1.

Introduction

Fluorescence and diffuse reflectance spectroscopy in the ultraviolet-visible (UV-VIS) spectrum, are increasingly being investigated for breast cancer diagnosis. ^{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15} This technology is cost-effective, and can be deployed through fiber optic probes to quickly, nondestructively, and quantitatively detect the structural, physiological, and biochemical changes that are associated with the progression of disease. Several studies have demonstrated that there are significant differences in the fluorescence ^{1, 2, 3, 5, 7, 9, 11} and diffuse reflectance ^{4, 13, 14, 15, 16, 17} spectra of normal, benign, and malignant breast tissues, which could be used for the diagnosis of breast malignancy.

In most previous studies, fluorescence ^{1, 2, 3, 5, 7, 9, 11} and/or diffuse reflectance spectra ^{1, 4, 5, 8, 13, 15} of breast tissues were analyzed using empirical methods, in which the spectral intensities and/or line shapes were examined to extract spectral patterns (such as intensities, ratio of intensities, or principal components) that consistently discriminate between malignant and nonmalignant tissues. For example, in one of the first reported studies, Alfano ¹¹ used the ratio of fluorescence intensities at 340 and $440\mathrm{nm}$ , excited at $300\mathrm{nm}$ , to distinguish malignant from normal breast tissues. While several studies ^{1, 2, 3, 4, 7, 11, 13} examined the spectral intensities, there are other studies where spectral line shapes were explored for breast cancer diagnosis. For example, Palmer ⁵ employed principal component analysis (PCA) to extract a set of principal components that dramatically reduced the spectral data dimension with minimal information loss, and showed that a subset of principal components obtained from diffuse reflectance spectra and fluorescence spectra at excitation wavelengths of 300, 400, 420, and $460\mathrm{nm}$ displayed the statistically most significant differences between malignant and non-malignant breast tissues. The preceding collection of studies examined spectral intensities or line shapes of turbid tissue fluorescence and/or diffuse reflectance spectra. Although empirical analyses can reveal important spectroscopic features for tissue characterization and disease diagnosis, they do not relate the measured spectra directly to the physically meaningful information that are contained in the spectra.

The diffuse reflectance spectrum reflects the absorption and scattering properties of the tissue. The absorption coefficient is directly related to the concentration of physiologically relevant absorbers in the tissue such as oxygenated and deoxygenated hemoglobin. The scattering coefficient reflects the size and density of scattering centers in tissue, such as cells and nuclei. The fluorescence spectrum contains biochemical information about a number of molecules that are relevant to the structural properties of tissues or respiratory metabolism of cells. However, there are distortions in the fluorescence spectra due to the interplay of tissue absorption and scattering, which makes it difficult to interpret the biochemically relevant information contained in the fluorescence. Extracting and quantitatively assessing the intrinsic and physically meaningful information from the fluorescence and diffuse reflectance measurements will provide insight into the biochemical, physiological and morphological changes that take place during the progression of disease.

Several studies have been reported in which the absorption and scattering properties were quantified from diffuse reflectance spectra to characterize breast tissues. Ghosh ¹⁶ used a model based on the standard diffusion approximation to extract the absorption and scattering coefficients from spatially resolved diffuse reflectance measurements of human breast tissue samples. They showed that malignant tissues had increased absorption and scattering relative to nonmalignant breast tissues. However, they did not report the classification accuracy for diagnosing breast cancer based on the parameters extracted from the model. More recently our group, in two independent studies^{17, 18} applied a Monte-Carlo-based inverse model for the extraction of absorption and scattering properties from the diffuse reflectance spectra of malignant and nonmalignant breast tissues. Both studies showed that malignant tissues have decreased hemoglobin saturation and increased mean reduced scattering coefficients compared to nonmalignant tissues. Classification based on machine learning using these absorption and scattering properties as inputs achieved an unbiased sensitivity and specificity of 82 and 92%, respectively, in the study by Palmer, ¹⁸ and 86 and 80%, respectively, in the study by Zhu ¹⁷, for discriminating between malignant and nonmalignant breast tissues.

