2.1.

Clinical Measurements of Fluorescence Spectra from Normal and Dysplastic Oral Mucosa

A point probe spectroscopic system was used to obtain depth-resolved fluorescence spectra at $350\text{-}\mathrm{nm}$ excitation of normal and neoplastic oral lesions in Institutional Review Board approved clinical studies carried out at the University of Texas M.D. Anderson Cancer Center.¹⁶ Lesions selected by the clinician and an anatomically similar contralateral normal appearing site were measured for each patient by placing the probe against the mucosal surface. To minimize artifact from exposure to room light, the measurements were performed in a darkened room. Wavelength calibration, power calibration, and standard measurements were performed for each set of clinical measurements. After spectroscopic measurements, biopsies were collected from the corresponding tissue site for histological examination, subject to prior patient consent and discretion of the clinician. Histologic results were reviewed by a collaborating pathologist.

2.2.

Tissue Geometry and Model Input Parameters

Oral mucosa is modeled as a multilayered medium with homogenous distribution of absorbers, scatterers, and fluorophores within each sublayer. As illustrated in Fig. 1a , the epithelium of normal oral mucosa can be divided into three layers with different fluorescence characteristics.²¹ The strongly scattering superficial epithelial layer includes cells that retain keratin, and the main fluorophore in this layer is assumed to be keratin. Immediately beneath this is a layer of intermediate nonkeratinized epithelial cells containing FAD as the dominant fluorophore. Beneath this is a layer of basal epithelial containing mitochondrial NADH as the dominant fluorophore. Similarly, the stroma can be divided into two layers with different optical properties: the superficial stroma and deeper stroma. Previous research shows that the main stromal fluorophores are collagen and elastin cross-links.²⁸ The thickness of each sublayer was derived from confocal fluorescence images of a representative normal tongue tissue slice that was used to determine the tissue geometry shown in Fig. 1a. In our model, the total thickness of the epithelial layer is $280\mu \mathrm{m}$ ; the superficial layer and the intermediate epithelial layer are each $80\mu \mathrm{m}$ thick, whereas the basal epithelial layer occupies $120\mu \mathrm{m}$ . The superficial stroma (stromal region 1) occupies an area $125\mu \mathrm{m}$ below the basement membrane, and deeper stroma (stromal region 2) is modeled as a semi-infinite medium.

Fig. 1

Baseline tissue geometry and input parameters. (a) Tissue geometry for normal oral tissue; superficial epithelial layer $80\mu \mathrm{m}$ , intermediate epithelial layer $80\mu \mathrm{m}$ , basal epithelial layer $120\mu \mathrm{m}$ , superficial stroma $125\mu \mathrm{m}$ , and deep stroma $1\mathrm{cm}$ . (b) Normalized fluorescence efficiencies. (c) Absorption coefficients. (d) Scattering coefficients.

In this model the fluorescence emission properties of a given sublayer are described by the fluorescence efficiency term $\mu \left({\lambda }_{\mathrm{ex}}\right)\cdot \phi ({\lambda }_{\mathrm{ex}},{\lambda }_{\mathrm{em}})$ of the dominant fluorophore within that sublayer, where $\mu \left({\lambda }_{\mathrm{ex}}\right)$ is the absorption coefficient and $\phi ({\lambda }_{\mathrm{ex}},{\lambda }_{\mathrm{em}})$ is the quantum efficiency of the dominant fluorophore. The relative magnitude of the fluorescence efficiency term for each sublayer was determined from the confocal images of the representative normal tongue tissue slice, used here as a baseline example. The spectral shape of the quantum efficiency for each sublayer was estimated from emission spectra at $350\text{-}\mathrm{nm}$ excitation of biological samples containing the predominant fluorophore and not directly from the confocal images of the representative oral slice. The spectrum of keratin was obtained by measuring the fluorescence of human nail,²⁸ which is dominated by keratin. The emission spectrum of an optically dilute solution of FAD was used to represent the shape of the fluorescence efficiency of the intermediate epithelial cells. The fluorescence from a suspension of SiHa cervical cancer cells²⁸ was used to describe the shape of the basal epithelium. To ensure that fluorescence from the cell suspension is a valid approximation for fluorescence from basal epithelium, SiHa cells were grown on coverslips and imaged under the same conditions as oral tissue slices. SiHa cells displayed similar autofluorescence patterns as basal oral epithelium: cellular autofluorescence that is bright at UV excitation, but dimmer at $488\text{-}\mathrm{nm}$ excitation. Next, a viable oral tissue slice was obtained that consisted of normal basal epithelium [the stroma and the rest of the epithelium were removed with a Krumdieck tissue slicer (Alabama Research and Development, Munford, Alabama)], and an emission spectrum of this sample was measured at $350\text{-}\mathrm{nm}$ excitation. Comparisons of the emission spectra from basal epithelium and SiHa cell suspension showed that the two samples have very similar spectral shapes and emission peaks. Previously, it was shown that the emission spectrum from collagen type-1 gel can be described as a linear combination of spectra from three main types of collagen cross-links.²⁸ The contribution from collagen cross-links with emission peaks at $410\mathrm{nm}$ was found to be dominant, and was used here as input for the shape of the stromal fluorescence efficiency. The relative fluorescence efficiencies for each sublayer are summarized in Fig. 1b.

