1.

Introduction

Biopsy followed by pathology is the gold standard and the most common approach to diagnose cancer. However, it is a time-consuming method and based on the pathologist’s expertise. Though diagnostic procedures for cancer have improved tremendously during the past decades, they are established as highly expensive techniques. Thus, there is the need to develop an accurate, fast, convenient, and inexpensive method for cancer diagnosis. Several research groups have investigated the application of near-infrared (NIR) spectroscopy as a noninvasive and time-saving method for cancer diagnosis.^{1, 2, 3, 4} Near-infrared spectroscopy is based on overtones and combinations of the molecular vibrations of C–H, N–H, and O–H groups, which are part of all biological molecules. Therefore, the absorption of near infrared radiation $(12000–4000{\mathrm{cm}}^{-1})$ leads to a complex spectrum, which provides qualitative and quantitative information about the chemical composition of the tissue with its main components of lipids, proteins, carbohydrates, and water. Any alteration in the chemical composition of the tissues can thus be detected by NIR spectroscopy and be evaluated with the help of mathematical nonsupervised methods such as cluster analysis, which is a pattern recognition technique. In the present study, we investigated the qualitative spectral differences between cancerous and its adjacent normal tissues in surgically resected pancreatic cancer specimens using NIR fiber-optic spectroscopy.

2.

Materials and Methods

3.

Results and Discussion

The difference in tissue composition between pancreatic cancer and normal tissues has been extensively analyzed using chemical, histochemical, biochemical, and immunohistochemical studies. Altered composition of isoferritin, glycosaminoglycans, glycoproteins, and glycolipids has been described in pancreatic adenomas and carcinomas.^{6, 7, 8, 9, 10} These differences in the tissue composition will contribute to the spectral variance between cancer and normal tissues. The near-infrared spectrum of the pancreatic tissue and the assignment of the peaks to various chemical substructures¹¹ in the spectrum are shown in Fig. 1 . The spectrum is a mixture of the spectral signatures of many tissue components including water, lipids, proteins, and carbohydrates. The drop of the $-\log$ (reflectance) around $4000{\mathrm{cm}}^{-1}$ is due to the low transmittance close to the interval end as shown in Fig. 1.

Fig. 1

Assignment of bands of various chemical substructures in the NIR spectrum of the tissue.

The mean of tumor and normal tissue measurements from one sample and their standard deviations are shown in Fig. 2 . The variance is essentially constant across the spectral range used. This suggests that the spectra are highly reproducible. The mean spectra from tumor and normal tissue measurements of all the samples and their standard deviations are shown in Fig. 3 . The relatively high variation between the samples could be due to different tumor stages of the tissues (Table 1 ) and differences in the tissue thickness of the samples.¹²

Fig. 2

Mean and standard deviation spectra of the tumor and normal tissue measurements of one patient.

Fig. 3

Population mean and standard deviation spectra of tumor and normal tissue measurements.

Table 1

Pathological stages of the tumors.

Since derivative techniques are normally used to remove or suppress constant background signals and to enhance the visual resolution of the spectra, we have used the first derivative spectra for the evaluation with cluster analysis. However, as derivation amplifies the spectral noise, we smoothed the data using the Savitzky-Golay algorithm. Further, to reduce the differences between each single measurement of the same sample, the spectra were vector normalized after derivation and smoothing. To emphasize the differences between the preprocessed cancer and normal spectra, the mean normal spectrum was subtracted from the mean cancer spectrum. The preprocessed mean spectrum of cancer and normal tissue and their differences are shown in Fig. 4 .

Fig. 4

First derivative and vector normalized mean spectra of pancreatic cancer and normal tissue and their difference spectra.

To assess whether significant differences existed between preprocessed normal and cancerous tissue spectra, paired t tests were applied for the absorbance values at each wave number and the resulting $p$ values were plotted. Several intervals of the resulting plot contained statistically significant $p$ values (Fig. 5 ). Of the total wave numbers, 76.2% were significantly different with $p<0.05$ . The result from a hierarchical cluster analysis using the wave numbers showing significant difference is presented in Fig. 6 . Five clusters can be found with a majority of cancerous tissue spectra in the first, second, and third clusters and the majority of the normal tissue spectra in the fourth and fifth clusters.

Fig. 5

P values of absorbances at wave numbers showing significant differences between pancreatic cancerous and normal preprocessed spectra.

Fig. 6

Hierarchical cluster analysis of the NIR preprocessed spectra of tumor and normal tissue measurements of the wave numbers showing significant differences $(p<0.05)$ .