For fluorescence, a number of studies ^{19, 20, 21, 22, 23, 24, 25} have been carried out to disentangle the effects of absorption and scattering from the measured fluorescence spectrum to recover the intrinsic fluorescence of the tissue. The intrinsic fluorescence is due only to fluorophores present in the medium, from which fluorophore concentrations can be extracted using simple analytical models. A majority of these studies used analytical methods to derive mathematical expressions for the intrinsic fluorescence, from which the intrinsic fluorescence spectrum can be calculated. ^{20, 21, 22, 24, 25} These methods are based on the diffusion equation and thus are valid for a limited range of absorption and scattering, not flexible in their applicability to various probe geometries or require extensive empirical calibration. Biswal ²³ proposed a method that was based on simultaneous measurements of polarized fluorescence and polarized elastic scattering spectra, and turbidity-free fluorescence was obtained by normalizing the polarized fluorescence spectrum by the polarized elastic scattering spectrum. Our group¹⁹ developed a fluorescence model based on scalable Monte Carlo simulations, which is capable of extracting the intrinsic fluorescence spectra of a turbid medium, and is applicable for a wide range of optical properties and, in principle, any illumination-collection geometry. The details of the model and its validation in synthetic phantoms are described in the preceding companion manuscript.²⁶ This paper uses the model to recover the intrinsic fluorescence of human breast tissues and to extract the relative fluorescence contribution of individual fluorescing components present in malignant and nonmalignant breast tissues.

The specific goals of this study were to (1) apply Monte-Carlo-model-based approaches developed by our group for the analysis of both fluorescence and diffuse reflectance spectra to extract the intrinsic fluorescence, absorption, and scattering properties of breast tissues and (2) test the diagnostic value of these intrinsic tissue properties for discriminating malignant from nonmalignant breast tissues. This work is distinct from previous publications by our group^{5, 17, 18} in that first, it reports on the application of a new fluorescence model to recover the intrinsic fluorescence and then extract the contributions of individual fluorescing components that may relate to intrinsic fluorophores in the breast; second, it compares the merits of the extracted fluorescence properties versus the tissue absorption and scattering properties for diagnosing breast cancer; and finally, it demonstrates the consistency of our model-based approach for the analysis of diagnostically useful information contained in fluorescence and diffuse reflectance spectra measured with two different instruments and probe geometries.

2.

Methods

3.

Results

Figure 1 shows the average absorption coefficient [Fig. 1a] and average reduced scattering coefficient [Fig. 1b] as a function of wavelength, for malignant $(n=37)$ , fibrous/benign $(n=8)$ , and adipose $(n=38)$ tissues in sample set 2. For all tissue types, significant hemoglobin absorption was observed at approximately 420, 530, and $575\mathrm{nm}$ . Adipose tissues displayed an additional absorption band in the wavelength range of 440 to $500\mathrm{nm}$ . This absorption band is primarily attributed to beta-carotene, which is present primarily in fat or adipose tissues. The scattering coefficients of all tissue types decreased monotonically with increasing wavelength. The average absorption and reduced scattering coefficients of the three tissue types in sample set 3 displayed similar features, but are not shown here.

Fig. 1

(a) Average absorption coefficient and (b) average reduced scattering coefficient as a function of wavelength, for malignant $(n=37)$ , fibrous/benign $(n=8)$ and adipose $(n=38)$ tissues in sample set 2.

Figure 2 shows the representative raw [Fig. 2a] and intrinsic [Fig. 2b] EEMs of a malignant tissue sample, each normalized to the integrated intensity of the EEM and plotted on a log scale. In comparing the raw and intrinsic EEMs, the most obvious change is the disappearance of the fluorescence peak at the wavelength pair around $(440,510)\mathrm{nm}$ in the intrinsic EEM. In addition, the dip shown in the raw EEM at around $420\mathrm{nm}$ , which is due to hemoglobin absorption, is also corrected for in the intrinsic EEM.

Fig. 2

(a) Raw and (b) intrinsic fluorescence EEMs of a representative malignant tissue sample, each normalized to the integrated intensity of the EEM and plotted on a log scale.

