2.1.

DLIPS System

The DLIPS system is shown schematically in Fig. 2. For all measurements, fluorescence excitation was accomplished using a Q-switched, frequency-tripled Nd: YAG laser, while the UV perturbation was accomplished using a 193 nm ArF excimer laser. The 355 nm fluorescence excitation beam was aligned coaxially with the 193 nm perturbation laser using a dichroic mirror (193 nm) as a beam combiner. Prior to the beam combiner, a variable attenuator was used to reduce the excimer laser to the desired perturbation intensity. The 355 nm beam diameter was about 30% less in diameter than the 193 nm beam (top-hat beam profile) at the target plane, to ensure complete perturbation of the entire fluorescence probe volume. The co-linear laser beams were passed through a pierced mirror positioned at 45 deg to the target plane. With the 355 nm beam incident on the target, fluorescence emission was redirected by the pierced mirror and then focused onto a fiber optic bundle using a combination of two 50-mm-diameter UV-grade lenses, where it was passed to a 0.3 m Czerny-Turner spectrometer and recorded with an intensified CCD (ICCD) array detector. Prior to entering the fiber optic, two sharp-edge filters were used in series: a high-pass filter designed to block the residual 355 nm light, and a low-pass filter designed to block any residual 532 nm light remaining from the frequency-tripling process. With this system, fluorescence emission was successfully collected in the spectral window between about 390 to 500 nm. A digital delay generator was also used to synchronize and control both the ICCD and the 355 nm laser, and a 3 μs detector gate was used to collect the fluorescence, set to the correspond to the leading edge of the 355 nm laser pulse (20 ns full width); hence all prompt and delayed fluorescence was recorded. Additional experimentation revealed that the majority of fluorescence emission corresponded to the first 100 ns of the detector gate.

Fig. 2

Schematic of the DLIPS system.

2.2.

Chemical Initiation and Promotion of Mouse Skin Tumors

Protocols were approved by the University of Florida Institutional Animal Care and Use Committee. Tumor formation on female athymic nude mice (Hsd: Athymic Nude-$\text{Foxn}{1}^{\text{nu}}$, Harlan Laboratories, Indianapolis IN), six to eight weeks old, was induced and promoted by 7,12-dimethylbenz(a)anthracene (DMBA) (Sigma-Aldrich, St. Louis, MO) in mineral oil (Fisher Scientific, Pittsburgh, PA) at a concentration of $0.5\%\mathrm{w}/\mathrm{w}$ applied topically to the dorsal skin according to previously described methods.^25–31 Application was repeated two to three times per week throughout the 11 week course of the experiment. DMBA application was discontinued if mice began exhibiting signs of systemic toxicity, particularly weight loss. In all cases, cessation of topical application resulted in alleviation of toxicity symptoms. Further, all of these mice showed visible lesions during week eight, consistent with those mice that continued receiving topical DMBA. Topical application of mineral oil alone was used on the control mice. The protocol was terminated at 11 weeks, prior to the emergence of squamous cell carcinomas (SCCs) as the goal of the study was not to investigate the biological difference between cancerous and normal tissue but rather to investigate DLIPS potential as a precancerous diagnostic technique. Thus tissue comparisons were made between healthy tissue and tissue during the course of SCC formation due to DMBA initiation and promotion.^26,27

2.3.

Histopathology

At weeks four and eight, one mouse was selected from the control group as well as the DMBA-treated group and euthanized using an intraperatoneal injection of Euthasol (Virbac AH, Ft. Worth TX). Likewise, the remaining mice were euthanized at the end of the study. Immediately following euthanasia, the dorsal skin was excised all the way down to the anteroposterior axis and fixed in formalin (Sigma-Aldrich, St. Louis, MO). Fixed samples were paraffin-embedded, sectioned at 5 μm, and stained with hematoxylin and eosin (H&E). Tissue preparation was performed at the Cell and Tissue Analysis Core, McKnight Brain Institute, University of Florida (Gainesville, FL). The sections were analyzed by a board-certified veterinary pathologist.

