2.1.

Mouse Tongue Carcinogenesis Model

Six-week-old female CBA/J mice were purchased from Jackson Laboratories (Bar Harbor, Maine). Animals were treated with 4NQO at $100\mu \text{g}/\mathrm{mL}$ in their drinking water for 16 weeks to induce epithelial carcinogenesis. In this murine model, over half of the specimens developed dysplasia by 12 weeks and squamous cell carcinoma (SCC) by 24 weeks.²⁸ Animals were euthanized at various time points during the treatment process (Table 1). Following sacrifice, the tongue was removed for imaging experiments as described below. All animal experiments were reviewed and approved by the Institutional Animal Care and Use Committee of the University of Texas School of Dentistry at Houston.

Table 1

Number of specimens imaged for each modality by time point.

2.2.

Widefield Imaging

White light and fluorescence widefield images were captured with a multispectral digital microscope (MDM), which is described in detail elsewhere.²⁹ Briefly, the system is an optical microscope that has been modified to capture digital white light images and fluorescence images at multiple excitation wavelengths. The MDM captures images with a field of view of $\sim 3\times 2\mathrm{cm}$. In this study, the MDM was used to capture autofluorescence images at 405 nm excitation with a 430 nm long-pass emission filter, as well as fluorescence images following topical application of two fluorescent contrast agents. Fluorescence images following topical staining with 2-NBDG were obtained with an excitation of 450 nm and a 550 nm long-pass emission filter; following topical staining with proflavine, images were obtained with an excitation of 450 nm using a 490 nm long-pass filter. The gain and exposure time were kept constant for each modality. Fluorescent standards and a frosted quartz disk were imaged during each imaging session/time point as positive and negative controls. Of particular importance, the green and blue fluorescent slides maintained a mean fluorescence intensity within $\pm 7\%$ over the course of the study, and the negative control showed minimal fluorescence above the background.

After each specimen was harvested, white light and autofluorescence images of the resected tissue were obtained. Specimens were then incubated in a 160 μM solution of 2-NBDG (Invitrogen, Carlsbad, California) in 1X phosphate-buffered saline (PBS) for 20 min at 37°C. Uptake of this fluorescent deoxyglucose derivative is associated with increased metabolic activity.^16,17 Specimens were washed once with PBS and then fluorescence images were obtained using the appropriate filters for 2-NBDG imaging. The specimen was next incubated in a 0.01% w/v solution of proflavine (Sigma Aldrich, St. Louis, Missouri) in 1X PBS for 2 min at room temperature. Specimens were washed once with PBS and then fluorescence images were obtained using the appropriate filters for proflavine imaging. The fluorescence signal from proflavine staining is much brighter than that of 2-NBDG, allowing for imaging of proflavine-stained tissue after 2-NBDG staining. Table 1 shows the number of samples imaged with each modality at each time point; white light and autofluorescence images were obtained from all specimens, but some specimens were not stained with 2-NBDG and proflavine but were reserved for confocal imaging as described below.

2.3.

Confocal Microscopy

Widefield fluorescence images characterize variations in tissue autofluorescence and uptake of the contrast agent across the epithelial surface. To qualitatively assess the distribution of fluorescence throughout the epithelium and stroma, confocal fluorescence images were collected from transverse fresh tissue slices of some specimens. Resected specimens were stored in iced phenol-free Dulbecco’s modified Eagle’s medium (Sigma-Aldrich, St. Louis, Missouri) until slicing. Approximately 300-μm-thick transverse tissue slices were prepared using a Krumdieck tissue slicer (Alabama Research and Development, Munford, Alabama). Slices were prepared from both unstained specimens and stained specimens, but were not fixed prior to imaging. Confocal images were acquired using a Zeiss LSM 5 Live confocal microscope (Carl Zeiss, Inc., Thornwood, New York) within 12 h of specimen collection. Images were obtained using a $20\times$ objective (NA 0.8). The focal plane was located between 15 and 25 microns below the surface of the tissue slice. Autofluorescence images were obtained at both 405 and 488 nm excitation, and collected with 420 and 505 nm long-pass filters, respectively. Images from proflavine-stained specimens were obtained at 488 nm excitation using a 505 nm long-pass filter. All images were taken with the same detector settings for each modality. Confocal images were not obtained from specimens stained with 2-NBDG due to photobleaching. 2-NBDG fluorescence is more prone to photobleaching than is autofluorescence or proflavine.

