1.

Introduction

Cancers of the brain and central nervous system (CNS) are recognized as global threats. According to the global cancer statistics, 238,000 cases of cancers of the brain and CNS have been recognized worldwide. The overall estimated mortality of 1.9%, with nearly an equal male to female ratio is caused by brain and CNS tumors.¹ The rate is higher in developed countries compared to developing or under-developed countries. The lower rate of mortality reported in the less-developed countries may be due to the lack of intensive diagnostic facilities which leads to lower incidence rates being reported.² Genetic susceptibility, exposure to ionizing radiation, dietary intake, lifestyle factors, family history, and immune factors are proposed as the main etiological factors.³ As the brain is the most sensitive organ in the human body, brain tumor patients require intensive medical care. The cost for brain tumor prognosis is nearly 25% of the entire annual cost of the care for overall cancer patients.⁴

Excision biopsy followed by histopathology is the widely accepted gold standard for tumor screening. But this technique is time consuming and restricted to limited statistical confidence level due to operator variability.⁵ Minimally invasive and ex vivo optical spectroscopic techniques based on fluorescence and absorbance of tissues such as fluorescence, diffuse reflectance, infrared, and Raman spectroscopy are the emerging tools in tumor diagnosis. These techniques give promising results in tumor differentiation.^6–10

Fluorescence and diffuse reflectance spectroscopic techniques are cost effective compared with other spectroscopic techniques, and hence affordable to financially needy patients. Early diagnosis and even minute variations in the biochemistry of the tissue can be easily detected with these techniques. Analysis of the fluorescence peaks corresponding to major fluorophores can be utilized for the effective monitoring of biochemical changes in the tissue.^8–13 These fluorophores include collagen, nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), phospholipids, and porphyrin. Additionally, the use of diffuse reflectance spectroscopy can monitor the variation in hemoglobin, which is an absorber within the tissue.^11,14–18

For many years, application of fluorescence spectroscopy has become an area of interest among many clinical and oncologic research groups. These studies demonstrate the variations in endogenous fluorophores of malignant and normal tissues using in vivo and ex vivo fluorescence spectroscopy.^9–11 But studies relating the fluorescence behavior of the fluorophores of brain tumor tissues are limited.^12,19 Available reports are on the fluorescence study of 5-aminolevunic acid (ALA) induced protoporphyrin synthesis in the brain tumor.^20,21 These studies are inadequate for judging the exact biochemical alterations that occur in brain tissues during malignant transformation.

Hemoglobin is a nonfluorescent metalloprotein.^22,23 Hemoglobin peak in the autofluorescence spectra of tissues has been considered as an anomaly by a majority of the researchers.^8,11,24,25 In order to correct these artifacts in fluorescence spectra due to blood absorption, many research groups have adopted mathematical correction factors assisted by the results of diffuse reflectance spectroscopic data.^11,25,26 Mallia et al. and de Veld et al. have reported a first-order approximation method that involves division of fluorescence spectral data with the diffuse reflectance data from the same tissue site to remove the hemoglobin absorption artifacts.^11,25 Thus, the possibilities of using fluorescence spectroscopy for evaluating the changes in the hemoglobin concentration have not been exploited in detail. In 2009, Liu and Vo-Dinh proposed a ratio-metric method based on spectral filtering modulation (SFM) for the estimation of total hemoglobin and oxygen saturation using synthetic tissue phantom models.²⁷ However, this has not been demonstrated experimentally on human tissue so far. In the present study, we have taken up this theory and tried to apply it practically and propose to evaluate the total hemoglobin concentration in brain tumor lesions. In addition, we also propose to evaluate the biochemical changes of prominent endogenous fluorophores such as collagen, FAD, phospholipids, and porphyrin within the brain tissue. For this, we have analyzed adjacent normal and tumor tissues from patients with intra-axial tumors like astrocytoma and glioma and extra-axial tumors like meningioma and schwannoma. The exact differentiation of the tumor and adjacent normal tissue among different groups was evaluated using linear discriminant analysis (LDA). Here, we report the efficacy of autofluorescence spectroscopy to serve as a single entity to discriminate tumor tissues from normal ones by analyzing the variations in both endogenous fluorophores and chromophores involved.

