Application of Raman spectroscopy for detection of histologically distinct areas in formalin-fixed paraffin-embedded glioblastoma

Abstract Background Although microscopic assessment is still the diagnostic gold standard in pathology, non-light microscopic methods such as new imaging methods and molecular pathology have considerably contributed to more precise diagnostics. As an upcoming method, Raman spectroscopy (RS) offers a “molecular fingerprint” that could be used to differentiate tissue heterogeneity or diagnostic entities. RS has been successfully applied on fresh and frozen tissue, however more aggressively, chemically treated tissue such as formalin-fixed, paraffin-embedded (FFPE) samples are challenging for RS. Methods To address this issue, we examined FFPE samples of morphologically highly heterogeneous glioblastoma (GBM) using RS in order to classify histologically defined GBM areas according to RS spectral properties. We have set up an SVM (support vector machine)-based classifier in a training cohort and corroborated our findings in a validation cohort. Results Our trained classifier identified distinct histological areas such as tumor core and necroses in GBM with an overall accuracy of 70.5% based on the spectral properties of RS. With an absolute misclassification of 21 out of 471 Raman measurements, our classifier has the property of precisely distinguishing between normal-appearing brain tissue and necrosis. When verifying the suitability of our classifier system in a second independent dataset, very little overlap between necrosis and normal-appearing brain tissue can be detected. Conclusion These findings show that histologically highly variable samples such as GBM can be reliably recognized by their spectral properties using RS. As conclusion, we propose that RS may serve useful as a future method in the pathological toolbox.


Application of Raman spectroscopy for detection of histologically distinct areas in formalin-fixed paraffinembedded glioblastoma
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
Spectroscopic techniques such as Raman spectroscopy recently gained more attention in tumor research. [1][2][3] The Raman effect is the process of inelastic scattering of photons, first described in 1928 by the Indian physicist, C.V. Raman. 4 The resulting possible differentiation of tissue types opens up opportunities for its diagnostic use in surgery and pathology. 1,5,6 Therefore, Raman spectroscopy is investigated as both an intraoperative tool due to its non-invasiveness and short integration time, 6 and for its use on processed tissue. The formalin fixation process causes, however, severe changes of the spectral properties that need to be taken into account when comparing fresh and processed tissue. 7,8 We went one step further and performed RS on formalinfixed, paraffin-embedded samples additionally offering precise histological information on adjacent stained serial sections. The general suitability of FFPE tissue for RS was demonstrated already on the small intestine of rats 9 as well as for the distinction of different tumor entities, 10 albeit without a focus on tissue heterogeneity. Changes in the Raman Spectra due to the influence of bound paraffin wax (certain points at 1063, 1133, 1296, and 1441 cm -1 in the Raman spectra can be assigned to contributing paraffin 11 ) make the analysis of the data more complex, even after an efficient dewaxing agent was applied. 11,12 Furthermore, the choice of slide substrate is crucial due to the small thickness of the sections; while silicon-based glass has a strong Raman signal, CaF 2 slides (Supplementary Figure 1) with only one single peak at 321cm -1 , are suitable for Raman Spectroscopy on thin sections. 13 While these obstacles make the conditions for spectroscopic examinations difficult, the possibility to obtain the precise orientation of the sample as well as the identification of certain areas, and therefore the possible option of accessing even a heterogenous tissue fragment (eg, heterogenous tumor entity) precisely, make this effort worthwhile.
Due to its high heterogeneity, which also reflects its previously used name "glioblastoma multiforme, " we chose glioblastoma (GBM) tumor tissue to analyze the suitability of classifying histological tumor areas using RS on FFPE tissue. GBMs are traditionally characterized and diagnosed according to histopathological features and more recently also genetic alterations. 14 Amongst multiple highly variable histological features, GBM usually show a vital, a necrotic, and an infiltrative peritumoral zone. [14][15][16][17] Apart from a traditional assessment of the histology in the light microscopic examination, new techniques and artificial intelligence can in the meanwhile diagnose GBM with a similar or even higher precision than experienced neuropathologists. [18][19][20][21] For this reason, data usually needs to be processed and classified by machine learning 22,23 algorithms, which have strongly influenced the pathology workflow in recent years. [24][25][26] Machine learning tools, as employed in this study, are additionally used for prediction of large vessel occlusion (LVO) stroke 27 and categorization of traumatological patients and prediction of bladder rupture. 28 This present study was designed to examine Raman spectra of different histopathological areas of GBM and to differentiate them with a MATLAB based classifier that we developed in order to test general suitability of RS as a further diagnostic tool in neuropathology.

