Open Access
Add to BibSonomy
Abstract
Delineating cancer tissue while leaving functional tissue intact is crucial in brain tumor resection. Despite several available aids, surgeons are limited by preoperative or subjective tools. Raman spectroscopy is a label-free optical technique with promising indications for tumor tissue identification. To allow direct comparisons between measurements preprocessing of the Raman signal is required. There are many recognized methods for preprocessing Raman spectra; however, there is no universal standard. In this paper, six different preprocessing methods were tested on Raman spectra (n > 900) from fresh brain tissue samples (n = 34). The sample cohort included both primary brain tumors, such as adult-type diffuse gliomas and meningiomas, as well as metastases of breast cancer. Each tissue sample was classified according to the CNS WHO 2021 guidelines. The six methods include both direct and iterative polynomial fitting, mathematical morphology, signal derivative, commercial software, and a neural network. Data exploration was performed using principal component analysis, t-distributed stochastic neighbor embedding, and k-means clustering. For each of the six methods, the parameter combination that explained the most variance in the data, i.e., resulting in the highest Gap-statistic, was chosen and compared to the other five methods. Depending on the preprocessing method, the resulting clusters varied in number, size, and associated spectral features. The detected features were associated with hemoglobin, neuroglobin, carotenoid, water, and protoporphyrin, as well as proteins and lipids. However, the spectral features seen in the Raman spectra could not be unambiguously assigned to tissue labels, regardless of preprocessing method. We have illustrated that depending on the chosen preprocessing method, the spectral appearance of Raman features from brain tumor tissue can change. Therefore, we argue both for caution in comparing spectral features from different Raman studies, as well as the importance of transparency of methodology and implementation of the preprocessing. As discussed in this study, Raman spectroscopy for in vivo guidance in neurosurgery requires fast and adaptive preprocessing. On this basis, a pre-trained neural network appears to be a promising approach for the operating room.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Around 300 000 new brain or central nervous system cancer diagnoses are made annually worldwide [1]. The diagnoses include both primary brain tumors and cancer that has metastasized to the brain causing secondary tumors. The first step of treatment is commonly surgery, followed by radio- and chemotherapy depending on the histology and molecular characteristics of the tumor.
During surgery, gross total resection, or removing as much tumor tissue as possible is linked to longer progression-free survival [2]. Yet, delineating brain tumor tissue while leaving functional tissue intact is difficult. Several aids are available, such as image-guided surgery with a contrast agent and fluorescence-guided surgery using either blue light microscopy [2] or spectroscopy [3,4]. These methods are all subject to limitations, e.g., tumor tissue is known to extend beyond the contrast-enhancing region in the images [5,6]. Fluorescence guidance requires oral administration of 5-aminolevulinic acid (5-ALA) before surgery that can metabolize to protoporphyrin IX (PpIX). The accumulation of PpIX is associated with high-grade tumor tissue. Generally, low-grade tumor tissue does not metabolize 5-ALA and the fluorescence technique may therefore not detect low-grade tumor boarders. A method to assist neurosurgeons with real-time detection of low-grade tumor tissue during resection, preferably without the need for additional drugs, is therefore sought after.
Raman spectroscopy is a label-free optical technique that measures molecular vibrations through the detection of inelastic scattering of light emitted from a monochromatic laser source. Raman spectroscopy has been demonstrated as a suitable method for cancer detection [7–11], by detecting molecular vibrations corresponding to proteins, lipids, and their intensity ratios. In addition to spectroscopy, coherent Raman imaging techniques are making advancements in histology [12–15]. With modern hand-held probe systems, Raman spectroscopy is feasible as a technique for in vivo classification of cancerous tissue [16–18]. In clinical settings, Raman measurements must be quick and accurate. The Raman signals are inherently weak, due to the low probability of inelastic scattering, and may be overwhelmed by intrinsic fluorescence and noise the signals require preprocessing. There are many recognized methods for preprocessing Raman spectra, however, there is no universal standard.
Preprocessing of Raman spectra usually consists of removing cosmic rays, smoothing, and baseline correction. Cosmic rays are commonly removed by taking the median of a measurement series. A popular smoother for noise reduction in Raman spectra is the Savitzky-Golay (SG) filter [19] where a polynomial fit is made locally to optimize signal reconstruction.
A literature search indicated that the common methods for baseline correction are based on either polynomial fitting, mathematical morphology, or signal derivatives. Both commercial software and individual implementations are available. The polynomial approach can be either direct, through least-squares fitting, or iterative. Morphology can be determined by a series of erosions and dilations of the signal. In contrast to these methods that have been designed to handle baseline correction specifically, we have previously shown that an artificial neural network trained on synthetic data can be used to perform both noise reduction and baseline correction in Raman preprocessing [20]. In this study, the identified methods were applied to the same dataset of Raman signals acquired from freshly resected brain tissue samples.