2.4.

Pathology Maps

Following all imaging experiments, tissue specimens were fixed in 10% formalin and returned to the Pathology Department of University of Texas School of Dentistry at Houston for hematoxylin and eosin (H&E) staining. To facilitate the comparison of widefield images to a histologic reference standard, a two-dimensional color-coded pathology map was created for each specimen. Histological sections were obtained for each section cut by the Krumdieck tissue slicer. The H&E slide from each section was reviewed by an oral pathologist (N.V.), who segmented the section into regions of normal/benign epithelium, mild, moderate, and severe dysplasia, and SCC using published criteria for grading oral epithelial dysplasia.³⁰ A color-coded two-dimensional pathology map was then created by registering the adjacent segmented sections; this map could then be superimposed on top of white light or fluorescence images from the same sample.

2.5.

Image Analysis

Confocal fluorescence images were qualitatively analyzed to identify key morphologic features associated with autofluorescence and proflavine uptake. These features correspond to those present in H&E-stained histology, including changes in keratin structure, nuclear crowding, and nuclear enlargement. Images of the epithelium were divided into three layers that are routinely used for histopathologic assessment to qualitatively describe spatial patterns in autofluorescence and proflavine staining throughout neoplastic progression.

The data used for quantitative image analysis were compiled from 12 of the 25 total specimens, which were selected because widefield images of autofluorescence, 2-NBDG, and proflavine were obtained for each of these specimens. Images from these three modalities were aligned with a built-in image registration algorithm and analyzed using MATLAB® software (The MathWorks, Natick, Massachusetts). A $25\times 25\text{pixel}$ grid was used to divide images into regions of interest (ROIs). Between 83 and 125 ROIs were created per specimen. The mean of the fluorescence intensity and the kurtosis, entropy, skewness, and standard deviation of the pixel histogram of fluorescence intensity were calculated for each ROI in each of the three imaging modalities. Each ROI was assigned a pathology read based on the corresponding pathology maps. ROIs without a corresponding pathology read due to missing epithelium in the histologic section were excluded from further analysis.

A two-class linear discriminant analysis-based classifier was developed to discriminate non-neoplastic from neoplastic ROIs based on quantitative feature values. The data set was split into a training and test set. The training and test groups were created by dividing the 12 specimens into two groups based on duration of carcinogen application to ensure both early and late time points were included into each set. All of the ROIs from an individual specimen were designated to either the training or test set to ensure all of the ROIs for an individual specimen were only in one of the two sets. The data sets contain a similar number of ROIs per diagnostic category (Table 2). For training and testing, moderate dysplasia, severe dysplasia, and SCC were grouped together as neoplastic samples. ROIs with mild dysplasia were excluded from the algorithm development. The algorithm generates a posterior probability that an ROI is neoplastic based on image features calculated from each imaging modality; samples with a posterior-probability exceeding the 0.39 threshold were classified as neoplastic. Performance measures, such as the area under the receiver operating curve (AUC), sensitivity, and specificity at the Q-point, were calculated for each of the individual image features using the pathology read as the gold standard. The two features with the best combined performance in distinguishing neoplastic tissue from non-neoplastic tissue were selected based on the AUC. The algorithm was then used to classify the held-out samples with a histologic diagnosis of mild dysplasia.

Table 2

Number of regions of interest (ROIs) and performance of the classification algorithm for the training and test groups by histologic diagnosis.

3.1.

