2.1.

Infinity-Corrected Imaging Setup

The system utilized the simplicity of infinity-corrected optics and the versatility of a commercial RGB camera [Fig. 1(a)]. Fluorescence signals were focused by a long working distance objective (Mitutoyo, Kawasaki, Japan) onto a 200-mm tube lens (#TTL-200A, Thorlabs), which were captured by a commercial Electro-Optical System (EOS) color camera (EOS 60D, Canon, Japan). Camera settings were initialized by the native EOS Utility, including exposure time, ISO levels, and image output format. The objective was able to fill up the Advanced Photo System type-C sensor size of the camera, given a field of view of $\sim 1.94\times 2.93\mathrm{mm}$. The imaging system was able to resolve Group 7 Element 6 on the U.S. Air Force Office of Scientific Research 1951 resolution target. The modulation transfer function (MTF) was computed, and resolution was found to be $0.5\mu \mathrm{m}$ at 10% contrast [Fig. 1(b)]. Raw image data were obtained by LabVIEW software readout and control.

Fig. 1

UV-fluor imaging system setup. (a) A schematic of the imaging system with camera tube lens and either $7.5\times$ or $10\times$ objective lens. Illumination sources were two UV-LEDs mounted symmetrically. (b) The imaging field of view and spatial resolution test are shown. (c) The illumination spot size generated by the UV-LEDs is shown with irradiance $10{\mathrm{mW}/\mathrm{mm}}^2$. (d) Emission spectra of UV-excited staining dyes when illuminated at 275-nm covered the detection range of the RGB camera.

2.2.

Ultraviolet Illumination Source

Open-faced, dark-field illumination was exploited using two 275-nm UV-LEDs (#M275L4, Thorlabs) mounted symmetrically to provide a more uniform light distribution. The LED emission was collimated by a 20-mm fused silica ball lens to provide an irradiance of $10{\mathrm{mW}/\mathrm{mm}}^2$, covering an area of $\sim 13\times 16\mathrm{mm}$ [Fig. 1(c)]. The LED drivers (#LEDD1B, Thorlabs) provided a current of 700 mA to each LED and a trigger signal for hardware synchronization. The LEDs functioned in trigger mode to minimize photobleaching and ensure temperature control. The short $50\text{-}\mu \mathrm{s}$ LED rise time as compared to 50-ms exposure time in addition to a couple of seconds of moving the translational table to a new spatial location provided adequate off time for the LEDs to avoid heading and any significant spectral changes.

2.3.

Animal Models and Tumor Sample Preparation

All animal procedures were conducted under the protocol approved by the Dartmouth Institutional Animal Care and Use Committee. A total of five athymic nude mice between the age of 6 to 8 weeks were used in this study. They were injected with human tumor cell line BxPC-3 (ATCC, Cat# CRL-1687). The pancreas was exposed, and tumor cells were injected with a $1:1$ ratio of Matrigel. BxPC-3 required 5 to 7 weeks for tumors to reach ideal imaging size of 1 cm in diameter. The mice were on purified diet to reduce autofluorescence from chlorophyll-based food consumption. After the tumors reached imaging size, the mice were anesthetized and sacrificed. Dextran Texas Red (TR) (Thermo Fisher Scientific, Cat# D1864) was intravenously injected 1-h before sacrifice to demonstrate the capability of perfusion imaging using the same imaging setup. Tumors were resected, embedded in 2.5% agar, and sliced in half. A layer of phosphate-buffered saline (PBS) was applied on the tumor surface to maintain proper hydration.

2.4.

Photodynamic Treatment

Two out of five mice were injected with $0.5\text{-}\mathrm{mg}/\mathrm{kg}$ Visudyne photosensitizer (1-h before PDT treatment with 690-nm laser at a dose of $75{\mathrm{J}/\mathrm{cm}}^2$ and irradiance of $100{\mathrm{mW}/\mathrm{cm}}^2$. The light was given by a fiber optic cable at the exposed tumor site. The mice were sacrificed two days after the treatment.