Patient Data
One hundred and seventeen FFPE tissue blocks from 59 tumors of 53 GBM patients were included in the study. For a more detailed description of the glioblastomas, see Supplementary Table 1. The patients underwent routine diagnostics at the National Center of Pathology (NCP) at the Laboratoire national de santé (LNS, Luxembourg) in the years 2018-2020 and were all part of the INSITU® study. The pathological diagnosis for each tumor was carried out by a board-certified neuropathologist (MM) based on histology, immunohistochemistry as well as on epigenetic and genetic results. Histological tumor areas that were analyzed in the present study were defined by means of light microscopic examination. The present INSITU® study, Nr. 201804/08, has been approved by the "Comité National d"Ethique de Recherche' (CNER). It is handled according to the "EU General Data Protection Regulation GDPR, 29 as well as the world medical declaration (WMA) Declaration of Helsinki as a statement of ethical principles for medical research involving human subjects. 30

Tissue Processing and Slide Preparation
The FFPE tissue blocks were cut with a microtome (thickness of 7μm was chosen, in order to receive a

Importance of the Study
In this study, we established an SVM (support vector machine)-based classifier, which can identify specific histological areas in glioblastoma by the use of spectral tissue properties. In contrast to other studies, our classifier can be used on aggressively chemically treated formalin-fixed paraffin-embedded tissue and also on small tissue fragments. Due to its unbiased approach to tumor diagnostics with very little requirements to the sample, Raman Spectroscopy may be a useful, relatively cheap and easy-to-apply future tool in the toolbox of pathology departments. In addition, we have shown that a heterogeneous tissue cannot be determined by a single spectral property. This must be taken into account when creating a global RS-based classifier to differentiate between several entities.

Neuro-Oncology Advances
sufficient amount of tissue as well as the possibility of a light microscopical examination) and two consecutive sections were put on two slides. The first section was placed on a glass slide and then HE (hematoxylin and eosin) stained for histological examination, the second section was placed on a CaF 2 (calcium fluoride; Crystran, Poole, UK) slide and left unstained to prevent spectral contamination. With a single peak 31 at 321 cm -1 (Supplementary Figure 1), CaF 2 slides have spectral properties that allow proper spectroscopic examination of the tissue on top. In order to reduce the impact of paraffin on the spectra, the slides were dewaxed. The slides were first heated for 60 min at 60°C and then placed in a bath of xylene for 2 × 15 min before being put in descending ethanol solutions for 3 × 2 min. In the next step, distinct areas of GBM (vital GBM core/ necrosis/peritumoral zone) were defined on HE stained slides. For easier optical handling, distinct areas were marked by circling them with a marking pen on the backside, before being matched and correlated with the corresponding areas on the CaF 2 slide ( Figure 1). The vital zone was defined as an area with increased cell density, nuclear pleomorphisms, increased mitotic rate as well as facultative hypervascularization. The necrosis zone was defined by eosinophilic areas within the tumor in which nuclei did not stain with hematoxylin anymore. The peritumoral zone was defined by only very mildly elevated cell density, rather containing reactive astrocytes and potentially very few tumor cells mainly infiltrating along blood vessels and axonal tracts. 14,32 Those areas were chosen because of their histological features, which could be used to provide the ground truth for the spectral classifier. For reuse, the CaF2 slides were washed after this process. To do so, the slides are placed in a bath of warm rinsing agent (consisting mainly of surface-active agents and ethanol) for 30 min and then cleaned manually, until tissue residues could no longer be seen visually. Finally, the slides are washed off with distilled water and adhering tissue residues were spectroscopically excluded.

Raman Spectroscopy
The ProRaman-L high-performance Raman spectrometer (TSI, Shoreview, USA) (Supplementary Figure 2) has an excitation laser at 785 nm and a CCD sensor with a cooling temperature of -60°C enabling repeatable measurements. 33 For data acquisition, a lens with 7 mm working distance and 100 μm spot size was used. The defined areas of the tumor tissue were measured repeatedly, but at different locations within the desired area on CaF 2 slides. The acquisition parameters of the measurement were set to 10 s, 30 averages, 90 mW. The data was recorded and saved as a.spc file using the ProRaman Reader software Version 8.3.6 (TSI, Shoreview, MN).

