1.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is the second leading cause of cancer-related deaths involving the digestive system after colorectal cancer among men and women in western countries.¹ Modern multimodal treatment plans including chemotherapy and surgical resection still do not improve overall survival beyond 4.5 years.^2,3 Curative pancreatic surgery for malignant neoplasms is highly demanding, with optimal results achieved when microscopically tumor-free resection margins are obtained. Therefore, differentiation of normal pancreatic tissue, inflammatory pancreatitis and PDAC is crucial. Reliable macroscopic tissue distinction is not possible for surgeons. This is especially true in the case of infiltrating tumor margins due to cellular spread into normal pancreatic tissue (Fig. 1).

Fig. 1

Intraoperative image of a pancreatic head resection for PDAC. The image displays the situs after having resected the tumor. The pancreatic tail (PT) leads to the resection margin (M) with the pancreatic duct (*) in the center. The liver (L) and the small intestine (SI) are located closely.

However, in most cases microscopic resection margin clearance of pancreatic cancer determines the patient’s survival rate.^3–5 Currently, the gold standard for tissue discrimination of pancreatic lesions is microscopic histopathological analysis of pre-stained samples. Frozen sections are sent to pathology intraoperatively to be microscopically analyzed for remaining tumor cells. This method is used to ensure tumor-free resection margins.

Several methods such as imaging mass spectroscopy and tumor-specific agents for fluorescence imaging have been investigated to accurately identify PDAC tissue.^6,7 Each method slightly increases the ability to discriminate between neoplasms, inflammatory tissue and adjacent normal tissue. However, the lack of molecular information and time-consuming measurements are great limitations of these methods. A straightforward approach for the identification of pancreatic tumors is biochemical analysis on the molecular level performed by vibrational spectroscopy. Recent reports have been largely dominated by studies focusing on this method to characterize brain tumors and distinguish between glioblastoma and normal brain cells even under in situ conditions or intraoperatively.^8–13 In particular, Fourier transform infrared (FT-IR) spectroscopic imaging is capable of showing changes in carcinogenesis-related vibrational modes in several human cancers, especially in larger areas or homogeneously structured tissue.^14–16 However, the evaluation of IR spectroscopic data captured from biological tissue or cells is often difficult due to the extensive amount of information.⁸ Therefore, IR spectroscopic imaging is often used in conjunction with multivariate methods such as principal component analysis (PCA) to compress data and maximize and highlight small area variations.

Considering the devastating mortality rate of pancreatic cancer, there is a comprehensive clinical need for fast and reliable differentiation of pancreatic tissue. IR spectroscopic imaging has already been used for the characterization and differentiation of human PDAC and pancreatic neuroendocrine tumors.¹⁷ In this study a pair-wise PCA was carried out to discriminate PDAC, pancreatic neuroendocrine tumor, dysplastic pancreatic, and normal tissue by FT-IR imaging spectroscopy. The authors found that the four histologic entities can be differentiated based on their IR spectra. The present study aimed to distinguish between normal pancreatic tissue, pancreatitis, and PDAC, using IR spectroscopic imaging in combination with PCA and supervised classification. IR spectroscopic imaging was performed to identify even small tumor clusters within the snap frozen tissues and to subsequently compare the spectral profiles in structural comparison to standard histological images.

2.

Material and Methods

3.

Results and Discussion

Tumor tissue samples were captured and analyzed in a visible transmission mode and an IR spectroscopic image (Fig. 2). HE-stained standard histopathological specimens served as control. The grid overlay [Fig. 2(b)] corresponds to the mapping matrix of the IR spectroscopic images. The sample was imaged by $4\times 12$ individual spectroscopic images containing a total of $4\times 12\times 1024=\mathrm{49,152}$ spectra. Binning reduces the spatial resolution and storage volume of the data by a factor of four, but increases the signal-to-noise ratio and significance of the spectral profile. Spectra obviously not associated with tissue or areas with dominant artifacts including a strong background signal were classified as outliers and removed from the data set. The image calculated from preprocessed spectra is represented in Fig. 2(d).