After preprocessing, the Raman peaks are generally assigned to specific molecular vibrations to classify the sample. This can be done either through individual peak assignment and ratios between proteins and lipids [21,22] or fitting of spectra from physiologically relevant biomarkers [23–25]. However, this can be a daunting task for large datasets with large variations and for the variety of central nervous system tumors. As can be exemplified by the ‘Central nervous system tumors, 5 ed’, that lists [26]. Signal analysis can be aided by data exploration techniques such as principal component analysis (PCA) in combination with t-distributed stochastic neighbor embedding (tSNE) [27,28]. For Raman spectroscopy, PCA transforms a set of spectra into a point cloud, where each spectrum is represented by a single point [27]. The location of the points in the PC-space is associated with the features in the data, such that two spectra that reside in close vicinity will have similar signals. If the dataset contains many different features, it can be difficult to visualize all significant planes. tSNE can be used to compute a projection of the points in PC-space where clustering is conserved [28]. For its benefits in visualizing high-dimensional data, the combination of PCA and tSNE were used in this study. To make an objective decision for models with multiple parameter choices, a parameter sweep was used for cross-validation with Gap-statistics for optimal clustering of the PCs. This study aimed to investigate whether the choice of preprocessing method affects what features appear in data and discuss the suitability of different methods for in vivo analysis of spectra.
2. Methods
2.1 Measurements
2.1.1 Study design
Fig. 1. Schematic of the methodology used in this paper. Measurements: Raman spectra of brain tissue samples were recorded using a 532 nm laser, microscope objective, and a spectrometer (not shown). Analysis: The tissue samples underwent neuropathological analysis. The spectroscopic data were preprocessed individually. Evaluation: Data exploration through feature evaluation and cluster techniques was used.
The study included 16 patients (8 females, mean age 58, range 27 to 80 years) referred for surgery at the Department of Neurosurgery, Linköping University Hospital. All patients had a suspected low- or high-grade tumor located in the cerebrum or pituitary. According to clinical protocol, patients with suspected high-grade tumors were given an oral dose of Gliolan (5-ALA, 20 mg/kg) 2-3 hours before the resection to utilize fluorescence guidance through blue light microscopy. Before surgery, the patients gave written informed consent (EPM 2020-01404). In total, 34 resected tissue samples were included in the study.
2.1.2 Raman setup and measurements
Raman measurements were performed using a continuous-wave diode-pumped solid-state laser of wavelength 532 nm laser (DPSS 532, Altechna, Lithuania). The sample was illuminated through a microscope objective (40x NA 0.6, LUCPLFLN, Olympus, Japan). The spectral signal from the tissue sample was separated from the laser line by dichroic filters (532 nm RazorEdge Dichroic and EdgeBasic, Semrock, MA, USA), before being guided into the spectrometer (Shamrock 303i, Andor Technology, Belfast, UK). The spectrometer had a slit opening of 50 µm, a grating of 1200 groves/mm, and the CCD (iDUS 401, Andor Technology, Belfast, UK) that recorded the spectra was cooled to -70°C. The system has a calculated spectral resolution of 0.2 nm. The setup was installed onto a cart and placed in a facility next to the operating room where measurements on fresh tissue samples were performed.
Each sample was placed on a cover glass and secured in front of the microscope objective. In each measurement point, three spectra with an integration time of either 5, 10, or 50 s each were obtained. Measurements were taken in a series of alternating fingerprint (350 cm-1 to 2230 cm-1) and high wavenumber (1870 cm-1 to 3430 cm-1) regions for three different laser powers (3, 9, and 18 mW). The region between 1800 cm-1 and 2200 cm-1 is commonly referred to as the silent region. It typically does not contain Raman peaks related to biomolecules but can reveal features introduced by the hardware or preprocessing. After each measurement, a new position was chosen. If the tissue sample was large enough, the entire procedure was repeated. A schematic of the in-house-built Raman acquisition set-up can be seen in Fig. 1(A)). An overview of the following analysis and data exploration workflow is found in Figs. 1(B)) and 1C).
2.2 Data analysis
2.2.1 Neuropathology
After the measurement series was acquired, the tissue samples were placed in formalin for neuropathological analysis and classification. The analysis included slicing, staining with hematoxylin and eosin (H&E), and if applicable, immunohistochemical and molecular markers were used to conclude patient diagnosis. Additionally, a reevaluation of diagnoses including a detailed subanalysis of each sample was conducted by a senior pathologist (MH) to conclude individual sample diagnosis. The percentages of low- or high-grade tumor, marginal zone, necrosis, and no-tumor tissue were evaluated. Whenever relevant, proliferation status (KI-67), isocitrate dehydrogenase (IDH) mutations, and presence and percentage of O6-methylguanine-DNA-methyltransferase (MGMT) methylation were noted. Thus each patient received an overall diagnosis according to CNS WHO 2021 [26] and in addition, each of the 34 samples received an individual description based on the specific growth pattern in that sample.