Mouse Tongue Carcinogenesis Model

A total of 25 mice were examined in this study. Nine animals served as controls and were not exposed to the carcinogen; the remaining 16 animals were exposed to the carcinogen and sacrificed at different time points (Table 1). The number of specimens imaged differed between imaging modalities because some specimens were saved for confocal imaging. Specimens destined for autofluorescence confocal imaging were not topically stained with 2-NBDG or proflavine to allow the autofluorescence signal to be imaged. Similarly, specimens destined for 2-NBDG confocal imaging were not topically stained with proflavine. In this model, dysplasia begins four weeks after exposure, and half of the specimens develop dysplasia by week 12.²⁸ SCC is seen 18 weeks postexposure, and half of the specimens develop SCC by 24 weeks, but the carcinogenesis process is multifocal with multiple lesions forming at different sites on the tongues.

3.2.

Widefield Imaging

Widefield white light, autofluorescence, and fluorescence images poststaining with 2-NBDG and proflavine from the course of the study are shown in Fig. 1. An increase in the blue channel of the autofluorescence images is seen at weeks 11 and 14. With increased carcinogen exposure, blue autofluorescence intensity decreases, but is never completely lost, as seen in images acquired at weeks 21 and 29. Punctate areas of increased autofluorescence in the red channel are visible after the 14-week time point. Areas of increased red autofluorescence are often found near ulcerated regions of tissue and have been associated with oral bacteria.³¹ For comparison, a control specimen that was not exposed to the carcinogen is shown in the column on the left, which shows minimal autofluorescence signal; this remained essentially constant at all time points (data not shown).

Fig. 1

Ex vivo widefield imaging of mouse tongues. White light and autofluorescence images of unstained mouse tongues are shown in the top two rows, respectively. Fluorescence images acquired after staining the specimens with 2-NBDG and proflavine are shown below. A control specimen that was not exposed to the carcinogen is shown in the left column. Specimens at 11, 14, 21, and 29 weeks after initial carcinogen exposure are shown from left to right. Dysplasia begins at 4 weeks postexposure with over half of the specimens developing dysplasia by 12 weeks. Squamous cell carcinoma (SCC) is seen 18 weeks postexposure with over half of the specimens developing SCC by 24 weeks.²⁸ Scale bar is 5 mm.

Focal areas of increased fluorescence intensity with 2-NBDG labeling are seen beginning at week 11 and increase over time throughout the study. The total area of increased fluorescence intensity per specimen also rises over time. High fluorescence intensity near the base of the tongue is due to nonspecific uptake of the cut edge. Removal of excess dye from the cut edge with exposed muscular tissue is difficult, leading to higher background; similar spreading of India ink occurs in muscular tissue for pathology. However, cut edges and ulcerations are detectable under white light examination, allowing these areas to be excluded from the analysis of fluorescence images.

Proflavine staining reveals changes in surface texture as a function of time. The fluorescence image of a normal specimen labeled with proflavine shows a regular periodic pattern. This pattern is disrupted beginning at week 11; there is a progressive increase in bright, punctate spots of proflavine staining with time across the course of the study. The number of focal areas of increased proflavine fluorescence per specimen rises over time with carcinogen exposure.

3.3.

Confocal Microscopy

Fluorescence confocal images of fresh transverse tissue slices were collected to improve the understanding of how changes at the cellular level affect widefield fluorescence images of intact tissue. Figure 2 shows autofluorescence confocal images of epithelial cross-sections. The blue channel utilized 405 nm excitation, the same as that used in widefield imaging. The green channel illustrates epithelial cell size and morphology. Images were collected as two separate channels and superimposed in Fig. 2. Corresponding H&E images from the same tissue slices are shown below [Figs. 2(f) to 2(j)]. The images are grouped by pathology grade ranging from normal [Figs. 2(a) and 2(f)], through multiple stages of dysplasia [Figs. 2(b) to 2(d) and 2(g) to 2(i)] and SCC [Figs. 2(e) and 2(j)]. The superficial layer of normal epithelium shows a bright layer of blue autofluorescence corresponding to superficial keratin. The pattern of autofluorescence follows the periodic filiform papillae structure of the underlying epithelium. The superficial keratin autofluorescence increases for the sample with mild dysplasia [white arrow, Fig. 2(b)]; this signal corresponds to the hyperkeratosis [black arrow, Fig. 2(g)] present in the H&E-stained section. In general, the autofluorescence from the keratin layer decreases as the thickness of keratin decreases with progression of the disease. However, these specimens contain keratin of highly variable thickness at all stages of disease. These findings are also seen in the corresponding H&E-stained sections.

