1.

Introduction

Nasopharyngeal cancer (NPC) is a nonlymphomatous squamous cell carcinoma that occurs in the epithelial lining of the nasopharynx.¹ The incidence of NPC occurs with much greater frequency in southern China, northern Africa, and Alaska. It is reported that the disease is the third most common malignancy among men, with an incidence about 50 per 100,000 in the Guangdong Province of China.² At present, the clinical diagnosis of NPC mainly depends on white-light endoscope technique and excisional biopsy. Endoscope technique relies on the observation of gross morphological changes of tissues, thus, it is difficult to identify the NPC in its early stages. Excisional biopsy remains the gold standard approach for cancer diagnosis, but it involves a high level of operational requirements for doctors and causes a lot of discomfort or may even be impractical for high-risk patients.³ Therefore, it is eminently desirable to develop a nondestructive technique to detect NPC at an early stage.

In recent years, the Raman spectroscopy technique has received a great deal of interest in cancer diagnosis.^4,5 Raman spectroscopy is a vibrational spectroscopic technique that can provide specific spectroscopic fingerprint information about the molecular composition, structure and content of constituents.⁶ Compared with other optical spectroscopic techniques, such as the fluorescence spectroscopy and infrared absorption spectroscopy, Raman spectroscopy has significant advantages. For instance, there is no photobleaching in Raman scattering, and Raman spectral peaks are narrow.^7,8 Especially, when the near-infrared excitation light is used, water exhibits very low absorption and tissues reveal far less autofluorescence compared with visible light excitation. This makes it easier to detect biochemical components within a deeper layer of the tissue. Over the past decade, Raman spectroscopy has been applied to classify normal and malignant tissues of various body sites, including the breast, bladder, lung, prostate, cervix, skin, and so on.^4,9–15 These studies show that Raman spectral features could be used to correlate with the molecular and structural changes associated with carcinomatous transformations, demonstrating the feasibility of early cancer detection by Raman spectroscopy.

The differences of Raman spectra among different tissue types are usually subtle because of obvious spectral overlapping. Thus, powerful and robust spectral data processing and diagnostic algorithms are required to extract significant Raman spectral features associated with the histopathology. In recent years, multivariate statistical algorithms such as principal component analysis (PCA), linear discriminant analysis (LDA) and partial least squares discriminant analysis (PLS-DA) have been widely applied for the classification of various tissues by Raman spectroscopy.^4,16 These statistical methods improve the precision and reliability of the procedure.¹⁷

PCA is a mathematical tool that reduces the dimensions of the dataset by orthogonally projecting data onto a lower dimensional linear space, such that the variance of the projected data is maximized. Combined with LDA and support vector machines, it has been applied to distinguish the biomedical Raman spectra.^18,19 PCA is efficient for data classification, but PCs have no definite physical meanings because they are extracted from a linear combination of original variables. Furthermore, PCA is generally calculated using the entire spectrum that contains much redundant data and noise, which have no contribution to the principal components and reduce the performance of the PCA.²⁰

Feature selection is another data reduction technique that mines feature subsets from the original data space. The classification attributes of the subset are, as much as possible, to remain consistent with the raw data. The main advantages of feature selection are to decrease the number of variables to help understand the discovered pattern by eliminating irrelevant features from raw dataset, and to increase the accuracy of classification. To improve the ability of finding a meaningful low-dimensional data structure in high-dimensional sample space, several feature selection strategies such as filter, embedded and wrapper have been developed according to different evaluation criteria.^21,22 The genetic algorithm (GA), based on Darwinian evolution and Mendelian genetics, has emerged as an adaptive heuristic search algorithm for efficient features selection.²³ Compared with other search methods, GA is robust, parallel, and has global search superiority. GA has displayed excellent ability in feature selection in microarray analysis, mass spectrometry, sequence analysis, and so on.^24–27 This study introduces GA combined with LDA method to select characteristic spectra for classification from Raman spectroscopy of NPC tissues. Furthermore, to compare the performance of GA-LDA, PCA-LDA algorithms are used to analyze the same Raman spectra of NPC tissues. The main purpose of this study is to explore the feasibility of classifying and diagnosing Raman spectra of NPC tissues with GA-LDA algorithms, and to provide an intuitive reference for the clinical diagnosis of Raman spectroscopy.

