Methods of spectral pretreatment can be adopted to reduce the effects of collineation, dimension, background, noise level and baseline drift of spectrum. Multiple Scatter Correction (MSC) [13] [14] is a kind of very efficient data-processing method in spectra analysis, which can be used to eliminate the scattering effect and enhance the absorption of spectral information that related to the components.

Principal component analysis (PCA) is one of the most commonly used unsupervised recognition methods, which means no prior knowledge is available about the group to which samples belong. PCA of a data matrix extracts the dominant patterns in the matrix, and represents it as new orthogonal variables (principal components), and display the pattern of similarity of the observations and the pattern of the variables as points in map [15] [16] . High dimensional data are often transformed into lower dimensional data via PCA method, and value characteristics are retained simultaneously.

Non-negative matrix factorization (NMF) is another unsupervised recognition method, which can divide large matrix into two small matrices, and the decomposition of the matrix does not contain negatives. The non-negativity makes the matrices easier to inspect. Also, non-negativity is meaningful for the data in actual applications [17] .

Partial least squares-discriminant analysis (PLS-DA) consists of a classical PLS regression wherein the response variable is a categorical one expressing the class membership [18] [19] [20] . It uses prior knowledge about the group to which samples belong. Instead of finding hyperplanes of minimum variance between independent variables and the response, it finds a linear regression model by projecting the predicted variables and the observable variables to a new space.

Linear correlation analysis (LDA) also named as FLD (Fisher Linear Discriminant), is a classical pattern recognition method that could put the high dimensionality of the data through maximum Fisher criterion to reduce into the low dimensional [21] . Through the observation of the relationship between the data points and the linear degree, the data are analyzed correctly.

Uncorrelated linear discriminant analysis (ULDA) is a kind of feature extraction and dimension reduction algorithm based on LDA, aiming to maximize separate different kinds of samples, which extracts the uncorrelated discriminant vector (UDV) with largest discriminant ability. At the same time, the vectors obtained are uncorrelated, which makes the information redundancy minimal [22] [23] [24] . ULDA could get the results that the linear correlation analysis method could not obtain. ULDA [25] [26] [27] was first put forward by Jin and others in the field of face recognition. Now it has been successfully applied in analysis of metabolomics, proteomics and gene expression profile. The specific steps of ULDA algorithm are referred to literature [28] .

Serum samples were provided by the Chinese people’s liberation army 252 hospital in Hebei province, which were composed of 68 patients with lung cancer (stage I) and 29 healthy people. The age of lung cancer group (65.32 ± 10.13) years old, including 40 males and 29 females. The age of the healthy group was (63.45 ± 8.32) years, including 15 males and 14 females. There was no statistically significant difference between the two groups in age and sex. The day before the blood collection, the relevant person cannot eat greasy food and drink alcohol. Study subjects were admitted to the hospital for the next day to take the morning fasting venous blood 4 ml, and then the blood were centrifuged for 10 minutes by 5000 rpm. Then the serums were placed into the test tube and stored in the refrigerator with −80˚C.

HORIBA Jobin Yvon micro-confocal Raman spectrometer (LabRAM XploRA ONE) was used. The serum samples were moved gradually from −40˚C - −20˚C ~ −4˚C - 24˚C (room temperature) for 30 minutes respectively. Then the thawed serums were taken out 200 μL in a liquid quartz cuvette and the cuvette was put horizontally on the stage. The laser wavelength, transmittance and grating were adjusted by univariate method, and the best experimental conditions were obtained. The laser wavelength was 532 nm, the transmittance was 50%, and the grating was 600 gr/mm. In the case of wavelengths of 100 - 5000 cm⁻¹, each sample was selected for three different angles to repeat the scan, resulting in three spectra, recording the average measured value. Finally, the spectra are smoothed for further calculation.

The first step is to preprocess the spectral data: compare the different spectral preprocessing methods, including normalization, scaling, concentration and MSC, and finally selecting the MSC to process the original spectral data. The second step is to model the data: lung cancer sample data and healthy human sample data PLS-DA, ULDA and LDA modeling two groups. 34 lung cancer samples and 14 healthy subjects were modeled, and the remaining 34 lung cancer samples and 15 healthy subjects were predicted.

Sensitivity and specificity are two statistical parameters used to measure diagnostic modeling accuracy. Ideally, the two parameters should be close to one.

