2.1.

Tissue Preparation

Colon tissue samples (biopsy) were obtained from the Endoscopic Department, Clinic for Internal Medicine II, Department of Gastroenterology, University Medical Center in Jena, Germany. Before the colonoscopy, patients gave informed consent for the study procedure prior to colonoscopy. Samples were taken from healthy control patients ($n=11$) undergoing a colonoscopy for surveillance purposes and from patients with IBD ($n=27$). In total, 14 patients with CD and 13 patients with UC were included. Biopsies were taken from endoscopic normal or moderately inflammed areas. The local ethics commission approved the study protocol.

For each specimen, fresh samples were taken and immediately frozen in liquid nitrogen. Samples were stored at $-80°\mathrm{C}$. On the day of the Raman measurements, tissue cryosections, 10 μm thick, were attached to calcium fluoride slides (${\mathrm{CaF}}_2$). Consecutive sections were cut from each sample, placed on a microscope glass slide, and stained with hematoxylin and eosin (H&E) for histopathological examination to define the Raman measurement region. The protocol for the H&E staining was as follows: fixation [methanol ($1\times 3^{\prime }$) and acetone ($1\times 3^{\prime }$)], rinsing ($1\times 2^{\prime }$ deionized water), the Hematoxylin staining {$1\times 5^{\prime }$ [Mayer’s hemalum solution (Sigma Aldrich)]}, rinsing ($1\times 10^{\prime }$ tap water, deionized water $1\times 1^{\prime }$), the eosin staining {$1\times 5^{\prime }$ [eosin Y solution 0.5% aqueous (Merck)]} and a dehydration (ethanol 80%, 98%, and 100% $1\times 1^{\prime }$). After this a xylene-fixation was carried out. The Raman measurement region was chosen in such a way that an epithelium layer consisting of 5 to 8 epithelial cells was located in the measured area. Subsequent to the Raman measurement, the tissue sections on the ${\mathrm{CaF}}_2$ slide were H&E stained and imaged (Zeiss-Axioplan 2 imaging HBO 100, objective: $40\times /0.75$ Plan-Neofluar) to reassess the Raman scanned area.

2.2.

Raman Instrumentation

Raman spectra were recorded with a Raman microspectrometer (Kaiser Optical Systems, Ann Arbor, MI) using a diode laser at 785 nm as excitation wavelength. The laser light was focused on the sample with a $100\times /\mathrm{NA}$ 0.9 objective (Nikon, Japan) to give $\approx 200\mathrm{mW}$ on the sample. The scattered Raman signal was detected on a Peltier-cooled, back-illuminated, deep-depletion CCD chip (Andor, Northern Ireland) after passing a holographic transmissive grating. Raman scans were acquired in the mapping mode using a motorized $xy$-sample stage with step-size 1.5 μm. The exposure time for each spectrum was 5 s, and an additional dwell time of 10 s was necessary to reduce the autofluorescence of the tissue. Cosmic ray removal and intensity correction were automatically performed by the Kaiser HoloSpec software.

2.3.

Chemometrics

The Raman spectra were subjected to a multivariate data analysis and the results were correlated with the histomorphology of the tissue. All of the computations were done using gnu R¹⁵ running on a commercial available PC system (Intel(R) Core(TM) 2Duo CPU, E6750 2.66 GHz, 1.97 GB RAM).

Prior to the chemometric analysis, the Raman spectra must be preprocessed to correct for corrupting signals due to the experimental setup and the sample itself.¹⁶ This Raman data preprocessing procedure is briefly summarized as follows. First, a wavenumber calibration has been performed using the $322{\mathrm{cm}}^{-1}$ band of ${\mathrm{CaF}}_2$, followed by truncating the wavenumber region to 250 to $3100{\mathrm{cm}}^{-1}$ and background removal applying a SNIP algorithm.¹⁷ Thereafter the spectral range was further narrowed to 675 to $3020{\mathrm{cm}}^{-1}$ and the wavenumber areas 944 to $996{\mathrm{cm}}^{-1}$ and 1725 to $2815{\mathrm{cm}}^{-1}$ were excluded due to artefacts and no Raman signals occurring in these regions, respectively. At the end of this preprocessing procedure a vector normalization was performed followed by a principal component analysis (PCA)^18,19 using only the first 50 PCs.