2.2.2 Preprocessing
The collected data were preprocessed in several steps. First, cosmic rays were removed by calculating the median of three consecutive measurements in each pixel along the spectral window [29]. To account for photobleaching, the measurement series were grouped by extended multiplicative scattering correction (EMSC) [30]. In the next step, either no smoothing or smoothing using the SG filter [19,31] was applied. Thereafter, the dataset was exposed to the six baseline correction algorithms (i)-vi)). Normalization was made before and after both smoothing and baseline correction by setting the mean to zero and dividing by the L2-norm. See the ‘Preprocessing’-box in Fig. 2 for an overview of the preprocessing schematic.
For baseline correction two polynomial methods that used a Chebyshev basis were investigated. The first (i) performs a direct polynomial fitting to the spectrum [32], and the latter (ii) fits a smooth curve below the measured signal and is henceforth called Vancouver [33]. The Rollingball algorithm (iii) was chosen for its basis on mathematical morphology [34]. Matlab’s msbackadj correction algorithm (iv, Bioinformatics Toolbox, Matlab R2021a, Mathworks) was chosen as representative of a commercial software. The final approach that was included was the second derivative (v). This approach utilizes that the baseline derivative is much smaller than that of the peaks in the Raman signal. Lastly, a neural network (vi) trained on synthetic data was applied [20]. The training data was generated from randomized baselines and Lorentz peaks covered with varying amounts of noise. The network was a 20-layer convolutional neural network. A Matlab (R2021a, Mathworks) script for the generation of synthetic data and training of the network can be found in the Supplement 1, which contain the exact settings used in this work, see Code 1 (Ref. [35].). A parameter sweep has been used for methods where multiple parameter choices are available. Note that due to the inherent difference in functionality and implementation of the methods their parameter ranges differ. An overview of the pipeline and evaluated parameter choices can be seen in Fig. 2.
Fig. 2. Preprocessing pipeline with parameter ranges for smoothing and baseline correction, clustering, and spectral feature estimation. The neural network was trained to perform both smoothing and baseline correction.
2.2.3 Data exploration and evaluation
After preprocessing, the data was evaluated with data exploration using PCA and tSNE. The resulting point cloud was clustered using k-means [36], where Gap-statistics was used to estimate the number of clusters [37], see ‘Evaluation’-box of Fig. 2. The preprocessing was repeated for a range of parameter choices (Fig. 2), where the variable settings that resulted in the largest elbow value in the Gap-statistics were kept for visualization of the distribution of the preprocessed Raman spectra. The Gap-values do not provide a qualitative method to compare the preprocessing methods and are solely used to find the optimal clustering for a particular method. The mean Raman spectrum for each cluster was analyzed to identify the molecular influences in the resulting clusters. Lastly, the clusters and corresponding spectral features were compared to those found in the literature.
3. Results
3.1 Neuropathology
Following neuropathological staining and relevant molecular analyses, the tissue samples were categorized into four CNS WHO groups [26]: adult-type diffuse gliomas, meningiomas, tumors of the sellar region, and brain metastases. Tissue samples without indication of tumor cells as well as background measurements were placed in separate categories. The cohort included tumors of CNS WHO grade 1-4 and three types of metastases. The largest sample group was adult-type diffuse gliomas with 379 Raman measurements (13 tissue samples) and tumors of the sellar region were the smallest group with 34 Raman measurements (3 samples). Necrosis was present in two glioblastoma and all the metastases samples. The majority of one glioblastoma IDH-wildtype and the granular cell tumor of the sellar region samples were non-tumor tissue. Samples in the non-tumor tissue group included gliosis and tissue associated with cysts. All sample categories and the detailed distribution of the 969 measurement points are found in Table 1.
Table 1. CNS WHO tissue sample classification and grade from neuropathology. IDH: isocitrate dehydrogenase, MGMT: O6-methylguanine-DNA-methyltransferase, meth: methylated, mut: mutated, wt: wild-type. For each group, the total number of patients and samples with that diagnosis are presented followed by the number of acquired spectra in the fingerprint and high wavenumber region, respectively.
3.2 Raman spectra
The resulting clusters varied in number, size, and features between preprocessing methods. The number of clusters ranged from eight to eleven in the fingerprint region and five to thirteen in the high wavenumber region. The preprocessed spectra are presented as points in the principal component space projected onto a plane by tSNE. The projections can be seen in Figs. 3 and 4, for the fingerprint region and high wavenumber region respectively. Each subfigure in the respective graphical objects is associated with a particular preprocessing method. The colors of the points/spectra are based on the sample diagnosis from the neuropathological analysis presented in Table 1. The parameters for the respective preprocessing methods can be found at the end of this section.
