1.

Introduction

Decreasing incidence and mortality rates have been reported for gastric cancer in most developed countries and also in selected developing countries over the last few decades. Still, 723,100 people died of gastric cancer, 951,600 new cases occurred worldwide in 2012,^1–3 and $>40\%$ of these patients were Chinese.³ Early detection and effective treatment of gastric cancer have a great impact on the 5-year survival rates of patients,⁴ but unfortunately, even in Japan (which is a highly developed Asian country), difficulties occur in early detection of gastric cancer because of a shortage of skilled endoscopists and limited availability of endoscopic devices,⁵ not to mention the problems and possibilities of nationwide screening for gastric cancer in China, which has the largest population in the world and great imbalances in regional development. Histological examination of gastric biopsy has been widely used since the popularization of endoscopy,⁶ which remains the gold standard for gastric cancer diagnosis. However, gastric biopsy under direct vision suffers from several disadvantages, such as lack of nondestructive character, extension of endoscopic examination time, and a delay of several days for the diagnostic result provided to patients. Consequently, additional efforts should be focused on the development of endoscopic diagnostic techniques that are real-time, nondestructive, and objective.

First introduced by Raman in 1928, Raman spectroscopy has shown great potential in diagnosis.^7,8 This technique detects an emission spectrum generated by high-frequency inelastic scattering of light resulting from the response of a molecule to incident rays.⁷ Each molecular motion creates its own individual bands in the Raman spectrum.⁷ Hence, Raman spectroscopy reveals biomolecular changes in the human body caused by malignancy and presents an opportunity for use as a promising tool for doctors in cancer diagnosis. In recent years, Raman spectroscopy has been used to diagnose different types of cancers and precancers in vivo and ex vivo, including glioma, nasopharyngeal carcinoma, Barrett’s esophagus, lung cancer, breast cancer, and melanoma.^9–14 The sensitivity and specificity for diagnosis of malignancies reported in these research studies are satisfactory and encouraging.^9–14

The Raman probe is sufficiently slender to freely pass through the instrument channel of an endoscope. Gastric biopsy is a routine process for patients suspected of suffering from gastric cancer, and this technique makes it easier to obtain samples compared with parenchymal viscera (e.g., brain, liver). Many doctors and optical scientists attempted to improve the diagnostic effect of Raman spectroscopy for gastrointestinal tumors both in vivo and ex vivo.^15–19 However, most of these studies focus on the fingerprint (FP) Raman spectroscopy range of 800 to $1800{\mathrm{cm}}^{-1}$, which presents biochemical information on the backbone structures of DNA, RNA, proteins, lipids, collagen, and other biological macromolecules. In 2014, we also used the FP region of Raman spectroscopy to explore the biochemical characteristics of molecules associated with the malignant transformation of gastric mucosa.²⁰ In our previous study, the great value of the high wavenumber (HW) region (2400 to $3800{\mathrm{cm}}^{-1}$) in diagnosis was not yet evaluated and discussed.²⁰ Compared with the main advantages of the HW region, in which the intense background signal of the silica fiber-optic probe is absent and more intense tissue Raman signals are generated, it appears that the FP region provides much more information on the molecular composition and structure of tissue.^21–27 However, the background-signal-free design of the Raman probe for FP Raman spectroscopy is much more complicated and expensive,²¹ which could further affect development and medical application of in vivo diagnostic tools based on Raman spectroscopy.²¹ Therefore, it is worthwhile to comprehensively investigate the diagnostic effect of both the FP and HW regions.

In this research, we compared the diagnostic potential of FP (800 to $1800{\mathrm{cm}}^{-1}$) Raman spectroscopy with that of HW (2800 to $3000{\mathrm{cm}}^{-1}$) Raman spectroscopy in gastric cancers. The biochemical changes of molecules found in both the FP region and the HW region were thoroughly reviewed as well. Multivariate statistical techniques, including partial least-squares-discriminant analysis (PLS-DA), were used to design diagnostic algorithms for discrimination between normal mucosa and gastric cancer.

2.

Materials and Methods

3.