Figure 3 displays the average intrinsic fluorescence spectra obtained at excitation wavelength of 340 [Fig. 3a] and $360\mathrm{nm}$ [Fig. 3b] for malignant $(n=37)$ , fibrous/benign $(n=8)$ , and adipose $(n=38)$ tissues in sample set 2. Malignant and fibrous/benign tissue samples share similar spectral profiles and displayed strong fluorescence at the emission wavelengths around $395\mathrm{nm}$ at $340\text{-}\mathrm{nm}$ excitation and around $450\mathrm{nm}$ at $360\text{-}\mathrm{nm}$ excitation. In adipose tissue samples, there was a weak presence of fluorescence at wavelengths around $395\mathrm{nm}$ , but strong fluorescence at the wavelengths between 450 and $500\mathrm{nm}$ . The shoulder at a wavelength of $420\mathrm{nm}$ in the spectra of malignant and fibrous/benign tissue samples are not fluorescence peaks but rather artifacts that may be introduced from the overcorrection of the Soret band of hemoglobin absorption. The average intrinsic fluorescence spectra of the three tissue types in sample set 3 (not shown here) displayed similar spectral features but the intensities were on a different scale.

Fig. 3

Average intrinsic fluorescence spectra obtained at excitation wavelengths of (a) 340 and (b) $360\mathrm{nm}$ , for malignant $(n=37)$ , fibrous/benign $(n=8)$ , and adipose $(n=38)$ tissues in sample set 2.

A three-component MCR analysis was carried out on the intrinsic fluorescence EEM to extract the relative contribution of the constituent fluorescing components, as well as their intrinsic spectral line shape. Figures 4a to 4c shows the EEMs of the three fluorescing components (denoted as $F_1$ , $F_2$ , and $F_3$ ) obtained from the MCR analysis on intrinsic EEMs of sample set 2. Each intrinsic tissue EEM is a linear combination of the component EEMs. These three components accounted for a vast majority of variance in the tissue EEMs, as the residual was only about 0.2% of the tissue EEM. The first component $F_1$ [Fig. 4a] displayed excitation-emission maximum at a wavelength pair of $\left(\mathrm{340,395}\right)\mathrm{nm}$ . The second component $F_2$ [Fig. 4b] displayed strong fluorescence at the emission wavelengths ranging from 450 to $500\mathrm{nm}$ over an excitation range of 360 to $420\mathrm{nm}$ . The third component $F_3$ [Fig. 4c] had peak fluorescence at an emission wavelength of $480\mathrm{nm}$ at 340 to $360\mathrm{nm}$ excitation. This component was found to be prevalent in adipose tissues.

Fig. 4

EEMs of fluorescing components (a) $F_1$ , (b) $F_2$ , and (c) $F_3$ obtained from a three-component MCR analysis on intrinsic EEMs of sample set 2.

Table 2 shows the correlations between the extracted scattering, absorption and fluorescence (marked with ${}^{*}$ ) properties and the constituent tissue composition, which were evaluated on breast tissues in sample set 1 (note: $+$ , positive correlation; $-$ , negative correlation; $\varnothing$ , no significant correlation). Correlations were considered significant for $p<0.01$ . The mean reduced scattering coefficient, ${\mu }_s^{\prime }$ and the fluorescence contribution from $F_1$ and $F_2$ were negatively correlated with %adipose, while positively correlated with %fibroconnective and %glandular tissue content. The $\beta$ -carotene concentration and the fluorescence contribution of $F_3$ both had significant positive correlations with %adipose content, while they showed negative correlations with %fibroconnective and %glandular tissue content.

Table 2

Correlations between the extracted scattering, absorption and fluorescence (marked with * ) properties and the histological tissue composition, evaluated on breast tissues in sample set 1.

Table 3 shows the pairwise cross-correlation between extracted scattering, absorption (including baseline absorption component) and fluorescence (marked with ${}^{*}$ ) properties of the combined sample sets (note: $+$ , positive correlation; $-$ , negative correlation; $\varnothing$ , no significant correlation). Correlation was evaluated using sample sets 2 and 3. Optical properties obtained from the diffuse reflectance did not show statistically significant correlations with each other. For fluorescence properties, the fluorescence contribution of $F_3$ was negatively correlated with that of $F_1(r=-0.67,p<1\times {10}^{-16})$ and $F_2(r=-0.70,p<1\times {10}^{-16})$ . In addition, the mean reduced scattering coefficient was positively correlated with the fluorescence contribution of $F_2(r=0.23,p<0.01)$ , while negatively correlated with the fluorescence contribution of $F_3(r=-0.26,p<0.005)$ . $\beta$ -carotene concentration had a strong positive correlation with the fluorescence contribution of $F_3(r=0.67,p<1\times {10}^{-16})$ , and negative correlations with that of $F_1(r=-0.56,p<1\times {10}^{-10})$ and $F_2(r=-0.38,p<0.00005)$ . There was also a positive correlation between the total hemoglobin concentration and $F_2(r=0.23,p<0.01)$ , and a negative correlation between hemoglobin saturation and $F_1(r=-0.30,p<0.001)$ . The baseline absorption component did not show statistically significant correlation with other extracted properties except for the fluorescence contribution of $F_3(r=-0.23,p<0.005)$ .

