Pseudo-HE images derived from CARS/TPEF/SHG multimodal imaging in combination with Raman-spectroscopy as a pathological screening tool

Mice with an intestinal epithelial cell (IEC)-specific deletion of the tumor suppressor gene p53 were received by crossing of floxed p53 mice with villin-Cre mice (Tp53 ^ΔIEC) [19]. As controls Tp53 ^F/F mice were utilized. These mice were treated intraperitoneally once a week for 6 weeks with the carcinogen azoxymethane (AOM; 10 mg/kg). For histological and Raman-spectroscopic investigations mice were sacrificed at different time points after the AOM application. The specific loss of p53 in the intestine markedly enhances the carcinogen-induced tumor incidence and leads to the development of invasive colorectal tumors beginning about 12 weeks after the first AOM treatment.

Full preparation of colon and rectum was performed for 47 individuals. Biopsies were acquired from 8 out of these 47 individuals. Another cohort of 22 individuals was divided into two groups of which biopsies were taken in an alternating fashion every two weeks. In total, 69 individuals contributed to the complete training Raman-dataset. For the present study, 6 individuals (2 males, 4 females) were selected randomly covering the whole spectrum of diagnoses. In total 34 Raman-maps (46240 spectra) and the respective non-linear multimodal images were acquired to investigate this reduced cohort.

Cryosections for the histo-pathological evaluation were obtained following the standard procedure of washing the specimens with phosphate buffered saline (PBS Buffer), fixation using paraformaldehyde, embedding in paraffin and cutting the sample in 10 μm slices which were stained with hematoxylin and eosin (HE staining). For the Raman-spectroscopic analysis, the specimens were washed with PBS buffer and immediately frozen in liquid nitrogen. Distilled water was used as embedding medium for the cryosectioning step resulting in 20 μm thick slices. Thus, a native sample - preserving the lipid distribution - is obtained.

Brightfield images of the parallel HE stained as well as Raman-spectroscopically examined sections were recorded using a halogen illuminated Leica DM2500 microscope (Leica Microsystems, Wetzlar, Germany) equipped with a Nikon Digital Sight (DS) camera system using the DS-Fi1 CCD camera head (Nikon Instruments Europe B.V., Germany). To superimpose the measured Raman-maps with the their corresponding histopathological assessment, all HE stained scan areas were imaged using five different Leica objectives in the range of 1.25 × to 40 × (1.25 × HCX PL Fluotar (NA 0.04), 5 × N Plan (NA 0.12), 10 × N Plan (NA 0.25), 20 × N Plan (NA 0.4), 40 × N Plan (NA 0.65)).

The experimental setup was presented elsewhere [20]. A schematic of the experimental setup is displayed in Fig. 1. Briefly, a Coherent Mira HP Titanium-Sapphire (Ti:Sa) laser (Coherent, Santa Clara, USA) is pumped by a continuous wave Neodymium-Vanadate laser with an average power of 18 W operating at 532 nm. The Ti:Sa-laser generates 2–3 ps pulses (FWHM) at 830 nm with a repetition rate of 76 MHz. The 3.5 W averaged output power of the Ti:Sa-laser is split into two parts. The first part is used directly, i.e., without frequency conversion, as the Stokes beam. The second part is coupled into an optical parametric oscillator (OPO, APE, Berlin, Germany) that allows to adjust the pump wavelength in the range from 500 to 1600 nm. To match the CH ₂ symmetrical stretching vibration at 2850 cm ⁻¹ for the CARS measurements, the OPO is tuned to 671 nm. Both beams, pump and Stokes, are temporally and spatially overlapped and coupled into a laser scanning microscope (LSM 510 Meta, Zeiss, Jena, Germany) and focused onto the sample with a 20 × (NA 0.8) achromatic objective (Zeiss). The optical non-linear response of the sample is wavelength filtered by means of various dielectric filters and detected by photomultiplier tubes (PMT, Hamamatsu Photonics, Hamamatsu, Japan) in forward (CARS, SHG) and backward direction (TPEF). Large area scans of the samples of up to 15 ×15 tile-scans, each having a size of 450 μm × 450 μm, were recorded. Every tile-scan was acquired with a resolution of 1,024 x 1,024 pixels and a pixel time of 1.6 μs. By averaging twice the total acquisition the time per tile-scan does not exceed 16 s for all three modalities, i.e. CARS, TPEF and SHG. Thus, the acquisition time for an image of 15 × 15 squares corresponding to a size of 6.75 mm × 6.75 mm is about 1 hour. The average power at the sample was 25 mW and 50 mW for the pump and Stokes beam, respectively. A discussion about the applied power and potential linear as well as non-linear tissue photo-damage can be found elsewhere [21].