Fig. 2

High-resolution, cross-sectional images of mouse tongue epithelium corresponding to areas imaged with widefield autofluorescence imaging. Confocal autofluorescence images [(a) to (e)] are matched with H&E-stained histology sections [(f) to (j)] from the same specimen. Images are grouped by pathology grade: normal [(a) and (f)], mild dysplasia [(b) and (g)], moderate dysplasia [(c) and (h)], severe dysplasia [(d) and (i)], and squamous cell carcinoma [(e) and (j)]. Arrows indicate hyperkeratosis. Qualitative features of the confocal autofluorescence images for the keratin, epithelial cell, and stromal collagen layers are listed in the table below the figure. Scale bars are 100 μm.

Beneath the superficial keratin, normal epithelial cells show weak green autofluorescence, which is strongest in the basal portion of the epithelium [Fig. 2(a)]. With progression to mild dysplasia, the cellular fluorescence encompasses the lower third of the epithelium [Fig. 2(b)]. Cellular fluorescence signal is observed in the lower two thirds of the epithelium in the specimen with moderate dysplasia [Fig. 2(c)]. The epithelial cells in severe dysplasia and cancer exhibit moderately intense fluorescence throughout the epithelium and in cellular regions of the tumor, respectively [Figs. 2(d) and 2(e)].

Proflavine staining highlights cell nuclei [Figs. 3(a) to 3(e)]. In keratinized tissue, the keratin is also stained by proflavine. A regular, periodic keratin pattern is seen in the fluorescence images of normal tissue [Fig. 3(a)]. This periodic pattern is disrupted and ultimately lost with progression from dysplasia to SCC. In normal tissue, strong epithelial cell nuclear staining is seen in the basal portion of the epithelium [Fig. 3(a)]. As expected, cell nuclei become enlarged and crowded with progression to dysplasia and cancer. The proflavine fluorescence signal from cell nuclei encompasses the lower half of the epithelium in mild dysplasia [Fig. 3(b)], the lower two thirds of the epithelium in moderate dysplasia [Fig. 3(c)], and the entire epithelium in severe dysplasia [Fig. 3(d)]. In cancer, enlarged cell nuclei are seen throughout the specimen [Fig. 3(e)]. Proflavine-associated fluorescence in areas of increased keratin is especially prevalent in the mild dysplasia example [Fig. 3(b)]. Corresponding H&E images from the same tissue slices are shown for comparison in Figs. 3(f) to 3(j).

Fig. 3

High-resolution, cross-sectional images of mouse tongue epithelium corresponding to areas imaged with widefield system following proflavine application. Confocal proflavine images [(a) to (e)] and H&E-stained histology sections [(f) to (j)] from the same specimen. Images are grouped by pathology grade: normal [(a) and (f)], mild dysplasia [(b) and (g)], moderate dysplasia [(c and (h)], severe dysplasia [(d and (i)], and squamous cell carcinoma [(e and (j)]. Qualitative features of the confocal proflavine images for the keratin and epithelial cell layers are listed in the table below the figure. Scale bars are 100 μm.

3.4.

