14289c03.PGS
6/27/06
12:11 PM
Page 348
Photomedicine and Laser Surgery Volume 24, Number 3, 2006 © Mary Ann Liebert, Inc. Pp. 348–353
Near-Infrared Raman Spectroscopy for Oral Carcinoma Diagnosis ANA PAULA OLIVEIRA, M.Sc.,1 RENATA ANDRADE BITAR, M.Sc.,2 LANDULFO SILVEIRA, JR. Ph.D.,1 RENATO AMARO ZÂNGARO, Ph.D.,1 and AIRTON ABRAHÃO MARTIN, Ph.D.2
ABSTRACT Objective: Fourier-transform (FT)–Raman spectroscopy has been used to explore the changes in the vibrational bands of normal, dysplastic (DE) and squamous cell carcinoma (SCC) tissues. Background Data: Raman spectroscopy has been applied as a diagnostic tool for the detection of cancers, due to its sensitivity to the changes in molecular composition and conformation that occurs in malignant tissues. The detection of weak Raman signals from biotissues becomes easier by FT-Raman due to fluorescence suppression. Methods: A carcinogen (7,12-dimethybenz[a]anthracene [DMBA]) was applied daily in the oral pouch of 21 hamster to induce oral carcinoma. After 14 weeks, the fragments of squamous cell carcinomas and oral normal tissue were collected and analyzed by FT-Raman spectroscopy, using a 1064-nm Nd:YAG laser line as an excitation source. A total of 123 spectra were obtained and divided in normal and malignant tissue groups, and analyzed statistically through principal components analysis (PCA) and classified using Mahalanobis distance. Results: Major differences between normal and malignant spectra seem to arise from the composition, conformational, and structural changes of proteins, and possible increase of its content in malignant epithelia. An algorithm based on PCA was able to separate the samples into two groups—normal and carcinoma. For the algorithm training group, 91% sensitivity and 69% specificity were observed, while the prospective group had 100% sensitivity and 55% specificity. Conclusion: The algorithm based on PCA has the potential for classifying Raman spectra and can be useful for detection of dysplastic and malign oral lesion.
INTRODUCTION
O
RAL CANCER is one among the 10 most common head and neck cancers, and its incidence is considerably significant. The death rate for oral cancer is higher than that of cervical cancer, Hodgkins disease, cancer of the brain, liver, testes, and kidney, or skin cancer (malignant melanoma).1 From 90% to 96% of oral cancers are squamous cell carcinomas.1 Survival rates for oral cancers after 5 years are reported to be about 55%. Twenty percent of mortality is reported to be a direct consequence of malignancy, and another 20% due to secondary cancers, even after apparent cure. The poor survival rate, despite advances in surgical and therapeutic modalities, is attributed to late detection of disease.1 Early detection of neoplastic changes using optical spectroscopy has been one of the most active areas of research in
recent years. Raman spectroscopy is one of these optical methods that could permit less invasive and nondestructive analysis of biological samples, allowing one to get precise information on biochemical composition from different types of human tissues. Raman spectra are obtained by exciting the molecules in the sample by a laser beam. The inelastic scattering light results in a frequency shift in the Raman spectra. Once these frequency shifts depends on the type of molecules, the Raman spectra holds important information on the different biochemical compounds. All stages of cancer are marked by fundamental changes in cellular morphology and/or tissue biochemistry, and biochemical tumor markers such as proteins, enzymes, and hormones could be detected by analyzing the differences of Raman spectra obtained from normal and pathologic tissues.2–12 Raman spectral differences could be used as a clinical method to provide real-time med-
1Grupo de Óptica Biomédica and 2Laboratório de Espectroscopia Vibracional Biomédica, Instituto de Pesquisa e Desenvolvimento (IP&D), Universidade do Vale do Paraíba (UNIVAP), São José dos Campos, SP, Brazil.