2.

Materials and Methods

2.1.

Patients and Tissue Sample

A total of 63 nasopharyngeal tissues were collected from 63 patients (17 women and 46 men with a median age of 44.6 years) who underwent endoscopic biopsies. Prior to research, all patients signed an informed consent to permit collection of excision specimens. After biopsies, the samples were divided into two parts, one for Raman measurements within 2 h and another for histopathologic examination conducted by a specialist nasopharyngeal pathologist after being fixed in 10% formalin solution. The results revealed that of the 63 nasopharyngeal tissue samples, 46 were dysplasia, and 17 were normal. Figure 1 shows the comparison of haematoxylin and eosin (H&E) stained tissue sections of normal and dysplastic nasopharyngeal tissues.

Fig. 1

(a) HE staining of normal nasopharyngeal tissue: a case showing nasopharyngeal pseudostratified ciliated columnar epithelium, no cell atypia. Magnification, $200\times$. (b) HE staining of nasopharyngeal carcinoma tissue: a case showing cancer cell nests (black line roped) with obvious nuclear atypia and abundant cytoplasm. Magnification, $200\times$.

3.

Results

Figure 2 shows the comparison of normalized average Raman spectra $-1$ standard deviations of NPC and normal tissues in the range from 720 to $1800{\mathrm{cm}}^{-1}$. Primary Raman peaks are observed in both NPC and normal tissue at the following peak positions with tentative assignments:^4,36–43 $788{\mathrm{cm}}^{-1}$ [$vs(\mathrm{O}─\mathrm{P}─\mathrm{O})$, cytosine, uracil ring breathing], $831{\mathrm{cm}}^{-1}$ [out of plane ring breathing tyrosine, $\mathrm{O}─\mathrm{P}─\mathrm{O}$ stretch], $855{\mathrm{cm}}^{-1}$ [$\nu (\mathrm{C}─\mathrm{C})$ ring of proline, ring breathing of tyrosine], $940{\mathrm{cm}}^{-1}$ [$\nu (\mathrm{C}─\mathrm{C})$ of proline, valine and protein backbone], $1004{\mathrm{cm}}^{-1}$ [$\nu s(\mathrm{C}─\mathrm{C})$ phenylalanine], $1033{\mathrm{cm}}^{-1}$ [$\mathrm{C}─\mathrm{H}$ in plane bending of phenylalanine], $1095{\mathrm{cm}}^{-1}$ [$vs$ (${\mathrm{PO}}_{2^{-}}$), $\nu (\mathrm{C}─\mathrm{N})$ of protein], $1128{\mathrm{cm}}^{-1}$ [$\nu (\mathrm{C}─\mathrm{C})$ of $\text{lipids}/\nu (\mathrm{C}─\mathrm{N})$ of protein], $1158{\mathrm{cm}}^{-1}$ [$\mathrm{C}─\mathrm{C}$/$\mathrm{C}─\mathrm{N}$ strething mode of proteins], $1208{\mathrm{cm}}^{-1}$ [$\nu (\mathrm{C}─{\mathrm{C}}_6{\mathrm{H}}_5)$) L-tryptophan and phenylalanine],⁴¹ $1339{\mathrm{cm}}^{-1}$ [${\mathrm{CH}}_3{\mathrm{CH}}_2$ wagging of collagen and polynucleotide chain], $1449{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$ of proteins and lipids], $1660{\mathrm{cm}}^{-1}$ [amide I $\nu (\mathrm{C}═\mathrm{O})$ collagen, $\alpha \text{-Helix}/\nu (\mathrm{C}═\mathrm{C})$ of lipid]. The strongest peaks are at 1004, 1339, 1449, and $1660{\mathrm{cm}}^{-1}$. The spectral differences between NPC and normal tissues are clearly displayed from the difference spectra of Fig. 2(c), implying the enormous potential to diagnose NPC with Raman spectroscopy.

Fig. 2

Average Raman spectra of NPC tissue in the range from 720 to $1800{\mathrm{cm}}^{-1}$. The solid lines indicate the average spectra and the shaded lines represent one standard deviation. (a) NPC tissue spectra, (b) normal tissue spectra, (c) cancer-normal difference spectra (the difference spectra intensity is enlarged five times for clear display).