Data Analysis and Machine Learning
We divided our measurements into two data sets: one set for training a classifier (471 measurements, 100 FFPE blocks from 48 tumors) and a separate, balanced set (99 measurements, 17 FFPE blocks from 11 tumors) for testing and validating the classifier with measurements from different samples to minimize the bias of tumor sample dependent measurement identification. The training data set contained 100 measuring points of necrosis and 105 measuring points of peritumoral zone and 266 measuring points of the vital zone, whereas the second independent data set contained 29 measuring points of necrosis, 37 measuring points of peritumoral area and 33 measuring points of the vital zone. An overview on the number of measurements in the individual histological areas for both data sets is given in Table 1. We used the software SPECTRAGRYPH 34 (Menges Friedrich, Oberstdorf, Germany) for the optical display of individual spectra as well as for processing and displaying of the average spectra of the three areas ( Figure 2). Data analysis was done using the MATLAB Toolbox (MathWork, Natick, MA) for Statistics and Machine Learning™. 35 The MATLAB Classification Learner App 36 is an interactive and user-friendly way to practice supervized machine learning. We trained a classifier in the Classification Learner App with the different areas as response variables, the intensities of the Raman shifts as predictor variables and 5-fold crossvalidation. The app trains a model for each fold and the average test error over all folds is calculated. 37 We chose SVMs (support vector machines) as a method of pattern recognition and trained the algorithm with our data. The resulting classifier was subjected to a hyperparameter optimization; an automatic selection and improvement of hyperparameters to tune our SVM. 38 We calculated the performance statistic based on the output of the SVM classifier. The post-test probabilities were calculated by the Classification Lerner App; the total accuracy was calculated as the percentage of correct classifications out of all classifications.

Classifier Sesults Show Division Into Three Different Spectral Groups
We established a multiclass categorical classifier with a total accuracy of 70.5% (332 correct vs 139 misclassified) and plotted the number of observations as well as the class prediction results in a confusion matrix ( Figure  3). Our classifier revealed a True Positive Rate (TPR) of 81% for the vital zone, 60% for necrosis and 54% for the peritumoral zone. The False Negative Rate (FNR) was 19% for the vital zone, 46% for the peritumoral zone and 40% for necrosis, respectively ( Figure 3). We obtained a Positive Predictive Value (PPV) rate for the vital zone of 76%, for necrosis of 66%, and for the peritumoral zone 58% as well as a False Discovery Rate (FDR) for the vital zone of 24%, for necrosis of 34%, and for the peritumoral zone of 42% (Supplementary Figure 3). In addition, the ROC (receiver operating characteristics curve) curve and the AUC (area under the curve) number can be used as an indicator of the classifier performance, even for multi-class classifications. 39 We received a ROC for each of the three classes. The corresponding AUC value for the necrosis class was 0.86 (Figure 4), for vital zone class 0.81 and for peritumoral zone class 0.80.

Good Discrimination Between Necrosis and Peritumoral Area Even in Unbalanced Training Data
As a result of the area distribution within the tumor tissue, the number of measuring points corresponding to one of the three areas is not distributed equally (refer to Table 1 3800  3600  3400  3200  3000  2800  2600  2400  2200  2000  1800  1600  1400  1200  1000  800  600  400  200   Intensity   95  90  85  80  75  70  65  60  55  50  45  40  35  30  25  20  15  10 average necrosis average peritumoral zone average vital zone Figure 2. Raman spectroscopy data of histologically defined areas in GBM. Average spectra of the three histologically defined areas in GBM (vital tissue / peritumoral tissue / necrosis). For a better overview, the CaF 2 peak at 321 cm -1 has been subtracted. SPECTRAGRYPH software was used for data analyses. Overview about the number of glioblastomas and measurements carried out in total as well as distribution of the measurements related to the two data sets. The external validation data set contains different tumors to avoid bias.

Neuro-Oncology Advances
with aminolevulinic acid) as well as visual and tactile impressions to resect tumor tissue, which, compared to peritumoral or healthy brain tissue, is resected in larger quantities. Of note, our classifier has been trained with an unbalanced data set reflecting this distribution. The detection of the vital zone revealed the highest TPR at 81% and PPV at 76%, which is dependent on the frequency of vital zones appearing in our data set. Most misclassifications occurred between the vital zone and the necrosis / peritumoral zone, respectively, while only a small percentage (in total 21 mispredictions of 471 measurements) is between necrosis and peritumoral zone ( Figure 5), suggesting a good separability of these two areas using our established classifier.

Classifier Validation
We tested our classifier with a balanced dataset of different tumors to exclude the imbalance of input data. Testing the classifier with an extra data set is also a protection against overfitting after the optimization process. 38 The true positive rate / true negative rate (TNR) in this classification was 59% (TPR) / 94% (TNR) for necrosis, 54% (TPR) / 94% (TNR) for peritumoral zone and 79% (TPR) / 58% (TNR) for vital zone (Supplementary Figure 4). Amongst the misclassifications, there was only one overlap (misclassification) between necrosis and peritumoral zone, all other misclassification occurred either between vital zone and necrosis (12 necrosis as vital zone, 3 vital zones as necrosis) or vital zone and peritumoral zone (16 peritumoral zones as vital zone, 4 vital zones as peritumoral zone). This mirrors the situation in the 5-fold cross-validation of the classifier training, suggesting a good separation between the peritumoral zone and necrosis. The vital zone is also a distinct spectral entity but more frequently misidentified as either of the other ones. This reflects the morphological arrangement of the zones. For a closer look at the sensitivity, specificity and positive predictive value/false discovery rate obtained from the data set for validation, see Table 2.