Baseline values for the absorption and scattering coefficients for each epithelial and stromal layer were obtained from previous studies of cervical, oral, and bronchial tissue. The scattering coefficients of the superficial and basal epithelial layer were derived from reflectance confocal measurements of cervical epithelium.²² In particular, the scattering coefficient extracted from keratinized cervical epithelium is used to represent superficial oral scattering, while the scattering coefficient from basal cervical cells is used here for the basal epithelial layer. The scattering coefficient of the intermediate epithelial layer was derived from reflectance confocal images of normal oral epithelium.²³ The stromal scattering coefficients for stromal layers 1 and 2 were assumed to be the same and were taken from a study on bronchial tissue.²⁹ Note that the scattering coefficients were extrapolated to lower wavelengths using the inverse proportionality relationship of cellular scattering to wavelength.³⁰ The absorption coefficient was assumed to be constant throughout the oral epithelium and was based on measurements from bronchial tissue.²⁹ Stromal absorption is assumed to be caused by collagen and oxy- and deoxyhemoglobin. The total hemoglobin absorption coefficient was calculated by the well-known expression for blood absorption published by Jacques.³¹ The extinction coefficients for oxy- and deoxyhemoglobin were taken from the same source, and the oxygen saturation of hemoglobin was assumed to be 80%. The concentration of hemoglobin per liter of blood used in this study was $150\mathrm{g}∕\mathrm{l}$ , the normal value reported by Jacques,³¹ and the baseline volume fraction of blood in stromal layers 1 and 2 was set to be 0.16%. This fraction is very similar to values used in a previous model of light propagation in cervical and colon tissue.^{17, 32} In addition, Jacques reported a typical value for the blood volume fraction in skin (dermis) to be 0.2%.³¹ To verify that a value of 0.16% is appropriate to describe blood absorption in oral mucosa, we have estimated the fraction of stroma occupied by the lumen of blood vessels from hematoxylin and eosin (H and E) slides of a representative normal and dysplastic oral tissue slice. Fractions derived from stromal areas $100\text{to}300\mu \mathrm{m}$ below the basement membrane are in agreement with the value used in this model. Stromal collagen absorption coefficients are based on measurements from bloodless skin samples.³³ Figures 1c and 1d summarize the baseline absorption and scattering coefficients for the stromal and epithelial layers used in this study.

2.3.

Description and Validation of the Monte Carlo Model of Fluorescence

A fixed weight, multilayered reflectance Monte Carlo code with a depth selective fiber optic probe has been modified to account for generation and propagation of fluorescent light. Details regarding the reflectance Monte Carlo code and the geometry of illumination and collection are described in a previous study.¹⁵ In the fluorescence Monte Carlo code used here, excitation photons are initially propagated using scattering and absorption coefficient values at the excitation wavelength. At each scattering event, the probability of the photon being absorbed is given by ${\mu }_a∕({\mu }_a+{\mu }_s)$ , where ${\mu }_a$ is the absorption coefficient and ${\mu }_s$ is the scattering coefficient for a specific layer. On absorption, the probability that a photon is absorbed by a fluorophore is calculated. In the epithelial layers, this probability is equal to one, because it is assumed that the only absorbers are also fluorophores. In the stromal layers, however, a photon can be absorbed by either collagen cross-links (a fluorophore) or by hemoglobin (an absorber). The probability that the photon is absorbed by a stromal fluorophore is given by ${\mu }_a\mathrm{coll}∕({\mu }_a\mathrm{coll}+{\mu }_a\mathrm{Hb})$ , where ${\mu }_a$ coll is the absorption coefficient of collagen and ${\mu }_a$ Hb is the absorption coefficient for a given stromal layer at the excitation wavelength. The probability of fluorescence emission at a particular emission wavelength is given by the fluorescence efficiency of each sublayer. After isotropic emission, fluorescence photons are further propagated using the scattering and absorption coefficient of each layer at the emission wavelength, until reabsorbed or remitted from the tissue surface. The Heyney-Greenstein approximation was used as the phase function.³⁴ The accuracy of the multilayered fluorescence Monte Carlo code for an infinite source and detector was compared to remitted fluorescence reported in the literature, and our results were found to be within 1% of published values.³⁵