2.4.

Collection of DLIPS and Fluorescence Spectra

Starting in week two of the study, two control mice and two DMBA mice were selected at the beginning of each week for DLIPS and laser-induced fluorescence spectral analysis of the skin. Prior to all DLIPS spectral measurements, each mouse was wiped clean with an alcohol wipe and anesthetized using a ketamine/xylazine solution (Phoenix Pharmaceutical, St. Joseph, MO; Lloyd Laboratories, Shenandoah, IA) at $10\mathrm{ml}/\mathrm{kg}$ of body weight. Each mouse was then positioned in an alignment holder and spectra were collected from the back of each mouse, two spots from each side of the mouse. For each spot, a fluorescence spectrum was recorded using a 200 shot ensemble average recorded at 5 Hz laser repetition rate, referred to as the pre-perturbation spectrum. Immediately following the pre-perturbation fluorescence measurement, the 193 nm excimer laser was used to deliver a total of 2500 perturbation pulses to the target spot with laser energy of $100\mathrm{\mu J}/\text{pulse}$ ($\text{fluence}=3\mathrm{mJ}/{\mathrm{cm}}^2$) and a laser repetition rate of 50 Hz. Following perturbation, a delay of 30 seconds was introduced, after which a second fluorescence spectrum was recorded using an identical 200 shot ensemble average, referred to as the post-perturbation spectrum. These two spectra were then subtracted to generate the difference spectrum at each location, per Eq. (1), with the difference divided by the pre-perturbation spectrum to generate the DLIPS response for the given surface site. In this manner, a total of eight DLIPS spectra ($4\text{spots}\times 2\text{mice}$) were recorded each week from each group (control and DMBA). Despite the emergence of some lesions on the treated mice by week eight, all spectra were collected from skin regions that appeared to be pathology-free under visual inspection (i.e. visible lesions were avoided). Finally, the pre-perturbation spectra alone were retained and averaged separately for each group (control and DMBA) and week, providing traditional fluorescence spectral data to be analyzed independently of the DLIPS approach.

2.5.

Multivariate Statistical Analysis

At the conclusion of the study, all spectra were visually inspected, and 22 individual spectra ($\sim 14\%$ of the total collected) were rejected due to (1) excessive spectral noise and/or no observable difference in the pre- and post-perturbation spectra; indicative of low Nd:YAG laser and/or low perturbation laser intensity during the collection process, or (2) due to atypical spectral appearance (e.g. unusually large negative or positive perturbation), which is indicative of mouse movement (e.g. twitching) between the pre- and post-perturbation measurements. Following this process, the remaining spectra (pre-perturbation fluorescence and DLIPS) for both the control and DMBA groups were imported into MATLAB (The Mathworks, Natick, MA) for processing. Difference spectra and fluorescence spectra were preprocessed using a five-point moving average filter to remove acquisition noise. Fluorescence and DLIPS spectra were then analyzed separately, but using parallel methods, to facilitate side-by-side comparison at each analytical level. Spectra from the treated and control groups were grouped together in several different ways and analyzed using principal component analysis (PCA). Clustering within the data sets was visualized using score plots and the validity of using only the first two principal components for analysis was verified using a scree plot of the principal components (cumulative variance explained greater than 95%). Subgroups from the entire data set were selected for individual PCA to highlight the emergence of differences between the DMBA-treated skin and the untreated skin. For each PCA analysis, Hotelling’s $T^2$ test was used to verify that the clusters from the treated and control samples were statistically different.³²

The component loadings from the first principal component were used to assess which spectral regions best explain variations within and between data sets.³³ Component loadings were squared to facilitate direct comparison of loading across wavelength regions between sub-populations of the data set (this casts each wavelength’s loading value in terms of percent variance explained in the particular component). The regions of the spectrum at higher component loading levels were considered to have the highest impact on the observed differences between DMBA treated and control mice. Analysis of variance (ANOVA) was used to relate variations observed in principal component space to changes in the discrete study parameters. The goal of this analysis was to determine which factors (age of target pathology, papilloma formation, etc.) significantly had an impact on either the DLIPS spectra or the laser-induced fluorescence spectra.