Raman-spectra of 20 μm thick cryosections from colon and rectum of mice as well as from biopsies on CaF ₂ slides were recorded with an upright micro-Raman-setup (CRM 300, WITec GmbH, Ulm, Germany) equipped with a 300 g/mm grating (7 cm ⁻¹ resolution) and a Deep Depletion CCD camera (DU401 BR-DD, ANDOR, 1024 × 127 pixels) cooled down to −65°C. The tissue Raman-spectra were recorded with a 785 nm diode laser which was focused through a Zeiss 50 × objective (NA 0.7) onto the sections. For the herein described study 34 Raman-scans were recorded in scanning mode with a step size of 5 μm and an integration time of 2 s per spectrum. The scan dimension was chosen to be around 34×40 pixels (170 μm × 200 μm), so every scan consists of approximately 1360 Raman-spectra. As a standard for the subsequent processing of the spectra a time series comprising 50 spectra of 4-acetamidophenol was collected with an integration time of 1 s per spectrum. Based on HE stained parallel sections a trained pathologist selected the areas to be measured. Here we measured the Raman-spectra before the multimodal images in order to check if burning effects occurred. In the presented data set burning effects were not observed.

All animal studies were approved by the governmental commission for animal protection (No. 02-007/13).

22 multimodal images were measured and the TPEF, CARS and SHG channel were combined in false-colors. Here we used green to represent TPEF, red for CARS and SHG was coded in shades of blue. The images were pre-treated separately: First the images were down-sampled by a factor of 4. Then the uneven illumination of the tile-scans was removed within the images and the contrast was adjusted (CARS:[0.05-0.015], TPEF:[0.05-0.04], SHG:[0.001-0.001]) [22]. The resulting images are given in Fig. 2 in row A. The workflow for pseudo-HE generation is sketched in Fig. 3. The pseudo-HE stained images are generated using a partial least squares regression (PLS) model with 3 components [23], which was trained with one image (data not shown). PLS is a multivariate regression method that estimates the relationship between two datasets and differs from traditional least squares regression in utilizing information of the independent and the dependent variables. The RGB-values of an HE stained image was modeled using the three color channels of the multimodal image. Afterwards, regions with a certain fingerprint of CARS, TPEF and SHG intensities were predicted to feature cell nuclei (dark violet) and thus this color was added. This procedure was based on a linear discriminant analysis model (LDA). The main idea of this classification method is to find the optimal linear combination of variables that maximizes the variations between different classes and minimizes the variations within these classes. This additional model was necessary as cell nuclei had negative contrast in the multimodal images. The result is referred to as computational HE stained or pseudo-HE image. Its generation is performed automatically and, therefore, allows for a fast screening. After the calculation of the computational HE stain was performed, the background was determined and set to white. The background was estimated by a sequence of steps. First the original multimodal image (RGB) was segmented by k-means clustering with k=6. The pixels of the darkest class, i.e. the class with the lowest value of sum over squared class center, are considered as background contribution. Based on this annotation a background mask was calculated. To remove noise contributions in the segmentation result, the estimated background mask was filtered by a median filter. Thereafter morphological closing was applied to fill gaps in the foreground and extending the foreground area. Finally, the background mask was morphologically opened in order to remove small foreground areas and smooth the background edges (or specimen contour). The resulting mask was mean filtered and used as a weighting mask allowing for a smooth removal of large background areas.

The data preprocessing in case of the Raman-spectra and statistical modeling were performed using the software package R [24]. The packages used were ‘MASS’ [25], ‘pls’ [23], ‘KKNN’ [26] and ‘Peaks’ [27]. Several multivariate statistical tools were applied, such as principal component analysis (PCA), k-means and Weighted k-Nearest Neighbors (KKNN), which shall be described briefly. For dimension reduction and data compression PCA is the most popular and useful tool. The PCA transforms the variables of a dataset into a new set of variables that are linear combinations of the former variables. The new values are called principal components and are ranked according to the variance contributions so that the first principal component provides the highest data variance. Choosing only the first few principal components allows a dimension reduction with a marginal loss of information. K-means is an unsupervised clustering method that arranges the unlabeled dataset into a given number of groups (clusters). It starts with a random distribution of cluster centers and iteratively moves them in order to minimize the total within-cluster variance [28]. KKNN is a non-parametric supervised classification method that is often used because of its simplicity and good performance. To assign a new observation, first the k observations in the training dataset have to be found, which are closest to the new observation. Then the new observation is classified through the majority vote among the k neighbors. KKNN - as an extension of k-Nearest Neighbors algorithm - also takes into account the individual distances of the nearest neighbors to the new observation in the form of weights [26]. The performance of the classification was verified through individual-out-cross-validation (IO-CV).