2.

Materials and Methods

2.3.

Data Processing and Analysis

2.3.1.

Processing of spectra

All spectra were baseline corrected and then normalized with respect to the maximum intensity of the peaks at $460\pm 10$ and $500\pm 10\mathrm{nm}$ for the 320 and 410 nm excitations, respectively. From the normalized data, peak intensity corresponding to the major peaks was extracted. Using SPSS-17 (SPSS Inc., Chicago, Illinois), one way analysis of variance (ANOVA) was carried out for checking the variations in the prominent fluorescence peak intensity between the adjacent normal and tumor tissues.

2.3.2.

Total hemoglobin concentration estimation

Schematic representation of SFM effect on autofluorescence spectra due to hemoglobin absorption from brain tumor tissues is given in Fig. 1. The total hemoglobin concentration was determined by considering the ratio of fluorescence emission intensities at 500 and 570 nm of the 410 nm excited spectra. The extinction coefficients of oxygenated and deoxygenated hemoglobin at these wavelengths are equal. Hence the ratio between these provides the total concentration of hemoglobin. The calculations were done using the following equations as per the method given by Liu and Vo-Dinh.²⁷

$$\text{Total hemoglobin concentration}={\mathrm{FI}}_{500}/{\mathrm{FI}}_{570}.$$

Fig. 1

Schematic representation of spectral filtering modulation (SFM) effect on autofluorescence spectra due to hemoglobin absorption in brain tissue.

2.3.3.

Multivariate analysis

LDA is a classical multivariate statistical analysis approach for dimensionality reduction. It computes an optimal projection by minimizing the “within-class distance” and maximizing the “between-class distance” simultaneously achieving maximum class discrimination. In this study, we performed LDA on normalized spectra in the range of 350 to 600 nm and 440 to 750 nm for 320 and 410 nm excitations, respectively, using SPSS. Discriminant analysis was carried out on the normalized datasets which transformed the scores into a single discriminant score. The classification based on “leave one out” method of cross-validation produces a confusion matrix that compares predicted versus actual group membership. This classification result was used to calculate sensitivity and specificity as per previously reported methods.^13,28

3.

Results

3.1.

Fluorescence Spectral Features

Autofluorescence spectra at two different excitation wavelengths, 320 and 410 nm, were collected and considered for the study. Broad peaks around 380, 460, and 560 nm were observed for the 320-nm excited spectra. Spectra corresponding to this excitation are given in Fig. 2. At 410-nm excitation, five distinct peaks around 500, 560, 590, 630, and 695 nm were observed for all the samples (Fig. 3). A marked difference in the fluorescence peaks were observed between tumor and adjacent normal tissue for both excitations. In both excitations the spectra were normalized with respect to the peaks at 460 and 500 nm, respectively, and hence these peaks are not considered for the analysis.

Fig. 2

Averaged fluorescence emission spectra of adjacent normal and tumor samples of (a) astrocytoma, (b) glioma, (c) meningioma, and (d) schwannoma at 320 nm excitation.

Fig. 3

Averaged fluorescence emission spectra of adjacent normal and tumor samples of (a) astrocytoma, (b) glioma, (c) meningioma, and (d) schwannoma at 410 nm excitation.

3.1.1.

Spectral features of 320 nm excited spectra

An intense peak at 380 nm is observed for both astrocytoma and glioma. A shoulder around 510 nm for the peak at 460 nm and a weak peak around 560 nm are also observed for these tumors. A similar pattern is observed for the adjacent normal tissue also. But the peak around 380 nm is more intense, whereas the peak around 510 nm is less intense compared to the tumor tissue. The 560 nm peak is more prominent in the case of tumor tissue compared to the adjacent normal. For glioma, the peak around 380 nm is weak for the tumor tissue compared to the adjacent normal. The shoulder at 510 nm is not well resolved for the tumor tissue compared to that of adjacent normal. But the peak around 560 nm appears to be more resolved than the normal tissues.