348
14289c03.PGS
6/27/06
12:11 PM
Page 349
Near-Infrared Raman Spectroscopy for Oral Carcinoma Diagnosis ical diagnosis less invasively, as well as, important information about the tissue biochemistry.4,8,9,12 Raman spectroscopy has been used in biology and biochemistry to study the structure and dynamic function of important biomolecules. A minor disadvantage of Raman spectroscopy is the autofluorescence emitted from biological tissues that could be almost completely suppressed by using excitation radiation in the near-infrared region. The main excitation source to obtain a fluorescence-free spectra in the FT-Raman is the Nd:YAG laser at 1064 nm. Researchers have applied multivariate statistics in the spectral data for tissue classification. Principal components analysis (PCA) is the most widely used statistical treatment and has been applied to classify spectroscopy data from a variety of human diseases such as Alzheimer,13 oral tissues,8 cervical precancer,14 human breast cancer,15 skin cancer,16 and human atherosclerotic lesions,17 detecting spectral alterations that occur after changes in the morphology and physiology of biological tissues. Within a set of spectral data, there are usually many different variations that make up a particular spectrum. Background, differences in constituents, instrument variations, sample handling, and others affect the appearance of the final spectrum. Yet with several changes occurring at the same time, there are only a few independent variables accounting for all spectral differences. The largest variation in the spectral data is due to differences in the constitution of the samples, meaning few variables. One can achieve reduction of variables by using PCA. PCA extracts the relevant information from the original data (A) and generates a new set of variables, called principal components (PC) and scores (S). The PC are related to the most important variation of all spectra; first PC account for most of the variation of the data, the last ones carry only noise, and S are related to the weight of each PC to reconstruct the original data18: A = S PC
(1)
where A is an m by n matrix of Raman spectra (m = wavenumber, n = number of spectra), S is the n by n matrix of reconstruction scores and PC is an n by m matrix of principal components vectors. Since S is the weight, meaning the importance of each PC to form the original spectrum, and the PC carries spectral information in the form of defined peaks and valleys, it could be used to build an algorithm based on the PCA scores to classify samples into well-defined categories depending of the histopathological findings and the occurrence of Raman features. Each principal component has a unique spectral characteristic, representing the new axis of maximum variance. The scores are the cosines (projections) of each spectrum in this new axis. The advantage of PCA is that, once a set of spectra have been modeled and classified as belonging to a particular class, a new spectrum can be compared to this model and be classified whether it fits into that class.8 The Mahalanobis distance (also called m-distance) is a very useful way to find the similarity of a set of values from one group of samples compared to another group for discriminant analysis. The m-distance (D2) is calculated as follow: D2i = (x µi)T V1 (x µi)
(2)
349
where i denotes the known sample group number, x the vector of the sample parameter, µ the mean vector for the specific group and V the covariance matrix of the group. The m-distance has advantages over Euclidean distance because it takes not only the relative distance from the sample to the mean of the group but also takes into account the covariance matrix of the data. It means that m-distance gives a statistical of how well the sample matches the grouped data. According to the equation (2) high m-distance means better group separation. The m-distance can be used to classify the sample into well-defined classes. In this study, a carcinogenesis model was developed in the hamster buccal pouch, and biochemical changes induced by the carcinogenic drug in the oral mucosa were analyzed by FT-Raman spectroscopy. The main goal was to distinguish malignant from normal oral tissues by analyzing the Raman spectral differences and correlating those spectral differences with the tissue histopathological analysis by modeling and testing prospectively a diagnostic algorithm based on PCA and m-distance.
METHODS Classical model of chemical cancerization (DMBA3) and tissue samples Twenty-one male Syrian Golden hamsters, 10–12 weeks old, weighing 100–120 g were used in this study. Oral carcinogen was induced using the model described by Salley in 195419 and standardized by Morris in 1961,20 which involves topical application of 0.1 mL of 0.5% 7,12-dimethybenz[a]anthracene (DMBA) in mineral oil in the right pouch of each animal three times a week. Their left oral pouches, used as a control, were treated in the same manner as the right pouch, but only using mineral oil. The animals were sacrificed at 14th weeks of treatment, after exophytic and endophytic tumors developed. The right and left oral pouches were exposed, and the whole mucosa was excised. The right oral pouch samples were taken from areas where lesions were visible. All the 42 excised samples were immediately snap-frozen and kept stored, immersed in liquid nitrogen (77K), until the Raman measurements were performed. This protocol was approved by the Ethics Committee of the University of Vale do Paraíba (no. L021/2002/CEP).
