2.1.

Patients and Tissues

In total, 119 patients were enrolled between 2008 and 2011. Of these, 102 patients underwent simple or radical gastrectomy for GISTs or adenocarcinomas, and the remaining 17 patients, underwent operations for other benign lesions, were obtained as the normal gastric mucosae. The formalin-fixed, paraffin-embedded tissue specimens were prepared as slides with a $3\text{-}\mu \mathrm{m}$ thickness and were subjected to Raman measurement. Before Raman measurement, the slides were treated with a dewaxing procedure with hexane for 5 min, ethanol for 3 min, methanol for 1 min, and then were dried with ${\mathrm{N}}_2$. The pathological diagnoses of these patients were reviewed by at least two experienced pathologists. Clinical data were retrospectively obtained from the patients’ medical records.

2.2.

Ethics Statement

The Institutional Review Board of Tungs’ Taichung MetroHarbor Hospital approved the study (approval No. 100006). All included patients provided written informed consent before the collection of specimens and clinical information.

2.3.

Raman Microspectroscopic Measurement

Confocal Raman microspectroscopy provides a platform for acquiring detailed Raman spectra from a small volume of specimens. Raman spectra of GIST, gastric adenocarcinoma, and normal gastric mucosa were obtained by using a confocal laser microRaman system (MploRA, Horiba Jobin-Yvon, France) with CCD detector. All Raman measurements were performed at room temperature. A Raman spectrometer with 532-nm laser excitation coupled with a confocal microscope (Olympus BX41, $\mathrm{NA}=0.9$) with $100\times$ objective lens was used. The solid-state diode laser with 8 mW was used as an excitation source. The laser power was attenuated after passing through optical elements (mirrors and filters) of light path, which would not degrade the sample. The excitation light beam was directly shining on the surface of specimen, and the backscattered light was collected, passed through the entrance slit ($100\text{-}\mu \mathrm{m}$ wide), dispersed by a diffraction grating ($1200\text{grooves}/\mathrm{mm}$), and detected by an air-cooled CCD detector. The exposure time for each spot was 100 s (integration time 1 s, accumulation: 100 scan), and 5 to 10 spots, depending on the sample size, were chosen for each specimen. The area was primarily diagnosed as a cancer abundant area by the pathologist. Raman spectra were produced over the Raman shift 500 to $3100{\mathrm{cm}}^{-1}$.

2.4.

Principal Component Analysis and Linear Discriminant Analysis

The spectra results were processed and analyzed with the ArrayTrack data analysis and interpretation tool at the U.S. Food and Drug Administration National Center for Toxicological Research.³³ Each spectrum from 500 to $3100{\mathrm{cm}}^{-1}$ was divided into 1022 segments and analyzed using principal component analysis (PCA) and linear discriminant analysis (LDA) to distinguish GISTs, adenocarcinomas, and normal mucosae. The first two principal components calculated (PC1 and PC2), which contained majority of the information, were plotted against each other for visualisation purposes. LDA with 20-fold cross-validation was used to evaluate the accuracy, sensitivity, and specificity of the differential diagnosis by confocal Raman microspectroscopy.

PCA is a multivariate technique used to classify and reduce the dimensionality of the spectral data. Orthogonal linear combinations to transform the original data into uncorrelated variables are termed PCs. LDA is another data reduction technique. The first few PCs were selected for LDA. PCA uses the most information from the original data and LDA maximizes the intergroup differences and minimizes the intragroup differences. Therefore, the eigenvectors of PCA and LDA are different^34,35. In 20-fold cross-validation, the original sample is randomly divided into 20 equal-sized subsamples. One of the 20 subsamples is retained as the validation data for testing the model, and the remaining subsamples are used as training data. The cross-validation process is then repeated 20 times, and the 20 results from the folds can then be averaged to produce a single estimation. Sensitivity (true positive rate) defined as measuring the proportion of positives that are correctly identified as such, and specificity (true negative rate) defined as measuring the proportion of negatives that are correctly identified as such. Receiver operating characteristic (ROC) curves were generated by successively changing the thresholds to determine correct and incorrect classifications for all subjects. The threshold of the sensitivity and specificity was defined as the maximum sum of the sensitivity and specificity (Youden index).^34–36 The ROC curve was made by PC1 using SigmaPlot software version 10.0 (Systat Software Inc. San Jose, California).

3.1.