Figure 2 shows a diagram of the depth selective probe used to measure tissue fluorescence spectra and the geometry used within the Monte Carlo code. In the probe, a ball lens is used to confine illumination and collection to a shallow area, approximately $300\mu \mathrm{m}$ beneath the tissue surface. In the Monte Carlo code, the ball lens was approximated using a half ball lens coupled to a flat window.¹⁵ To validate the depth selectivity of the Monte Carlo code, simulations of a thin fluorescent target were performed as a function of the distance between the detector and the target. Results were compared to experimental measurements. The experimental measurements were performed in water to better simulate the refractive index mismatch between the probe tip and tissue. In the simulations, the fluorescence target was modeled as a very thin $\left(10\mu \mathrm{m}\right)$ strongly fluorescent layer, with a large absorption coefficient and a very small scattering coefficient, and a refractive index equal to 1.4. The gap between the probe and the target was represented by a layer with the optical properties of water. Simulations of measurements at different probe-to-target distances were performed by increasing the thickness of the water layer in increments of $50\text{to}100\mu \mathrm{m}$ .

Fig. 2

(a) Diagram of the depth selective fiber optic probe geometry used in the Monte Carlo forward model. S represents the source fiber and D the detector fiber. An air gap separates the fibers and the half ball lens. The distance between the source and detector fibers is $0.072\mathrm{cm}$ . The window is placed in contact with tissue. (b) Detected fluorescence intensity from a thin fluorescent layer, as the separation between the target and window is varied. The dotted line represents experimental data and the solid line represents the Monte Carlo predictions.

2.4.

Predictions of Tissue Fluorescence Using the Monte Carlo Model

After validation, the Monte Carlo model was then used to explore the predicted fluorescence spectra of tissue. All simulations were done at an excitation wavelength of $350\mathrm{nm}$ . A single simulation output consists of the detected fluorescence intensity (total number of detected fluorescence photons) at emission wavelengths of 400, 420, 440, 450, 460, 480, 500, 520, 540, and $560\mathrm{nm}$ . Each simulation was performed with ${10}^8$ excitation photons. Simulations were performed first for a baseline case using the optical properties of normal oral mucosa. Then a sensitivity analysis was performed by varying the optical properties of individual layers that are expected to change with neoplasia. In the baseline case and the model sensitivity analysis of stromal input parameters, simulations were repeated three times. Simulations for the model sensitivity analysis of epithelial parameters were performed only once.

For each detected photon, the depth at which fluorescence emission occurred was recorded. The mean depth of fluorescence was calculated by averaging these depths for all detected photons. To analyze the origin of the detected fluorescence signal at each emission wavelength, we calculated the fraction of detected photons with a depth of emission originating from the epithelium and the stroma, and also the fraction of detected photons originating from the superficial oral mucosa (the epithelium and stromal layer 1).

3.1.

Validation of Model with Probe-to-Target Measurements in Water

Figure 2 shows simulated fluorescence measurements for a thin fluorescent target as the distance between the probe tip and the sample is varied, and compares the performance of the modeled probe to the actual probe used in clinical studies. The relative error between predictions and experimental data is on average small, varying from 0.01 to 0.24 in the $0\text{to}400\mu \mathrm{m}$ range, and illustrates the depth selective nature of the probe, which collects fluorescence generated within $300\text{to}400\mu \mathrm{m}$ of the probe tip.

3.2.