Fourier-transform–Raman spectroscopy High-quality, fluorescence-free Raman spectra were obtained from carcinogen-induced lesions and control samples, using a FT-Raman spectrometer (RFS 100/S; Bruker Inc., Karlsruhe, Germany) excited by a Nd:YAG laser at 1064 nm. The laser power at the sample, the signal integration time, and the spectrometer resolution were set to 100 mW, 100 scans, and 4 cm1, respectively. At the time of the spectroscopy study, tissue samples were brought to room temperature and kept moist by defrosting with 0.9% saline solution. For each sample, two to three FT-Raman spectra were obtained, depending on the sample size. After all Raman measurements were performed, the tissue samples were fixed in 10% formalin solution. Samples were
6/27/06
12:11 PM
Page 350
350
Oliveira et al.
processed by routine histopathology: embedded in paraffin, sectioned, and stained with hematoxylin and eosin for further histopathological analysis. Based on these histopathological results, the samples were classified into three main tissue types: (a) squamous cell carcinoma; (b) dysplastic epithelium; and (c) normal oral tissue. The right oral pouch samples were diagnosed as Squamous cell carcinoma (SCC) in 19 samples and dysplastic epithelium (DE) in two samples. All the 21 left oral pouch samples were classified as normal oral epithelium. In total 123 Raman spectra were obtained: 61 were from normal tissue, 56 from SCC, and six from DE.
box. High m-distance means that the data does not belong to that particular group. The 51 PC vectors from the training spectra were then used to calculate the scores of the 72 prospective data. Each PC was fitted with each one of the prospective spectral data using a Least-Square minimization function from Matlab© Optimization toolbox. The resulting scores (S) were plotted in a binary scatter plot, keeping the discriminant Mahalanobis curves the same as the training model.
RESULTS
Spectral data analysis
Significant differences between normal and malignant oral pouch tissues were observed in the Raman spectra. Figure 1 shows a typical spectrum of normal, dysplastic and carcinoma oral samples, in the range of 800–1800 cm1. The spectrum of normal tissues shows strong peaks in 858 and 941 cm1 assigned to the molecular vibrational modes of proteins (mainly due to Try and C-C stretch). The vibrational band at 1007 cm1 in normal and malignant tissues is assigned to the phenylalanine in proteins.8 The Raman spectra of SCC has a peak at 1560 cm1 that can be assigned to the nucleic acids.21–23 In normal tissue spectra, this peak has been shifted down to 1555 cm1. The bands at 1263 and 1337 cm1 were assigned to the vibration of carbohydrate-amide III and C-H bending mode of proteins, respectively. In the Raman spectra of normal, BCC, and DE tissues, peaks were observed around 1454 and 1663 cm1, which corresponds to the deformation of CH2 of proteins and lipids, and the vibration of carbohydrate-amide I, respectively.3,8 PCA reduced the number of variables from the training group, the PC, which are obtained on the basis of highest spectral variance. It was calculated that the first six PC represent
1263
(c)
1663
1337
1263
1007
945
1454
All spectra were recorded in binary format and converted to ASCII format for post-processing. In order to pre-process the data to perform PCA, all spectra were normalized using the most intense peak in the region of 800–1800 cm1 to obtain scale-free intensity data. All spectra were randomly divided into two groups, the first one with 51 spectra treated as training (modeling) group and the second one with 72 spectra treated as prospective (testing) group. The training group for the classification algorithm contained 22 spectra of normal oral tissues, six spectra from DE, and 23 spectra of SCC, whereas the prospective group contained 72 spectra, with 39 from normal tissue and 33 from SCC tissue. In order to develop an algorithm for tissue classification, the Raman spectra were first analyzed through PCA technique. The principal components (PC) and scores (S) were obtained using a routine written in Matlab© software (Mathworks, Inc., CA) with the NIPALS-PCA algorithm.18 Since S is the weight (or the importance of each PC to form the original spectrum), it can be used to build an discriminant algorithm to correlate the histopathological findings to FT-Raman features. The discriminator among the histopathological groups was the m-distance, which has been calculated by using the Matlab© Statistics tool-
Intensity (arb. units)
941
(b) 858
14289c03.PGS
(a)
FIG. 1. Fourier transform (FT)–Raman spectra collected from mucosa of oral pouch of hamster. (a) Normal. (b) Dysplastic. (c) Squamous cell carcinoma.
