Open Access
Add to BibSonomyAbstract
In this paper, we investigated the feasibility of using surface enhanced Raman spectroscopy (SERS) of blood serum to discriminate liver cancer and liver cirrhosis patients from normal people. Serum taken from 44 healthy people, 45 liver cancer patients, 42 post-treatment liver cancer patients and 45 liver cirrhosis patients was measured. SERS peaks from these groups were compared and the assignments and biomedical meanings were analyzed and explained. In addition, support vector machine (SVM), partial least square-discriminant analysis (PLS-DA) and artificial neural networks (ANN) was used on the obtained SERS spectra to identify its diagnostic potential for liver diseases. PLS-SVM, PLS-DA and PLS-ANN indicated 91.5%, 89.2% and 90.3% accuracy, respectively. This preliminary study demonstrates that serum SERS can be used for liver cancer screening.
© 2015 Optical Society of America
1. Introduction
The incidence of liver cancer ranks fifth among all cancers worldwide and, of all cancers, it is the second leading cause of death [1]. Hepatitis virus infection (including type B and C) is the dominant cause of liver cancer, and cirrhosis can also enhance the development of liver cancer to a significant degree [2]. About 10% of hepatitis B infections (HBV) will convert to liver cirrhosis, and about 10% of liver cirrhosis will transform to liver cancer [3]. In China, the high incidence of HBV has resulted in liver cancer accounting for more than 50% of the world’s new liver cancer cases and deaths in 2012 [4]. The situation is even worse in Taiwan, where ten thousand deaths are caused by liver cancer every year. Since 1987, China has implemented a neonatal immunization program, and there was also a replant for children and adolescents under the age of 15. Because the liver does not contain any nerves, liver cancer and other liver diseases are usually asymptomatic at earlier stages. This has led to more than 60% of liver cancer patients receiving terminal diagnoses when they visit the hospital. HBV is capable of mutating into different quasi-species, allowing the virus to survive in adverse environments and become more drug resistant. The 5-year survival rate is 28% for patients with localized liver cancer and only 3% for those with distant stage cancer [5]. Imaging tests such as x-rays, ultrasound and CT scans are the current means for diagnosing liver cancer. But, these methods are highly dependent on accurate interpretation by very careful inspection. Besides, these processes are time-consuming and unfit for large screenings. Thus, a sensitive and early liver cancer detection method is very essential for future liver cancer patients.
Surface enhanced Raman spectroscopy (SERS) is a highly sensitive detection method due to the enhancement effect caused by its physical and chemical mechanism [6], SERS can enhance the normal Raman spectroscopy signals by 4 to 14 orders of magnitude [7]. In addition to its high sensitivity, SERS can also effectively quench the interference of fluorescence. The enhancement feature make SERS especially suitable for the identification of biomarkers by specific binding of biomolecules [8] and the detection of biofluids via dispersal of enhancing nanoparticles [9]. For cancers which have molecular changes as their precursors, SERS can detect the changes at an early stage. Besides, the features of non-invasive and rapidity of SERS make it an ideal tool for large-scale screening. At early stages, liver cancer or other diseases often cause structural changes of the biomolecules in blood [10]. The resulting differences in the SERS spectra of biofluids can therefore indicate changes in corresponding tissues and allow for disease diagnosis.
There are plenty of biomarkers in blood. With the development of immunomics and proteomics, increasing amounts of biomarkers have been identified. There are three main objectives for SERS blood analysis: to detect certain substances [11,12], to detect particular cells [13,14], and to detect whole blood (blood [15], plasma [16] or serum [17]). Among sample types, serum can retain the most substances in blood and is not subject to blood cell interference. SERS of serum albumin and globulin has been proved has prediction ability for hepatocellular carcinoma [18]. The components in serum change to accompany the development of diseases, and these changes make a fingerprint-like identity for each individual sample. Thus, in our study, whole serum was selected as the detection target and SERS was selected as the diagnostic method for liver cancer and liver cirrhosis.
The combination of SERS spectroscopy with chemometric analysis is routine technique for disease diagnosis. Three different types of analysis – support vector machine (SVM), partial least square-discriminant analysis (PLS-DA) and artificial neural networks (ANN) – were applied to the SERS data.