2.5.

Tumor Staining

Conventional pathological fluorophores such as Hoechst 33342, eosin, and rhodamine B (Sigma) were used as exogeneous stains on fresh and fixed tissue samples. Theses fluorophores could be excited by a deep UV illumination to highlight tissue morphology [Fig. 1(d)]. In this study, a combination of Hoechst 33342 ($0.5\mathrm{mg}/\mathrm{ml}$ in PBS), eosin ($1\mathrm{mg}/\mathrm{ml}$ in PBS), and rhodamine ($0.2\mathrm{mg}/\mathrm{ml}$ in PBS) was adequate for PDAC stromal imaging. The tissue sample was submerged in this dye combination for 30 s then washed off by PBS for 1-min.

2.6.

Image Acquisition, Depth of Field Correction, and Stitching

Image acquisition was automated by LabVIEW to allow for synchronization of illumination sources, camera shutter, and a motorized three-dimensional (3D) stage. At each $xy$ coordinate, a series of images taken at seven $6.3\text{-}\mu \mathrm{m}$ vertical increments provided an image stack for depth of field (DOF) correction [Fig. 2(a)]. While $6.3\text{-}\mu \mathrm{m}$ step size was the limit of our current vertical stage, smaller vertical step-size and more acquisitions for each location would result in higher image quality. Due to the high stiffness nature of PDAC tumors, tissue surface irregularity became a significant issue. Therefore, optical sectioning at multiple $z$-locations was necessary to maximize the number of in-focus pixels. Then, the DOF correction algorithm written by Aguet et al.²¹ was executed in MATLAB. Image stitching was implemented by Microsoft Image Composite Editor Software with 10% overlap to produce a whole tissue sample field of view [Fig. 2(c)]. After imaging, the tumors were formalin fixed and prepared for staining with Masson’s trichrome (MT) to visualize collagen and Hematoxylin and Eosin (H&E) to verify necrosis areas, which was essential for tumors that were treated with PDT to evaluate treatment effects. Both the UV-fluor data and the MT data acquired for collagen analysis were imaged with a 10× magnification lens.

Fig. 2

Schematic of primary image processing steps. (a) At each location, an image stack was acquired by moving tissue samples on a motorized translation stage. High-pass filtering was applied to reduce out of focus information. DOF correction was then applied. $\text{Scale bar}=100\mu \mathrm{m}$. (b) Nuclei and collagen were identified by color segmentation in HSV space. Fiber analysis was done on segmented collagen to yield information such as fiber length, thickness, orientation, and crosslinking profile. UV-fluor images were also color-map transferred to mimic the color palette of MT staining (blue collagen, dark purple nuclei, and light pink cytoplasm). The color remapping process was executed in L*a*b space, utilizing the mean and standard deviation of each channel to create color a scaling factor. (c) Image stitching was performed with 10% overlap to create a whole-tumor view. Rigid image co-registration was performed on the whole-tumor size UV-fluor images, histology images, and brightfield photos of freshly resected tumors to facilitate comparison across imaging modalities. $\text{Scale bar}=2\mathrm{mm}$.

2.7.

Tumor Parameter Identification

2.7.2.