Results

Comparison of the mean intensities of FP and HW Raman spectra with their standard deviations (SD) is shown in Fig. 1. The prominent tissue Raman peaks in the FP region are observed at 829, 853, 879, 937, 1003, 1033, 1086, 1127, 1157, 1173, 1207, 1319, 1338, 1448, 1526, and $1658{\mathrm{cm}}^{-1}$. Tentative vibrational mode assignments for the human gastric tissue Raman peaks in the FP region are summarized in Table 1 of our previous work.²⁰ Moreover, intense Raman peaks are also found in the HW region, e.g., $2849{\mathrm{cm}}^{-1}$ (symmetric ${\mathrm{CH}}_2$ stretching of lipids), $2875{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_2$ symmetric stretch of lipids and ${\mathrm{CH}}_2$ asymmetric stretch of lipids and proteins), and $2932{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_3$ stretching of proteins).^21,33 Independent-sample $t$-tests showed that Raman peak intensities are notably higher for gastric cancer at $1157{\mathrm{cm}}^{-1}$ (polyene chain of carotenoids), $1173{\mathrm{cm}}^{-1}$ [$\delta \mathrm{C}-\mathrm{H}$ in-plane bending of phenylalanine, tyrosine (protein)], 1319 and $1338{\mathrm{cm}}^{-1}$ ($\upsilon {\mathrm{CH}}_2{\mathrm{CH}}_3$ and $\delta {\mathrm{CH}}_2{\mathrm{CH}}_3$ of proteins and nucleic acids), and $1526{\mathrm{cm}}^{-1}$ ($\upsilon \mathrm{C}-\mathrm{C}$ of carotenoids), but notably lower differences are significant at $829{\mathrm{cm}}^{-1}$ ($\upsilon \mathrm{as}$ ${\mathrm{PO}}_2^{-}$ of nucleic acid), $853{\mathrm{cm}}^{-1}$ [$\upsilon \mathrm{C}-\mathrm{C}$ of proline (collagen)], $879{\mathrm{cm}}^{-1}$ [$\upsilon \mathrm{C}-\mathrm{C}$ hydroxyproline (collagen)], $1003{\mathrm{cm}}^{-1}$ ($\upsilon \text{ring}$ breathing of phenylalanine), $1033{\mathrm{cm}}^{-1}$ ($\delta \mathrm{C}-\mathrm{H}$ plane bending of aromatic compounds), $1448{\mathrm{cm}}^{-1}$ ($\delta {\mathrm{CH}}_2$ of proteins and lipids), $1658{\mathrm{cm}}^{-1}$ [amino compounds I ($\alpha$-helix)], and $2932{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_3$ stretching of proteins). To better reflect the differences in the main Raman peaks between normal tissue and gastric cancer, a comparison of the mean intensities and SDs is shown in Fig. 2.

Fig. 1

(a) Mean FP Raman spectra of 38 normal and 37 gastric cancer tissues. (b) Mean HW Raman spectra of 38 normal and 37 gastric cancer tissues. The Raman spectra for normal tissues (blue line) are shifted for better visualization. The red-shaded and blue-shaded areas represent the respective SD of the means. Difference spectra (normal-gastric cancer) calculated from the average Raman spectra between normal tissue and gastric cancer are shown with a green line, and the main Raman peaks are marked.

Fig. 2

Comparison of the mean intensities and SD of selected Raman peaks showing significant differences in independent-sample $t$-tests between normal tissues and gastric cancer.

The PLS-DA with leave-one-sample-out cross validation was applied to generate diagnostic algorithms for differentiation between normal tissue and gastric cancer. As shown in Fig. 3, six LVs were found to be the optimal number of retained components for PLS-DA modeling of both the FP region and HW region as defined by the minimum error rate of cross validation. To demonstrate the use of LVs for diagnosis and the intrinsic clustering of normal mucosa and gastric cancer, two three-dimensional (3-D) scatterplots are presented in Fig. 4 using the three LVs (LV1, LV2, and LV3) of the FP and HW regions, respectively. The LV points are divided into two distinct groups, suggesting that both FP and HW Raman spectroscopy have potential for use as a clinical diagnosis tool for gastric cancer. The developed PLS-DA algorithm of the FP region achieves diagnostic sensitivities of 94.59%, specificities of 86.84%, and accuracies of 90.67%. The HW modeling achieves diagnostic sensitivities of 83.78%, specificities of 71.05%, and accuracies of 77.33%. The ROC curves (Fig. 5) are also generated for the two spectral datasets (FP, HW) to further evaluate the diagnostic performance. The integration areas under the ROC curves for FP modeling are 0.9495, and for HW modeling, the number is 0.8254. The diagnostic performance comparisons (i.e., sensitivities, specificities, accuracies, and AUCs) of the FP and HW PLS-DA modeling are summarized in Table 2.