SVM was a classification algorithm introduced by Vapnik and Burges, and has high discriminant power due to its nonlinear features and discrimination through hyperplanes [19]. PLS-DA is a supervised regression discriminant method. It is particularly suited for our high-dimensional spectral data where the numbers of variables exceed sample size. PLS-DA has been preferred over principal component analysis (PCA) because it focuses on the spectral variation that is relevant for the discrimination between classes [20]. Research has shown that PLS is a better alternative than other popular classification techniques such as PCA [21]. ANN is a machine learning method which imitates the functioning of animal central nervous systems [22]. It can classify by adjusting the highly nonlinear topology. By weighting and transmitting input variables between neurons repeatedly, ANN sorts input observations into different outputs. Both methods are commonly used for spectral analysis, and have shown their efficiency in data discrimination [23,24]. For these reasons, various techniques including PLS-DA and ANN were investigated to build classification models for disease diagnostics.
In this paper, we intend to measure the differences of serum SERS spectra between current and former liver cancer patients, liver cirrhosis patients, and healthy controls. First, we will analyze the differences of SERS peaks to determine the variation of serum components between groups. Then, PLS-SVM, PLS-DA and PLS-ANN will be utilized to understand the diagnostic results and to form a diagnostic algorithm. Because liver cirrhosis is often a precursor of liver cancer, the detection of liver cirrhosis via this method can give notice to the possibility of liver cancer even before the cancerization. If applied successfully, this technique will greatly lower the death rate of liver cancer.
2. Materials and methods
2.1 Serum sampling
Blood samples were drawn from donors who signed an informed consent form in accordance with the ethical guidelines published by council for international organizations of medical sciences (CIOMS) [25]. Blood samples were taken from 44 healthy people, 45 liver cancer patients, 42 post-operative liver cancer patients and 45 liver cirrhosis patients provided by Shengjing Hospital of China Medical University. The serum was collected at the first time of definite diagnosis before any treatment. For the post-operation samples, serum was taken 7 – 8 days after operation (this is before chemotherapy, if any). Table 1 provides the demographic distribution and cancer stages of the subjects. All samples were phlebotomized before breakfast to avoid the interference of food. Approximately 2mL venous blood was collected and no anticoagulant was added. Each blood sample was centrifuged at the speed of 3000 rpm for 10 min to isolate serum. Serum samples were refrigerated (temperature: −80°C) hermetically for later investigation.
Table 1. Demographics of study population
2.2 SERS
SERS spectroscopy data in the range of 400 cm−1 - 1800 cm−1 was collected from an inverted microscope (British Renishaw). Radiation of 632.8 nm created by a He-Ne laser operating at 3.5 mW was used for excitation. The exposure time was 10 s. Silver hydrosol was synthesized using the deoxidizing method reported by Leopold and Lendl [26]. The obtained Ag nano-particles have a mean diameter of about 80 nm (Fig. 1). Samples were prepared by ultrasonically mixing 2μl serum with 2μl Ag colloid in a centrifuge tube, and the mixture remain its liquid state under spectroscopy-taking. SERS was measured via 180° back-scattering.
Fig. 1 UV absorbance spectroscopy and transmission electron microscopy (TEM) photograph of silver colloid.
2.3 Pre-treatment of spectroscopy
Because spectroscopy is inclined to be contaminated by factors such as fluorescence background, laser power fluctuation, and differences in sample concentration, we must eliminate noise and other variations which disturb the following multi-variant analysis. Therefore, the spectral preprocessing of smoothing and normalization was performed on the raw spectral data. Smoothing was carried out using the pspline package for R. Normalization was fulfilled through the hyperSpec package. Such preprocessing is highly advantageous since raw Raman spectra is noisy and may contain variations in signal. To lessen the influence of harmful factors such as electronic noise, light scattering and laser power fluctuation, every sample was scanned three times, and an average spectroscopy was calculated and used in further analysis.
2.4 Statistical analysis
SVM is a young classification method based on the mechanism of separating classes with a hyperplane between them. In the process, input vectors are mapped to a newly constructed high dimensional space, and then two parallel hyperplanes are constructed to separate the data by maximizing the inter-plane distances. This method has been recently proven to exceed LDA’s capabilities in SERS spectra discrimination for disease diagnosis due to its hyperplane property [19]. In this study the most frequently used radial basis function (RBF) was used as the kernel function. The radial basis kernel function is defined as exp(-γ |u-v|2) (where u and v are the two generic sample data vectors). The optimal values for parameter γ and penalty factor C, which was used for preventing over-fitting, were found using grid searching.