Observe that the position of each point/spectrum in the tSNE-space depends upon their respective features. The spectral features from the pattern of Raman peaks were compared to the scientific literature to find associated molecules. In the fingerprint region carotenoids, hemoglobin, and neuroglobin were found. For the high wavenumber region, the clusters could be associated with variations in features related to water, as well as the existence of protoporphyrin. Table 2 lists the molecule or -groups, their corresponding Raman peaks found in literature, references, and their associated cluster [21,22,38–43]. Cluster labels are based on their spectral signature; tissue (T), hemoglobin (H), neuroglobin (N), background (B), carotenoid (C), water (W), and protoporphyrin (P). The spectra varied within the given labels to form multiple clusters due to differences in signal-to-noise ratio, peak proportions, or the addition of some otherwise missing Raman peaks. To account for these variations, sub-labeling was introduced by adding asterisks (*) or subscripts (ex. WA). Note that no cluster should be considered more important than another, regardless of asterisks or sub-labels. The tissue (T) clusters are associated with lipids and proteins or varying signal quality. Hemoglobin (H), neuroglobin (N), carotenoid (C), and PpIX-fluorescence peak from metabolized 5-ALA (P) have multiple clusters that are distinguished by a significant reduction in signal-to-noise ratio. Peaks that appear above 3100 cm-1 in tissue are often considered to be water, therefore different spectral signatures were labeled as WA, WB, and WC, while asterisks indicate a lower signal-to-noise ratio.
Fig. 3. Clustering results from Raman data in the fingerprint region, when having preprocessed the data with different methods, subfigures i) to vi), before computing the principal components and the two-dimensional projection using tSNE. Each dot represents a spectrum, colored to indicate the neuropathological labeling. The clusters are labeled with a letter to indicate the primary spectral feature observed in the cluster mean (T: tissue, C: carotenoid, N: neuroglobin, H: hemoglobin, B: background, *: variations within label, SG: Savitzky-Golay).
Fig. 4. Clustering results from Raman data in the high wavenumber region when having preprocessed the data with different methods, subfigures i) to vi), before computing the principal components and the two-dimensional projection using tSNE. Each dot represents a spectrum, where they are colored to indicate the neuropathological labeling. The clusters are labeled with a letter to indicate the primary spectral feature observed in the cluster mean (T: tissue, W: water, B: background, *: variations within label, SG: Savitzky-Golay).
Table 2. Molecules and groups, their associated wavenumber found in literature, references, and associated clusters from the processed Raman signals. Cluster labels are associated with tissue (T), hemoglobin (H), neuroglobin (N), background (B), carotenoid (C), water (W), and PpIX-fluorescence from metabolized 5-ALA (P). Asterisks (*) indicate variations within label.
Figures 5 and 6 show the average preprocessed spectra from each preprocessing method for three clusters in the fingerprint and high wavenumber region, respectively. The remaining spectra from cluster means can be found in Supplementary Figures S1 – S7. In Fig. 5, spectra from the fingerprint region can be seen. Figure 5(A) shows clusters labeled as tissue (T) that have features associated with proteins and lipids. The vibrational modes seen at 1312 cm-1 and 1467 cm-1 can be assigned to lipids [21]. Carbon stretching modes of proteins can be assigned to the peak at 1147 cm-1 [21]. The peaks at 1560 cm-1, 1645 cm-1, and 1673 cm-1 can be assigned to amide I and amide II, respectively [21]. 1617 cm-1 can be assigned to the C2 stretch in proteins [38]. There are variations in the composition of tissue and the corresponding spectrum, this is seen in Fig. 5(B) which shows the spectra from the cluster mean of T*. The peaks at 960 cm-1, 995 cm-1, 1287 cm-1, 1555 cm-1, and 1460 cm-1 can be assigned to collagen, glucose, amide II and lipids, respectively [21]. The most pronounced feature in the fingerprint region is the three carotenoid peaks at 1001 cm-1, 1155 cm-1, and 1511 cm-1 [41], which can be seen in Fig. 5(C).
Fig. 5. Preprocessed Raman spectra from fresh brain tissue samples covering the fingerprint region. Six different methods were used for smoothing and baseline correction: Polynomial, Vancouver, msbackadj, Rollingball, second derivative, and a neural network trained on synthetic data. Spectral features of a cluster with the influence of A) proteins and lipids (T in cluster analysis), B) proteins and lipids, but at different peak intensities (T* in cluster analysis), and C) carotenoid signature (C in cluster analysis). SG: Savitzky-Golay.
In the high wavenumber region, the clusters labeled tissue (T) can be seen in Fig. 6(A). In this cluster, the peaks at 2850 cm-1, 2890 cm-1, and 2940 cm-1 are believed to have significant importance for cancer identification [21]. The peak at 2850 cm-1 can be associated with CH2 symmetric stretch [21]. Similarly, 2890 cm-1 is the vibrational mode of CH2 asymmetric stretch [21]. These peaks reappear in the clusters that have additional peaks beyond 3100 cm-1, which can be associated with water (W). In Fig. 6(B) the cluster labeled as WA can be seen. In samples from patients that had been given 5-ALA for fluorescence-guided resection a pronounced peak at 3060 cm-1 could be seen, see Fig. 6(C).
Fig. 6. Preprocessed Raman spectra from fresh brain tissue covering the high wavenumber region. Six different methods were used for smoothing and baseline correction: Polynomial, Vancouver, msbackadj, Rollingball, second derivative, and a neural network trained on synthetic data. Spectral features of a cluster with the influence of A) proteins and lipids (T in cluster analysis), B) proteins, lipids, and water (WB in cluster analysis), and C) PpIX-fluorescence peak from metabolized 5-ALA (P in cluster analysis). SG: Savitzky-Golay.