Table 3

Pairwise cross-correlation between extracted scattering, absorption (including baseline absorption component) and fluorescence (marked with * ) properties.

Figure 5 shows the bar graphs of $\beta$ -carotene concentration [Fig. 5a], mean reduced scattering coefficient (mean ${\mu }_s^{\prime }$ ) [Fig. 5b], hemoglobin saturation $({\mathrm{H}}_{\mathrm{b}}\mathrm{O}2\%)$ [Fig. 5c] and relative fluorescence contribution of individual fluorescing components [Fig. 5d], as a function of tissue type. Data shown here were obtained from the combined sample sets (sets 2 and 3). $\beta$ -carotene concentration [Fig. 5a] was statistically higher in adipose breast tissues compare to that in malignant $(p<1\times {10}^{-10})$ and fibrous/benign breast tissues $(p<0.001)$ . The mean reduced scattering coefficients ${\mu }_s^{\prime }$ [Fig. 5b] of malignant tissues were higher than that of fibrous/benign $(p<0.05)$ and adipose breast tissues $(p<1\times {10}^{-9})$ . Hemoglobin saturation [Fig. 5c] was higher in both fibrous/benign and adipose breast tissues in comparison to that in malignant tissues $(p<0.0001)$ . In Fig. 5d, it is shown that relative fluorescence contribution of $F_1$ and $F_2$ were higher in malignant and fibrous/benign tissues compare to that in adipose tissues $(p<0.0001)$ , while the inverse trend was observed for $F_3$ $(p<1\times {10}^{-7})$ .

Fig. 5

Bar graphs of (a) $\beta$ -carotene concentration, (b) mean reduced scattering coefficient (mean ${\mu }_s^{\prime }$ ), (c) hemoglobin saturation $(\mathrm{HbO}2\%)$ , and (d) relative fluorophore contribution of individual fluorophores ( $F_1$ , $F_2$ , and $F_3$ ), as a function of tissue type for samples combined from sets 2 and 3.

Table 4 lists the results from Wilcoxon rank-sum test on extracted absorption, scattering, and fluorescence properties for malignant versus nonmalignant and malignant versus fibrous/benign breast tissues (properties are listed in the order of decreased $p$ value for differentiating between malignant and non-malignant tissues, and fluorescence properties are marked with ${}^{*}$ ). The difference in data distribution is considered statistically significant for $p<0.05$ . All extracted properties, except for total hemoglobin concentration and the baseline absorption component, displayed statistically significant differences between malignant and non-malignant breast tissues $(p<0.0001)$ . However, only hemoglobin saturation and the mean reduced scattering coefficient displayed statistically significant difference between malignant and fibrous/benign breast tissues. This suggests that the fluorescence contrast shown between malignant and nonmalignant breast tissues may be primarily associated with the differences between malignant and adipose tissues.

Table 4

Results from Wilcoxon rank-sum test on extracted absorption, scattering, and fluorescence properties for malignant versus nonmalignant and malignant versus fibrous/benign breast tissues (properties are listed in the order of decreasing p value for differentiating between malignant and nonmalignant tissues; fluorescence properties are marked with * ).