A supervised PLS-DA technique was first used on the pre-processed data to generate classification rules. The goal of PLS is to conduct a dimensional reduction. PLS inherits the fundamental principle of PCA, but uses components that are related to y instead of principal component scores [27]. The main difference between these two methods is that PLS tries to relate the variables and the categorical variables. PLS-DA separates groups by rotating PCA components so that distances among classes are maximized. It is one of the most commonly used multivariate calibration tools, and can be applied even when there are many more predictor variables than observations [28]. The typical use case of PLS-DA is a PLS regression with a categorical variable. PLS-DA discriminates by replacing independent y-data with categories. The x-data was set as preprocessed spectral data in our experiment, and the y-data were the original sample groups. This takes the dependent variables into account when building latent variables, which may include more information. Additionally, high correlation variables with the response variables are given extra weight which makes the prediction more effective [29]. After the data reduction, only a few of the linear combinations of input variables were used in the discriminant analysis.
Finally, ANN was applied to the latent variables (LVs) obtained from PLS regression. Artificial neural networks (ANN) are algorithms inspired by the working of central nervous systems. ANN can be used for both machine learning and pattern recognition. Typically, a neural network contains three layers: an input layer, single or multiple hidden layers and an output layer [29]. After input of the spectra data into the network, samples can be grouped at the output layer. For the PLS-ANN models, PLS first extracted information from the whole spectral regions, and the first few LVs were used as neurons on the network input layer.
The neural networks can yield solutions to relationships without having knowledge on the physics of the data. It allows nonlinear behaviors in the data to be considered when identifying differences [30]. The neural nets are formed by neurons that process entrance group and output results. Feed-forward training was used to train the neural networks. The back propagation (BP) network architecture is formed by three layers: the input layer has 2 neurons corresponding to the 2 first LV components, the hidden layer contains a varied number of neurons for training, and there are 4 neurons representing 4 classification groups in the output layer.
To fully assess the suitability of SERS spectroscopy for distinguishing between groups, the data set was split into a training data set and validation data set. The purpose of the split was to simulate the clinical scenario where a blind sample serum was predicted using SERS measurements. In our work, two thirds of the data samples were used for training and the rest were used for testing of the network. The leave-one-out method was applied in order to check the performance of PLS-SVM, PLS-DA and PLS-ANN. This method divided the data into equally large 10 subsets and used each subset as the validation data until no subsets remained [31].
All spectra data analysis was performed in the statistical environment R version 3.0.3 (http://www.R-project.org). PLS-DA analysis was performed using the DiscriMiner package. SVM analysis was carried out by e1071 package. ANN analysis was fulfilled using the nnet package. ANN network figures were drawn using functions provided by Marcus Beck (beckmw.wordpress.com). All classifiers were applied using a 10-fold cross validation method.
3. Results
3.1 SERS spectra
There were 27 liver cancer patients, 23 liver cirrhosis patients and 31 healthy volunteers who participated in our experiment. For each sample, three measurements were performed and the averaged spectrum was recorded. Spectra were preprocessed by smoothing, baseline correction and area normalization. Each sample was measured three times and the averaged spectra were used in further analysis. Figure 2 shows the average SERS spectra of the three groups of samples. Peaks at wavenumbers of 590, 636, 727, 820, 886, 1021, 1073, 1132, 1211, 1323, 1355, 1446 and 1582 cm−1 can be consistently observed in all four groups. Peaks at 590 (C-S twist of Amide-VI), 820 (ring breathing of tyrosine), 886 (ring bending of tryptophan), 1021 (C-H stretch of phenylalanine), 1073 (C-C stretch of phospholipids), 1132 (C-N stretch of D-mannos), 1323 (CH3CH2 twisting of collagen, tryptophan), 1355 (CH3CH2 wagging of tryptophan adenine and guanine) and 1582 cm−1 (C-C bending of phenylalanine, acetoacetate of riboflavin) are the most distinct between groups. In comparison with the controls, the intensities at 820, 1132, 1323 and 1355 cm−1 are lower for diseased samples, while bands at 590, 886, 1021, 1073 and 1582 cm−1 are more intense. Comparisons of these characteristic peaks can be viewed more clearly in the box-and-whisker plot (Fig. 3). All the peaks noted have statistical significance (with p values less than 0.001, Table 2). To better understand the molecular basis of the SERS peaks, dominant SERS bands were assigned tentatively (Table 2).
Fig. 2 Average SERS spectra of four groups (normal, liver cancer, liver cancer after operation and liver cirrhosis group).