Sensitivity of the Fluorescence Model to Variations in Epithelial Input Parameters

The development of neoplasia in the oral cavity is accompanied by a series of physiological changes in the epithelium such as hyperkeratosis and higher cellular metabolic activity. These physiologic changes lead to changes in optical properties, including increased epithelial scattering and an increase in the NADH associated fluorescence from epithelial cells. The effect of each of these changes on the detected fluorescence was studied individually by varying a particular input parameter while holding all other factors constant. Figure 3 summarizes the predicted fluorescence spectra for three different sets of simulations designed to model hyperplasia, hyperkeratosis, and dysplasia in oral mucosa. Predictions in Figs. 3a, 3c, 3e were normalized to the maximum intensity value of the baseline spectra (black line) to emphasize the effect of varying a particular input parameter on the overall intensity of the fluorescence. Predictions in Figs. 3b, 3d, 3f were normalized by dividing each individual spectrum by its maximum intensity value to emphasize changes in spectral shape. In all sets of simulations, the baseline fluorescence spectrum is plotted as a black line. Table 1 summarizes the fraction of detected fluorescence photons that originate in the epithelium at $440\text{-}\mathrm{nm}$ emission for each predicted spectrum.

Fig. 3

Effect of epithelial input parameters, associated with (a) and (b) hyperplasia, (c) and (d) hyperkeratosis, and (e) and (f) dysplasia on the intensity (left column) and spectral shape (right column) of the predicted fluorescence spectra.

Table 1

Effect of input parameters associated with hyperkeratosis, hyperplasia, and dysplasia on the fraction of photons detected from the epithelium and the fraction of photons detected from the superficial oral mucosa (the epithelium and superficial stroma) at emission wavelength of 440nm .

The effect of hyperplasia is simulated by changing the thickness of the nonkeratinized epithelium (keeping the thickness of the superficial layer and the ratio of the intermediate epithelium to the basal epithelium thickness constant) in increments of $50\mu \mathrm{m}$ . Figures 3a and 3b show that, as the thickness of the nonkeratinized epithelium is increased, the overall fluorescence intensity drops, accompanied by a shift of the spectra to longer emission wavelengths and an increase in the fraction of detected photons generated in the epithelium. For example, increasing epithelial thickness by $150\mu \mathrm{m}$ results in a drop of fluorescence intensity at $440\text{-}\mathrm{nm}$ emission by about a factor of 1.5, and an increase in the fraction of epithelial photons by about a factor of 1.6. Examination of the number of epithelial and stromal photons indicates that the relative increase in the fraction of epithelial photons is due more to a drop in stromal photons than to an increased number of epithelial photons. Increasing the thickness of the nonkeratinized epithelium decreases excitation light penetration to stromal areas, and thus increases sampling of epithelial fluorescence. Decreased epithelial thickness results in the opposite trend. It is interesting to observe that even small changes $\left(50\mu \mathrm{m}\right)$ in the thickness of nonkeratinized epithelium have a significant effect on both the intensity and shape of the predicted spectra.

The effect of hyperkeratosis on the spectral shape and fluorescence intensity was studied by changing the thickness of the strongly scattering, keratinized superficial layer (and hence changing the total thickness of the epithelium) by $50\mu \mathrm{m}$ . Results are illustrated in Figs. 3c and 3d. A thicker keratin layer results in a drop of fluorescence intensity by a factor of 1.3 without a significant shift in spectral shape and fraction of epithelial photons. However, a thicker keratin layer causes a small increase in photons detected from the keratin layer, and a large decrease in the numbers of photons from the underlying basal epithelial layer and stroma. Decreasing the thickness of the superficial layer has the opposite effect, due to deeper light penetration and increased detection of stromal fluorescence.

The development of dysplasia is associated with an increase in epithelial scattering and an increase in NADH-based cellular fluorescence. We have modeled these changes by increasing the thickness of the basal epithelial layer without changing the overall thickness of the epithelium. As shown in Figs. 3e and 3f, an $80\text{-}\mu \mathrm{m}$ increase in the thickness of the basal epithelium leads to higher total fluorescence intensity and a shift of the spectral shape to the red. These changes are accompanied by a significant increase (a factor of 1.4) in the fraction of detected epithelial photons, due an increase in the number of epithelial photons, but no significant change in the number of detected stromal photons. Thus, as the fraction of cells with higher scattering and NADH-based fluorescence increases, fluorescence from the epithelium starts to dominate the total detected fluorescence.