Fig. 2

(a) Histopathologic HE-stained image served as area recognition and standard tissue classification. (b) Microscopic transmission image of a cryo-preserved specimen of representative tissue section with an overlaid grid of the area mapped by IR spectroscopic imaging. (c) IR spectroscopic raw data bright field image (integrated absorbance [int. Abs.] between 950 and $1800{\mathrm{cm}}^{-1}$) with bar representing width of the image. (d) Image of selected preprocessed spectra from (c). Spectra of white pixels were identified as outliers and removed from the data set.

In the first step, the mean spectrum and standard deviations of the data set shown in Fig. 2 were calculated and applied (Fig. 3). The spectrum is dominated by the amide I band at $1650{\mathrm{cm}}^{-1}$ ($\mathrm{C}═\mathrm{O}$ stretching) and amide II band at $1550{\mathrm{cm}}^{-1}$ (N-H bending) of the amide groups comprising the peptide linkages of proteins (Table 1).

Fig. 3

Mean spectrum ($\mu$, black curve) and standard deviation area ($\delta$, gray band) of the sample represented in Fig. 2. Mean spectrum was calculated from the preprocessed data set of the image shown in Fig. 2(d).

Table 1

Assignment of vibration modes. The symbols refer: ν, stretching mode; δ, bending mode; s, symmetric, as antisymmetric.

Two more significant bands appear around 1453 and $1400{\mathrm{cm}}^{-1}$, which were assigned to ${\mathrm{CH}}_3$ bending vibrations of methyl groups. The band at $1235{\mathrm{cm}}^{-1}$ corresponds to the amide III band and the group between 1000 and $1150{\mathrm{cm}}^{-1}$ is mainly composed of absorption bands of $\mathrm{C}─\mathrm{O}$ and ${\mathrm{PO}}_{4^{-}}$ groups of nucleic acids, phospholipids, and carbohydrates. The most important spectral bands assigned to vibrational modes are summarized in Table 1. Of note is the standard deviation is 0.1 au. This indicates a large variance in biochemical composition, therefore the average considered in isolation can be misleading. Furthermore, the most common types of PDAC are histologically heterogeneous with different tumor tissue compartments such as tumor cell clusters, stromal desmoplastic tissue, and infiltration zones.

To inspect the biochemical variations across the sample, a more sophisticated data analysis is required to extract signals from different types of tissue and cells. The mathematical investigation of spectral signals to distinguish cancer from normal tissue or to detect tumor cells infiltrating into normal tissue was performed by PCA. It was used because it is a quick way to assess variations within the spectral data set. It does not assume any explicit statistical model underlying the variance of the original spectra. The results of the first three PCs are shown in Fig. 4.

Fig. 4

The first three PCs of the spectroscopic image from Fig. 1. (a) Maps of the score values are displayed in the centered rainbow color code. (b) Corresponding loading plots to PCs.

The loading plots correspond to spectral bands with the highest variations and weight the signals in the positive and negative directions. The score map reveals the weight of the loading plot for each pixel of the image. PCA is applied to the baseline corrected and area normalized spectra. The first PC represents the mean spectrum of the sample (Fig. 3). The corresponding score map only shows a few weak variations in intensity. The loading plot of the second PC shows two positive signals located at 1516 and $1618{\mathrm{cm}}^{-1}$. Both signals are associated with vibrational modes of proteins and can be assigned to collagen. Another positive broad signal occurs between 1000 and $1080{\mathrm{cm}}^{-1}$ with a maximum located at $1025{\mathrm{cm}}^{-1}$, representing collagen as well. The weak signal between 1680 and $1750{\mathrm{cm}}^{-1}$ is composed of amide I vibrational modes and modes of $\mathrm{C}═\mathrm{O}$ groups. Therefore, the second PC is interpreted to indicate collagen, which is supported by the histological image (Fig. 2). The red and orange pixels in the score map correspond to blood vessel structures. The loading plot of the third PC displays strong positive signals at 1080 and $1655{\mathrm{cm}}^{-1}$ assigned to phosphate groups of DNA and alpha-helical substructures of proteins, respectively. The red and orange pixels of the loading plot correspond to tumor infiltrating tissue as indicated in the histopathological image (Fig. 2). A series of negative signals represents vibrational modes of ${\mathrm{CH}}_x$ groups and other protein structures. In particular, the pattern of signals at 1200, 1340, and $1516{\mathrm{cm}}^{-1}$ points to collagen. The blue and turquoise pixels in the corresponding score map show a similar pattern to that of the red and orange pixels in the score map of the second PC. The second and the third PC are the most important for classification purposes. Higher PCs predominantly represent signals without any impact on the classification, noise, or atmospheric signals.