Fig. 3 Mean intensities and standard deviations of peaks with the most distinguishable differences between groups.
Table 2. Tentative peak assignments [19, 32-34]
3.2 Statistical analysis
Support vector machine (SVM) was utilized on the two parameters LV1 and LV2 that have the best prediction performance. In the SVM model, C-classification was used as the SVM type for the purpose of classification. Radial Basis Function (RBF) was used as the kernel function.
The optimum parameters of C and γ in the RBF kernel were found using grid searching. The search range for cost C was performed from 100 to 102, and from 10−12 to 1012 for parameter γ. Figure 4 shows the SVM performance as a function of cost C and γ. The darkness represents the performance of SVM, with deeper colors corresponding to higher diagnostic accuracy. The optimum C and γ were found to be 10 and 0.1, respectively. These C and γ values were then input into the SVM model for classification. The SVM classification plot is shown in Fig. 5(a). Sample represented by solid symbols are misclassified ones.
Fig. 4 SVM performance as a function of penalty factor C and parameter γ. Deeper color represents better performance.
Fig. 5 Class prediction plot using PLS components: (a) PLS-SVM, and (b) PLS-DA. Solid symbols represent misclassified samples.
Four groups of SERS spectra from the liver cirrhosis group, pre- and post-operative liver cancer group, and the control group were input in the PLS-DA model for analysis. The PLS-DA model was built using 2 latent variables (LVs). LVs are linear combinations of wavenumbers that represent the spectral changes. Once the optimal number of LVs for the calibration was selected, LVs were employed to build a PLS-DA model for the classification.
The variable of importance for projection (VIP) with more than a certain value should be retained for further analysis [21]. Here, we selected 0.8 as the value, so components t1 and t2 were used for the following comparison. A 2D plot can be drawn using those two components. Variable t1 was selected as the x-axis, and t2 was selected for the y-axis (Fig. 5(b)). The meanings of the numbers and colors in this figure are identical to those in the PLS-SVM scatter plot (Fig. 5(b)). As illustrated in the figure, samples from different groups scattered apart and can be easily distinguished from each other. Sample misclassification was likely due to the similar band intensities of the input spectral data.
Artificial neural networks was also used for the quantitative analysis of the PLS components obtained in the above processing. In this study, we selected a network with 3 layers – an input layer, a hidden layer, and an output layer. For the input layer, each neuron is one PLS component, and we took the first two scores as inputs. The output layer is the predicted groups which in this case are the control group, liver cirrhosis group and liver cancer group. We initialize ANN by first extracting the PLS components from the prior analysis and then applying those components to the ANN [30]. A large number of input variables may cause ANN to be over-fit, so compressed PLS components were used for this analysis. It was impractical to introduce all of the spectra data into the ANN anyways due to the high dimensionality. Additionally, full SERS spectra often contain much redundant information. Therefore, we applied the LVs obtained from the PLS model to reduce the input variables. Here, we chose only the first 2 LVs which matched the VIP value of more than 0.8.
Feed-forward networks was chosen for this neural network, and the training was performed by error backpropagation. As ANN is inclined to over-fit, it is best to use the smallest possible hidden layers. So, only one hidden layer was used to build the network. Thus, a 3-layer network was constructed which includes an input layer and an output layer. Figure 6 shows the completed ANN networks. I1 and I2 represent the input variables, H1 – H18 are the neurons of hidden layers, and O1, O2, O3 and O4 are the output groups. The hidden layer was composed of an adjustable number of neurons interconnecting input and output layers. Nodes from 2 to 20 were tested for the classification of the LVs in the hidden layer. Eighteen neurons (H1 – H18) were used for best performance. The output layer contained four neurons giving the predicted groups. The samples were separated into training set (114 samples), and test set (62 samples) when being processed in the neural network.