800
1000
1200
1400
1600 -1
Raman shift (cm )
1800
6/27/06
12:11 PM
Page 351
351
1263
Near-Infrared Raman Spectroscopy for Oral Carcinoma Diagnosis
1560
1337
1007
PC3
800
1000
1200
1400
1600
1800
-1
941
858
1555
1263
PC6
1337
Raman shift (cm )
800
1000
1200
1400
1600
1800
-1
Raman shift (cm )
major spectral variations (>90%). By visual inspection of the first six PC, the 3rd and 6th ones were shown to be responsible for the major spectral features, differentiating among the biotissues analyzed in this study (Fig. 2), and thus could provide better classification. Scores were then used to separate the spectra into one of these three categories according to the differences found in each PC. A scatter plot of PC3 versus PC6 scores was made, and a diagnostic line based on the Mahalanobis distance among normal and diseased tissue was drawn (Fig. 3). This curve fits the average m-distance from one group data point to the mean of the other group.
FIG. 2. Principal component 3 and principal component 6 obtained from the 51 Raman spectra of training group.
In the second spectral group (72 samples), a prospective, double-blind analysis was undertaken to test the algorithm’s ability to classify (diagnose) a spectrum that did not participate in the training. For so, each of the prospective spectrum was fitted to the score of each PC from the training group, based on Least-Square minimization. These prospective scores were then plotted, keeping the same diagnostic line based on mdistance calculated in the training group (Fig. 4). The algorithm’s sensitivity and specificity indices for detecting abnormal tissue were calculated for both training and prospective groups. For the first group, 91% sensitivity and
0.6 Normal Squamous Cell Carcinoma
0.4 Principal Component 6
14289c03.PGS
Mahalanobis Dysplastic
0.2 0.0 -0.2 -0.4 -0.6 -0.4
-0.3
-0.2
-0.1
0.1
Principal Component 3
0.2
0.3
FIG. 3. Scores of principal component 3 versus 6 for the 51 spectra using the training group.
6/27/06
12:11 PM
Page 352
352
Oliveira et al. 0.6 Normal
0.4 Principal Component 6
14289c03.PGS
Squamous Cell Carcinoma Mahalanobis
0.2 0.0 -0.2 -0.4
FIG. 4. Scores of principal component 3 versus 6 for the 72 spectra using the prospective analysis group.
-0.6 -0.4
69% specificity were found, while in the prospective group 100% and 55%, were found, respectively.