Spectra of Raman Microspectrospy

Raman spectroscopy provides a lot of molecular information and is powerful in characterizing molecular structure. Varieties of Raman techniques, such as Fourier transform (FT)-Raman, resonance Raman, surface-enhanced Raman scattering, tip-enhanced Raman, coherence antistoke Raman, and confocal Raman microscope, have been developed for biological and biomedical applications. Each Raman technique exhibits its unique preference in a certain usage. Confocal Raman microscope offers decisive advantages in contrast enhancement, rejection of stray light, and discrimination of a well-defined spatial region in a complex multiphase specimen. Confocal microscope provides an efficient way to obtain interference-free Raman spectra. The optical microscope focusing laser point onto a diffraction-limit spot on the specimen improves significantly in the lateral resolution and depth discrimination. Zhou et al.²⁷ successfully demonstrated the usage of a confocal microRaman system with 532-nm excitation to study human brain cancer. The magic laser wavelength at 532 nm would cause resonance Raman scattering associated to vibration in the brain tissues. The strength of the resonance Raman signal excited by 532 nm is stronger than that by 732 nm because 532 nm falls in the special wing of the flavin fluorescence spectrum, and the background noise from the wing in Raman scattering is nearly gone.³⁷ Raman scattering with excitation in the infrared wavelength such as 1064 nm may eliminate fluorescence interference; however, it is not suitable for a CCD detector due to a dramatic drop in the quantum efficiency at 1064 nm. In contrast, coupling with the advances in the CCD detector, confocal Raman microcope provides a noninvasive, interference-free, less sample preparative, and spatially resolving method for studying the molecular composition of gastric tumors.

Figure 1 shows the microspectroscopic images of normal mucosae, GIST, and adenocarcinoma. The area was primarily diagnosed as a cancer abundant area by a pathologist. The arrow indicated the spot on the specimen exposing to the laser beam. It is difficult to differentiate those tissues based on the microscopic image (Fig. 1) without sample staining with hematoxylin and eosin (H&E) stain. However, Raman profiles exhibited significant difference in those gastric tissues as shown in Fig. 2. Raman spectra of normal mucosae, GISTs, and gastric adenocarcinomas were demonstrated in Figs. 2(a), 2(b), and 2(c), respectively. Characteristic Raman peaks at 861, 1004, 1098 to 1128, 1240, 1342, 1442, 1584, and $1655{\mathrm{cm}}^{-1}$ were assigned to the $\mathrm{C}─\mathrm{C}$ stretching, $\mathrm{C}─\mathrm{C}$ symmetric stretching, $\mathrm{C}─\mathrm{N}$ stretching, $\mathrm{C}─\mathrm{N}$ stretching and $\mathrm{N}─\mathrm{H}$ bending, ${\mathrm{CH}}_3{\mathrm{CH}}_2$ wagging, ${\mathrm{CH}}_2$ and ${\mathrm{CH}}_3$ bending, $\mathrm{C}═\mathrm{C}$ bending, and $\mathrm{C}═\mathrm{O}$ stretching, respectively.³⁸ These peaks represented for proline, phenylalanine, phospholipids, amide III, collagen, phospholipids, phenylalanine, and amide I, respectively.³⁸ There are significant differences in the characteristic Raman bands and the relative Raman intensity. Spectral difference among the GIST, gastric adenocarcinoma, and normal gastric mucosa was observed in three segments: 1098 to $1128{\mathrm{cm}}^{-1}$, 1240 to $1342{\mathrm{cm}}^{-1}$, and 1584 to $1655{\mathrm{cm}}^{-1}$. For example, the Raman signal at $1098{\mathrm{cm}}^{-1}$ of adenocarcinoma splitted into two peaks of 1098 and $1128{\mathrm{cm}}^{-1}$ for the GISTs and normal mucosae. This suggests that the character of phospholipids of the adenocarcinomas is different from that of the normal mucosae and GISTs. The intensity ratio of Raman peaks at 1240 and $1342{\mathrm{cm}}^{-1}$ exhibited differently too. In the GIST, the intensity of the $1240{\mathrm{cm}}^{-1}$ peak was almost equal to that of the $1342{\mathrm{cm}}^{-1}$ peak; however, in the adenocarcinoma and normal mucosa, the intensity of the $1240{\mathrm{cm}}^{-1}$ peak was obviously higher than that of the $1342{\mathrm{cm}}^{-1}$ peak. This implies that the proportion of amide III and collagen in these three types of gastric tissue is different. Another difference is in the relative Raman intensity between 1584 and $1655{\mathrm{cm}}^{-1}$. GIST and normal mucosa tissue had higher Raman intensity at $1655{\mathrm{cm}}^{-1}$, but the $1584{\mathrm{cm}}^{-1}$ peak was absent in the adenocarcinoma spectrum. Difference in the intensity ratio of 1655 to $1584{\mathrm{cm}}^{-1}$ reflects to the different proportions of amide I to phenylalanine. Apparently, Raman spectra reflect faithfully to the biochemical and biomolecular characteristics.

Fig. 1

Microspectroscopy images of gastric tissues ($100\times$ objective lens): (a) normal mucosae, (b) GIST, and (c) adenocarcinoma. The arrows pointed to the tumor cell.

Fig. 2

Raman spectra, in the range from 500 to $1800{\mathrm{cm}}^{-1}$, of gastric tissues: (a) normal mucosa, (b) GIST, and (c) adenocarcinoma. Data acquisition conditions: excitation wavelength, 532 nm; laser power, 8 mW; and integration time 100 s.

3.2.