4. Discussion
SERS spectrum of serum is the fingerprint of that certain serum. It contains the information about present biomolecules (such as proteins, nucleic acids, and lipids) which can predict component changes, and assists in the diagnosis of diseases. The general shape and trend of SERS spectra from the three groups are approximately the same, which suggests that the overall serum composition for the three groups is not entirely different. The decrease of band intensity at around 828 and 1132 cm−1 is caused by a decrease in concentration of the corresponding amino acids [16] and may also be due to the vigorous metabolism of tumors [32,35]. The peak at 1073 cm−1 is caused by C-C stretching of phospholipids. The increase of this peak is consistent with the Raman spectroscopy report of human hepatocytes, which indicated the phenomenon as a characteristic of neoplastic cells [36]. The increases of other peaks which are caused by amino acids and nucleic acids, etc. are reported to be the result of pathological lesions of cells [37]. The intensity differences of SERS peaks comparing with normal group enlarged with the development of liver diseases (liver cancer is more severe than liver cirrhosis). This shows that serum SERS spectra reflect molecular component changes in blood caused by liver diseases. And the similarities in SERS spectra between liver cancer patients and liver cirrhosis patients demonstrate that the two diseases share some pathology. Even though different peaks may be assigned to the same biomolecules, they can undergo different changes. For example, the three Raman shifts at 886, 1323, and 1355 cm−1 caused by tryptophan have different changes in the liver cancer and the liver cirrhosis group. This is because the peaks are assigned to different vibrational modes of that particular molecule [38]. The three SERS peaks were contributed to by ring bending, CH3CH2 twisting and CH3CH2 wagging of tryptophan separately. In comparison with previous studies, peak at 886 cm−1 can also been found in serum globulin SERS but has different changes (decrease with hepatocellular cancer) [18], this maybe because of the contribution of other proteins existed in our experiment. The Raman intensities at 820, 886, 1073 and 1132 cm−1 have also been found show similar changes during the detection of other cancers such as colon, stomach, cervix, and nasopharyngeal cancers [16,32,35,38–40]. This suggests that some spectra changes common results of several cancers. Trends of some of the peaks are different from our results, and we think that are specific for liver diseases, such as peak at 886, 1021 and 1323 cm−1 for colorectal cancer serum protein [38], and peak at 1323 and 1355 cm−1 for gastric cancer serum [41].
In the SVM processing, the radial basis function (RBF) was selected because linear kernel is a special case of RBF and RBF has fewer numerical difficulties when compared to polynomial kernel. Many new kernels have been studied and put forward; future research will explore other various kernels for classification. Two parameters - penalty factor C and γ - in the RBF kernel can significant influence the performance of the SVM model. Grid searching was applied on the two parameters to find the optimum pairs of (C, γ) values were found.
PLS-DA is a multivariate least-squares regression discrimination method that is commonly used as a classifying method for spectroscopy analysis. In our experiment, scatter plot by PLS component t1 and t2 provided good discrimination between different groups. The differences captured by the PLS-DA model are basically due to the variation in serum SERS between groups. This shows that PLS-DA is a valid method for discriminating SERS of serum for the diagnosis of liver diseases.
For ANN classification, a total of two neurons were chosen for the hidden layer, because of its good performance and because too many neurons lead to over fitting [42]. First, the original spectra data is too large, and would make the construction of an ANN model too time-consuming. Second, with more input variables, the ANN model will likely have convergence problems with the training algorithm. One problem with ANN is that its need for a large amount of training spectra makes it too targeted. If there are other interferences to the spectra, a retraining may be needed. Diverse functions and parameters need to be tested to find a best-trained network, and thus a lot of time is required.
To test model performance of PLS-SVM, PLS-DA and PLS-ANN, we applied a leave-one-out cross-validation procedure. Table 3 summarizes the results. For a total of 176 samples, PLS-SVM had an accuracy of 91.5%, and PLS-DA model achieved a classification accuracy of 89.2%. For the PLS-ANN testing set, the prediction accuracy was 90.3%. From the scattering plot of both PLS-SVM and PLS-DA, we can see than sample points of different groups cluster distinctly. This indicates the potential usefulness of the LVs obtained from PLS for the differentiation between serum SERS [21]. SVM performed the best in terms of accuracy, sensitivity and specificity, possibly due to its hyperplane classification mechanism which have a better performance than linear discriminant classification methods [19,43] (PLS-DA in this study). ANN classified LVs slightly more accurate than PLS-DA in most of the prediction values, while PLS-DA only has more accurate performance for the discrimination of normal cases (represented by specificity). This maybe because ANN can model complex nonlinear systems by interconnected mathematical neurons network [44]. We can conclude from our work that SERS of serum is a powerful analytical tool for the fast prediction of liver diseases.