An entirely different spectral pattern is observed for extra-axial tumor tissues. For meningioma and corresponding adjacent normal tissues at 320-nm excitation, the peak around 380 nm is many fold more intense than the other peaks. For the tumor tissue, this peak is considerably weaker compared to that of normal tissue. The peak around 510 nm is absent, and the peak at 560 nm is very weak for the tumor tissues, whereas well-resolved peaks are observed around 510 and 560 nm for the adjacent normal tissue. Similar to meningioma, spectra of schwannoma also have the most intense peak around 380 nm. Here also, this peak is much weaker in tumor compared to normal tissues. Well-resolved peaks are observed around 510 and 560 nm for adjacent normal compared to the tumor tissues.

ANOVA was performed on the maximum peak intensity values and the results are given in Table 2. A significant difference ($p<0.05$) in the intensity is observed for 380 nm peak of glioma and adjacent normal. But in the case of astrocytoma ($p=0.142$), meningioma ($p=0.085$), and schwannoma ($p=0.120$) intensity difference for this peak is not significant. Differences among other peaks of 320 nm excited spectra are also found to be insignificant.

Table 2

One way analysis of variance (ANOVA) on comparison of intensity values between adjacent normal and tumor (p < 0.05 are given in bold letters).

Tumor type	320 nm excitation			410 nm excitation
	380	510	560	560	590	630	695
Astrocytoma	0.142	0.543	0.125	$\mathbf{0.000}$	$\mathbf{0.023}$	0.979	0.964
Glioma	$\mathbf{0.001}$	0.895	0.122	$\mathbf{0.000}$	$\mathbf{0.000}$	0.786	0.720
Meningioma	0.085	0.165	0.343	$\mathbf{0.002}$	0.057	0.893	0.281
Schwannoma	0.120	0.997	0.656	$\mathbf{0.024}$	0.069	0.273	0.534

3.1.2.

Spectral features of 410 nm excited spectra

Here, adjacent normal tissue showed well-defined peaks when compared to the tumor tissue in all the four cases. A varied intensity pattern is observed for the different peaks among different tissues. Considering the peak around 500 nm, a blue shift is observed in the case of tumor tissue compared to the normal. This is true for all the tumor types considered. When the whole spectrum is analyzed, the spectral intensity around 560 and 590 nm is much less compared to corresponding adjacent normal tissues. For astrocytoma and glioma, the peak around 560 nm is more prominent whereas it is not resolved for meningioma and schwannoma. Well-defined highly resolved peaks are observed around 590 nm for all the adjacent normal tissues. Interestingly, this peak disappears from the tumor spectra. The peaks at 630 and 695 nm are of nearly equal intensity for both adjacent normal and tumor tissues in all cases. But a significant difference in peak intensity is observed for these peaks among different tumor types considered (Fig. 3). For extra-axial brain tumors and its adjacent normal tissues, the peak at 695 nm is not resolved and appears as a valley whereas characteristic medium intense and well-resolved peaks are observed in intra-axial brain tissues.

ANOVA gives a significant difference in the peak intensity ($p<0.05$) for all the four tumor types when the peak centered at 560 nm is considered. The 590 nm peak shows significant difference in the peak intensity for astrocytoma ($p=0.023$) and glioma ($p=0.000$), but not for meningioma ($p=0.057$) and schwannoma ($p=0.069$). The intensity variation of the peaks at 630 and 695 nm was also found to be insignificant among the four types of tumor.

3.2.