Table 5 shows the results from the (a) holdout validation and (b) leave-one-out cross validation of a linear SVM classification on the combined data sets for discriminating malignant from non-malignant breast tissue samples using: (1) absorption and scattering properties only; (2) fluorescence properties only; and (3) combination of fluorescence, absorption, and scattering properties that show statistically significant differences between malignant and non-malignant breast tissues. The classification rate (C rate), sensitivity and specificity were averaged over repeated training-and-testing trials. In holdout validation [Table 5(a)], classification using absorption and scattering properties alone provided an average sensitivity and specificity of 82.6 and 90.1% respectively in training, and 81.7 and 88.6%, respectively, in testing for discriminating malignant from nonmalignant breast tissues. The classification using fluorescence properties alone yielded slightly higher sensitivity but relatively lower specificity (average SE and SP of 83.4 and 78.7%, respectively, in training, and 82.6 and 75.7%, respectively, in testing) compared to that achieved using absorption and scattering properties alone. It was also shown that combining fluorescence properties with absorption and scattering properties for classification slightly improved the diagnostic accuracy, and the average sensitivity and specificity were improved to 87.6 and 90.6% respectively in training, and 83.9 and 88.6% in testing, for discriminating malignant from non-malignant breast samples. In the leave-one-out cross-validation [Table 5(b)], the classification accuracy for the three cases were comparable to those yielded from the holdout validation. To summarize, the classification using absorption and scattering properties alone and that using combined fluorescence, absorption and scattering properties yielded similar classification accuracy, and both outperformed the classification using fluorescence properties alone.

Table 5

Results from the (a) holdout validation and (b) leave-one-out cross validation of a linear SVM classification on the combined data sets (sets 2 and 3) for discriminating malignant from nonmalignant breast tissue samples using (1) absorption and scattering properties only, (2) fluorescence properties only, and (3) combination of fluorescence, absorption and scattering properties that show statistically significant differences between malignant and non-malignant breast tissues.

Table 6 lists the number of tissue samples from each tissue type that were frequently (repeatedly) misclassified in the SVM classification using the three sets of inputs: (1) absorption and scattering properties only; (2) fluorescence properties only; and (3) the combination of fluorescence, absorption, and scattering properties. The sample number for each of the misclassified tissue samples is listed in parentheses to show samples that were misclassified by more than one algorithm. The majority of misclassified malignant tissues had IDC. DCIS, ILC, and LCIS constitute 25 to 40% of the misclassified malignant samples. This may be partly attributed to the fact that these subtypes of malignancy only account for a very small portion of the malignant samples and may be underrepresented in this data set. Among the misclassified malignant samples, the same LCIS and two IDCs were consistently misclassified using both the fluorescence only and the reflectance only algorithms. The ILCs were misclassified in the reflectance only algorithm, while DCIS was misclassified in the fluorescence only algorithm. More fibrous and benign samples were misclassified when using fluorescence properties alone for classification versus using reflectance properties, and these fibrous or benign samples accounted for the majority of misclassified non-malignant samples. Two of the benign samples, two of the fibro glandular samples and one of the adipose samples were consistently misclassified in reflectance only and fluorescence only classifications. For both malignant and non-malignant cases, most samples misclassified using combined tissue properties overlapped with those misclassified using the reflectance only algorithm.

Table 6

Number of tissue samples and the sample numbers of each tissue type that were misclassified in the SVM classification using the three sets of inputs: (1) absorption and scattering properties only; (2) fluorescence properties only; and (3) combination of fluorescence, absorption, and scattering properties.

4.

Discussion

In this paper, we explored the use of Monte-Carlo-model-based approaches to extract a set of tissue properties, including intrinsic fluorophore contributions and absorption and scattering properties, which include $\beta$ -carotene concentration, total hemoglobin concentration, hemoglobin saturation, and mean reduced scattering coefficient from fluorescence and diffuse reflectance measurements of malignant and nonmalignant breast tissues. For normal samples obtained from breast reduction patients, the mean reduced scattering coefficient decreased with the %adipose tissue content, while it increased with the %fibroconnective and %glandular tissue content. This agrees with findings from previously published studies. For example, Peters ¹⁴ estimated the reduced scattering coefficients at $540\mathrm{nm}$ for normal fibrous and normal adipose breast tissues and showed that fibrous tissues had higher reduced scattering coefficients as compared to adipose tissues. Cerussi ³³ and Durduran ³⁴ also showed in their studies that the scattering coefficients decreased with increasing body mass index (BMI). $\beta$ -carotene concentration was positively correlated with the %adipose tissue content, while negatively correlated with %fibroconnective and %glandular tissue content. This agrees with the fact that the majority of body reserves of $\beta$ -carotene are thought to be in adipose tissues.³⁵ In comparing the malignant, fibrous/benign, and adipose breast tissues, $\beta$ -carotene had an increased concentration in adipose breast tissues compared to that of malignant $(p<1\times {10}^{-10})$ and fibrous/benign breast tissues $(p<0.001)$ . This is expected since $\beta$ -carotene is primarily stored in fatty tissues.³⁵ The hemoglobin saturation showed a significant decrease in malignant tissues relative to that in nonmalignant (including both fibrous/benign and adipose) tissues $(p<0.0001)$ , which is likely due to the oxygen extraction of rapidly proliferating tumor cells. This property agreed with the findings from our previous studies,^{17, 18} and has also been reported in several previous studies using other techniques such as ${\mathrm{pO}}_2$ measurements^{36, 37} and near-IR (NIR) diffuse optical spectroscopy.^{33, 38, 39} Malignant breast tissues had an increased mean reduced scattering coefficient compared to that of nonmalignant breast tissues, which is also consistent with the findings from earlier studies published by Peters ¹⁴ and Ghosh ¹⁶