Table 3. Diagnostic evaluation of the PLS-SVM, PLS-DA and PLS-ANN technique
5. Conclusion
In summary, we investigated the feasibility of using SERS of serum combined with PLS-SVM/DA/ANN for discriminating liver cancer patients (before and after operation), liver cirrhosis patients, and healthy people. Differences at SERS peaks indicated composition changes in serum for different groups of samples. The intensity increase of most Raman peaks as a result of the deterioration of the liver due to disease can be explained by the high proliferation of cancer cells, which is the feature of carcinogenesis. By using PLS-SVM/DA/ANN, features of the original SERS spectra were reserved and spectra from different groups were classified. The diagnostic accuracies of PLS-SVM, PLS-DA and PLS-ANN were 91.5%, 89.2% and 90.3%, respectively. This preliminary study demonstrates the potential for the clinical use of serum SERS for liver disease diagnosis. Further research will be focused upon the improvement of pre-treatment (purification) of blood samples, optimization of algorithms, and the inclusion of large-scale samples.
Acknowledgments
The authors would like to acknowledge the financial support of the National Natural Science Foundation of China (No. 11074029), Science and Technology Planning Project of Liaoning Province (No. 201302739), and Science and Technology Foundation of Shenyang City (No. F14-231-1-34).
References and links
1. A. Jemal, F. Bray, M. M. Center, J. Ferlay, E. Ward, and D. Forman, “Global cancer statistics,” CA Cancer J. Clin. 61(2), 69–90 (2011). [CrossRef] [PubMed]
2. A. Forner, J. M. Llovet, and J. Bruix, “Hepatocellular carcinoma,” Lancet 379(9822), 1245–1255 (2012). [CrossRef] [PubMed]
3. S. Yu, “An evaluation of hepatitis B vaccination program strategies,” Doctorate, Fudan University (2003).
4. International Agency for Research on Cancer, World Cancer Report 2014 (World Health Organization, 2014).
5. R. Siegel, D. Naishadham, and A. Jemal, “Cancer statistics, 2013,” CA Cancer J. Clin. 63(1), 11–30 (2013). [CrossRef] [PubMed]
6. M. Moskovits, “Surface‐enhanced Raman spectroscopy: a brief retrospective,” J. Raman Spectrosc. 36(6-7), 485–496 (2005). [CrossRef]
7. C. L. Haynes, A. D. McFarland, and R. P. V. Duyne, “Surface-enhanced Raman spectroscopy,” Anal. Chem. 77(17), 338A–346A (2005). [CrossRef]
8. H. T. Ngo, H.-N. Wang, T. Burke, G. S. Ginsburg, and T. Vo-Dinh, “Multiplex detection of disease biomarkers using SERS molecular sentinel-on-chip,” Anal. Bioanal. Chem. 406(14), 3335–3344 (2014). [CrossRef] [PubMed]
9. J. M. Reyes-Goddard, H. Barr, and N. Stone, “Photodiagnosis using Raman and surface enhanced Raman scattering of bodily fluids,” Photodiagn. Photodyn. Ther. 2(3), 223–233 (2005). [CrossRef] [PubMed]
10. J. Kewal, The Handbook of Biomarkers (Humana Press, 2010).
11. K. E. Shafer-Peltier, C. L. Haynes, M. R. Glucksberg, and R. P. Van Duyne, “Toward a Glucose Biosensor Based on Surface-Enhanced Raman Scattering,” J. Am. Chem. Soc. 125(2), 588–593 (2003). [CrossRef] [PubMed]
12. G. Wang, R. J. Lipert, M. Jain, S. Kaur, S. Chakraboty, M. P. Torres, S. K. Batra, R. E. Brand, and M. D. Porter, “Detection of the potential pancreatic cancer marker MUC4 in serum using surface-enhanced Raman scattering,” Anal. Chem. 83(7), 2554–2561 (2011). [CrossRef] [PubMed]
13. M. Y. Sha, H. Xu, M. J. Natan, and R. Cromer, “Surface-Enhanced Raman Scattering Tags for Rapid and Homogeneous Detection of Circulating Tumor Cells in the Presence of Human Whole Blood,” J. Am. Chem. Soc. 130(51), 17214–17215 (2008). [CrossRef] [PubMed]