2.6.

Detection Theory Analysis of Classification Performance

To investigate the ability of the DLIPS technique to discriminate between healthy and pathological tissue, a likelihood ratio classifier was implemented, as this has been shown to be the optimal detector to discover a signal in noise, irrespective of the type of data passed into it.^34,35 Further, this classifier implementation facilitates direct performance comparison between laser-induced fluorescence and DLIPS without the need for the subjective evaluation of classifier appropriateness for one spectral technique versus another.

The spectral data set was first processed using PCA to reduce the dimensionality of the data and consolidate the majority of the variation in the first few resolved components. PCA was used because this change of basis maximizes the variance across all of the data and does not require a priori knowledge of class associations. Due to the large percentage of variance explained by the first principal component for both the DLIPS and fluorescence spectra, as well as the limited size of the data set, detection theory modeling was restricted to decisions made using this first component. In this study, the null hypothesis ($H_0$) was considered to be the absence of DMBA-induced pathology plus some noise (e.g. biological background, shot noise in the ICCD), while the alternate hypothesis ($H_1$) was that a signal due to pathology is present in a noisy background, as shown in Eq. (2).³⁴

However, the overall inter-sample variance is assumed to be dominated by the presence or absence of the pathological state, not the noise. No prior knowledge of the data distribution under $H_0$ or $H_1$ was assumed, and the distributions were approximated using kernel density estimation (bandwidth set using a Gaussian approximation) as $P(X|H_0)$ and $P(X|H_1)$ respectively.³⁶ The probability density functions were combined into a likelihood ratio and compared against a range of operating points ($\tau$) to compute the receiver operating character (ROC) curves, as shown in Eq. (3). For each operating point, the null hypothesis was rejected if the likelihood ratio was less than or equal to $\tau$. The classification rule was applied across the entire data subset under consideration to compute the false positive rate (FPR) and true positive rate (TPR) for the corresponding operating point, given by:

To compare the inherent performance of the two spectral techniques, the area under the ROC curve (AUC) was used.³⁷ The maximum Matthews correlation coefficient (MCC) was evaluated as a basis for comparing “optimal” operating points between the two techniques as it provided equal weight to minimizing both false positive and false negative results.

Finally, a multi-modal pathology detector was designed using decision fusion in order to show the ability of DLIPS to complement traditional spectral techniques and boost the overall detection performance. Liao³⁸ and Jesneck³⁴ provide excellent reviews of the details for applying this technique to fuse multiple detectors as well as the benefits and limitations. The implementation employed here follows directly from Liao³⁸ without significant modification. Operating points for the local detectors were optimized by choosing the threshold that maximized the MCC. In this case, only the decisions from the DLIPS detector and the laser-induced fluorescence detector were used to create a fused detector. Performance was evaluated by calculating the AUC of the ROC for the fusion processor and comparing that with the performance of a single detector.

3.1.

Histopathology

Figure 3(a) through 3(d) are representative photomicrographs of the H&E sections of the skin at four, eight, and eleven weeks of the study. The top left panel [Fig. 3(a)] is from a control mouse after week four of the study. There is normal epidermal thickness with several layers in the stratum corneum. Hair follicles and sebaceous glands are in normal density within dermal collagen and subcutis. Skin samples from control mice in subsequent weeks were comparable with the sample at week four. Figure 3(b) shows the skin after four weeks of DMBA treatment, which displayed moderate epidermal changes including acanthosis and orthokeratotic hyperkeratosis, characterized by thickening of the stratum spinosum, stratum granulosum, and stratum corneum. Hair follicles were reduced in density, and the dermis in treated areas was almost devoid of sebaceous glands compared to skin from control mice. At the same time dermal collagen density was increased. After eight weeks of topical DMBA application, skin was characterized by more severe acanthosis, orthokeratotic hyperkeratosis, and dermal fibrosis than in skin after four weeks, as shown in Fig. 3(c). Hair follicles were infrequently found in treated areas, and sebaceous glands were almost completely absent. Squamous cell papillomas were occasionally present in the epidermis (although not shown in this section). At the conclusion of the study, week 11, the DMBA treated skin [Fig. 3(d)] showed more severe epidermal acanthosis and hyperkeratosis. There was also an increased, but still low density, scattering of squamous cell papillomas throughout the epidermis. Dermal fibrosis was slightly more severe than in samples from mice at eight weeks of DMBA treatment.