Three fluorescence components were extracted from the MCR analysis of the intrinsic fluorescence spectra. The first component $\left(F_1\right)$ has peak fluorescence at an excitation-emission pair of $(340,395)\mathrm{nm}$ , which coincides with the primary excitation-emission maximum of collagen.⁴⁰ The EEM of this component is similar to that reported⁴¹ for collagen type 1. Therefore, this component is likely attributed to collagen fluorescence, and in this study we assign $F_1$ as the collagen component. Collagen is the structural protein present in tissue extracellular matrix. Certain collagens, especially type 1 collagen, is a major constituent of the dermis and fibrous stroma of breast.⁴² It was shown that in normal breast tissues the fluorescence contribution of the collagen component decreased with increasing adipose tissue content, while it increased with increasing fibro-connective and glandular tissue content (as seen in Table 2). This agrees with the fact that collagen resides primarily in the dense fibrous stroma of the breast.⁴² Taroni ⁴³ measured absorption of collagen in healthy breast tissues using time-resolved transmittance spectroscopy, and showed that less fatty tissues are characterized by higher collagen concentration. Results of our study also showed that the collagen component has statistically higher contribution to the fluorescence of malignant and fibrous/benign breast tissues, as compared with that of normal adipose tissues $(p<1\times {10}^{-6})$ . Similar observations were also reported in previous studies using other optical spectroscopy techniques for breast tissue characterization.⁴⁴ Haka ⁴⁴ used Raman spectroscopy for the diagnosis of breast cancer, and found that there was increased contribution from collagen in benign and cancerous breast tissues, relative to that in normal breast tissues.

The second component $\left(F_2\right)$ displayed strong fluorescence at the emission wavelengths ranging in 450 to $500\mathrm{nm}$ at excitation wavelengths of 360 to $420\mathrm{nm}$ . The fluorescence feature is spectrally very similar^{40, 41, 45} to that of NADH, thus, the primary fluorophore that contributes to this fluorescent component is likely NADH. This component $F_2$ is thus referred as the NADH component. Like the collagen component, the fluorescence contribution from the NADH component was also statistically higher in malignant and fibrous/benign breast tissues than that in adipose breast tissues. Since NADH is one of the important coenzymes for a large number of metabolic activities in cells, the relative concentration of this fluorophore varies according to the oxidative metabolic status of the cells. The higher fluorescence contribution from NADH component in malignant tissues may result from the increased metabolic activity and oxygen depletion in cancer cells, and thus accumulation of NADH. Uppal and Gupta⁴⁶ carried out an enzymatic measurement of NADH concentration in malignant and normal breast tissues, and found that NADH is significantly higher in malignant tissues as compared with normal breast tissues. In the normal breast, NADH may also increase with the increased metabolism associated with the duct proliferation and secretory activities of the breast lobules, while adipose cells have an indolent metabolism.⁴² Therefore, fluorescence of adipose tissues would be expected to have lower contribution from NADH than that of fibrous/benign breast tissues. This explains the negative correlation observed between the fluorescence contribution of this component and adipose tissue content, as well as the positive correlation observed between this fluorescence component and fibroconnective and glandular tissue content in healthy breast tissues.