Fig. 7. Final parameter choices for six preprocessing methods and the raw data of Raman spectra from fresh brain tissue samples. Following a parameter sweep, the combination with the highest Gap-value was chosen. For each method, the corresponding variables, number of clusters, and Gap-values are given. The values are presented for the fingerprint and high wavenumber region, respectively. Clustering was performed using PCA (principal component analysis) and tSNE (t-distributed stochastic neighbor embedding). SG: Savitzky-Golay.
Observe that the preprocessing methods, with some notable exceptions required user input parameters to dictate the performance of the algorithms. The choices were made with a parameter sweep that found the optimal clustering by calculating Gap-values. A summary of the maximum Gap-values for the respective preprocessing methods employed in this study is found in Fig. 7. Note that the outcome when omitting either the smoother or the baseline correction has been included in this summary.
4. Discussion
4.1 Raman measurements
The measurements in this study were acquired from two different spectral regions (fingerprint [350-2230 cm-1] and high wavenumber [1870-3430 cm-1]). Both regions contained information related to potential biomarkers. It is therefore suggested that a Raman system with clinical relevance should have a spectral window that is sufficiently large to include both the fingerprint and high wavenumber region, i.e., approximately 400–3500 cm-1 as to perform measurements in both regions in the same measurement point. This can be achieved by changing the grating, which would not alter the complexity of the setup. There is however a trade-off between the size of the spectral window and resolution, which can affect the outcome.
In Figs. 5 and 6, there is the appearance of an oscillatory pattern in the residual baseline. The oscillations are likely due to chromatic imperfections in the optics in combination with the strong fluorescent baseline, which have been discussed in the works of Alleon et al. [44]. A dichroic filter transmits the Raman scattering and the fluorescence with close to 100% transmittance. However, the transmission profile of the filter is not flat and modulates the signal. The strong fluorescent background seen in our measurements was often close to the detection limit of the CCD, which results in the visible oscillatory pattern from the filter. In those cases where the Raman signal is significantly stronger, the oscillatory pattern becomes less prominent. The effect can be seen by comparing the tissue spectra of Fig. 5(A)) - 5B) with the significantly stronger carotenoid signature in Fig. 5(C)). These effects do affect the clustering, especially in those cases where the Raman signal is weak. Spectra that have significantly different Raman signals can reside close to one another in the PC-space because the residual baseline is the dominant feature. However, clusters for different spectral patterns can still be separated because the Raman signals are placed on top of the oscillations that remain constant throughout the dataset.
4.2 Preprocessing
4.2.1 Outcome of methods
The choice of preprocessing methods and parameters will affect which spectral features appear in a data set. As the features vary, so does the clustering, as can be seen in the results of data exploration by using different preprocessing methods on spectra from brain tumor tissue samples (Figs. 3 and 4). Spectra from biological samples often contain many molecules, which in turn results in multiple vibrational modes that can add up to create large spectral features with multiple peaks. Two commonly discussed points of apprehension when preprocessing Raman spectra with that level of difficulty are that a smoother can average peaks and a baseline correction algorithm can cut into wide structures of the spectrum.
In Figs. 5 and 6, it can be seen how the methods used in this study performed regarding deletion of key information or creation of false artifacts. The SG filter appears to have deleted peaks in Fig. 5(A)) and Fig. 6(A)), both in combination with baseline correction from msbackadj in Matlab. Classification could still be possible if a smoother has transformed a cluster of peaks into a single peak and if ratios between those peaks hold key information. When a baseline correction algorithm cuts into clustered peak formations, it can alter the proportions required for classification. For example, in the high wavenumber region (Fig. 6(A)) and Fig. 6(B))) the Rollingball algorithm can be seen to cut into the triple peak from carbon-hydrogen bonds (2850, 2890, and 2940 cm-1). False spectral features can appear because of artifacts from the preprocessing, such as corner effects from baseline correction. In the fingerprint region in the outcome from msbackadj, Fig. 5(A))-5C), the resulting baseline has an artificial steep downward trend on the left side of the spectrum. Similarly, in the high wavenumber region of the polynomial-based methods (Polynomial and Vancouver), Fig. 6(C)), a wide peak towards the right side of the spectrum can be seen, located beyond the protoporphyrin peak from 5-ALA (3060 cm-1). This is an effect that appears when the estimated baseline finds a shortcut along the path of the data, rather than tracing the outline of the curve. Many baseline correction algorithms suffer from a trade-off between cutting into wide structures that should be preserved and removing the full extent of fluorescence. One potential consequence of cutting into spectral features is that peak-to-peak proportions are altered. The second derivative avoids that problem by not computing an estimate of the baseline at all. The estimation of the derivative does however amplify the noise in the signal. Assuming normally distributed noise, the standard deviation of the noise is increased with a factor of $2\sqrt 2 $ by computing the second derivative.