Fig. 3

H&E stained histology sections of skin from: (a) control mouse, week four; (b) DMBA mouse, week four; (c) DMBA mouse, week eight; and (d) DMBA mouse, week 11; (▴) denotes thickened regions of the stratum corneum, symptomatic of orthokeratotic hyperkeratosis; (⊗) highlights the thickening of the stratum spinosum, characteristic of acanthosis; ($*$) denotes areas of increased dermal collagen density; (d) is a section through a developed papilloma.

3.2.

DLIPS and Laser-Induced Fluorescence Spectra

Figure 4(a) through 4(c) show the average DLIPS spectra of the DMBA treated skin as compared with the control samples at weeks 2, 6, and 8. Figure 4(d) through 4(f) show the average of the laser-induced fluorescence spectra at the same time points, again comparing DMBA and control groups. Over the weeks of DMBA application, noticeable changes in shape of DLIPS spectra emerge (DMBA versus control), while these changes are not directly observable in the fluorescence spectra alone. Figure 5 shows individual (i.e. single mouse) spectral samples at the same time points of the study, illustrating the ability of the DLIPS technique to exploit internal normalization of the spectra relative to the tissue, and thus revealing information about the underlying pathology. Both figures show the emergence of a noticeable band in the lower wavelengths of the DLIPS spectra throughout the pathological progression of the DMBA model. This information is either obscured or not detected using fluorescence alone, as the spectra from the skin (both control and DMBA-treated) are dominated by the broad, correlated fluorescence peak centered on 460 nm.

Fig. 4

Average (a) through (c) DLIPS and (d) through (f) normalized fluorescence spectra at weeks 2 $(n=6,7)$, 6 $(n=5,4)$, and 8 $(n=7,8)$ for both DMBA-treated (▴) and untreated skin.

Fig. 5

(a) through (c) DLIPS and (d) through (f) fluorescence spectra at weeks 2, 6, and 8 from the same spot on the backs of DMBA-treated (▴) and untreated mice.

3.3.

Multivariate Statistical Analysis

Figure 6(a) and 6(b) show the score plots from the entire data set for the DLIPS data as well as the laser-induced fluorescence spectra. In both cases, 68 individual spectra were used from the DMBA-treated mice and 80 from the control mice throughout the eleven weeks of the study. For the entire DLIPS spectral data set, the cumulative variance explained by PC1 and PC2 was 95% while these components explained 99% of the variance for the fluorescence spectra. For both data sets Hotelling’s $T^2$ test returned a $p$-value less than 0.001, confirming that the spectra collected from untreated mice are statistically unique from those collected from the DMBA-treated mice.

Fig. 6

Score plots for the principal component analysis of the (a) DLIPS and (b) laser-induced fluorescence spectra from week 2 to 11 of the study. Control (•); DMBA-treated weeks 2 and 3 (▪); weeks 4 and 5 (♦); weeks 6 and 7 (◂); weeks 8 and 9 (▸); weeks 10 and 11 ($*$).