14. X. Wang, X. Qian, J. J. Beitler, Z. G. Chen, F. R. Khuri, M. M. Lewis, H. J. C. Shin, S. Nie, and D. M. Shin, “Detection of Circulating Tumor Cells in Human Peripheral Blood Using Surface-Enhanced Raman Scattering Nanoparticles,” Cancer Res. 71(5), 1526–1532 (2011). [CrossRef] [PubMed]
15. M. Casella, A. Lucotti, M. Tommasini, M. Bedoni, E. Forvi, F. Gramatica, and G. Zerbi, “Raman and SERS recognition of β-carotene and haemoglobin fingerprints in human whole blood,” Spectrochim. Acta A Mol. Biomol. Spectrosc. 79(5), 915–919 (2011). [CrossRef] [PubMed]
16. S. Feng, R. Chen, J. Lin, J. Pan, Y. Wu, Y. Li, J. Chen, and H. Zeng, “Gastric cancer detection based on blood plasma surface-enhanced Raman spectroscopy excited by polarized laser light,” Biosens. Bioelectron. 26(7), 3167–3174 (2011). [CrossRef] [PubMed]
17. R. Liu, X. Zi, Y. Kang, M. Si, and Y. Wu, “Surface-enhanced Raman scattering study of human serum on PVA-Ag nanofilm prepared by using electrostatic self-assembly,” J. Raman Spectrosc. 42(2), 137–144 (2011). [CrossRef]
18. J. Wang, S. Feng, J. Lin, Y. Zeng, L. Li, Z. Huang, B. Li, H. Zeng, and R. Chen, “Serum albumin and globulin analysis for hepatocellular carcinoma detection avoiding false-negative results from alpha-fetoprotein test negative subjects,” Appl. Phys. Lett. 103(20), 204106 (2013). [CrossRef]
19. S. X. Li, Q. Y. Zeng, L. F. Li, Y. J. Zhang, M. M. Wan, Z. M. Liu, H. L. Xiong, Z. Y. Guo, and S. H. Liu, “Study of support vector machine and serum surface-enhanced Raman spectroscopy for noninvasive esophageal cancer detection,” J. Biomed. Opt. 18(2), 027008 (2013). [CrossRef] [PubMed]
20. P. de Peinder, M. J. Vredenbregt, T. Visser, and D. de Kaste, “Detection of Lipitor counterfeits: A comparison of NIR and Raman spectroscopy in combination with chemometrics,” J. Pharm. Biomed. Anal. 47(4-5), 688–694 (2008). [CrossRef] [PubMed]
21. M. Pérez-Enciso and M. Tenenhaus, “Prediction of clinical outcome with microarray data: a partial least squares discriminant analysis (PLS-DA) approach,” Hum. Genet. 112(5-6), 581–592 (2003). [PubMed]
22. D. Graupe, Principles of Artificial Neural Networks, 3rd ed. (World Scientific Publishing Company, 2013).
23. A. M. T. Monfared, V. S. Tiwari, M. M. Tripathi, and H. Anis, “Raman spectroscopy for clinical-level detection of heparin in serum by partial least-squares analysis,” J. Biomed. Opt. 18(2), 027010 (2013). [CrossRef] [PubMed]
24. R. S. Uysal, I. H. Boyaci, H. E. Genis, and U. Tamer, “Determination of butter adulteration with margarine using Raman spectroscopy,” Food Chem. 141(4), 4397–4403 (2013). [CrossRef] [PubMed]
25. Council for International Organizations for Medical Sciences, International Ethical Guidelines for Biomedical Research Involving Human Subjects (World Health Organization, 2002).
26. N. Leopold and B. Lendl, “A New Method for Fast Preparation of Highly Surface-Enhanced Raman Scattering (SERS) Active Silver Colloids at Room Temperature by Reduction of Silver Nitrate with Hydroxylamine Hydrochloride,” J. Phys. Chem. B 107(24), 5723–5727 (2003). [CrossRef]
27. K. Varmuza and P. Filzmoser, Introduction to Multivariate Statistical Analysis in Chemometrics (CRC Press, 2009).
28. Y. Li, G. Du, W. Cai, and X. Shao, “Classification and quantitative analysis of azithromycin tablets by raman spectroscopy and chemometrics,” Am. J. Anal. Chem. 2(02), 135–141 (2011). [CrossRef]
