorescence excitation scans from 390 to 450 nm of normal mucosa (solid red line) and adenocarcinoma (dashed blue line) from four patients. The total fluorescence excitation intensity (magnitude) of adenocarcinomas is lower than normal mucosa. In addition, adenocarcinomas present several characteristic spectral features, which normalized to a peak value of unity, including a local peak at 400 nm and a local minimum at 430 nm. These differences in spectral signature (fingerprint) offer a potential approach for discriminating cancerous and noncancerous tissues with high accuracy. (a–d) Raw spectra, corrected to NIST-traceable response. (e–h) Spectra normalized to a peak value of unity.

It is currently not clear which molecules contributed to changes in the fluorescence excitation spectrum of adenocarcinomas. Zângaro et al.⁶⁶ reported a similar peak excitation wavelength of 440 to 450 nm but did not attribute this to a specific molecule. Wagnieres et al.³⁷ summarized fluorescence excitation data, indicating that flavins have peak excitation wavelengths at 380 and 460 nm. In addition, lipopigments were reported to have a secondary excitation peak at 440 nm. By contrast, NADH was reported to have lower peak excitation wavelengths, at 260 and 350 nm. However, the fluorescence excitation spectrum of tissues has been reported only in a small number of studies, and even fewer studies have measured the fluorescence excitation spectrum from the UV through the visible range. Hence, future work is needed to study autofluorescence molecules under a broad range of fluorescence excitation wavelengths.

In the long term, a spectral library of fluorescence excitation spectra could be used to assist physicians in the diagnostic process in several ways: (1) by providing an estimate of tissue composition for specified areas of interest, (2) by providing molecular maps of a tissue using linear unmixing^67,68 or other spectral analysis technique,^69–71 (3) by allowing calculation of derived maps (such as redox ratio^72,73 or blood oxygen saturation^74–76), which may correlate with cancer development. If unmixing or spectral analysis is performed, it may be advantageous to employ more sophisticated nonlinear unmixing approaches, especially if reflectance and fluorescence image data can be combined to better estimate tissue parameters, such as scatter.⁷⁷ Regardless of the algorithm employed, the ability to visualize spectrally analyzed “molecular maps” during endoscopic navigation could improve the ability to detect small lesions and delineate the margins of larger lesions.

An advantage of fluorescence excitation scanning is that it can offer a large increase in signal strength.⁵⁵ As shown in Fig. 1(a), there is no need to spectrally resolve the fluorescence emission, as long as there is clean separation of the emission from excitation light using an appropriate long-pass fluorescence emission filter. In this configuration, all of the fluorescence emission is detected in a single long-pass emission channel for each excitation wavelength band, greatly increasing the signal strength. The increased signal strength can, in turn, allow proportional increases in imaging speed. In addition, the microscope configuration used for this study lends itself to an endoscope design in which fluorescence and reflectance may be measured simultaneously (Fig. 6). A future endoscope system using this design would provide identical fluorescence information as the microscope lightpath used in this study, but would be limited to measuring spectral reflectance instead of transmission. However, due to the increased signal strength provided by the fluorescence excitation-scanning approach and due to the ability to simultaneously measure fluorescence and reflectance, it is likely that this technology could be implemented in real time for screening during an endoscopy procedure.

Fig. 6

Potential optical design for a future spectral imaging endoscope based on the fluorescence excitation-scanning technology. Both fluorescence excitation spectral images and reflectance spectral images may be acquired simultaneously. Note that a future endoscope design would likely be limited to measuring spectral reflectance, instead of transmittance, due to the inability to place the illumination and detector on opposite sides of the lumen wall.

It should be noted that images of transmission or reflectance represent combined characteristics of the tissue, at the point of measurement. Even in current endoscopes, reflected light images are likely affected by several terms, including wavelength-dependent variations in refractive index of the tissue. Hence, for the case of transmission- or reflectance-based spectral imaging, the measured spectrum represents the composite effects of absorption, scatter, anisotropy, and wavelength-dependent refractive index changes. Many of these parameters have previously been shown to be heterogeneous across tissue types.⁷⁷ Hence, the basis for detecting colorectal lesions using reflectance-based spectral imaging would incorporate these variations, even if only from a lumped-parameter perspective.

Tissue preparation and/or fixation may affect the spectral properties and hence, the perceived feasibility of excitation-scanning hyperspectral imaging for colorectal screening. As an example, we compared spectra from the tissue specimens used in this study to similar colon tissues isolated from rats, which were discarded upon completion of an unrelated study. The tissue specimens used in this study were resected during surgical procedure, screened by pathology, flash-frozen in liquid nitrogen for transport, and then thawed in cold PBS prior to imaging. By contrast, the fresh rat colon tissues were obtained immediately after euthanizing the animal and stored in cold PBS for no more than 3 h prior to imaging. While the tissues were obtained from different sources, results indicate that there may be additional spectral components present in nonfrozen tissues that are lost during the tissue freezing and thawing process (Fig. 7). Further studies are warranted to quantify the effects of tissue freezing, as well as tissue fixation, on autofluorescence spectra.

Fig. 7

Spectral artifacts may be introduced during flash freezing and thawing, resulting in reduced molecular discrimination. (a) Example of a false-colored spectral image of flash-frozen human colorectal tissue specimen from this pilot study (transverse view); (b) false-colored spectral image of fresh (nonfrozen) rat colon from an ongoing animal study (en-face view); (c) average spectra from cancerous and normal tissue from two patients; (d) average spectra from three ROIs selected from the image in panel b. The red arrow indicates additional spectral components that are present in the fresh tissue and were likely lost during the flash freezing and thawing process.