The third component $\left(F_3\right)$ displayed an excitation-emission maximum at (360, 480) nm. This fluorescence peak or shoulder was observed in the intrinsic fluorescence spectra of the majority of adipose breast tissues. However, to our knowledge there is very limited work discussing the source of this fluorescent component. Possible sources for this fluorescence feature can be FAD or retinol/vitamin A, both reported to have fluorescence emission maxima at $520\mathrm{nm}$ . We measured the fluorescence EEMs of commercially available FAD (F-6625, Sigma-Aldrich Co.) and retinol acetate (R-7882, Sigma-Aldrich Co.). By comparing the extracted EEM of the third component with those of FAD and retinol, we found noticeable similarity between the third component EEM and that of retinol (Fig. 6 ). Also the relative fluorescence contribution of the third component was highly correlated with the percent adipose tissues content in the breast tissues ( $r=0.83$ , $p<1\times {10}^{-18}$ , evaluated on sample set 1, and $r=0.76$ , $p<1\times {10}^{-16}$ , evaluated on samples set 2, not evaluated on sample set 3 since the percent tissue composition was not available for this sample set), as well as the $\beta$ -carotene concentration ( $r=0.67$ , $p<1\times {10}^{-16}$ , evaluated on combined data sets), which was derived from the Monte-Carlo-inverse-model-based analysis of concomitantly measured diffuse reflectance. $\beta$ -carotene is a precursor of vitamin A, and is primarily stored in adipose tissues.⁴⁷ Based on these observations and facts, we have tentatively attributed this component to retinol, or vitamin A, and referred to $F_3$ as the retinol component. A systematic investigation of this fluorescence component may be required in the future in order to fully identify the source and also explore the potential confounding effects of FAD or other factors that may be present in the tissue.

Fig. 6

Fluorescence EEM of retinol acetate (R-7882, Sigma-Aldrich Co.)

Results from our study suggest that there is a statistically significant increase in fluorescence contribution from retinol component in normal adipose breast tissues compared to that in other tissue types, which may indicate the prevalence of retinol content in normal adipose tissues. Since retinol is fat-soluble, it is primarily present in adipose tissues, and most women carry stores of retinol or vitamin A in their fat cells. Lunetta ³⁵ have estimated and compared the retinol concentrations in normal and cancerous breast tissues in breast cancer patients, and found that the retinol concentration in normal breast tissues was slightly higher (on average) than that in malignant breast tissues, however the differences were not statistically significant.

The negative correlation observed between the fluorescence contribution of collagen and NADH components versus the $\beta$ -carotene concentration may be attributed to fact that collagen and NADH is found predominantly in fibroglandular tissue while $\beta$ -carotene is found in fatty tissues. The mean reduced scattering coefficient has a positive correlation with the fluorescence contribution of NADH component, while it has a negative correlation with the fluorescence of the retinol component. This may be due to decreasing scattering,^{14, 34, 48} decreased NADH fluorescence,⁴² and increased retinol fluorescence in adipose tissues. The physiological basis for the negative correlation observed between fluorescence contribution of the collagen component and hemoglobin saturation is not clear. However, fluorescence of the collagen component was found to be prevalent in malignant tissues, which has lower hemoglobin saturation than normal tissues. This may contribute to the correlation observed between the two variables.

In this study, we extracted the contribution of absorbers, scatterers, and fluorophores in breast tissue and correlated them with histological changes in the normal breast as well as histopathology of the breast. The trends observed in these extracted tissue properties with respect to different tissue types and histopathology findings are consistent with some known facts (such as prevalence of $\beta$ -carotene and fat in breast tissues) or results reported in other studies, which indicates that the optically determined tissue properties reflect the intrinsic changes in biochemical and physiological properties in breast tissues. However, the limitation of this study is that it was not possible to directly compare the extracted tissue properties to tissue constituent concentrations, because this is a study involving clinical samples, on which disruptive biochemical assays cannot be performed. However, we are currently doing a study to compare the extracted tissue properties to immunohistochemical assays (for example, hypoxia, microvessel density, collagen distribution, etc.) to demonstrate that the optical technique is nondestructively able to quantify biological and morphological information from breast tissues. This will provide a means to directly validate the quantitative physiological aspects of this technique.