3.3.

Sensitivity of the Fluorescence Model to Variations in Stromal Input Parameters

The fraction of epithelial photons, calculated for different sets of simulations and summarized in Table 1, indicates that the modeled probe detects a substantial number of stromal photons. For example, in the baseline case, more than half of the detected photons originate from the stroma. In tissue geometries representing hyperplasia and dysplasia, the fraction of epithelial photons rises to 0.60 to 0.66; however, complete rejection of stromal photons is not achieved, suggesting that variations in stromal parameters can also influence simulations. Here we have modeled stroma as a two layered medium, which allows us to study variations in input parameters within each individual stromal sublayer.

Figure 4 illustrates the effect of changing the magnitude of the stromal fluorescence efficiency on the total intensity and spectral shape of the predictions. Benign and dysplastic alterations in oral tissue were found to be characterized by a large drop of fluorescence intensity in the superficial stroma.²¹ We simulated the effect of changes in the magnitude of the fluorescence efficiency of superficial stromal layer 1 [Figs. 4a and 4b], separately from changes in the deeper stromal layer 2 [Figs. 4c and 4d]. Decreasing the fluorescence efficiency of stromal layer 1 leads to a drop in the total fluorescence intensity and a shift of the spectrum to the red. These changes are also associated with an increased epithelial fraction of photons, due to a decrease in the number of stromal photons. Note that decreasing the fluorescence of stromal layer 1 by 40% leads to a significant decrease in total fluorescence (by a factor of 1.4 at $440\mathrm{nm}$ ) and a relatively large increase in the fraction of epithelial photons (by a factor of 1.3). Previous research indicates that with inflammation and dysplasia, fluorescence from the stroma can drop by more then a factor of 3.²¹ Thus, changes in the magnitude of the fluorescence efficiency of stromal layer 1 can have a potentially more significant effect than shown in Fig. 4. Decreasing the fluorescence efficiency of stromal layer 2 by 40% has only a small effect (less than 10%) on the spectral shape, fluorescence intensity, or the fraction of detected epithelial photons. This indicates that only changes in the magnitude of the fluorescence efficiency of stromal layer 1 lead to changes in predicted spectra collected with the depth selective probe.

Fig. 4

Effect of variations in the magnitude of the fluorescence efficiency of the stroma on (a) and (c) the intensity and (b) and (d) spectral shape of the predicted fluorescence spectra: (a) and (b) show changes in superficial stromal layer 1, and (c) and (d) in deeper stromal layer 2.

The effect of changes in stromal hemoglobin absorption on predictions is illustrated in Fig. 5 . An increase in blood vessel density by about a factor of 1.2 to 3 is commonly associated with premalignant progression in oral mucosa.³⁶ Here the volume fraction of hemoglobin in stroma was varied, which changes the stromal absorption coefficient. Simulations show that an increased volume fraction of hemoglobin in stromal layer 1 leads to a decrease in the total fluorescence intensity, mainly in the $400\text{-}\text{to}440\text{-}\mathrm{nm}$ emission region. These changes are also associated with small differences in the epithelial fraction of photons, also limited to the $400\text{-}\text{to}450\text{-}\mathrm{nm}$ emission range (data not shown). Changes in stromal layer 2 have a minimal effect on predicted spectra. Figure 6 summarizes the effect of decreasing scattering in stromal layers 1 and 2. Results indicate that variations in the scattering coefficient in stromal layers 1 and 2 do not change the spectral intensity, shape, or epithelial photons fraction by more then 10%. It should be noted that varying both stromal scattering and hemoglobin absorption had a less pronounced effect on predictions compared to changes associated with variations in the stromal fluorescence and epithelial optical parameters.

Fig. 5

Effect of variations in the volume fraction of hemoglobin on (a) and (c) the intensity and (b) and (d) spectral shape of the predicted fluorescence spectra: (a) and (b) show changes in superficial stromal layer 1 and (c) and (d) in deeper stromal layer 2.

Fig. 6

Effect of variations in the stromal scattering on (a) and (c) the intensity and (b) and (d) spectral shape of the predicted fluorescence spectra: (a) and (b) show changes in superficial stromal layer 1 and (c) and (d) in deeper stromal layer 2.

3.4.