Principal Component Analysis and Linear Discriminant Analysis

To enhance the application of Raman microspectrocopy in differentiating, the Raman spectral data were further analyzed by PCA and LDA. By using 20-fold cross-validation, the PC1 component accounted for the greatest Raman spectral variance of 90.838%, and the PC2 component for 6.93%. Plotting PC1 against PC2 (Fig. 3) showed the separation of GISTs, gastric adenocarcinomas, and normal gastric mucosae. For GISTs, the sensitivity, specificity, and accuracy were 100%, 97.50%, and 99.16%, respectively. The sensitivity, specificity, and accuracy for adenocarcinomas were 100%, 94.74%, and 97.48%, respectively, and for normal gastric mucosae, they were 99.02%, 94.12%, and 98.32%, respectively (Table 1). Clearly, each evaluated tissue sample fell within the correct gastric tissue region on the plot. In comparison, the diagnosis of GISTs with EUS-fine needle aspiration had an accuracy of 82% and a specificity of 100%.³⁹ Previously, a study evaluating intraoperative frozen sections to determine the margin of gastric cancer and to distinguish gastric adenocarcinomas from normal mucosae in the proximal end showed a sensitivity of 77.8%, specificity of 100%, and accuracy of 97%.⁴⁰ Another study also showed that intraoperative frozen sections determined the distal margin in gastric adenocarcinoma with a sensitivity of 100%, specificity of 100%, and accuracy of 100%, and the proximal margin with a sensitivity of 67%, specificity of 100%, and accuracy of 93%, when performed by an expert gastrointestinal pathologist.⁴¹ In Fig. 4, we demonstrated the ROC curve of gastric disease mucosa (GIST and adenocarcinoma) versus normal mucosa [Fig. 4(a)] and of GIST and normal mucosa versus adenocarcinoma [Fig. 4(b)], the area under the ROC curves is 0.99 and 0.88, respectively. Apparently, the results obtained by confocal Raman microspectroscopic measurement are quite promising. Raman measurement requires a very small sample and only takes a few minutes without the need for tissue pretreatment such as immunohistochemical staining. By using confocal Raman microspectroscopy coupling with PCA analysis, we reported an effective method for differentiating GISTs from adenocarcinomas and normal gastric mucosae. It was also suggested to be applicable in in situ measurements as combined with endoscopy.^42,43 To analyze the subgroups of patients with adenocarcinomas and GISTs, a larger sampling size including every subgroup of adenocarcinomas and GISTs would be necessary for further study of the gastric tumors.

Fig. 3

PCA of the Raman spectral characteristics for gastric tissues. The first two principal components (PC1 and PC2) from the spectroscopy, which contained most information, were plotted against each other. The plot shows the separation between gastrointestinal stromal tumors, adenocarcinomas, and normal mucosae. (Closed circle: gastrointestinal stromal tumors; open circle: adenocarcinomas; and triangle: normal mucosae).

Table 1

Performance report of sample classification using 20-fold cross-validation.

Fig. 4

(a) The ROC curve of disease (GIST and adenocarcinoma) versus normal mucosa, the area under the ROC curve is 0.99. (b) The ROC curve of GIST and normal mucosa versus adenocarcinoma, the area under the ROC curve is 0.88.

Optical histopathology fast emerges as a potential tool in the diagnosis of various cancers. Fresh tissue is the best choice and serves as an ideal sample for optical histopathology. However, due to severe constraints associated with handling fresh tissues and inevitable problems in the time lapse between tissue excision and spectral analysis, tissue samples are usually preserved in formalin for postspectroscopic analysis. For example, in this study, we cannot investigate simultaneously all samples at one time, and it takes times to complete all Raman measurements of 119 samples, particularly as 5 to 10 spots of Raman measurement for each sample. Therefore, fixation of tissue samples is an inevitable process for protecting tissue from spoilage. The effect of formalin fixation on spectroscopic detection of normal and cancerous tissues has been reported.^44–47 Xu et al.⁴⁴, using two-photon fluorescence spectroscopy to study mouse skeletal muscle, indicate that formalin is a better fixative than methanol for preserving samples in in vitro research. Huang et al.⁴⁵ used a dispersive-type near-infrared Raman system with an excitation wavelength of 785 nm to study whether formalin fixation would affect the potential diagnostic ability for the detection of lung cancer. Although the overall intensities of Raman peaks decreased, the characteristic Raman peaks were found in both fresh and formalin-fixed tissues. In the earlier study on Raman image of eye lenses,⁴⁶ it also showed that the overall intensity of Raman peaks decreased; however, there was no significant difference between the Raman profile of fresh and fixed lenses. Krishna et al.⁴⁷ also indicated feasibility of using formalin-fixed tissues in optical histopathology, especially from Raman spectroscopy point of view. Resonance enhancement by using 532-nm excitation in Raman measurement as described by Alfano et al.^24,29,37 may offset Raman signal attenuation caused by fixation of tissue. Hence, in this study, although fixation of the tissue may attenuate Raman intensity, Raman signals of various sample tissues (Fig. 2) are quite promising and differentiation of normal and cancerous tissues can be easily achieved. We successfully demonstrate that confocal Raman microspectroscopy is powerful to differentiate GISTs, adenocarcinomas, and normal gastric mucosae. To enhance the application of confocal Raman microspectroscopy in gastric tumors, a study combining endoscopy or EUS in situ would much help to further evaluate the characteristics of gastric tissues and may allow an intraoperative measurement.