Discussion
Our data shows that RS can distinguish histologically defined zones in GBM even upon aggressive chemical   treatment leading to a considerable change of the biochemical tissue composition as in FFPE samples.
A considerable number of spectra acquired from the vital zone shows spectral similarities to either necrosis or infiltrative peritumoral areas, however, most spectra of this area are still recognized as a distinct spectral entity by the classifier. The clearest separation is achieved between the necrotic and the peritumoral areas of the tumor. This is in line with previous findings on frozen tumor samples. 40,41 In our SVM classifier, all three histological areas are presented as distinct spectral entities. This could be confirmed with a balanced, independent data set revealing a good differentiation between necrosis and peritumoral zone. This confirms that GBM cannot be sufficiently determined by a single type of spectral property, which is an important fact to be considered for global RS-based classifiers aiming at distinguishing different tumor entities. For a global classifier approach, eg, using RS as an intraoperative diagnostic tool to identify the adequate resection margin, future studies may use our findings to distinguish the zones in fresh, unprocessed tissue as well. However, the translation of spectra across different levels of processing is challenging and will require additional analytic tools. The Raman shift values provide information about the underlying biochemical properties of the sample. However, the approach used in this study does not make use of such information but treats the spectra as fingerprint-like patterns. This leaves several potential sources of influence such as the formalin fixation or individual residues of paraffin wax sticking to the tissue, difficult to access. Despite optimization protocols, we thus cannot exclude residue spectra.
While necrosis and peritumoral zone can be well distinguished by our classifier, we see some misclassification between the vital zone and of both adjacent areas. There are several possible explanations for this phenomenon. The classifier was trained based on the histological assignment of areas as one of the three classes, however, the precision of this assignment is naturally limited. In representative tumor tissue, the three areas are in tight morphological association and the identification of the histological ground truth was not done on the measured section but on the adjacent one, which was stained. In the small deviation between these two sections, some spatial shifts may explain the observed differences between the assigned class and the one predicted by the classifier. Even within the same section, the intimate intertwining of features may contribute to a heterogenous spectral signal within the spot size of the excitation laser, which is 100 µm in diameter. It would be possible to increase the resolution with smaller spot sizes, however, we chose a spot-measurement-based method that, in contrast to microscopic Raman scanning, aims to provide a fast readout with a single measurement that is largely insensitive to sampling errors. Furthermore, it could be assumed that necrotic areas may still contain biochemical constituents that appear within the spectral range of the vital zone. As proposed in a rat model, central necrosis (no signs of proliferation) and peripheral necrosis (associated not only with pyknotic but also pseudopalisading cells) display different spectral identities, attributed to a different concentration of plasma proteins or high lipid content. 42 Therefore our necrotic area itself may display a merging zone, of which some spectral overlap with the vital zone occurs.
Taken together, the classification of tumor areas is based on fundamentally different properties of the tissue, on the one hand, the morphological appearance, versus on the other hand the biochemical composition. Given the good discrimination between the peritumoral zone and necrosis and the identification of a distinct but overlapping spectral class for the vital zone, this tool presents a useable method for the identification of tissue even when not all required features for diagnostics are present in a sample. This taken into account, our method of measurement and classification does provide a good way to identify the different zones in an unstained section of an FFPE sample. Even if only one of the histological areas would be present in a given sample, the classifier could be used for diagnostics, in contrast to histopathological Overview of sensitivity, specificity and positive predictive value when using SVM classifier for the validation cohort (n = 99). On the one hand, measuring points can be divided into three different groups. On the other hand, the number of incorrect classifications between the peritumoral zone and the necrosis is visualized.

Neuro-Oncology Advances
assessment, which requires all three. Further studies are needed to determine if this also holds true against similar areas of different origin, such as necrosis in a metastasis. Further studies with complementary analyzes (eg, mass spectroscopy) could offer the possibility of determining changed biochemical processes in the course of the previous fixation and paraffin wax embedding.
In conclusion, this study demonstrates the possibility of classifying different tumor areas on FFPE GBM tissue with the help of Raman spectroscopy. This may be a useful, relatively cheap and easy-to-apply tool to complement histopathological and molecular diagnostics. Like other qualitative methods, it provides an unbiased approach to tumor diagnostics with very little requirements to the sample.

Supplementary Material
Supplementary material is available at Neuro-Oncology Advances online.