A standard spectral pre-treatment was applied [29]. First a background was subtracted using the SNIP algorithm [30] followed by a vector normalization. The spectra were projected on a PCA, minimizing the necessary dimensions and computational time. The annotation of a pathologist was transferred to a computer model by a k-means-cluster-analysis with k ranging from 9–15 based on the number of scans of one individual (mouse). The annotation was performed in a blinded manner meaning that the pathologist was not biased by the k-means-cluster-analysis plots and utilized only the HE stained images for diagnostics. Following the annotation, every Raman-spectrum was linked to the pathologist’s diagnosis as the major reason for measuring scans and not single Raman-spectra. This procedure is called spectral histopathology (SHP) and described elsewhere in full detail [31, 32]. The Weighted k-Nearest Neighbors (KKNN) model was trained with one neighbor to be used for prediction and the Minkowski distance in combination with the kernel ‘optimal’. The evaluation was done by an individual-out-cross-validation. This approach incorporates the biological variance in the performance estimation. Only with such an approach a robust estimate of how the model will perform in future applications, e.g. the generalization performance of the model, can be derived. Shortly, the Raman-spectra of one individual were excluded from the model building and a classification model for the groups ‘adenoma’, ‘carcinoma’ and ‘normal’ was constructed. The Raman-spectra of the individual currently excluded were predicted using the classifier and the outcome was stored. Accordingly, the Raman-spectra of all individuals were once used as a test set. The result is put together in Table 1. It should be noted that hyperplasia occured only once and the corresponding region was excluded from the dataset.

The first step of the proposed diagnostic workflow (Fig. 3) consisted of recording a multimodal image, which is utilized to derive a pseudo-HE image by multivariate statistics. For exemplification, the first row (A) in Fig. 2 displays multimodal images of mouse colon sections (for sample details see “Methods” Section, sample preparation). False-colors code CARS in red, TPEF in green and SHG in blue. CARS was adjusted to map the CH ₂-distribution outlining mostly lipids, while the CARS signal originating from proteins is weaker as their molecular mass-to-methylene group ratio is comparably decreased. The TPEF signal was collected in the spectral range between 426–490 nm, thereby, highlighting the distribution of strong auto-fluorophores such as elastin, NAD(P)H and keratin, while SHG collected at 415 nm visualizes the fibrous collagen network.

Proving the similarity of the information content of non-linear multimodal and classical HE stained images, we applied multivariate statistics in order to generate a computational HE stain image out of the multimodal dataset. The results are exemplified in row B of Fig. 2. For the prediction of pseudo-HE images, we utilized a model combining partial-least-square (PLS) regression with a linear-discriminant-analysis (LDA) that was trained using a single HE image. Subsequent to all multimodal measurements, the samples were HE stained and imaged. The corresponding HE images are depicted in row C of Fig. 2. Apparently, morphological information are readily retrieved from both set of images, i.e. in the pseudo-HE images derived from the multimodal images (Fig. 2 b) and the real HE images (Fig. 2 c). Some information, however, is missing in the pseudo-HE images (Fig. 2 b). In particular the cell nuclei are at certain areas not optimally resolved possibly due to signal intensity variations resulting in a non-uniform brightness of the images. It shall be noted that nuclei may be resolved in future pseudo-HE images if stimulated Raman scattering at DNA specific Raman resonances is applied featuring the required high signal-to-noise ratio [33]. From Fig. 2 b and c differences in the color composition and brightness are visible. For example, the submucosa appears darker within the pseudo-HE images than in the corresponding HE images. Further deviations result from the ethanol washing step during the staining procedure removing soluble components. Consequently, some soluble components were imaged by non-linear multimodal microscopy but were removed in advance to the acquisition of the HE images. The advantage of the generated pseudo-HE images is that the measurement is non-invasive, therefore allowing for further analysis. Here, a HE stain was applied afterwards for comparison, but the employment of other stains or measurement modalities is also possible. Due to its speed and non-invasiveness, the pseudo-HE stain can be applied in a cryosection analysis setting or potentially even in-vivo. Based on the pseudo-HE images a pathologist can identify or define suspicious areas (red flags). These small red flag areas can be further investigated by various approaches like e.g. immunostains, or by other label-free spectroscopic techniques featuring a higher molecular sensitivity than non-linear multimodal imaging. In this contribution we applied Raman-spectroscopy as a second diagnostic technique, offering molecular fingerprint information. The results of the Raman-study are summarized in the following section.