Fig. 3

Relationship between the numbers of PLS factors-LVs and the error rate of cross validation for correct diagnosis of normal mucosa and gastric cancer. The error rate shows that the optimal number of retained components for PLS-DA modeling of both the FP and HW regions is 6 LVs.

Fig. 4

(a) Three-dimensional scatter plots spanned by the three LVs (LV1, LV2, and LV3) of the FP region demonstrating the discrimination between normal mucosa and gastric cancer. (b) Three-dimensional scatter plots spanned by the three LVs (LV1, LV2, and LV3) of the HW region demonstrating the discrimination between normal mucosa and gastric cancer.

Fig. 5

ROC curves of diagnostic performance for the FP and HW Raman spectra for gastric cancer diagnosis via Raman spectroscopy and PLS-DA together with the leave-one-sample-out cross-validation method. The integrated areas under the ROC curves are 0.9495 and 0.8254 for the FP and HW Raman spectra, respectively.

Table 2

Comparison of diagnostic performance of different Raman spectral regions (FP, HW) for differentiation of gastric cancer from normal mucosa.

4.

Discussion

Raman spectroscopy facilitates the collection of specific molecular information for tissue diagnosis and characterization in a nondestructive manner. This distinctive vibrational spectroscopic technique allows researchers to study the tissue in its natural condition without the need for tissue preparation.¹⁸ Currently, Raman studies of gastric cancer have evaluated Raman spectra mostly in the FP region.^16–19,34 Only a few articles have reported Raman measurements from both the FP and HW regions,³⁴ but multivariate statistical analysis was not performed in these studies to evaluate the diagnostic potential of the FP and HW regions. In this work, we compared the diagnostic potential of FP Raman spectroscopy in gastric cancers with that of HW Raman spectroscopy and investigated the biochemical information contained in both spectral regions for the purpose of tissue characterization.

When normal tissues experience the process of malignant disease initiation, its structural conformation and relative content of macromolecules are altered. The main advantage of using the FP region is that it provides highly specific information on proteins, lipids, and nucleic acids associated with malignancy.³¹ For instance, in a previous study, we reported that the peak at $1526{\mathrm{cm}}^{-1}$ was significantly stronger in gastric cancer tissue than normal tissue ($p<0.05$), indicating that the carotenoid content in cancer is increased.²⁰ In this work, another peak at $1157{\mathrm{cm}}^{-1}$ (polyene chain of carotenoids) was also found to be significantly stronger in gastric cancer tissue ($p<0.05$). These observations provide further evidence for our assumptions that cancer cells probably evolve an enhanced ability to resist damage because of the antioxidant function of carotenoid.²⁰ In addition, the peak at $853{\mathrm{cm}}^{-1}$ [$\upsilon \mathrm{C}-\mathrm{C}$ of proline (collagen)] was observed to be lower for gastric cancer tissue. A possible reason for this result is that cancer cells synthesize and secrete matrix metalloproteinase that degrades matrix proteins such as collagen to assist cancer metastasis.²⁰ Huang et al.³⁵ also observed a decrease of collagen in gastric cancer in vivo using a novel image-guided Raman endoscopy technique. In addition, the protein and nucleic acid Raman signals at 1319 and $1338{\mathrm{cm}}^{-1}$, which most likely arise from the chromosomes in the nucleus, have been found to be significantly higher for gastric cancer tissues compared with normal tissue. This observation might be explained by an increase in nuclear chromatin and excessive DNA replication during malignant transformation. Therefore, the distinct differences in the FP region between normal tissues and gastric cancer reinforce the hypothesis that the FP region can be used to interpret the biochemical transformation of molecules associated with gastric mucosa malignancy.