DISCUSSION Raman spectra of normal tissue present peaks assigned to lipids, with intensity significantly higher when compared to SCC. According to the studies conducted by Venkatakrishna et al.8 and Yang et al.,24 where spectral differences in malignant and normal oral tissues were analyzed, peaks were found with different intensity in the shift region of 1454 and 1663 cm1 depending on tissue type. However, in this study, those peaks showed no notably significant intensity differences. The positions of the amide I and amide III bands at 1650 and 1270 cm1, and the presence of a well-developed band around 935 cm1, suggested a helical protein structure. In addition to collagen, the SCC tissue contains a matrix of other proteins and polysaccharides. The phenylalanine band at 1007 cm1 and other very weak bands presented in the difference spectrum may arise from these species. In normal tissue spectra, this peak has decreased intensity. The intensity ratio of 1447, 1268, and 1246 cm1 bands compared to the 1670 cm1 amide I band in the SCC spectrum were very close to those in Type I collagen. The shift region of 1555 to 1560 cm1 can distinguish malignant from normal tissue and has been attributed to vibrational nucleic acids.2,8,10 These differences could be related to the increased concentration of nucleic acids relative to proteins of cell nucleus.8 Yang et al.24 related the increased concentration of nucleic acids only in SCC. According to Mahadevan-Jansen and Richards-Kortum,21 biological molecules such as nucleic acids have distinct Raman lines that could provide information of the cellular structure and development. Also, major changes that occur within cancer and pre-cancer development can be linked to the increased content of nucleic acids in the cell. This can be verified as the 1560 cm1 peak has equal intensity in the spectra of SCC and dysplasia. In the dysplastic tissue spectrum, the shift region of 1263–1337 cm1 was more representative, showing higher in-
-0.3
-0.2
-0.1
0.1
0.2
0.3
Principal Componente 3
tensity than normal and malignant tissues; nevertheless, only a small number of samples was analyzed. In general, the spectrum of normal oral tissue is closer to the spectrum of lipid, while the spectrum of malignant tissue resembles the spectrum of proteins, which was also observed by Venkatakrishna et al.8 The PCA algorithm applied to Raman spectra was shown to be effective for classification of cancerous diseases, detecting biochemical changes that occur within the cells during the development of cancer. The spectra can be classified as normal or malignant (or anywhere in between), correlated with the histopathology. In our case, the best separation was obtained between tissues using PC3 and PC6. The significance of such decision can be quantified by various parameters. A diagnostic line based on average Mahalanobis distance where used to assist on the separation. It was found that, for the prospective group, the spectra of SCC stay almost totally apart from, the spectra of normal tissue, giving 100% and 55% sensitivity and specificity, respectively. It was also found that spectra of dysplastic tissue belong to the same group separation of malignant tissues. This result could be explained by the low variation found in normal tissues samples compared to the dysplastic or carcinoma tissues. The low specificity found for the prospective group is probably due to a broad spectrum of abnormalities in growth and differentiation for the pathological samples.
CONCLUSION The present study showed a model to successfully diagnose normal, dysplastic, and squamous cell carcinoma samples obtained from a classical model of chemical cancerization. We could demonstrate the possibility to obtain reproducible highquality FT-Raman spectra from biotissues. The advantage in the optical method, in addition to the tissue evaluation without staining, labeling, or any other processing, is that an automatic scanning system could be set-up to record spectra at several sites with a standard calibration dataset prepared earlier, and give statistical probability of the sample belonging to a particular calibration set. PCA associated with discrimination by Ma-
14289c03.PGS
6/27/06
12:11 PM
Page 353
Near-Infrared Raman Spectroscopy for Oral Carcinoma Diagnosis halanobis distance showed to be an appropriated data analysis method that could be employed for clustering. Thus, Raman spectroscopy, as an analytical and non-destructive technique, could provide information about the molecular structure of the tissue under investigation, and may be used to assist the histopathological analysis in the detection of malignant oral lesions.
ACKNOWLEDGMENTS A.A.M. would like to thank CNPq (30 2393/2003-0) and FAPESP (2001/14384-8) for providing financial support.
REFERENCES 1. American Cancer Society. (2002). Cancer Facts & Figures New York: American Cancer Society. 2. Choo-Smith, L.P., Edwards, H.G.M., Endtz, H.P., et al. (2002). Medical application of Raman spectroscopy: from proof of principal to clinical implementation. Biopolymers 67, 1–9. 3. Wu, J.G., Xu, Y.Z., Sun, C.W., et al. (2001). Distinguishing malignant from normal oral tissues using FTIR fiber-optic techniques. Biopolymers 62, 85–192. 4. Krishna, C.M., Sockalingum, G.D., Kurien, J., et al. (2004). Micro-Raman spectroscopy for optical pathology of oral squamous cell carcinoma. Appl. Spectrosc. 58, 1128–1135. 5. Schut, T.C.B., Witjes, M.H.J., Sterenborg, H.J.C.M., et al. (2000). In vivo detection of dysplastic tissue by Raman spectroscopy. Anal. Chem. 72, 6010–6018. 6. Stone, N., Starroulaki, P., Kendall, C., et al. (2000). Raman spectroscopy for early detection of laryngeal malignancy: preliminary studies. Laryngoscope 110, 1756–1763. 7. Stone, N., Kendall, C., Shepherd, N., et al. (2002). Near-infrared Raman spectroscopy for the detection of ephitelial pre-cancers and cancers. J. Raman Spectrosc. 33, 564–573. 8. Venkatakrishna, K., Kurien, J., Pai, K.M., et al. (2001). Optical pathology of oral tissue: a Raman spectroscopy diagnostic method. Curr. Sci. 80, 665–669. 9. Kartha, V.B., Lakshmi, R.J., Mahato, K.K., et al. (2003). Raman spectroscopy in clinical investigations. Indian J. Physics 77B, 113–123. 10. Majumder, S.K., Mohanty, S.K., Ghosh, N., et al. (2000). A pilot study on the use of autofluorescence spectroscopy for diagnosis of the cancer pf human oral cavity. Curr. Sci. 79, 1089–1094. 11. Manjunath, B.K., Kurein, J., Rao, L., et al. (2004). Autofluorescence of oral tissue for optical pathology in oral malignancy. J. Photochem. Photobiol. B 73, 49–58.