Ex Vivo Widefield Image Processing

Figure 4 shows a representative pathology map for one of the specimens in this study. Images from each modality were registered and a pathology map was created based on a detailed pathology review. The colored regions overlaid on the white light image indicate the pathology read. Qualitative comparison of the fluorescence images and the gold standard pathology map demonstrates that there is greater fluorescence intensity in areas of higher degrees of dysplasia and cancer. For example, the area of tissue marked as cancer (purple) on this specimen shows increased fluorescence intensity in the same region in the autofluorescence as well as the 2-NBDG and proflavine-stained images. Additionally, Fig. 4 emphasizes the multifocal lesions formed in this model and the spatial heterogeneity in the histology across a single tissue specimen.

Fig. 4

A representative specimen with the corresponding pathology map. All images were registered in MATLAB® and a pathology map was created based on a pathology review. The color scale on the bottom shows the key for the pathology map. Scale bar is 2.5 mm.

While 2-NBDG images highlight areas of increased fluorescence intensity, the proflavine images accent the heterogeneity of the surface texture across the tissue. The changes in keratin surface structure with dysplastic progression are highlighted by the widefield proflavine images shown in Fig. 5(a). ROIs corresponding to different pathology diagnoses are shown at higher magnification in Fig. 5(c). The regular, repeated structure of the papilla and surface keratin is visualized in the normal region [Fig. 5(b)]. Progression through dysplasia leads to a disruption of this pattern, which is completely lost with progression to cancer.

Fig. 5

Changes in keratinized tongue texture versus pathology grade. Structural changes in the keratinized surface are apparent with progression of disease in a carcinogen-exposed proflavine stained specimen (a). A control proflavine stained specimen (b) is shown for comparison. Scale bar is 2.5 mm for widefield images. Enlarged areas of interest corresponding to each of the five diagnostic categories are below (c). Scale bar is 0.5 mm for insets.

3.5.

Linear Discriminant Analysis

The ROIs were divided into a training set and test set. Table 2 details the number of ROIs by diagnostic category in each set. One hundred ninety-nine ROIs with a diagnosis of mild dysplasia were withheld from the initial analysis. The image features in the training set were used to select the two top performing image features and train a classification algorithm to differentiate neoplastic from non-neoplastic ROIs. Proflavine standard deviation and 2-NBDG mean fluorescence intensity were the two top performing features to discriminate neoplastic and non-neoplastic tissue with an area under the receiver operating characteristic curve of 0.96.

A scatter plot of the proflavine standard deviation and 2-NBDG mean fluorescence intensity is shown for the 415 ROIs in the test set in Fig. 6(a). In the test group, 89% ($116/130$), 100% ($30/30$), and 100% ($5/5$) of the moderate dysplasias, severe dysplasias, and SCCs were correctly classified as neoplastic, respectively. Of the normal ROIs, 91% ($228/250$) were correctly classified by the algorithm as non-neoplastic. The algorithm yields a sensitivity and specificity of 91% at the Q-point, the location on the curve with the shortest Euclidean distance to the upper left-hand corner of the plot. Mild dysplasias were excluded from the training algorithm due to the difficulty in accurately making this histological diagnosis and its uncertain clinical potential.³² When the classifier is applied to ROIs diagnosed histologically as mild dysplasia, 42% ($84/199$) are classified as non-neoplastic and 58% ($115/199$) classified as neoplastic. Interestingly, this is similar to recent results reported for the diagnostic accuracy of optical imaging in patient samples.³² Although intriguing, this does not indicate that these are two distinguishable populations. Figure 6(b) shows a scatter plot of the proflavine standard deviation versus the mean 2-NBDG fluorescence intensity for the samples diagnosed histologically as mild dysplasia.

Fig. 6

Classification of ROIs using proflavine standard deviation and 2-NBDG mean fluorescence intensity. Dashed lines represent the linear threshold values to discriminate normal sites from moderate/severe dysplasia or cancer. Plot of wide-field values of proflavine standard deviation and 2-NBDG mean fluorescence intensity for test regions of interest (ROIs) with a pathology diagnosis of normal, moderate dysplasia, severe dysplasia, and squamous cell carcinoma (a). Plot of values for test set ROIs with a pathology diagnosis of mild dysplasia (b). The color scale in upper right corner shows the histology key.