Depth Sensitivity Analysis for Fluorescence Light

Results in Figs. 4, 5, 6 indicate that the fluorescence output is most affected by variations in input parameters from stromal layer 1. This suggests that the modeled probe geometry limits the penetration of excitation light to the superficial layers of the oral mucosa and detects fluorescence that originates mostly from the epithelium and superficial stromal areas. The fraction of superficial fluorescence photons (the number of fluorescence photons originating from the epithelial and superficial stromal layers divided by the total fluorescence photons) was calculated for the baseline case and for simulations with varied epithelial and stromal fluorescence parameters. As shown in Table 1, in the baseline case, 90% of detected fluorescence photons originate from the epithelium and superficial stromal layer 1. While variations in the input parameters influence the value of the superficial photon fraction, results consistently indicate that most (86 to 93%) of the detected fluorescence originates from the epithelium and superficial stroma of the oral mucosa for all sets of simulations.

3.5.

Comparison of Average Clinical Fluorescence Spectra from Normal Oral Mucosa and Oral Lesions with Moderate to Severe Dysplasia to Monte Carlo Predictions

The performance of the Monte Carlo code in predicting normal and abnormal fluorescence spectra was validated by comparing simulation results to average spectra from normal and dysplastic sites measured from the buccal mucosa, the tongue, the lip, and the floor of the mouth at $350\text{-}\mathrm{nm}$ excitation. Depth-resolved fluorescence spectra from 281 normal oral sites and 27 oral sites diagnosed with moderate to severe dysplasia were used to calculate the average normal and dysplastic spectra.

Normal oral tissue was simulated with the input parameters and tissue geometry used in the baseline predictions. Dysplasia is characterized by a series of physiological changes, including an increase in the total epithelial thickness (hyperplasia), thickening of the keratinized superficial epithelium, and spreading of abnormal cells with increased scattering and metabolic activity throughout the epithelium. In addition, the fluorescence intensity of superficial stroma was found to decrease in oral lesions diagnosed with dysplasia. Other stromal changes, such as increased Hb absorption and decreased stromal scattering, were not considered because the model sensitivity analysis shows that these parameters have only a small effect on the intensity and shape of predictions. To model dysplastic tissue, the combined effect of changes in the thickness of epithelial sublayers and the fluorescence intensity of subepithelial stoma were considered. Several simulations were performed by varying the thickness of individual epithelial sublayers (data not shown). We find that the set of changes that most closely matches the measured dysplastic fluorescence spectrum is a combination of hyperkeratosis, an increase in the thickness of the basal epithelium to simulate the spreading of metabolically active and highly scattering cells, and a drop of fluorescence intensity of the superficial stroma. In particular, hyperkeratosis was modeled by a small increase in the superficial thickness $\left(70\mu \mathrm{m}\right)$ , a small increase in the thickness of the nonkeratinized epithelium $\left(50\mu \mathrm{m}\right)$ , and an increased fraction of the epithelium occupied by the basal sublayer. The fluorescence intensity of the superficial stromal layer 1 was reduced by a factor of 1.6.

Figure 7 compares the average clinical spectra at $350\text{-}\mathrm{nm}$ excitation to simulated spectra from normal and dysplastic sites. The error bars in the figure represent the standard deviation of the clinical average spectra at $420\mathrm{nm}$ and indicate the degree of variability in the datasets. Results show that the Monte Carlo model is able to predict the drop in intensity and the shift of spectral shape to longer emission wavelengths that accompany dysplastic progression in the clinically measured data. However, it should be mentioned that both the normal and dysplastic simulations are slightly shifted to the red emission region compared to clinical data, suggesting the need for a more precise estimation of the input fluorescence properties for each epithelial sublayer. Bar graphs representing the total number of detected photons from the epithelium and stroma are displayed in Figs. 7b and 7c. The drop in fluorescence intensity can be attributed to a decrease in the total number of photons originating from the stroma, whereas the shift of the emission peak is due to the increased fraction of epithelial photons of the total detected signal.

Fig. 7

Comparison of average clinical fluorescence spectra from normal oral mucosa and lesions diagnosed with moderate to severe dysplasia to Monte Carlo predictions. (a) Average clinical spectra and predictions. Error bars represent the standard deviation in the clinical data for each diagnostic category. Number of detected photons from the epithelium and stroma in (b) normal and (c) dysplastic predictions.