Sensitivity = true positive/(true positive + false negative) × 100%

Specificity = true negative/(true negative + false positive) × 100%

Analysis of the spectral data using self-compiled program to complete, the calculation software used for the Matlab 7.0.

It is investigated that the first two principal components (PC1 and PC2) had explained most variances of the data sets. The score values of lung cancer and healthy group were plotted against the sample index in Figure 2. It can be clearly observed that, the data points of two groups are seriously dispersed and overlapped, and it is hard to separate the two groups. Even though more principal components were used for the data, the classification results were not significantly improved. So the patients and healthy people could not be distinguished by MSC-PCA method.

The classification results of patients of lung cancer and healthy people by MSC-NMF were also showed in Figure 2. The values of non-negative factorization NNF1 and NNF2 were plotted against the sample index. As can be seen in Figure 2, the results are similar with those of PCA, so the patients and healthy people also could not be distinguished by MSC-PCA method.

Above all, when no prior knowledge is available about the group to which samples belong, the two common unsupervised recognition methods are not suitable for extracting diagnostic information on the data in this study. Next, supervised recognition methods were adapted to analyze and classify the data.

In this study, the number of n_f in the PLS-DA modeling was set at 13 after optimization from 5 - 15. As a supervised recognition method, the lung cancer group was marked as [1, 0] and the healthy group was marked as [0, 1] in advance. The prediction results of MSC-PLSDA are given in Figure 3. It was found that the samples were classified with 52.94% (18/34) sensitivity and 33.33% (5/15) specificity. Though the obtained prediction results were more intuitive than those of PCA and MMF, the lower prediction sensitivity and specificity could not meet the demand of the clinical diagnosis by a great deal.

It is estimated that PLS-DA is essentially a feature transformation way, the new variables are some type of combination of the original variables. The variables with large variance or high covariance may affect the modeling, although those variables contain little or even no message contributing to the discrimination of samples, which may result in the loss of optimal features in some case [31] [32] . So it seemed that MSC-PLSDA was not a brilliant method on Raman spectroscopy that could help us to distinguish and identify the samples consisted of lung cancer and healthy people.

At first, the data was not disposed by MSC method, the classification results by sole LDA and ULDA individually were shown in Figure 4. Due to the number of

categories used in this study was two (malignant and healthy), only one discriminant vector (DV) or uncorrelated discriminant vector (UDV) was obtained, which results in the decreased complexity of the model and the enhanced explain ability of the model. It can be clearly observed from Figure 4, although the cancer and healthy groups cannot completely be distinguished, the values of DV and UDV in the vertical axis had shown obvious clustering. In the case of the dotted lines in the figures as the classification boundaries, the prediction sensitivity of LDA reached 88.24% (30/34) and the prediction specificity of LDA reached 80.00% (12/15). It was also found with the same sensitivity of 88.24% and the specificity of 80.00% in ULDA method.

The basic idea of LDA is to project the high dimensional pattern samples to the best discriminant vector space, and to make the maximum of between-class scatter matrix and the smallest of within-class scatter matrix. That is to say, the projection patterns in the new space are ensured with the minimum between-class distance and the maximum within-class distance, and accordingly the patterns in the space have the best separability. ULDA is developed base on LDA algorithm, and considers no correlation between column vectors in the transformation matrix; therefore, it can reduce the data redundancy after dimension reduction and simultaneously avoid the shortage of sample lacking [33] . Probably because of above causes, the classification performances of LDA and ULDA far exceed those of PCA, NMF and PLS-DA when they are applied to data of high dimensional pattern samples.

Then the data were further pre-processed with MSC method, and the classification results by MSC-LDA and MSC-ULDA were given in Figure 5 separately. It can be clearly observed from Figure 5 that cancer group and healthy group were all completely distinguished and clustered with 100% sensitivity and 100% specificity. Compared with the results of sole LDA and ULDA method, sensitivity and specificity of distinguishing were significantly improved, which reveals MSC is a very effective preprocessing method for extracting illness characterize information on serum Raman spectra. The results of the above methods used were list in Table 1. Therefore, it can be concluded that MSC-LDA and MSC-ULDA may be more feasible methods which could distinguish serum samples that consisted with lung cancer and healthy people.