For the chemometric analysis, a $k$-means-cluster-analysis (KM)^20,21 and support-vector-machines (SVM)^22,23 were applied. In the following a brief description of both methods is given. KM is an unsupervised algorithm, which aims to group the dataset into a given number of groups (cluster). It starts with a random distribution, and calculates the means of the groups and the distance of all Raman spectra to the means. After resorting, every spectrum is placed in its nearest cluster, the means are updated, and this is done until a stable arrangement is reached. Every cluster is color-coded and the spatial arrangement of these clusters is displayed in a false-color-plot. SVM is a classification technique used to build up a model based on group differences. This method searches for the hyperplane, which separates two groups in the best way. Here the radial-base kernel is utilized and the SVMs for multiclass tasks are combined by the one-against-one scheme. At the end the performance of the SVM is evaluated by a 10-fold cross-validation²⁴ in order to check the prediction ability of the model.

The dataset of the present study consist of total 38 scans with 80,356 Raman spectra containing 13 scans emanated from donors with UC (30,008 Raman spectra), 14 scans from patients with CD (26051 Raman spectra), and 11 scans from healthy control patients (25,279 Raman spectra). The dimension of the scans ranges from $46.5\times 40.5{\mu \text{m}}^2$ to $135\times 60{\mu \text{m}}^2$.

3.1.

First Level—Morphology Visualization

First we checked to determine whether the pathological assignment is reflected by the Raman data. Therefore, the mean Raman spectra [Fig. 2(a)] of the five groups specified above were inspected and differences between these mean Raman spectra were calculated [Fig. 2(b)]. While the mean Raman spectrum of blood is dominated by resonance Raman contributions from hemoglobin,²⁵ the other four mean Raman spectra are dominated by bands, which can be assigned to proteins, lipids, and nucleic acids to a small extent. Figure 2(b) shows the difference spectra of the C, N, and M groups with the epithelium group (E): the negative bands at 743, 1245, and $1580{\mathrm{cm}}^{-1}$ in the N-E and C-E difference spectra can be assigned to heme vibrations.²⁵ This finding is reasonable because IBDs are chronic inflammatory diseases in which the colonic mucosa often appears with a lot of bleeding signs. The mucus-epithelium difference spectrum shown in Fig. 2(b) is characterized by difference bands at 835, 1041, 1140, 1280, and $1380{\mathrm{cm}}^{-1}$, which arise due to the different composition of mucus (M) as compared to epithelium (E).

Fig. 2

(a) Mean Raman spectra of the morphological tissue arrangement: The spectral differences between the blood Raman spectrum and the other mean spectra are obvious, which is due to the resonant contribution of hemoglobin. Even though the differences between the E, C, M, and N spectra are smaller, they were proven to be significant by a $t$-test. For details see text. (b) Difference spectra of the mean Raman spectra. The difference of E-C and E-N are dominated by heme-vibrations (743, 1245, and $1580{\mathrm{cm}}^{-1}$), while the bands observed in the difference spectrum E-M (835, 1041, 1140, 1280, and $1380{\mathrm{cm}}^{-1}$) are caused by the different biochemical constitution of mucus.

In addition to wavenumber positions, we evaluated the band intensities of the following characteristic vibrations: $1003{\mathrm{cm}}^{-1}$ (phenylalanine ring breathing), $1450{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_2$-deformation), $1305{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_2$-twisting), $1245{\mathrm{cm}}^{-1}$ (amide III), and $1660{\mathrm{cm}}^{-1}$ (amide I and $\mathrm{C}=\mathrm{C}$-stretching), which can be used to characterize the protein-lipid composition of biological tissue:^26,27 while the the first band at $1003{\mathrm{cm}}^{-1}$ can be used as marker for proteins, the vibrations at $1450{\mathrm{cm}}^{-1}$ and $1305{\mathrm{cm}}^{-1}$ display biomaterial and lipid content, respectively. The two other bands at $1245{\mathrm{cm}}^{-1}$ and $1660{\mathrm{cm}}^{-1}$ can be assigned to heme and also to lipids and proteins. These five bands form a so-called metric²⁸ and are used throughout this paper. This metric was chosen according to Mahadevan-Jansen et al.²⁹ who discriminated between cervical epithelium, inflammation, metaplasia, low-grade and high grade lesions utilizing Raman bands at 1246, 1330, 1454, and $1656{\mathrm{cm}}^{-1}$ that are similar to our metric. After calculating the band heights of these five Raman bands, a (two-sided) Student’s $t$-test^30,31 was performed between all combinations of groups (B, E, C, M, N) and for all band heights. The obtained $p$-values were all smaller than $10^{-19}$ and therefore the significance level was ***. Here the common definition of the significance levels (* $p<0.05$; ** $p<0.01$; *** $p<0.001$) was utilized. This result permits the training of a classification technique.