As an extension of this dimension reduction and remapping of the spectral data set into principal component space, the component loadings were used to identify the important spectral regions in explaining the variance across the entire data set as well as several subpopulations of the data set. Figure 7 shows the proportion of variance in a given data set as a function of wavelength. PC1 was the only component considered since greater than 90% of the total variance in principal component space was accounted for in this component for all cases. Figure 7(a) through 7(c) show the proportion of variance using the DLIPS technique for the entire data set, the DMBA data subset, and the control data subset, respectively. Across all acquired spectra, the most important regions in explaining the differences are below 400 nm, a peak at 420 nm, and a region of increasing importance at wavelengths above 460 nm. Additionally, when the control spectra and DMBA-treatment spectra are analyzed separately, unique band regions emerge to potentially explain the observed pathological progression to lesion formation in the DMBA model as opposed to the skin from the control mice. Primarily, the DMBA-treated skin has an important band between 400 and 420 nm, while the control skin’s band is below 400 nm. Although neither of these sub-populations shows much loading above 460 nm, the entire data set shows significant variability above this wavelength. This implies that this region may specifically account for variability between control and DMBA-treated skin using the DLIPS technique.

Fig. 7

Proportion of variance in principal component 1 accounted for by each wavelength from (a) through (c) DLIPS spectra; (d) through (f) laser-induced fluorescence spectra; and (g) through (i) post-perturbation spectra. In each row, the leftmost plot considers the variation across the entire data set; the middle plot represents the variation from only from DMBA-treated mice; and the right plot represents variation within the control data.

Remarkably, the region around the peak fluorescence emission is of minimal importance using the DLIPS technique, a dramatic contrast with the component loadings for the traditional laser-induced fluorescence spectra across the same data sets [Fig. 7(d) through 7(f)]. The fact that the shape of the component-loading map closely resembles that of the fluorescence spectra is indicative of a strong dependence on intensity fluctuations at the peak emission wavelength in describing variations across the data set. Further, there are no unique spectral regions within either the DMBA-treated or control groups when compared with the loadings from the entire data set. This sensitivity to signal intensity highlights an important limitation of traditional in vivo fluorescence techniques; namely, that subject-to-subject variability can often obscure the relatively much smaller variability due to changes in tissue autofluorescence.¹³ Finally, Fig. 7(g) through 7(i) shows the same technique applied to the spectra collected after the perturbation laser pulse. This analysis was applied to this data to assess whether any new information gained from applying the differential technique is unique to DLIPS, or merely inherent to the post-perturbation fluorescence spectra. Not unlike the component loadings for laser-induced spectra, these resemble the shape of the acquired fluorescence spectra, indicating that the most significant feature is the variability of the intensity of the broad fluorescence peak fluorescence centered at 460 nm. Again, the component loadings are the same across all three figures, indicating limited coupling of these spectra to underlying tissue biomolecular structure.

One-way analysis of variance (ANOVA) was used to determine which experimental factors were significant in explaining the observed differences in the spectral data set. The analysis was restricted to a one-way test because none of the two-way or three-way combinations of factors were significant, and there were missing factor combinations in the data set. The experimental factors that were controlled and accounted for in this study were treatment group, study week, spot number, acquisition order, and mouse number. There were only two treatment groups, either DMBA-treated or control. Spot number refers to which of the up to four measurements taken on a mouse on a particular day is being referenced. Acquisition order refers to whether the spectra were taken from the first or second mouse measured within a treatment group on a particular day. Finally, mouse within treatment group is a unique identifier for each of the five mice in each treatment group. Table 1 shows the $p$-values returned for each of the independent variables against PC1 and PC2 using DLIPS and laser-induced fluorescence. A value of 0.05 or less was considered a significant result and thus an underlying factor in explaining the variation of the data set. Under that threshold, the only experimental factors significantly reflected in the DLIPS spectra are study week (i.e. progression of DMBA-induced pathology or aging) and treatment group. While the raw fluorescence spectra show those factors to be significant, there are also significant contributions from acquisition order and mouse number, which are attributed primarily to random fluctuations in the absolute fluorescence signals.