Total Hemoglobin Estimation

Total concentration of hemoglobin values calculated for astrocytoma, glioma, meningioma, schwannoma, and the corresponding adjacent normal tissues are given in Fig. 4. It is clear that there is considerable increase in the total hemoglobin concentration for all tumor tissues compared to the normal ones. This increase is found to be statistically significant ($p<0.005$) irrespective of the tumor type.

Fig. 4

Total hemoglobin concentration of adjacent normal and tumor tissue. Data shown are in the form $\text{mean}+\mathrm{SD}$. Significance (**$p<0.005$) of differences between adjacent normal and tumor tissues.

3.3.

Linear Discriminant Analysis for Tissue Classification

The pairwise discrimination score for astrocytoma, glioma, meningioma, schwannoma, and corresponding adjacent normal tissues are given in Figs. 5 and 6. From these plots, it is clear that the intergroup distance for tumor and adjacent normal is very well separated among the various groups. It is noted that the intergroup distance between the glioma and adjacent normal is highly significant compared to astrocytoma, meningioma, and schwannoma. This is more prominent in the case of data based on the 410 nm excitation. A discrimination line drawn at 0 gives a good differentiation. In the case of 320 nm excited spectra, the discrimination score for tumor is observed considerably above the discrimination line, and for adjacent normal it is below the discrimination line. But for the 410 nm excitation, the pairwise discrimination score for tumor is considerably below the discrimination line, and for adjacent normal tissue it is considerably above. Using the position of pairwise discrimination scores, the number of true positive and negative values as well as the number of false positive and negative values was obtained. Results from this binary classification are utilized for the discrimination of tissues (Table 3).

Fig. 5

Pairwise discriminant plot based on linear discriminant analysis (LDA) for adjacent normal and tumor samples for 320 nm excitation.

Fig. 6

Pairwise discriminant plot based on LDA for adjacent normal and tumor samples for 410 nm excitation.

Table 3

Classification efficacy obtained using linear discriminant analysis (LDA) for the differentiation of tumor and adjacent normal tissues.

Tumor type	320 nm excitation		410 nm excitation
	Sensitivity (%)	Specificity (%)	Sensitivity (%)	Specificity (%)
Astrocytoma	93.33	95	93.33	95
Glioma	100	93.33	100	93.33
Meningioma	83.33	91.31	100	91.31
Schwannoma	91.67	78.57	91.67	92.86

4.

Discussions

In this study, we have attempted to differentiate tumor and adjacent normal brain tissue based on the changes observed in the autofluorescence spectra. We have also utilized diagnostic algorithms to interpret the wealth of spectral information for the exact differentiation of brain tumor tissues from normal. Preprocessing methods like baseline correction and normalization are also done in all cases to get better efficiency in the classification. The importance of preprocessing of spectral data has been reported earlier.¹³ We have adopted the method of normalization with respect to specific peak intensity which gives better results in classification. Hence, the relative concentration of fluorophores is considered throughout this study. An excitation wavelength of 320 nm is ideal for picking up the fluorophores collagen, NADH, and FAD whose emissions are expected around 380, 460, and 510 nm, respectively. Excitation at 410 nm gives the emission values of FAD and phospholipids around 500 and 590 nm, respectively. This excitation also gives the emission of porphyrin around 630 and 695 nm.^8–13 The fluorescence peaks of this study are assigned accordingly.

Variation of collagen in tumor, especially in brain tumor, is a hot area of discussion.⁶ In the present study, the collagen level as observed by the peak around 380 nm is less intense for tumor tissues than the normal one in all the cases. Also among the four tumor types considered, extra-axial tumors and its adjacent normal show a high level of collagen compared to intra-axial tumor tissues. The reason for this is intra-axial brain tissues have only 20% of extracellular matrix (ECM) compared to its total volume, whereas extra-axial brain tissues roughly have a balance of 50% to 50% between ECM and cells.²⁹ When the collagen level between the tumor and adjacent normal is considered, a decrease in collagen level is observed in all the four types. Using protein secondary structure determination of infrared spectra, Noreen et al. have observed that the level of fibrillar collagen (collagen I, III, and V) decreases and nonfibrillar collagen (collagen IV and VI) increases in brain tumor compared to healthy tissues.⁶