For classification, a SVM with the PCA scores as input was applied. The resulting morphology classification model can be visualized by the score plot of the decision values [see Fig. 3(a) where three out of 10 decision values (C/E, N/E, and E/M) are displayed]. The separation of the clouds in Fig. 3(a) is a clear indicator for a valid model. The prediction ability of the model was evaluated by a 10-fold cross-validation and is shown in Table 1. From this table the correct classification rate can be calculated to be (91.85%). Also the sensitivity, specificity, and selectivity for finding the epithelium (E) were computed, which were 93.84%, 90.30%, and 88.26%, respectively. The relatively low classification rate of only 91.85% is not a crucial problem, because this low value results from the notation procedure, which will be discussed below.

Fig. 3

(a) Decision values of the first level SVM (morphology): The separation of the four groups (E, C, M, N) is obvious. The blood group is not visible because the Raman spectra of that group are hidden by the other Raman spectra. (b) Decision values of the second level classification (disease recognition) are visualized. The three groups HC, CD, and UC show good separated data clouds.

Table 1

Confidence table of a SVM, which learned the morphological notation. The classification rate of only 91.85% is no problem, because it was calculated with the notation of the pathologist. It is of much more importance that the morphological features were reproduced correctly compared with the H&E-stained image (see Figs. 4 and 5).

Panels A of Figs. 4 and 5 show two H&E images of colon tissue sections where the area investigated by means of Raman microspectroscopy is marked by a black rectangle. In panels F the results of a $k$-means-cluster-analysis with $k=10$, which has been further used to generate the notation plot presented in panels E, is shown. It should be noted that the colors used in panels F are chosen arbitrarily, while the colors used in the notation plot (panels E) reflect the morphological features described above (i.e., $B=\text{blue}$, $E=\text{cyan}$, $C=\text{green}$, $M=\text{orange}$, $N=\text{brown}$). Panels D of Figs. 4 and 5 display the SVM-based notation and can be seen as the sample’s spectral histopathology. Panels C highlight the wrongly classified pixels that are pixels marked in red where the notation (E) and the SVM notation (D) differ. Most of the misclassifications can be found at the borders between the different regions and where the pathologist was uncertain and assigned it to the N-group. The latter can be found, for example, in Fig. 4 at the top side edge: a cluster [yellow in panel Fig. 4(f)] that was labeled as N by the pathologist has been classified by the SVM to the epithelium group [cyan in Fig. 4(d)], which seems to fit better to the H&amp;E image [Fig. 4(b)]. However, these misclassification errors are not a problem as long as the SVM-learned images [Figs. 4(d) and 5(d)] display the same properties as the corresponding H&E images [Figs. 4(b) and 5(b)], which is fulfilled here for all of the 38 scans. For both examples shown in Figs. 4 and 5 the epithelium, connective tissue, and mucosa regions, visible in the H&E images [Figs. 4(b) and 5(b)], are reproduced well in the corresponding SVM-learned images [Figs. 4(e) and 5(e)]. It should be noted that empty goblet cells were assigned to the N-group due to their weak Raman intensity. In contrast goblet cells filled with mucus were grouped as “M” (see Fig. 5).

Fig. 4

Results derived from a Raman scan from the UC group (one example out of 13); (a) H&E-stained image where the Raman scanned region is marked by a rectangle; (b) H&E zoomed area investigated by Raman spectroscopy; (f) false-color-plot of a $k$-means-cluster-analysis of the Raman map; (e) (morphological) notation by a pathologist; (d) SVM based notation; (c) difference of both notations. For details see text.

Fig. 5

Results derived from a Raman scan from the control group (one example out of 11); (a) H&E-stained image where the Raman scanned region is marked by a rectangular; B: H&E zoomed area investigated by Raman spectroscopy; (f) false-color-plot of a $k$-means-cluster-analysis of the Raman map; (e) (morphological) notation by a pathologist; (d) SVM-based notation; (c) difference of both notations. For details see text.