29. J. Miller and J. Miller, Statistics and Chemometrics for Analytical Chemistry, 4 edition (Prentice Hall, 2001).
30. J. El Haddad, F. de Miollis, J. Bou Sleiman, L. Canioni, P. Mounaix, and B. Bousquet, “Chemometrics Applied to Quantitative Analysis of Ternary Mixtures by Terahertz Spectroscopy,” Anal. Chem. 86(10), 4927–4933 (2014). [CrossRef] [PubMed]
31. H. Mohamadi Monavar, N. K. Afseth, J. Lozano, R. Alimardani, M. Omid, and J. P. Wold, “Determining quality of caviar from Caspian Sea based on Raman spectroscopy and using artificial neural networks,” Talanta 111, 98–104 (2013). [CrossRef] [PubMed]
32. D. Lin, S. Feng, J. Pan, Y. Chen, J. Lin, G. Chen, S. Xie, H. Zeng, and R. Chen, “Colorectal cancer detection by gold nanoparticle based surface-enhanced Raman spectroscopy of blood serum and statistical analysis,” Opt. Express 19(14), 13565–13577 (2011). [PubMed]
33. S. Feng, D. Lin, J. Lin, B. Li, Z. Huang, G. Chen, W. Zhang, L. Wang, J. Pan, R. Chen, and H. Zeng, “Blood plasma surface-enhanced Raman spectroscopy for non-invasive optical detection of cervical cancer,” Analyst (Lond.) 138(14), 3967–3974 (2013). [CrossRef] [PubMed]
34. S. Li, Y. Zhang, J. Xu, L. Li, Q. Zeng, L. Lin, Z. Guo, Z. Liu, H. Xiong, and S. Liu, “Noninvasive prostate cancer screening based on serum surface-enhanced Raman spectroscopy and support vector machine,” Appl. Phys. Lett. 105(9), 091104 (2014). [CrossRef]
35. S. Feng, R. Chen, J. Lin, J. Pan, G. Chen, Y. Li, M. Cheng, Z. Huang, J. Chen, and H. Zeng, “Nasopharyngeal cancer detection based on blood plasma surface-enhanced Raman spectroscopy and multivariate analysis,” Biosens. Bioelectron. 25(11), 2414–2419 (2010). [CrossRef] [PubMed]
36. S. R. Hawi, W. B. Campbell, A. Kajdacsy-Balla, R. Murphy, F. Adar, and K. Nithipatikom, “Characterization of normal and malignant human hepatocytes by Raman microspectroscopy,” Cancer Lett. 110(1-2), 35–40 (1996). [CrossRef] [PubMed]
37. S. Luo, C. Chen, H. Mao, and S. Jin, “Discrimination of premalignant lesions and cancer tissues from normal gastric tissues using Raman spectroscopy,” J. Biomed. Opt. 18(6), 067004 (2013). [CrossRef] [PubMed]
38. J. Wang, D. Lin, J. Lin, Y. Yu, Z. Huang, Y. Chen, J. Lin, S. Feng, B. Li, N. Liu, and R. Chen, “Label-free detection of serum proteins using surface-enhanced Raman spectroscopy for colorectal cancer screening,” J. Biomed. Opt. 19(8), 087003 (2014). [CrossRef] [PubMed]
39. D. Lin, J. Pan, H. Huang, G. Chen, S. Qiu, H. Shi, W. Chen, Y. Yu, S. Feng, and R. Chen, “Label-free blood plasma test based on surface-enhanced Raman scattering for tumor stages detection in nasopharyngeal cancer,” Scientific Reports 4, (2014).
40. S. Feng, J. Pan, Y. Wu, D. Lin, Y. Chen, G. Xi, J. Lin, and R. Chen, “Study on gastric cancer blood plasma based on surface-enhanced Raman spectroscopy combined with multivariate analysis,” Sci. China: Life Sci. 54(9), 828–834 (2011). [CrossRef] [PubMed]
41. J. Lin, R. Chen, S. Feng, J. Pan, Y. Li, G. Chen, M. Cheng, Z. Huang, Y. Yu, and H. Zeng, “A novel blood plasma analysis technique combining membrane electrophoresis with silver nanoparticle-based SERS spectroscopy for potential applications in noninvasive cancer detection,” Nanomedicine (Lond. Print) 7(5), 655–663 (2011). [PubMed]
42. R. Wehrens, Chemometrics with R - Multivariate Data Analysis in the Natural Sciences and Life Sciences (Springer, 2011).
43. B. Yan, Y. Li, G. Yang, Z. N. Wen, M. L. Li, and L. J. Li, “Discrimination of parotid neoplasms from the normal parotid gland by use of Raman spectroscopy and support vector machine,” Oral Oncol. 47(5), 430–435 (2011). [CrossRef] [PubMed]
44. K. Raj, A. Kardam, J. K. Arora, and S. Srivastava, “Artificial Neural Network (ANN) design for Hg–Se interactions and their effect on reduction of Hg uptake by radish plant,” J. Radioanal. Nucl. Chem. 283(3), 797–801 (2010). [CrossRef]