The decrease in collagen level may be due to the degradation of ECM which occurs during tumor cell invasion process to facilitate migration of neoplastic cells to the surrounding tissues.^30,31 Specifically in the case of brain tumors, overexpression of matrix metalloproteinase enzyme will degrade fibrillar and nonfibrillar collagens leading to the degradation of ECM.³² Due to the changes in the ECM protein components, tumor angiogenesis and tumor cell proliferation and invasion occur in brain.^30,31,33 The current findings lead to the conclusion that total collagen level in brain decreases as tissue transforms from normal to tumor.

In the 410 nm excited spectra, the phospholipid peak at 590 nm is less intense for tumor tissue than for the normal, indicating a decrease in the lipid level in the tumor tissue. A reduction in the total lipid level in brain tumor has been reported earlier by Köhler et al.³⁴ A possible reason for this can be lipid peroxidation that occurs in brain tumor as a result of the decrease in antioxidant enzymes such as catalase, superoxide dismutase, and glutathione reductase.^35,36 Decrease in such antioxidant enzymes induces oxidative stress in brain tissues and results in lipid peroxidation.

The fluorescence peaks of porphyrin are observed around 630 and 695 nm in the 410 nm excited spectra. In this study, the level of porphyrin is found to be highest for astrocytoma, followed by glioma, meningioma, and schwannoma. A similar pattern is observed for the normal tissues as well. These peaks are very weak and not resolved in the case of extra-axial tissues compared to intra-axial tissues.

Comparison of Fig. 3(a) and 3(b) to that of 3(c) and 3(d), it is clear that the level of porphyrin is less in extra-axial tissues compared to intra-axial tissues. It is also seen that the porphyrin level does not vary much between the tumor and normal tissue, irrespective of the tissue types. But an increase in porphyrin level in brain tumor has been reported in ALA-induced fluorescence method. This is due to the fact that ALA gets converted to proptoporphyrin by the ferro chelates present in the tumor.^20,21 No other reports are available on the level of porphyrin in brain tumors.

For biological samples, the emission of FAD is expected in the 500 to 530 nm range for both excitations considered.^10–13 But in the 320-nm excited spectra, we observed a blue shift for this peak from the expected value of 520 nm. The same is observed again in the 410-nm excited spectra, but only for the tumor samples. In the study using synthetic tissue phantoms, Liu and Vo-Dinh have reported a blue shift in the FAD peak on increasing the concentration of hemoglobin.²⁷ In the present study it is also very clear from the 410-nm excited spectra that the blue shift is noticeable in cases of higher hemoglobin concentration (Fig. 3). So, we ascribe the observed shift as due to the increase in hemoglobin concentration in the tumor tissue.

The peak appearing at 560 nm is assigned to the hemoglobin absorption. So far, this peak has not been considered seriously by any research groups for differentiating tissue biochemistry. Many research groups have reported this peak as an anomaly.^8,11,24,25 But, Liu and Vo-Dinh, using synthetic fluorescent tissue phantoms, demonstrated that this peak can be used to estimate the total hemoglobin concentration.²⁷ This theory has been later demonstrated experimentally on tissue phantoms with varying FAD concentrations and glioma-induced rats.³⁷ In this study, it is demonstrated that the emission path length of fluorescence light is a crucial factor in the determination of total hemoglobin concentration. Emission path length of fluorescence light depends on the illumination and collection geometry of the optical probe. This path length must be long enough, so that the effect of absorption on fluorescence intensity due to hemoglobin is significant. It is also found that SFM effect on the fluorescence spectra due to hemoglobin and its variation is clearly identifiable around 560 nm when excited around 410 nm. This justifies the selection of 410 nm excitation in our present study for the estimation of total hemoglobin concentration.