The GA-LDA algorithms are developed to search for the significant Raman spectral features that are relevant to different nasopharyngeal tissue pathologies. The algorithm parameters are adjusted repeatedly to obtain the best subset of Raman variables for tissue differentiation. Figure 3 displays the best and $\mathrm{mean}\pm 1\mathrm{SD}$ area under ROC curve (AUC) of the individuals in 50 generations with GA-LDA algorithms. There is a significant increase of AUC by the GA-LDA algorithm. In order to pick out diagnostic feature information contained in Raman spectra, the GA-LDA algorithms are repeated over 100 runs for selecting seven spectral bands. Figure 4 is the cumulative counts of Raman bands chosen with GA-LDA algorithms. It reveals the most useful spectral feature located in the regions of 792–805, 867–880, 996–1009, 1086–1099, 1288–1304, 1663–1670, and $1742–1752{\mathrm{cm}}^{-1}$. The corresponding tentative assignments of significant Raman bands in nasopharyngeal tissue are listed in Table 1.^{37,38,41–43} The intensity differences of these Raman features between normal and cancerous tissues are also verified to be significant ($p<0.005$, unpaired Student’s t-test, 2-sided, equal variances for classification and diagnosis of nasopharyngeal tissues.

Fig. 3

The mean area under ROC curve (AUC) of the individual $\pm 1$ SD versus the best performance individuals in 50 generations with GA-LDA (20 individual, $\mathrm{cross}\mathrm{rate}=0.7$, $\mathrm{mutation}\mathrm{rate}=0.05$).

Fig. 4

The cumulative counts of Raman bands chosen with GA-LDA in 100 runs.

Table 1

Tentative assignment of significant Raman bands identified by GA-LDA algorithm.

The scatter plot of the linear discriminant scores of normal and NPC tissue using GA-LDA diagnostic model based on the seven significant Raman bands together with the leave-one-out cross-validation is displayed in Fig. 5. The dividing line produces diagnostic sensitivity of 69.2%, specificity of 100.0% and overall accuracy of 76.9% for discriminating NPC from normal nasopharyngeal tissue. Hence, the results of classification suggest that the GA-LDA algorithms develop a novel way to diagnose NPC from normal nasopharyngeal tissues by searching significant Raman features to build the discriminant model.

Fig. 5

Scatter plot of the linear discriminant scores of normal and NPC tissue using GA-LDA. The separate line produces diagnostic sensitivity of 69.2% ($117/169$) and specificity of 100.0% ($56/56$) for discriminating NPC from normal nasopharyngeal tissue.

To further evaluate the performance of the diagnostic model developed by GA-LDA algorithms, the ROC curve is generated (Fig. 6) from the scatter plot in Fig. 5. The ROC curve is an intuitive method of reflecting the relationship of sensitivity and specificity. It is obtained by calculating the sensitivity and specificity of different diagnostic thresholds. The integration area under the ROC curves is positively correlated with the diagnostic accuracy. To comparatively assess the effective diagnostic performances of GA-LDA algorithms, the ROC curve of PCA-LDA algorithms are also generated. It is illustrated by the improvement in the diagnostic sensitivities and specificities that GA-LDA algorithms give more effective diagnostic performances for differentiation of NPC from normal tissues. The integration areas under the ROC curves of GA-LDA algorithms and PCA-LDA algorithms are 0.956 and 0.924, respectively. These results further confirm that GA-LDA algorithms yield a better diagnostic accuracy than the PCA-LDA algorithms.

Fig. 6

Receiver operating characteristic (ROC) curves of Raman spectra discrimination results from NPC and normal tissue with PCA-LDA and GA-LDA algorithms together with the leave-one-out cross-validation method. The integration area under the ROC curves for PCA-LDA and GA-LDA are 0.924 and 0.956, respectively, illustrating the efficacy of GA-LDA in NPC diagnosis with Raman spectroscopy.

4.

