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 β
-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 et al.14
estimated the reduced scattering coefficients at 540 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 et al.33
and Durduran et al.34
also showed in their studies that the scattering coefficients decreased with increasing body mass index (BMI). β
-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 β
-carotene are thought to be in adipose tissues.35
In comparing the malignant, fibrous/benign, and adipose breast tissues, β
-carotene had an increased concentration in adipose breast tissues compared to that of malignant (p
) and fibrous/benign breast tissues (p
<0.001). This is expected since β
-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 pO2
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 et al.14
and Ghosh et al.16
Three fluorescence components were extracted from the MCR analysis of the intrinsic fluorescence spectra. The first component (F1
) has peak fluorescence at an excitation-emission pair of (340, 395) 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 F1
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 ). This agrees with the fact that collagen resides primarily in the dense fibrous stroma of the breast.42
Taroni et al.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
). Similar observations were also reported in previous studies using other optical spectroscopy techniques for breast tissue characterization.44
Haka et al.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 (F2
) displayed strong fluorescence at the emission wavelengths ranging in 450 to 500 nm at excitation wavelengths of 360 to 420 nm. The fluorescence feature is spectrally very similar40,41,45
to that of NADH, thus, the primary fluorophore that contributes to this fluorescent component is likely NADH. This component F2
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 (F3
) 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 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 (). Also the relative fluorescence contribution of the third component was highly correlated with the percent adipose tissues content in the breast tissues (r
, evaluated on sample set 1, and r
, 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 β
-carotene concentration (r
, evaluated on combined data sets), which was derived from the Monte-Carlo-inverse-model-based analysis of concomitantly measured diffuse reflectance. β
-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 F3
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.
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 et al.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 β
-carotene concentration may be attributed to fact that collagen and NADH is found predominantly in fibroglandular tissue while β
-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 β-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 () 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.