In brain tumor, the total hemoglobin concentration is an index of angiogenesis.^38,39 To date, no reports are available on the evaluation of total hemoglobin concentration of different types of brain tumor lesions using fluorescence spectroscopy. Using diffuse reflectance spectroscopy studied ex vivo and in vivo, various groups have reported that total hemoglobin concentration increases when the tissue becomes malignant.^40,41 Our results on SFM effect on the fluorescence spectra due to hemoglobin absorption show significant increase in hemoglobin concentration in the tumor tissue which is in agreement with these findings. From the observations, we suggest that fluorescence spectroscopy is an ideal tool for the effective quantification of total hemoglobin concentration within brain tissue.

Although the explained method is promising, we envisage some limitations that would have affected the results of the study due to certain factors which need to be discussed. As is seen from Fig. 4 where hemoglobin concentration is demonstrated, the error bars appears to be high. This may be because of the fact that we have considered adjacent tissue of the tumor as normal ones for the analysis, as it is impossible to have actual normal tissue of the human brain for the study. Hence, there is very high chance that the adjacent tissue considered here had pathological transformations, at least in some cases, which could be the reason for the large deviations observed in these results. Moreover, the deviations observed in hemoglobin concentration is not affecting the classification of the tissue as is evidenced by the significant difference in the ANOVA values between adjacent normal and tumor tissues.

The differentiation of normal from tumor tissue using LDA yielded a sensitivity of 83% to 100%, 93% to 100% and overall specificity of 91% to 95%, respectively, for 320 and 410 nm excited spectra. A direct correlation between the intergroup distance between tumor and normal tissue, and the grading of the tumor is very clear from Figs. 5 and 6. Intergroup distance is maximum between glioma and adjacent normal, whereas it is minimum for schwannoma and corresponding normal. The discrimination is clearer with the dataset of 410 nm excited spectra. This further projects the efficacy of excitation wavelength around 410 nm in the evaluation of hemoglobin concentration than that of 320 nm. As by Liu and Vo-Dinh, excitation wavelength around 410 nm is the most suitable for evaluation of hemoglobin concentration than any other excitation wavelength.¹⁷

Previously, Butte et al. have reported an overall sensitivity of 25% to 100% and specificity of 78% to 98% to differentiate normal brain tissues, low grade glioma and high grade glioma using fluorescence lifetime spectroscopy.⁴² In a study using fluorescence and diffuse reflectance spectroscopy, Lin et al. have reported an overall sensitivity and specificity of 76% to 100% to discriminate normal from brain tumor tissues.¹⁹ Another study by Saraswathy et al. reported a sensitivity and specificity of 86% to 100% and 66% to 100% using principal component analysis in differentiating adjacent normal from brain tumor tissue.¹² In the present study, we report a better sensitivity and specificity than the previously reported ones. Moreover, we have considered not only the endogenous fluorophores but also the chromophore, hemoglobin using the single technique in achieving this discrimination. The current method provides high sensitivity and specificity compared to other studies reported earlier. These results can be considered further in the development of minimally invasive methodology based on fluorescence spectroscopy which will help to differentiate brain tumor, in the future.

5.

Conclusions

Emissions from endogenous fluorophores and hemoglobin can be utilized for the effective characterization and classification of different types of brain tumors and normal tissues. Inexpensive and easy to handle techniques of autofluorescence spectroscopy can be utilized for the evaluation of both fluorophores and chromophores simultaneously. As this method can directly be used to monitor hemoglobin variation, separate equipment and measuring protocols are not required for diffuse reflectance and fluorescence measurements. The variations in the collagen and lipid levels due to the degradation of ECM and lipid peroxidation in tumor tissues can also be used to discriminate the tissue biochemistry. Additionally, data analysis methods like LDA further support the discrimination between normal and tumor tissues. A study focused on larger sample set is necessary to facilitate the use of this technique in clinical set up. Further studies in this direction are in progress, and we believe that the application of this technique in clinical practice will be a reality in the very near future.