As mentioned in the previous section certain areas (red flags) were further characterized by means of Raman-spectroscopy. These red flag regions are marked with a white arrow in Figs. 2 a and 4 a. To prove whether Raman-spectroscopy can be utilized as a molecular selective diagnostic platform, the Raman-spectroscopically characterized regions were annotated and diagnosed by an experienced pathologist in a blinded manner. The adjective ’blinded’ means here that the multimodal images and the Raman-spectroscopic generated image were unknown to the pathologist. The test of the diagnostic value of Raman-spectroscopy was achieved by application of a recently reported workflow termed as spectral-histo-pathology (SHP) [31, 32, 34]. In that approach every Raman-scan is clustered and the diagnosis of the pathologist is transferred to a computer model. Figure 4 d displays such a cluster-analysis. This cluster scan is subsequently annotated using the groups displayed in Fig. 4 e. Based on this annotation mean Raman-spectra of different regions can be derived. Figure 4 f shows as an example the mean Raman-spectrum of normal epithelial tissue. Thereafter, a supervised classification can be applied in order to discriminate between normal, adenoma and carcinoma tissue. A hyperplastic area was only present once in the data set, therefore the corresponding region was excluded from further analysis. In the same manner Raman-spectra of other morphologic areas (muscle, connective tissue) or artefacts (background, spikes) were excluded.

First, a univariate statistical test was applied to verify whether the groups feature statistical significant differences. The statistical significance of the Raman-spectra for the differentiation of tumor against normal tissue and carcinoma against normal tissue were investigated by applying a two-sample Wilcoxon test. The principal component (PC) scores of the components 3, 4 and 5 were proven to be significant for the task normal against tumor tissue. The p-values were 0.001005, 0.019 and 0.003, respectively. The scores of the fourth and fifth component were significant for the task ’normal against carcinoma’ (p-value 0.027 and 0.00039). A boxplot of the scores of PC 3 is visualized in Fig. 5 together with the mean Raman-spectra of the groups. The band assignment for interpreting these Raman-spectra can be found in a number of publications [35, 36]. Here, we assigned the bands for visualization of the biggest differences and marked them in Fig. 5. The band at 785 cm ⁻¹ can be attributed to the phosphate backbone vibration of DNA [35]. The other three bands at 1003, 1449, 1657 cm ⁻¹ can be assigned to protein vibrations [35, 36]. The band at 1003 cm ⁻¹ originates from the symmetric ring breathing of phenylalanine [35], while the Raman-resonances at 1449 and 1657 cm ⁻¹ can be attributed to the CH ₂ deformation vibration [36] and Amide I vibration [35, 36], respectively.

The mean Raman-spectra were subsequently tested for their diagnostic value. Therefore, the Raman-spectra were pre-treated and a Weighted k-Nearest Neighbors (KKNN) classifier was trained and evaluated. In order to estimate the generalization performance of the classifier, the biological variations between the different mice has to be accounted for. To do so, we used an individual-out-cross-validation scheme (IO-CV), where all Raman-spectra of one individual were excluded from training and then predicted. This procedure was iterated and the result is shown in Table. 1. The confusion table shows that the diagnosis of tumor regions, e.g. the classification between tumor regions and normal regions, is accomplished with 100 % mean sensitivity. The differential diagnosis, e.g. the discrimination between adenoma and carcinoma regions, is also possible, but with a lower sensitivity. Here, mis-classifications between adenoma and carcinoma regions occurred. Nevertheless, the overall mean sensitivity for the differential diagnosis is 80.16 % and may be increased in future experiments by improving the detection scheme using, e.g. shifted-excitation Raman-difference-spectroscopy (SERDS) [37]. In a nutshell: the presented workflow allows for a robust and objective diagnosis at least for tumorous and normal epithelial tissue.