As cancerous tissues mainly differ from normal tissues in glycolipids, glycoproteins, and glycosaminoglycans composition^{6, 7, 8, 9, 10} and the CH vibrational regions represent these classes,^{3, 4, 13, 14, 15} we decided to use the CH stretching bands for the evaluation by a cluster analysis. As a result, the spectra were separated with high sensitivity and specificity using the CH stretching first $(5951–5608{\mathrm{cm}}^{-1})$ and second overtone regions $(8605–7938{\mathrm{cm}}^{-1})$ . The peaks in the cancer tissue spectrum between the wave numbers 8584–8497, 8263–7938, 5795–5729, and $5678–5605{\mathrm{cm}}^{-1}$ were reduced in intensity compared with those in the normal tissue spectra. The band intensities between the wave numbers 8605–8585, 8497–8263, 5951–5795, and $5728–5678{\mathrm{cm}}^{-1}$ were strongly in relation to those in the normal tissue spectra (Fig. 7 ). The dendrograms of the cluster analysis using CH first and second overtone regions are shown in Fig. 8 and the classification results are summarized in Table 2 .

Fig. 7

Spectral differences in the mean spectra of cancerous and normal tissues in CH stretching second $(8605–7938{\mathrm{cm}}^{-1})$ and first overtone $(5951–5608{\mathrm{cm}}^{-1})$ regions.

Fig. 8

Hierarchical cluster analysis of the NIR preprocessed spectra of tumor and normal measurements in CH stretching first $(5951–5608{\mathrm{cm}}^{-1})$ and second overtone $(8605–7938{\mathrm{cm}}^{-1})$ regions showing dendrograms with separate clusters for tumor and normal tissue.

Table 2

Classification results of the hierarchical cluster analysis of NIR preprocessed spectra of pancreatic tumor and normal tissues using data from CH stretching first and second overtone regions.

When evaluating the spectral interval data of $8605–7938{\mathrm{cm}}^{-1}$ , the hierarchical cluster analysis classified the spectra into two groups with 88.9% tumor and 27.8% normal spectra in first group and 72.2% normal and 11.1% tumor spectra in the second group. The heterogeneity between the two groups was 2.387. With the evaluation of the spectral region $5951–5608{\mathrm{cm}}^{-1}$ , the spectra were classified into two groups with 83.3% tumor and 16.7% normal spectra in the first group and 83.3% normal and 16.7% of the spectra in the second group. The heterogeneity between the two clusters was 7.402. With the evaluation of the CH second overtone region, three spectra from adenocarcinoma and five spectra from normal were misclassified, and with the evaluation of the CH first overtone region three spectra from adenocarcinoma and three spectra from normal were misclassified, respectively. The comparison of the misclassified spectra with the mean spectra of normal tissue and cancer are shown in Figs. 9 and 10 . All the normal spectra that were misclassified exhibit spectral features similar to the cancer spectra [Figs. 9(A), 9(B)] and all the cancer spectra that were misclassified exhibit spectral features similar to the normal spectra [Figs. 10(A), 10(B)]. The sensitivity of the method in CH first overtone region $(5951–5608{\mathrm{cm}}^{-1})$ was 83.3% with a specificity of 83.3%. With the CH second overtone region $(8605–7938{\mathrm{cm}}^{-1})$ data we could achieve a sensitivity of 88.9% and specificity of 72.2%. The accuracy of the method using CH first and second overtone region data was 83.3% and 80.5%, respectively.

Fig. 9

Comparison of the misclassified normal spectra with the mean spectra of cancer and normal tissue measurements in CH stretching second (A) and first overtone (B) regions. For clarity the individual spectra are displayed by using different offsets.

Fig. 10

Comparison of the misclassified cancer spectra with the mean spectra of cancer and normal tissue measurements in CH stretching second (A) and first overtone (B) regions. For clarity the individual spectra are displayed by using different offsets.

4.

Conclusion

Cluster analysis of the NIR fiber-optic spectra of surgically resected human pancreatic tumor tissues showed high sensitivity and specificity in discriminating the spectra obtained from tumor tissues from the spectra measured from normal tissues. With a sample population size of 18 patients, we achieved highest accuracy with a sensitivity of 83.3% and a specificity of 83.3% using data from CH first overtone region of the NIR spectrum, which suggests a potential spectral region to distinguish tumor from normal tissue. The advantage of this method is that it is a very fast, simple, and relatively inexpensive technique. These findings suggest that NIR fiber optic spectroscopy offers the potential for a minimally invasive in-vivo diagnosis of pancreatic tumors.