353
12. Gniadecka, M., Philipsen, P.A., Sigurdsson, S., et al. (2004). Melanoma diagnosis by Raman spectroscopy and neural networks: structure alterations in proteins and lipids in intact cancer tissue. J. Invest. Dermatol. 122, 443–449. 13. Hanlon, E.B., Itzkan, I., Dasari, R.R., et al. (1999). Near-infrared fluorescence spectroscopy detects Alzheimer’s disease in vitro. Photochem. Photobiol. 70, 236–242. 14. Mahadevan-Jansen, A., Mitchell, M.F., Ramanujam, N., et al. (1998). Near-infrared Raman spectroscopy for in vitro detection of cervical precancers. Photochem. Photobiol. 68, 123–132. 15. Manoharan, R., Shafer, K., Perelman, L., et al. (1998). Raman spectroscopy and fluorescence photon migration for breast cancer diagnosis and imaging. Photochem. Photobiol. 67, 15–22. 16. Nunes, L.O., Martin, A.A., Silveira, L.J., et al. (2003). FT-Raman spectroscopy study for skin cancer diagnosis. Spectroscopy 17, 597–602. 17. Nogueira, G.V., Silveira, L.J., Martin, A.A., et al. (2005). Raman spectroscopy study of atherosclerosis in human carotid artery. J. Biomed. Optics 10, 031117. 18. Geladi, P., and Kowalski, B.R. (1986). Partial least square regression: a tutorial. Analyt. Chim. Acta 185, 1–17. 19. Salley, J.J. (1954). Experimental carcinogenesis in the cheek pouch of the Syrian hamster. J. Dent. Res. 33, 253–262. 20. Morris, A.L. (1961). Factors influencing experimental carcinogenesis in the hamster cheek pouch. J. Dent. Res. 40, 3–15. 21. Manadevan-Jansen, A., and Richards-Kortum, R. (1996). Raman spectroscopy for the detection of cancers and precancers. J. Biomed. Optics 1, 31–70. 22. Yazdi, Y., Rananujam, N., Lotan, R., et al. (1999). Resonance Raman spectroscopy at 257-nm excitation of normal and malignant cultured breast and cervical cells. Appl. Spectrosc. 53, 82–85. 23. Grasselli, J.G., and Bulkin, B.J. (1991). Raman spectroscopy for biological applications, in: Chemical Analysis—Analytical Raman Spectroscopy. New York: Wiley, pp. 397–423. 24. Yang, Y., Liu, C.H., Savage, H.E., et al. (1998). Optical fluorescence and Raman biopsy of squamous cell carcinoma from the head and neck, in: Proceedings of Optical Biopsy II. San Jose, CA, pp. 68–71.
Address reprint requests to: Dr. Airton Abrahão Martin Universidade do Vale do Paraíba (UNIVAP) Instituto de Pesquisa e Desenvolvimento (IP&D) Av. Shishima Hifumi, 2911 Urbanova, 12244-000 São José dos Campos, SP, Brazil E-mail:
[email protected]