The score maps show tumor-budding areas of PDAC that correspond very well with the histologic image (Fig. S2 in the Supplemental Material). Although these results suggest that PCA alone is a suitable tool to identify and highlight pancreas tumor tissue, it should be noted that a precise delineation of cancer is not always possible by PCA due to the high intrinsic variability of pancreatic tissue. Since tumor-specific spectral features may be weak and overlaid by all other common tissue signals, a supervised classification has to be performed. This requires a defined training data set before performing the classification of unknown spectra. Each class of the training set is used as a reference for the classifier. Since the quality of training data determines the accuracy of classification, it is necessary to select representative spectra for each class that cover both the general spectral feature of the class and the natural variability of these features. In particular, tumor tissue has a heterogeneous appearance with areas of different cell density and stages of development. Therefore, two more PDAC samples with histologically distinct visible tumor areas were analyzed by PCA, and score maps and loading plots calculated (Fig. 5). The corresponding histological images are represented in the supplementary part (Fig. S2 in the Supplemental Material). The spectral features of both loading plots are very similar to each other and to those of the third PC in Fig. 4. In accordance with the histological images, only spectra indicated by red or orange pixels in the third PC were selected for creation of the training set. A total of 55,000 spectra representing tumor tissue were selected.

Fig. 5

Third PC of two additional tumor samples with (a) score values and (b) loading plots. Only spectra with a score value $>0.005$ were selected for the training set.

Since it is often difficult to distinguish between tumor and pancreatitis tissue, a second class of the training set was defined by spectra of pancreatitis. A third class of the training set was defined by physiological pancreatic tissue. The spectra of two representative samples of pancreatitis and normal tissue were taken for creation of the training set. Pancreatic tumor tissue is often surrounded by connective, inflammatory, and stromal tissue and associated with cellular inhomogeneity. To select only spectra of pancreatic tumor cells, a fuzzy k-means cluster analysis was performed to discriminate between tumor cells and nontumor cells. The image of the cluster assignment was compared to the histological stained micrograph of a consecutive section to identify areas of tumor cells. An example of cluster analysis and the corresponding histological image are represented in the supplementary part (Figs. S2 and S3 in the Supplemental Material). Reddish and orange pixels correspond to dense tumor clusters in the histological image. Only spectra belonging to the red and orange clusters were used as reference tumor spectra for the training set. Normal tissue and pancreatitis tissue samples were considered uniform and were not differentiated. To compare classes and find unique features, the average spectra were calculated from more than 55,000 spectra for each tissue class. The averaged spectra and standard deviations of normal, tumor, and pancreatitis tissue are shown in Fig. 6.

Fig. 6

Mean spectra (solid curves) and standard deviations (dotted curves) of (a) normal tissue, (b) tumor tissue, and (c) pancreatitis tissue. Mean spectra and standard deviation were calculated from training data sets of each tissue class.

Although the spectral profiles may appear similar, there are a few differences between the classes. While the plot of normal tissue [Fig. 6(a)] exhibits an overall low standard deviation, the spectra of tumor tissue [Fig. 6(b)] and pancreatitis tissue [Fig. 6(c)] show stronger variation, as indicated by the higher standard deviation. Another conspicuous difference is the absorption profile between 1050 and $1150{\mathrm{cm}}^{-1}$. Nucleic acids and carbohydrates as well as their derivatives give rise to prominent bands in this spectral region. Figure 7 highlights the spectral changes and also shows the difference spectra of tumor tissue [Fig. 7(a)] and pancreatitis [Fig. 7(b)] compared to normal tissue.