Collagen analysis

Collagen analysis was performed in MATLAB. The binary mask of segmented collagen in the previous step was refined prior to collagen analysis. In cases in which nuclei were positioned on top of a collagen, their superficial position severely interfered with skeletonization and crosslinking analysis, therefore, they needed to be removed initially using function bwareaopen.m to perform hole removal on the collagen binary mask. After that, collagen thickness map was computed using the function bwdist.m, which calculates the distance of all positive pixels to their nearest background pixel [Fig. 3(a)]. Thresholding was applied to the UV-fluor collagen thickness map for two purposes: (1) to eliminate small collagen areas that were the cross-sections of long strands due to the orientation of tissue cuts and (2) to reduce the blur effect due to out of focus pixels that could inaccurately clump the nearby strands together. After these refinements, skeletonization was performed on the collagen thickness map using bwskel.m. Intersections of collagen strands were identified using branchpoints operation with bwmorph.m, and these intersection points were used to break down the collagen network into separate individual strands [Fig. 3(b)]. For each collagen strand, average fiber thickness was determined by twice the average distances of all pixels of the skeletonized centerline to their nearest background pixel. Calling regionprops.m with relevant measurement properties provided analysis for fiber count, length, and orientation [Fig. 3(c)]. The whole algorithm was applied to MT data as validation testing to ensure accurate fiber analysis on UV-fluor data. Manual segmentation and quantification of fibers were performed on a panel of 10 test images for both UV-fluor and MT image sets, which were chosen to cover a wide range of collagen fiber lengths, thicknesses, and orientations. After the algorithm was validated, auto-segmentation was performed with manual checking of key intermediate results such as the collagen thickness maps and skeletonized maps, for all regions of interest (ROIs) reported in Sec. 3.

Fig 3

Demonstration of collagen segmentation and quantification methods for UV-fluor data with validation testing on trichrome data, summarized in three main steps. (a) Step 1 generated collagen thickness maps using the MATLAB function bwdist.m performed on segmented collagen in HSV space. (b) Step 2 resulted in skeletonized maps of collagen in combination with thickness maps. Necessary refinements on UV-fluor data were highlighted. A thresholding value was applied on the UV-fluor collagen thickness map to eliminate inaccurate connections of clusters and small cross-sections of collagen due to tissue cut orientation. Skeletonization using bwskel.m and cluster breakdown using bwconncomp.m yielded individual collagen strands. (c) Fiber analysis was performed on each collagen strand to measure the length, thickness, and orientation. Statistical analysis to compare fiber analysis between UV-fluor and trichrome data was conducted to validate the feasibility of using UV-fluor imaging data to obtain quantified collagen information.

2.8.

Color Remapping

After structural segmentation described in the previous step, color remapping was implemented to adjust from UV-fluor color palette to a more conventional color scheme of MT staining. The color transfer process described in Reinhard et al.²² was executed in MATLAB. Image data were converted to L*a*b space (L for lightness, $a$ for red to green color values, and $b$ for yellow to blue color values). For each structure (collagen, nuclei, or cytoplasm), the mean and the standard deviation of each color channel were computed to create a scaling factor between the UV-fluor and the MT data sets [Fig. 2(b)]. Output images resembled the color scheme from MT staining.

2.9.

Statistical Analysis

A two-sample student’s $t$-test was employed in MATLAB without the assumption of equal variances between the tested samples to determine the difference in means using a two-tailed analysis and $\alpha =0.05$.

3.1.

Stromal Content of Fresh PDAC Tumors can be Visualized Microscopically and Macroscopically Using UV-fluorescence Imaging

Figure 4(a) shows the contrast enhancement of illuminating samples below 300 nm. For the same tissue specimen, 275-nm excitation reveals morphological features that are undetectable at 340-nm excitation. Collagen strands in yellow-green color at both peritumoral and intratumoral regions are observable, suggesting that UV-fluor imaging is capable of detecting thin collagen strands with diameters down to $5\mu \mathrm{m}$. Since image acquisition required optical sectioning of the specimens at different depths, the data after being DOF corrected also yielded semi-3D morphology [Fig. 4(b)]. Structural components such as blood vessels, collagen crosslinking, and collagen bundling can also be visualized with this imaging technique. Due to rapid imaging, a translation stage coupled with the imaging system could produce whole-specimen images with microscopic resolution. The stitching algorithm provides macroscopic image data, from which regions of necrosis could be easily identified as observed in Fig. 4(c). The light yellow/orange regions in the UV-fluor images are well-aligned with regions of necrosis inferred from H&E staining. A closed-up look at the images revealed collagen content in the tumors, as shown in Fig. 4(d). Side-by-side comparison of UV-fluor data and corresponding MT staining from the same specimens shows great agreement of where tumor collagen is located and the complexity of collagen network.