4.2.2 Parameter choices
The outcome of preprocessing from methods that rely upon user-defined parameters will depend on the choices made. The artificial neural network has the benefit of being adapted for a variety of signal-to-noise ratios without changing the parameters [20]. This allows for rapid preprocessing without the need for parameter optimization through means such as cross-validation which is computationally demanding. In this study, the neural network resulted in fewer clusters in the data exploration than the other methods (only compared with results from combined smoothing and baseline correction). This is because raw spectra that have seemingly similar features, but variations in signal-to-noise, ended up in different clusters. One drawback of the neural network is that if spectra contain more variations than those included in the training, the quality of the preprocessing is reduced [20]. The 5-ALA-induced peak from protoporphyrin is an example of a situation that goes beyond the range of the training data, the peak can be seen in Fig. 6(C)). In the raw data, this peak consumes the entire spectrum where all details from other Raman peaks are lost. The neural network transformed the peak to resemble a Lorentz peak, which was the model for the peaks during training. Alternatively, the network considered the 5-ALA peak to be part of the background and completely removed it. The latter example could be an explanation for why there are so many samples labeled as glioblastoma in the cluster with spectra that look solely like background for the neural network, see Fig. 4(vi).
As presented in Fig. 7 the largest Gap-values were found for baseline correction using Vancouver in the fingerprint region and Polynomial in the high wavenumber region, respectively, both in combination with smoothing. Gap-values were used to determine the optimal clustering for each method. There exists a potential bias for the number of clusters using Gap-values, as compared to variance or other decision-making parameters. Further, the Gap-value is not a measure of the quality of the preprocessing, but rather a measure of how much variation is explained by the clustering. An objective measure of the quality of preprocessing would require us to perform tissue classification – more on this in the next section.
4.3 Tissue classification
From a clinical perspective, it would be preferable to label brain tissue as either healthy or tumor in vivo before the tissue is resected. Due to ethical considerations, no samples were taken from tissue that was considered healthy by the surgeon, thus the discussion of healthy versus tumor tissue could not be investigated based on this dataset. Brain tumors are inherently heterogenous, often with a diffuse border to the surrounding tissue, thus tissue samples taken from different locations in the same patient often show different tissue morphology. As a result, this cohort was found to be inhomogeneous. This can result in many different spectral features appearing in a sample with a single histopathological evaluation. The way this study was conducted makes it impossible for the neuropathologist to correlate an exact histopathological evaluation to each of the Raman measurement positions. For example, one sample that contained 90% non-tumor cells was, according to clinical standard, classified as meningioma. Note that Raman spectra were collected at different positions on this sample, some of which may correspond to non-tumor tissue. This standard can lead to ambiguous sample labels in the clusters. Further, the sample sizes are often skewed with unevenly distributed group sizes due to the frequency of each tumor type. In this study, data sets from tumors of the sellar region were significantly smaller than those of meningioma or glioblastoma, see Table 1.
To be able to perform in vivo classification of brain tumor tissue using Raman spectroscopy it is currently necessary to have fast and reliable preprocessing. There are studies where a classifier was built using machine learning techniques that do not rely on extensive preprocessing [45]. It is, however, difficult to acquire sufficient data with accurate labeling. Classifiers based on PCA, or similar methodologies can yield fair classifiers with around 100 independent samples [46], but they require the data to be preprocessed. At this stage, no classification of tissue has been attempted. However, such methods perform classification on patterns that are similar to those done by clustering using PCA and tSNE. The clustering shown in Figs. 3 and 4 does appear to show some tendencies that might be useful for classification, but we urge caution in this regard. For example, the existence of neuroglobin or carotenoids may be correlated with meningioma judging from the current dataset, see Fig. 4. However, to verify this further measurements are necessary. For most methods used in the preprocessing, the number of clusters exceed the labels from the neuropathology. This could imply that there are insufficient neuropathological labels, but many clusters have similar spectral signatures but at different signal-to-noise ratios. Such an outcome is expected for preprocessing with fixed parameters, that do not adapt to variations in individual spectra. See Supplementary Figures S1 – S7 for the average spectral tendency for each cluster.
A pretrained neural network shows promise to be compatible with an in vivo classification system because it allows fast online preprocessing and, as suggested in this study, it adapts to variations in signal quality without user interference. An argument can be made for the speed of the neural network compared to other methods. All methods, except the neural network, require computations to find suitable parameters for adaptive processing. However, for the neural network, all these computations are performed prior to when the preprocessing is needed.
5. Conclusions
In this study, more than 900 spectra from 34 fresh brain tumor tissue samples were investigated with Raman spectroscopy. It has been shown how different preprocessing procedures result in varying numbers of clusters by comparing the principal components following preprocessing of spectra. The features seen in the data set are considered constant following reasonable preprocessing, however their spectral appearance does change when employing different methods. With differences in the appearance of features and how they can be clustered it becomes difficult to compare different Raman spectroscopic studies. To enable Raman spectroscopy for guidance in neurosurgery, measurements must be performed in vivo, and the analysis must be automatic and quick. A neural network trained from synthetic data to do both noise reduction and baseline correction appears to be a promising approach to enable adaptive preprocessing in the operating room. The quality of the preprocessing for the clinical perspective requires that the accuracy of classification is studied.