To summarize, the analysis of Raman maps by chemometric approaches allows one to generate images providing the same information as H&E images (“computational H&E stain”), which is why this approach can be referred to as spectral histopathology, which—in principle—can be applied under in vivo conditions. In other words, all features of the H&E images [Figs. 4(b) and 5(b)] can be also found in the SVM-based notation [Figs. 4(e) and 5(e)]. Such a morphological presorting step is necessary, if the epithelium is of diagnostic relevance (as, for instance, in the diagnosis of cancer or IBD). In case such a morphological presorting step is skipped, the observed spectral differences might be caused by different measured areas and not by a different biochemistry of the sample.

3.2.

Second Level—Disease Recognition

In the previous section it has been shown that it is possible to establish a statistical model for a histopathological classification of Raman maps. In the following section, the Raman spectra of only the epithelium region identified by this approach will be used to build up a classification model for the differentiation of HC from IBD. Therefore, as for the morphology classification reported in the previous section, a SVM was applied. By doing so the SVM was trained by the Raman spectra of the epithelium region to achieve a separation between epithelium Raman spectra from healthy control patients, patients with CD, and patients with UC. The corresponding mean Raman spectra of these three groups are plotted in Fig. 6(a). It can be seen that the three Raman spectra are rather similar, characterized by only small subtle spectral differences [see difference plot in Fig. 6(b)]. However, this similarity is not surprising since the epithelium region for HC, CD, and UC are roughly of the same biochemical composition. Nevertheless, the difference spectrum of UC-HC and CD-HC [see Fig. 6(b)] exhibits some differences, mainly at 743, 1245, and $1580{\mathrm{cm}}^{-1}$ where prominent Raman bands of heme can be found. The different heme contributions can be explained by the inflammation of the colonic mucosa in patients with IBD as compared to HC and by a characteristic hyperemic colonic mucosa in UC in contrast to CD. Beside these major spectral differences, some minor differences for protein and lipid Raman bands also exist. Now the question arises of whether these spectral differences in the epithelium Raman spectra of the three classes HC, CD and UC are significant and enable a reliable IBD diagnosis. This question will be pursued in the following.

Fig. 6

(a) mean Raman spectra of the epithelium of the three groups (HC, UC, CD). (b) Corresponding difference Raman spectra. The main differences can be assigned to bands of heme (743, 1245, and $1580{\mathrm{cm}}^{-1}$ ), which can be explained by characteristic hyperemic colonic mucosa in UC in contrast to CD and HC. For details see text.

To start with, a $t$-test^30,31 was carried out, in order to check if these subtle spectral differences are significant. In doing, so the band heights of the metric defined above (i.e., of the Raman bands at 1003, 1450, 1305, 1245 and $1660{\mathrm{cm}}^{-1}$) were calculated. For all $p$-values of every band and for every combination of groups (UC, CD, HC), significance level *** was obtained, clearly demonstrating that the spectral differences are significant. This finding agrees with the results of Bi et al.¹⁴ However, the significance level of $p<0.001$ (***) found here is better than the one of $p<0.05$ (*) found by Bi et al.¹⁴ Furthermore, we identified a significant difference between IBD and HC, which is a key feature for establishing an IBD diagnostic procedure.

After the significance was proven, a classification model (SVM) was trained using the PCA scores of the epithelium Raman spectra as input. The prediction quality of the model was evaluated by a 10-fold cross-validation. The results are summarized in a confidence table (Table 2), clearly demonstrating that the three groups (HC, UC, and CD) are well separable, resulting in a very good classification rate of 98.90%. The sensitivity and specificity for detecting the HC-group are with 99.07% and 98.81%, also very high. Overall, the performance of the trained SVM model is extremely good, showing that the spectral differences between the epithelium Raman spectra of HC, UC, and CD are significant. These spectral differences point toward different chemical compositions of HC, UC, and CD, e.g., different protein patterns of epithelium cells in HC, UC, and CD. This finding is additionally confirmed by the decision value plot of the SVM [Fig. 3(b)], where the three groups form well-separated clusters. To sum up it can be said that the presented results can be seen as a first promising step toward a Raman spectroscopy–based IBD diagnostic method.

Table 2

Second-level classification based on epithelium Raman spectra evaluated by 10-fold cross-validation. The correct classification rate was 98.90%, while the sensitivity and specificity for detecting the HC-group are 99.07% and 98.81%.