The intrinsic fluorescence spectra of each tissue sample were normalized to the integrated intensity over the entire EEM prior to MCR analysis. This normalization removed the intersample variations in the spectral intensity so that the extracted fluorescence contribution only reflected the fractional contributions of individual fluorophores to the fluorescence of each tissue sample. MCR analysis was also carried out on un-normalized intrinsic fluorescence spectra from set 2 only, and it was observed that the relationships of the extracted fluorophore contribution with different tissue types were the same as those observed with the normalized spectra.

The relative fluorescence contribution of all three fluorescing components displayed statistically significant differences between malignant and nonmalignant breast tissues, these differences, however, may be associated with the differences between malignant and adipose tissue only, as a large portion of non-malignant breast samples are adipose tissues. The fluorescence spectra of malignant and fibrous/benign tissue samples shared similar spectral line shapes, and the Wilcoxon rank-sum test did not show statistically significant difference between malignant and fibrous/benign breast tissues in the fluorescence contributions of individual components. In the holdout validation, the classification based on fluorescence contributions alone achieved an average sensitivity and specificity of 83.4 and 78.7%, respectively in training, and 82.6 and 75.7%, respectively, in testing for discriminating malignant from nonmalignant breast tissues. The separation boundary yielded from this classification may discriminate primarily between malignant and normal adipose tissues and a retrospective look at the misclassified samples indicate that most of the misclassified non-malignant samples are fibrous or benign tissues.

Classification using absorption and scattering properties that displayed statistically significant differences between malignant and non-malignant breast tissues provided a slightly decreased sensitivity (82.6% in training and 81.7% in testing using holdout validation) but higher specificity (90.1% in training and 88.6% in testing using holdout validation) for discriminating breast malignancy, as compared to that achieved using fluorescence properties alone. The improvement to the specificity may primarily be attributed to the inclusion of hemoglobin saturation and mean reduced scattering coefficient, especially the former, which displayed statistically significant differences between malignant and fibrous/benign breast tissues.

The comparable classification performance using fluorescence properties, absorption and scattering properties alone, or both suggest that fluorescence and diffuse reflectance spectra both yield diagnostically useful information for the discrimination of breast malignancy. The retrospective review of the misclassified tissue samples (Table 6) shows that most samples misclassified using fluorescence properties were different from those misclassified using absorption and scattering properties, suggesting the diagnostically useful information yielded from reflectance and fluorescence could be complementary. Although in this study a direct combination of the two sets of tissue properties did not significantly improve the overall classification accuracy, a strategy using absorption and scattering properties and fluorescence properties separately (i.e., sequentially) may have the potential to improve the overall classification.

The model-based approach for feature extraction enables a full exploration of the physiologically relevant information contained in the fluorescence and diffuse reflectance spectra, thus providing a comprehensive understanding of the biochemical, physiological, and morphological changes that take place in the tissue, which has implications in applications such as monitoring tumor response to therapy where both metabolism and hemoglobin oxygenation, for example, are important. Besides providing physiologically meaningful information about the tissue composition and pathology, another advantage of the model-based approach for the analysis of tissue spectra is that it can eliminate the instrument dependency, therefore results obtained using different instruments and probes are directly comparable. In our study, we combined the fluorescence, absorption, and scattering properties extracted independently from two sets of tissue spectra, which were measured with two different instruments and fiber optic probes. The extracted properties were comparable and a statistical $t$ test showed that there was no statistically significant difference between the two data sets. This enabled an increased sample size by combining the two data sets.

In this study, the sensitivity and specificity for discriminating breast malignancy were evaluated using two validation methods, i.e., holdout validation and leave-one-out cross-validation, both of which provide an unbiased evaluation of the classification accuracy and the robustness of the classification algorithm. The classification accuracies were consistent across the repeated trials in which training samples were randomly selected from the entire sample set (combined sets 2 and 3). The classification accuracy achieved using holdout validation and leave-one-out cross validation were also comparable.

In conclusion, we presented Monte-Carlo-model-based approaches for the analysis of fluorescence and diffuse reflectance spectra of breast tissues, which enable the extraction of intrinsic fluorescence, absorption, and scattering properties that provide the biochemical and morphological information about the tissue. The diagnostically useful absorption/scattering properties and fluorescence properties can be used alone or in combination for the discrimination of breast malignancy. It was also demonstrated that the model-based approach has the advantage of eliminating the dependency of fluorescence and diffuse reflectance measurements on the instrument and probe geometry, which makes it a more generalized approach for the analysis of tissue spectra.