Fig. 7

Difference spectra of the calculated mean spectra: (a) pancreatitis tissue–normal tissue, (b) tumor tissue–normal tissue.

The difference spectrum between tumor and normal tissue shows three prominent features. While the spectral profile between 1050 and $1150{\mathrm{cm}}^{-1}$ is higher for tumor tissue compared with physiological pancreatic tissue, both the amide II (1510 to $1590{\mathrm{cm}}^{-1}$) and amide I bands (1610 to $1690{\mathrm{cm}}^{-1}$) are clearly weaker. The stronger absorption between 1050 and $1150{\mathrm{cm}}^{-1}$ arises mainly from the greater amount of carbohydrates due to the enhanced and altered metabolism of the tumor cells. Although the cell density in PDAC is lower compared to normal pancreatic tissue, tumor cells are characterized by a high proliferation rate. This leads to an increased amount of DNA as well as enhanced metabolic activity with a higher gene transcription rate and consequently elevated RNA levels. Hence, PDAC tissue shows a stronger absorption of nucleic acids, which also contribute to the positive difference signal. In addition, PDAC cells show altered extracellular matrix protein production, which is associated with myofibroblastic stellate cells and a remarkable density of collagen fibers.^19,20 This desmoplastic reaction is a surrogate marker of PDAC tissue, which leads to a high fraction of connective tissue. Consequently, the intensity of the amide I and amide II bands is supposed to be higher than in normal pancreatic tissue. Due to the selection process for tumor spectra using only dense PDAC cell cluster areas, connective tumor tissue is underrepresented in the mean tumor spectrum, thereby resulting in lower amide I and amide II bands.

Pancreatitis tissue shows a lower intensity in the spectral range from 1050 to $1150{\mathrm{cm}}^{--1}$ compared to normal tissue but stronger signals of proteins [Fig. 7(b)]. The amide III band ($1235{\mathrm{cm}}^{-1}$), parts of the amide II band ($1522{\mathrm{cm}}^{-1}$), and also components of the amide I band ($1675{\mathrm{cm}}^{-1}$) are higher. The negative signal at $1590{\mathrm{cm}}^{-1}$ as well as the slightly lower differences at 1208, 1346, and $1650{\mathrm{cm}}^{-1}$ can be assigned to collagen. Pancreatitis is an inflammatory biological process within the pancreatic tissue. Different triggers (e.g., chronic alcohol consumption, biliary system alteration, and hereditary mutations) cause the migration of various cell types associated with the immune system (macrophages, lymphocytes, etc.) into the tumor tissue, where they may initiate and maintain the immune response through the secretion of mediators. Macroscopically visible variations of pancreatitis are fibrosis of the tissue and atrophy of the parenchyma. Hence, the entire biochemical profile is altered compared to normal tissue. In particular, cell atrophy and retained pancreatic secretion in combination with dilated ductal systems cause a lower density of cells, thereby resulting in weaker absorption signals between 1050 and $1150{\mathrm{cm}}^{-1}$. It should be noted that the explicit molecular mechanisms of pancreatitis are complex and may cause interactions within the biochemical profile which are not yet well understood and need to be examined further.

The finding of spectral differences distinguishing the class average spectra leads to the hypothesis that it might be possible to discriminate between normal, pancreatitis, and pancreatic tumor tissue. The success of a spectral classification method is dependent on the ability to factor out possible variability among the spectra within each class. In particular, supervised classification methods enable the identification of spectral patterns that are valid within a class even with intraclass variability.