Funding
Swedish Foundation for Strategic Research (RMX18-0056).
Acknowledgements
The authors would like to thank the staff at the Department of Neurosurgery and the Department of Clinical Pathology at Linköping University Hospital.
This project is financially supported by the Swedish Foundation for Strategic Research (RMX18-0056).
Disclosures
KW: FluoLink AB (I, P). JW, EK, MH, JH, and KR declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper is not publicly available due to privacy regulations. Synthetic training data and settings for the Raman preprocessing artificial neural network that was used in this study is available in Code 1 [35].
Supplemental document
See Supplement 1 for supporting content.
References
1. J. Ferlay, M. Colombet, I. Soerjomataram, D. M. Parkin, M. Pineros, A. Znaor, and F. Bray, “Cancer statistics for the year 2020: An overview,” Int. J. Cancer 149(4), 778–789 (2021). [CrossRef]
2. W. Stummer, U. Pichlmeier, T. Meinel, O. D. Wiestler, F. Zanella, and H.-J. Reulen, “Fluorescence-guided surgery with 5-aminolevulinic acid for resection of malignant glioma: a randomised controlled multicentre phase III trial,” The Lancet Oncology 7(5), 392–401 (2006). [CrossRef]
3. N. Haj-Hosseini, J. Richter, S. Andersson-Engels, and K. Wardell, “Optical touch pointer for fluorescence guided glioblastoma resection using 5-aminolevulinic acid (in eng),” Lasers Surg. Med. 42(1), 9–14 (2010). [CrossRef]
4. J. C. O. Richter, N. Haj-Hosseini, M. Hallbeck, and K. Wardell, “Combination of hand-held probe and microscopy for fluorescence guided surgery in the brain tumor marginal zone,” Photodiagnosis Photodyn Ther. 18, 185–192 (2017). [CrossRef]
5. J. Coburger, J. Engelke, A. Scheuerle, D. R. Thal, M. Hlavac, C. R. Wirtz, and R. König, “Tumor detection with 5-aminolevulinic acid fluorescence and Gd-DTPA-enhanced intraoperative MRI at the border of contrast-enhancing lesions: a prospective study based on histopathological assessment,” Neurosurg Focus 36(2), E3 (2014). [CrossRef]
6. P. Schucht, S. Knittel, J. Slotboom, K. Seidel, M. Murek, A. Jilch, A. Raabe, and J. Beck, “5-ALA complete resections go beyond MR contrast enhancement: shift corrected volumetric analysis of the extent of resection in surgery for glioblastoma,” Acta Neurochir 156(2), 305–312 (2014); discussion 312. [CrossRef]
7. D. DePaoli, É. Lemoine, K. Ember, M. Parent, M. Prud’homme, L. Cantin, K. Petrecca, F. Leblond, and D. Côté, “Rise of Raman spectroscopy in neurosurgery: a review,” J. Biomed. Opt. 25(5), 050901 (2020). [CrossRef]
8. M. Jermyn, J. Desroches, K. Aubertin, K. St-Arnaud, W. J. Madore, E. De Montigny, M. C. Guiot, D. Trudel, B. C. Wilson, K. Petrecca, and F. Leblond, “A review of Raman spectroscopy advances with an emphasis on clinical translation challenges in oncology,” Phys. Med. Biol. 61(23), R370–R400 (2016). [CrossRef]
9. C. Kallaway, L. M. Almond, H. Barr, J. Wood, J. Hutchings, C. Kendall, and N. Stone, “Advances in the clinical application of Raman spectroscopy for cancer diagnostics,” Photodiagn. Photodyn. Ther. 10(3), 207–219 (2013). [CrossRef]
10. A. Mahadevan-Jansen and R. Richards-Kortum, “Raman spectroscopy for the detection of cancers and precancers,” J. Biomed. Opt. 1(1), 31–70 (1996). [CrossRef]
11. J. L. Yan, C. Li, A. V. Hoorn, N. R. Boonzaier, T. Matys, and S. J. Price, “A neural network approach to identify the peritumoral invasive areas in glioblastoma patients by using MR radiomics,” Sci. Rep. 10(1), 9748 (2020). [CrossRef]
12. T. C. Hollon, B. Pandian, A. R. Adapa, et al., “Near real-time intraoperative brain tumor diagnosis using stimulated Raman histology and deep neural networks,” Nat. Med. 26(1), 52–58 (2020). [CrossRef]