Fig. 4

UV-fluor imaging is capable of capturing stromal content in PDAC tumors. (a) Deep UV-illumination below 300 nm provides excellent image contrast, revealing morphology at the microscopic level. Collagen content from peritumoral and intratumoral regions are illustrated, $\text{scale bar}=100\mu \mathrm{m}$. (b) Optical sectioning and DOF correction provide semi-3D depth information, which highlights 3D structures, such as vessel, collagen crosslinking, and bundling. (c) Rapid imaging allows tumor visualization at macroscopic level, $\text{scale bar}=2\mathrm{mm}$. (d) Demonstration of strong yellow-green collagen signal was obtained from an RGB camera of fresh tissue imaging, and this is compared to MT staining data.

3.2.

UV-fluorescence Image Data with a 3-Dye Staining Technique Can Mimic Masson’s Trichrome Staining

Image data acquired by UV-fluor imaging shows that collagen network and viable tumor cells are distinct from each other, due to different emission peaks of Hoechst 33342 and eosin. Figure 5(a) highlights the feasibility of converting UV-fluor data into MT equivalent color scheme, i.e., collagen as blue, nuclei as dark purple, and cytoplasm as light pink. In Fig. 5(b), color remapping process was described for each main structure, i.e., collagen and nuclei. The $L$ channels for those structures were inverted to reflect the color intensity inversion, so the bright purple nuclei in the original UV-fluor data appear dark just as shown in MT staining. As for collagen, all of the channels were flipped to reflect not just the inversion in intensity, but also the hues to truly reflect the color transition in MT staining. The histograms of each channel ($L$, $a$, and $b$) in the L*a*b color space show that the distribution of intensity and hue information was preserved after the remapping process.

Fig. 5

Color remapping of UV-fluor images allows visualization using MT staining color scheme. (a) UV fluorescence image after color remapping showed collagen in blue and nuclei in dark purple, similar to MT staining results, $\text{scale bar}=30\mu \mathrm{m}$. (b) Illustration of color remapping in L*a*b color space, based on structural segmentation. Corresponding histograms of reference, input, and output images show that the distribution of color intensities and hues was preserved.

3.3.

Classification of Stroma in PDAC Tumors is Feasible in Fresh Tissue via UV-fluorescence Imaging

Heterogeneity of collagen formation in PDAC tumors was captured and classified in Fig. 6. Collagen fiber visualization general agreement between UV-fluor and trichrome data was visualized in Fig. 6(a). Exact matching between two data sets was not expected due to tissue deformation and tissue changes due to pathology staining process to create trichrome-stained samples. However, direct comparison of the same tissue regions confirms the capability of UV-fluor imaging to visualize collagen in ex vivo samples. Quantification of the fluorescence signals also shows a strong agreement with results obtained from trichrome data. Figure 6(b) shows the outputs of the fiber strand analysis based on fiber thickness maps for both data sets.

Fig. 6

Quantification of collagen content in PDAC tumors. (a) Different types of collagen observed in both trichrome and UV-fluor data sets. (b) Quantification of collagen fibers. (c) Distribution of fiber length, fiber thickness, and fiber orientation obtained from both data sets (d) Statistical analysis was performed to show that there is no significant difference in means in terms of tumor collagen content, fiber length, and fiber orientation between UV-fluor and trichrome data. However, there was a statistical difference between two data sets for fiber thickness. Mean and standard deviation of UV versus trichrome data, respectively, for each category are $18.8\pm 4.2$ versus $15.6\pm 8.1$ (collagen content %), $15.9\pm 7.8$ versus $7.9\pm 7.3$ (collagen thickness in $\mu \mathrm{m}$), $48.5\pm 36$ versus $49.3\pm 47$ (collagen fiber length in $\mu \mathrm{m}$), $2.53\pm 51$ versus $-0.85\pm 51.3$ (collagen fiber orientation in degrees). Data were analyzed for five tumors, necrosis areas were excluded, 50 ROIs were randomly selected with ROI size of $0.6\times 0.4\mathrm{mm}$ each, total collagen strands from UV $\text{data}=6442$, from trichrome $\text{data}=6606$.