Table 1

P-values from an ANOVA-based significance test of the controlled variables in the spectral dataset acquired from the mice. Significance with respect to each of the first two principal components using a particular spectral technique was assessed using one-way ANOVA (constrained sum of squares). Bold p-values are considered significant under a p<0.05 criteria.

Table 2

Detection performance metrics using DLIPS, laser-induced fluorescence, or a decision fusion approach.

3.4.

Pathology Detection Performance

Using the same compressive transformation as above, sub-populations of the data set were compared to describe the ability of each technique to discriminate the pathological (or pre-pathological) state from normal skin. In this case, only the first principal component was considered due to the small size of the sub-populations used in classifier construction. Specifically, the spectral data was partitioned into “early” (weeks 2 to 4), “middle” (week 5 through 8), and “late” (week 9 through 11) segments, based upon the visual observation of similar pathological states within those time periods for the DMBA-treated mice. The “early” DMBA-treatment group was considered pre-pathological as both observational and histological evidence showed only the earliest stages (inflammation and hyperplasia) of development. The “middle” and “late” periods were considered definite-pathological because dysplastic lesions and papillomas were visible during those periods of the study. PCA was performed independently on all of the data (DMBA treated and control) for the given time periods. Figure 8 shows the ROC curves derived from the likelihood ratios computed using the DLIPS technique, and Fig. 9 shows the curves for laser-induced fluorescence. With the exception the “early” time period, detection performance using the DLIPS technique was considerably better than using laser-induced fluorescence, as measured by calculating the AUC for each ROC curve. Even during the “early” stage of pathology development, which we note is considered pre-pathological, the DLIPS technique only lagged fluorescence by 2.5%. However, during the “middle” and “late” stages, DLIPS showed a 29.79 and 5.01% improvement in detection performance, respectively. In addition, Table 2 shows the sensitivity and specificity at the “optimal” operating points for these pathology detectors, assuming an equal weighting for false negatives and false positives. This table shows that for DLIPS, as the pathology matures, the specificity of detection stays at a high level while the sensitivity improves. On the other hand, when laser-induced fluorescence is used the sensitivity still improves, but at the expense of specificity. Thus, even when the detection performance of these techniques is optimized using an off-the-shelf classification scheme (i.e.: artificial neural network, linear discriminant analysis, etc.), the information encoded using DLIPS is more robust than traditional fluorescence spectroscopy and offers improved discrimination when detecting pathology.

Fig. 8

ROC curves for the detection of DMBA-induced skin pathology using the DLIPS technique. The data set was partitioned into early (weeks 2 to 4); middle (weeks 5 through 8); and late (weeks 9 through 11) time periods of papilloma development resulting from topical DMBA application.

Fig. 9

ROC curves for the detection of DMBA-induced skin pathology using laser-induced fluorescence. The data set was partitioned into early (weeks 2 to 4); middle (weeks 5 through 8); and late (weeks 9 through 11) time periods of papilloma development resulting from topical DMBA application.

As evidenced above using component loadings, the DLIPS spectra and fluorescence spectra emphasize unique spectral regions and features. Assuming this means that these spectra reflect different underlying biological changes, or at the very least some amount of statistical independence between the two techniques, the DLIPS technique was combined with laser-induced fluorescence to improve the overall detection performance.^34,38 Figure 10 shows the ROC curve for this fused classifier as well as the ROC curves for the local detectors with the operating points used to create the fused classifier labeled. When used to evaluate the “late” data set, the AUC is improved by 11.70% over detection using laser-induced fluorescence alone. Based upon these results, DLIPS may serve as a new dimension in the composite evaluation of pathology for in vivo detection and diagnosis applications.

Fig. 10

ROC curves for the fused decision incorporating both fluorescence and DLIPS spectra to boost the optimal performance in detecting DMBA-induced skin pathology in the “late” timeframe, weeks 9 to 11. The other ROC curves are from the DLIPS and fluorescence based techniques alone, with the selected operating points noted ($*$).