Discussions

Raman spectroscopy is a unique noninvasive detection technique that can produce abundant information about molecular composition and structure of biological tissue. It may become a promising clinical diagnostic tool by probing subtle molecular changes relevant to tissue pathology. The Raman signal is very weak, but with the development of confocal Raman microscopy spectroscopy techniques, the development of Raman spectroscopy in biomedical field is greatly enhanced.^44–47 NPC is a malignancy of the head and neck with a marked racial and geographic distribution. NPC detection in the early stage is often difficult due to the nonspecific symptoms. Several groups have been studying the diagnosis of NPC with Raman spectroscopy technique. For example, Lau et al. implemented a preliminary study of normal and NPC tissue from six patients with Raman spectroscopy, they found consistent differences in three bands 1290–1320, 1420–1470, $1530–1580{\mathrm{cm}}^{-1}$.⁴⁰ Feng et al., researched surface-enhanced Raman scattering of human nasopharyngeal tissue with gold nanoparticle. They found that Raman spectral imaging based on three characteristic peaks at 962, 725, and $1366{\mathrm{cm}}^{-1}$, confirmed there were vigorous metabolism and strong enzymatic activities in tumor tissues.⁴¹ This study demonstrates the great potential of Raman spectra for probing the biochemical heterogeneities of nasopharyngeal cancerous tissues using gold nanoparticles. But there are fluctuations depended on the excitation light intensity in the Raman scattering signals at the tissue level. Later Feng et al., measured blood plasma surface-enhanced Raman spectroscopy of NPC patients and normal volunteers. They successfully separated the two groups’ spectra with sensitivity of 90.7% and specificity of 100% by PCA-LDA technique.³⁷

In this work, we have studied Raman spectral properties of NPC and normal tissue with confocal Raman microscopy and explored the potential to improve the diagnostic performance with feature selection technique based on GA-LDA algorithms. The distinct spectra differences of normal and NPC tissue are observed from difference spectra in Fig. 2(c). It indicates that there is a great potential of differentiating dysplasia from normal nasopharyngeal tissue with Raman spectroscopy. The main intensity differences are located at around $788{\mathrm{cm}}^{-1}$ [$vs(\mathrm{O}─\mathrm{P}─\mathrm{O})$, cytosine, uracil ring breathing], $1004{\mathrm{cm}}^{-1}$ [$\nu s(\mathrm{C}─\mathrm{C})$ phenylalanine], $1339{\mathrm{cm}}^{-1}$ [${\mathrm{CH}}_3{\mathrm{CH}}_2$ wagging of collagen and polynucleotide chain], $1449{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$ of proteins and lipids], $1660{\mathrm{cm}}^{-1}$ [amide I $\nu (\mathrm{C}═\mathrm{O})$ collagen, $\alpha \text{-}\text{Helix}/\nu (\mathrm{C}═\mathrm{C})$ of lipid]. These results are basically in agreement with the other reports about NPC.^5,40,41 In the same range of 950 to $1650{\mathrm{cm}}^{-1}$, our Raman spectral shape is the same to the Lau et al.’s, results, but different with the Feng et al.’s. The changes of Raman spectral shape are due to the use of surface enhanced Raman spectroscopy technology which only enhances some peaks.^40,41

However, the Raman spectral pattern of normal and cancer tissue is very similar. There is an urgent need to develop the efficient diagnostic algorithms to interpret these tiny spectral changes. To solve the above problems, we employ feature selection technique based on GA-LDA algorithm to correlate significant spectral bands with tissue pathology. The GA-LDA results display that diagnostically important variables are limited to several Raman spectral bands that are related to proteins, nucleic acids and lipids. The important seven spectral bands are distributed in the regions around 792–805, 867–880, 996–1009, 1086–1099, 1288–1304, 1663–1670, and $1742–1752{\mathrm{cm}}^{-1}$. These Raman bands mainly correlate with dysplasia progression. For instance, the Raman peak $867–880{\mathrm{cm}}^{-1}$ ($\mathrm{C}─\mathrm{C}$ stretching mode of collagen) is found to decrease significantly, indicating that there is a relative reduction of collagen content in cancer tissue. The main cause is that cancer cells spread to underlying stromal layer and express a class of metalloprotease, leading to an overall decrease of collagen content in cancer tissue.⁴⁸ The decrease of Raman bands $1288–1304{\mathrm{cm}}^{-1}$ (CH3CH2 bending of lipid) and $1742–1752{\mathrm{cm}}^{-1}$ ($\mathrm{C}═\mathrm{O}$ stretching of phospholipids) associated with lipids illustrates that the proportion of fat content in the cancerous tissue is greatly reduced. The likely reason is that amplified cancer cells consume a lot of fat, resulting in decrease of lipid molecules.^49,50 Nucleic acid-related Raman bands of $1086–1099{\mathrm{cm}}^{-1}$ (${\mathrm{PO}}_2^{-}$ Symmetric breathing of nucleic acids) are significantly enhanced, as cancer cells proliferate indefinitely, and DNA replicates greatly, leading to increased content of cells DNA.^41,42 In the Raman bands of $1086–1099{\mathrm{cm}}^{-1}$ and $1530–1560{\mathrm{cm}}^{-1}$, our difference spectra are consistent with Lau et al.’s, results. While in the Raman bands of 1420 and $1470{\mathrm{cm}}^{-1}$, our difference spectra are different from their results. This may be attributed to the measurement conditions and amount of sample. We measured 63 patients with the beam spot diameter about 5 μm, and Lau measured 6 patients with the beam spot diameter of 3.5 mm.