In the first step, 25,000 spectra of each tissue class were randomly selected for the training set. PCA was applied to the entire data set to reduce its dimension while still retaining variability. The number of PCs included in the classification model was determined based on the classification accuracy criterion. The plot of classification accuracy in relation to PC is presented in the supplementary material for all three classes (Fig. S1 in the Supplemental Material). The optimal number of PCs was defined if a global accuracy of 95% was attained. This optimal number value of reduced dimensions of PCA is reached when the score values of the first 30 PCs are used for the classification. The results of the reclassified training set are presented in Fig. 8.

Fig. 8

False color images of the reclassification of samples from which selected spectra were used for the training set. Reclassification of (a) normal tissue samples, (b) tumor tissue samples, and (c) pancreatitis tissue samples. $N$, normal tissue; P, pancreatitis; T, tumor tissue.

Both normal (N1, N2) and pancreatitis (P1, P2) tissue samples exhibit an overall high classification accuracy. Tumor tissue sections are known to be heterogeneous and usually display infiltrating tumor cell areas but also desmoplastic stroma reaction and normal pancreatic tissue. Therefore, the classification of these samples shows quite heterogeneous RGB assignments compared to normal and pancreatitis tissues. Consequently, we have defined the term “accuracy” in the clinical context primarily to identify and to discriminate areas and clusters of tumor tissue. The tumor samples (T1 to T3) show red pixels, thus indicating tumor tissue. However, they also show green and in the case of T1 even blue pixels, corresponding to normal and pancreatitis tissues, respectively. A closer inspection of the tissue sections reveals that most of the areas classified as normal tissue within the tumor samples do not correspond to tumor cell clusters but rather to stromal tissue. Structures associated with blood vessels in sample T1 were predominantly classified as “pancreatitis.” Since the spectral selection process for the tumor category mainly used dense tumor cell cluster areas, the tumor connective tissue is underrepresented in the classifier, thus leading to misclassification in these histological areas. Physiological pancreatic features may outweigh malignant characteristics in regions of stromal tissue. The comparison of the classified samples with their histopathological HE-stained images does not show any tumor cell clusters in the normal (green) pixel areas, thus supporting this hypothesis (Fig. S2 in the Supplemental Material).

The results obtained show that the classification model is successful in recognizing normal and pancreatitis tissues with an overall accuracy of close to 100% for the reclassified training set. Based on this training set and classification model, a number of “unknown” tissue samples from all three classes were analyzed (Fig. 9). Similar to the reclassification of the training set, nearly all of the spectra of normal and pancreatitis tissue samples were classified correctly. While most tumor samples were identified properly, they showed a more inhomogeneous false color image including different pixel types. Comparing the histological images shows that predominantly stromal tissue was misclassified, thus supporting the selection process hypothesis. Moreover, in the last several years pancreatic cancer research has focused on the heterogeneity of PDAC tissue. Different molecular alteration patterns have been recognized and classified into three common subtypes (exocrine-like, quasi-mesenchymal, and classical) with clinical implications and varying degrees of chemosensitivity.²¹ To date, subtype-specific spectral characterization has not been developed. With regard to the molecular pattern, there might be subtypes with normal-like spectra, which may lead to partial misrecognition due to the little number of training set samples. However, the general assignment to the class “tumor” can be defined for all five test set samples. For a further automatic classification system, a categorizing misclassification threshold would have to be evaluated based on clinical needs and safety issues. Nevertheless, IR spectroscopy is a suitable analytical method to differentiate pancreatic tissue classes in a fast and marker-free manner.

Fig. 9

False color images of the test samples classification results. (a) Normal tissue samples, (b) tumor tissue samples, and (c) pancreatitis tissue samples. $N$, normal tissue; P, pancreatitis; T, tumor tissue.

4.

Conclusion

IR spectroscopy imaging combined with PCA and supervised classification methods have demonstrated the ability to clearly distinguish between benign and malignant pancreatic tissue in a fast and label-free way. However, further investigation of the biochemical profile of pancreatic cancer is needed to improve our understanding of tissue characteristics. Therefore, IR spectroscopy opens up a wide range of possibilities for characterizing solid human cancer entities in a markerless manner and may also lead to changes in intraoperative strategy that help achieve the best outcome for patients.