However, the HW region (2800 to $3000{\mathrm{cm}}^{-1}$) appears to contain less biochemical information than the FP region. Raman peaks at 2849, 2875, and $2932{\mathrm{cm}}^{-1}$ are observed in the HW region. The protein signal at $2932{\mathrm{cm}}^{-1}$ was significantly stronger in normal tissue than in the gastric cancer tissue ($p<0.05$). The other two peaks showed no significant difference. However, this result does not mean that the HW region is less important for diagnosis. The HW region contains information related to ${\mathrm{CH}}_3$ stretching of proteins, which are complementary to the properties of the FP region.³⁶ In addition, the biochemical information provided by the HW region is sufficient for cancer diagnosis.²³ Nijssen et al. reported that a linear discriminant analysis (LDA)-based tissue classification model using the HW region (2800 to $3125{\mathrm{cm}}^{-1}$) discriminated between basal cell carcinoma (BCC) and normal skin with high accuracy.²³

We further used PLS-DA with leave-one-sample-out cross validation to generate diagnostic algorithms for gastric cancer using the FP region and the HW region. The PLS-DA modeling of the FP region, which achieved diagnostic accuracies of 90.67% ($68/75$) [sensitivities of 94.59% ($35/37$) and specificities of 86.84% ($33/38$)] performed better than the HW region modeling. The HW modeling misclassified 17 samples out of the total 75 samples, but the FP modeling reduced the number of misclassified samples to 7. Similar results were also observed in cervical precancers that were investigated using simultaneous FP (800 to $1800{\mathrm{cm}}^{-1}$) and HW (3100 to $3700{\mathrm{cm}}^{-1}$) confocal Raman spectroscopy.³¹ The ROC curves (Fig. 5) further validate the better performance and robustness of the FP region modeling. Fortunately, compared with the accuracy of gastric biopsy under direct vision, which was reported to vary from 50% to 96%,⁶ the performances of both FP modeling and the HW modeling are satisfactory for diagnosis of gastric cancer. Furthermore, we also performed the joint analysis of these two regions, but the diagnostic power is the same with the FP modeling. The FP + HW modeling also achieves diagnostic sensitivities of 94.59%, specificities of 86.84%, and accuracies of 90.67%.

Although our work showed that FP region modeling provided diagnostic advantages over HW region modeling, the development of both FP region and HW region diagnostic technologies should be encouraged because the HW region contains information related to lipids, proteins, and water bands that does not show up in the FP region.³⁶ Another reason is that the great difference in composition of human tissue (e.g., brain and bone) makes it worthwhile to find a suitable region of Raman spectroscopy for use in building diagnostic algorithms. In addition, it has been reported that diagnostic algorithms developed ex vivo might not be directly used for in vivo applications in the colon.³⁷ The performance comparison of the FP and the HW diagnostic algorithms developed for gastric cancer in vivo is not complete, therefore, further investigations are needed to determine whether the performance of the FP region modeling is still better than the HW region modeling in vivo. Another limitation of this study was the constraint on experimental conditions which means that the sample size of our study might not be sufficiently large to provide better PLS-DA modeling of both the FP region and HW region. We plan to enlarge the sample size to improve the diagnostic algorithms in our future studies. Currently, we are working on prototype spectrometers with a Raman probe that is compatible with the instrument channel of the endoscope for in vivo applications. The ultimate goal of our work is to develop a Raman spectroscopy diagnosis system that can be used as a diagnostic tool during screening for gastric cancer and will also provide an accurate gastrointestinal tumor margin assessment during endoscopic submucosal dissection. It is difficult to predict whether Raman spectroscopy can replace histopathology, but we do believe that this approach will provide good clinical tools for endoscopists in the future.

5.

Conclusions

We compared the diagnostic potential of FP Raman spectroscopy in gastric cancers with that of HW Raman spectroscopy ex vivo and uncovered the biochemical information contained in both spectral regions for the purpose of tissue characterization. Better differentiation between normal and gastric cancer tissues can be achieved using FP Raman spectroscopy and PLS-DA techniques. However, the HW region contains complementary spectral information that might have a positive effect for gastric cancer diagnosis. We expect that both FP Raman spectroscopy and HW Raman spectroscopy will become useful clinical diagnostic tools for the early detection of gastric cancer in the future.