The same algorithm was showed to work well for both. The distribution of collagen fibers in terms of length, thickness, and orientation was shown in Fig. 6(C) for both data sets to validate the feasibility of extracting quantitative information from UV-fluor imaging on fresh tissues. Results from statistical analysis in Fig. 6(d) shows that there was not a significant difference in means between the two data sets in terms of collagen content percentage, collagen fiber length, and orientation. About 20% difference in means (${\mu }_{\mathrm{UV}}=18.8\%$ versus ${\mu }_{\mathrm{MT}}=15.6\%$) reported from comparing collagen content data was largely due to the differences in fiber thickness analysis (${\mu }_{\mathrm{UV}}=15.9\mu \mathrm{m}$ versus ${\mu }_{\mathrm{MT}}=7.9\mu \mathrm{m}$), which could be largely attributed to the difficulty of accurate segmentation and separation of fibers from imaging bulk tissue. Another reason for thickness inconsistency was the thresholding applied on the collagen thickness map mentioned in Sec. 2. Better image quality with minimized out-of-focus pixels would significantly minimize the inaccuracy from thresholding and close the gap of fiber thickness reported by these two imaging modalities. Total count of collagen strands from both data sets is $<3\%$ different (6442 strands from UV data and 6606 strands from trichrome data). Fiber length on average is accurately reported with a discrepancy of less than 2% (${\mu }_{\mathrm{UV}}=48.5\mu \mathrm{m}$ versus ${\mu }_{\mathrm{MT}}=49.3\mu \mathrm{m}$) while orientation shows a difference of 3.4 degrees on average.

3.4.

UV-Fluorescence Imaging Could be Utilized as an Assay Platform to Evaluate the Effects of Photodynamic Priming on Collagen Modulation

Figure 7 showcases additional features of UV-fluor imaging in fresh tissues that can be utilized in targeted therapy response assessments. Figure 7(a) shows the capability of imaging perfusion using the same simple imaging optics. Emission from TRexcited by 275-nm illumination was captured by the red channel on the RGB camera. Dextran-perfused samples could then be stained to obtain structural information without interference from TR signals. In Fig. 7(b), yellow outline was drawn to identify the tissue imaging surface while red outline locates the regions of necrosis due to photodynamic priming effect. This illustration highlights the capability of visualizing necrosis areas in fresh tissues, which normally is hard to delineate in brightfield images therefore in need of pathological confirmation. Furthermore, the capability of imaging collagen in situ demonstrated in this study could be utilized to assess the collagen modulation effect as an outcome of acute photodynamic priming, as reported in Obaid et al.²³ Preliminary data acquired by UV-fluor imaging have showed possible reduction in desmoplasia for PDAC treated tumors in Fig. 7(c), which suggests that collagen content was reduced by 13%.

Fig. 7

UV-fluor imaging as a fresh tissue assay platform to evaluate photodynamic priming responses. (a) Perfusion imaging of dextran tagged with TR in kidney samples. UV-excited TR shows perfused dextran in kidney samples (middle) as compared to endogenous fluorescence without any dextran injection (left). Dextran-perfused tissues could then undergo pathological staining (right) to obtain structural information, $\text{scale bar}=500\mu \mathrm{m}$. (b) Tissue imaging surface was outline in yellow. Necrosis outlined in red is hard to distinguish in brightfield image (left), but observable under UV-fluor imaging (middle), which is confirmed by trichome staining (right), $\text{scale bar}=1\mathrm{mm}$. (c) A comparison of collagen content in control and PDP treated tumors. Collagen modulation effect was observed, in which the PDP treated tumors show a 13% reduction in collagen content ($n=2$ animals per group, control $\mathrm{ROIs}=23$, treated $\mathrm{ROIs}=19$), $\text{scale bar}=100\mu \mathrm{m}$.