There are several indicators to evaluate the diagnostic capabilities of a particular algorithm such as sensitivity, specificity, accuracy and AUC, in which AUC can be a comprehensive reflection of the diagnostic algorithm to identify the disease. Other teams have also chosen accuracy as evaluation standard.⁵¹ In this study, AUC is adopted as the fitness function of GA and has obtained better search results in Fig. 3. The best AUC is 0.956, corresponding diagnostic accuracy 0.769. We also attempt to use the diagnostic accuracy as fitness function to search the characteristic Raman bands, and get the best AUC 0.944, accuracy 0.780. Although the results of the evaluation functions are close, we believe that the AUC is more suitable to assess the diagnostic model.

The diagnostic model built by GA-LDA technique searching important Raman feature spectra of nasopharyngeal tissue produces a sensitivity of 69.2%, specificity of 100.0% and overall accuracy of 76.9%. The area of 0.956 under the ROC curve further affirms the efficient diagnostic performances of GA-LDA-based algorithms for classification NPC from normal tissue. To compare the performances of GA-LDA algorithms with conventional multivariate statistical technique, PCA-LDA algorithms are applied to classify the same Raman dataset. The cumulative contribution rate in the first 20 PCs is 95.3% of total variance, corresponding diagnostic accuracy of 73.3%, sensitivity of 63.3% and specificity of 94.6% with unpaired two-side t-test ($p<0.05$) together with the leave-one-out cross-validation. The integration area of ROC curve is 0.924 for PCA-LDA algorithms as display in Fig. 6. Although PCA-LDA algorithms yield good classification effects, it can’t explain the contributions to discriminant results come from what biochemical component associated with tissue malignancy because PCA variables are extracted from entire Raman spectra of nasopharyngeal tissue. The GA-LDA algorithms provide a simplified diagnostic model with improved diagnostic accuracy by the selection of significant Raman feature bands related to biochemical components such as proteins, nucleic acids and lipids. The possible reason of GA-LDA algorithms improving diagnostic performances is that GA is a heuristic optimization technique that can search characteristic Raman spectra with distinct classification attributes.³⁰ Other groups have also applied feature selection technique based on GA with Raman spectroscopy for the classification. For example, Duraipandian et al., use GA partial least squares discriminant analysis and Raman spectroscopy to identify biomolecular changes of cervical tissue associated with dysplastic transformation. They achieved a diagnostic accuracy of 82.9%, sensitivity of 72.5% and specificity of 89.2% for precancer detection.²⁷ Lavine et al., successfully classified several wood types by the application of GA to extract features of woods’ Raman spectroscopy.⁵² All of above results demonstrate that the GA-LDA algorithms are very attractive techniques that are expected to provide a more accurate diagnostic model for the clinical application of Raman spectroscopy.

5.

Conclusion

In conclusion, an efficient diagnostic model based on GA-LDA algorithms picking out the characteristic Raman bands associated with biochemical components is developed and applied to classify the NPC from normal tissue. An improved diagnostic accuracy of 76.9%, sensitivity of 69.2% and specificity of 100.0% is obtained. Compared with the PCA-LDA algorithms, the GA-LDA technique can build a simpler diagnostic model with clearer physical meaning and higher diagnostic efficiency. This work demonstrates that the Raman spectroscopy associated with GA-LDA diagnostic algorithms has enormous potential to detect and diagnose NPC.