13. M. Ji, S. Lewis, S. Camelo-Piragua, S. H. Ramkissoon, M. Snuderl, S. Venneti, A. Fisher-Hubbard, M. Garrard, D. Fu, A. C. Wang, J. A. Heth, C. O. Maher, N. Sanai, T. D. Johnson, C. W. Freudiger, O. Sagher, X. S. Xie, and D. A. Orringer, “Detection of human brain tumor infiltration with quantitative stimulated Raman scattering microscopy,” Sci. Transl. Med. 7(309), 309ra163 (2015). [CrossRef]
14. M. Ji, D. A. Orringer, C. W. Freudiger, S. Ramkissoon, X. Liu, D. Lau, A. J. Golby, I. Norton, M. Hayashi, N. Y. R. Agar, G. S. Young, C. Spino, S. Santagata, S. Camelo-Piragua, K. L. Ligon, O. Sagher, and X. S. Xie, “Rapid, label-free detection of brain tumors with stimulated Raman scattering microscopy,” Sci. Transl. Med. 5(201), 201ra119 (2013). [CrossRef]
15. D. A. Orringer, B. Pandian, Y. S. Niknafs, T. C. Hollon, J. Boyle, S. Lewis, M. Garrard, S. L. Hervey-Jumper, H. J. L. Garton, C. O. Maher, J. A. Heth, O. Sagher, D. A. Wilkinson, M. Snuderl, S. Venneti, S. H. Ramkissoon, K. A. McFadden, A. Fisher-Hubbard, A. P. Lieberman, T. D. Johnson, X. S. Xie, J. K. Trautman, C. W. Freudiger, and S. Camelo-Piragua, “Rapid intraoperative histology of unprocessed surgical specimens via fibre-laser-based stimulated Raman scattering microscopy,” Nat. Biomed. Eng. 1(2), 0027 (2017). [CrossRef]
16. J. Desroches, M. Jermyn, M. Pinto, F. Picot, M. A. Tremblay, S. Obaid, E. Marple, K. Urmey, D. Trudel, G. Soulez, M. C. Guiot, B. C. Wilson, K. Petrecca, and F. Leblond, “A new method using Raman spectroscopy for in vivo targeted brain cancer tissue biopsy (in eng),” Sci. Rep. 8(1), 1792 (2018). [CrossRef]
17. J. Desroches, É. Lemoine, M. Pinto, E. Marple, K. Urmey, R. Diaz, M. C. Guiot, B. C. Wilson, K. Petrecca, and F. Leblond, “Development and first in-human use of a Raman spectroscopy guidance system integrated with a brain biopsy needle,” J. Biophotonics 12(3), e201800396 (2019). [CrossRef]
18. N. Lakomkin and C. G. Hadjipanayis, “The use of spectroscopy handheld tools in brain tumor surgery: current evidence and techniques,” Front. Surg. 6, 30 (2019). [CrossRef]
19. A. Savitzky and M. J. E. Golay, “Smoothing and differentiation of data by simplified least squares procedures,” Anal. Chem. 36(8), 1627–1639 (1964). [CrossRef]
20. J. Wahl, M. Sjödahl, and K. Ramser, “Single-step preprocessing of Raman spectra using convolutional neural networks,” Appl. Spectrosc. 74(4), 427–438 (2020). [CrossRef]
21. I. Anna, P. Bartosz, P. Lech, and A. Halina, “Novel strategies of Raman imaging for brain tumor research,” Oncotarget 8(49), 85290–85310 (2017). [CrossRef]
22. Y. Zhou, C.-H. Liu, Y. Sun, Y. Pu, S. Boydston-White, Y. Liu, and R. R. Alfano, “Human brain cancer studied by resonance Raman spectroscopy,” J. Biomed. Opt. 17(11), 116021 (2012). [CrossRef]
23. M. Bergholt, W. Zheng, K. Lin, Z. Huang, K. Y. Ho, K. G. Yeoh, M. Teh, and J. Yan So, “Characterizing variability in in vivo Raman spectra of different anatomical locations in the upper gastrointestinal tract toward cancer detection,” J. Biomed. Opt. 16(3), 037003 (2011). [CrossRef]
24. N. Bergner, A. Medyukhina, K. D. Geiger, M. Kirsch, G. Schackert, C. Krafft, and J. Popp, “Hyperspectral unmixing of Raman micro-images for assessment of morphological and chemical parameters in non-dried brain tumor specimens,” Anal. Bioanal. Chem. 405(27), 8719–8728 (2013). [CrossRef]
25. X. Feng, A. J. Moy, H. T. M. Nguyen, J. Zhang, M. C. Fox, K. R. Sebastian, J. S. Reichenberg, M. K. Markey, and J. W. Tunnell, “Raman active components of skin cancer,” Biomed. Opt. Express 8(6), 2835–2850 (2017). [CrossRef]
26. WHO Classification of Tumours Editorial Board, Central Nervous System Tumours, 5 ed. (WHO Classification of Tumours Series) (International Agency for Research on Cancer, 2021).
27. I. T. Jolliffe and J. Cadima, “Principal component analysis: a review and recent developments,” Phil. Trans. R. Soc. A. 374(2065), 20150202 (2016). [CrossRef]
28. L. van der Maaten and G. Hinton, “Viualizing data using t-SNE,” J. Mach. Learn. Res. 9, 2579–2605 (2008).
29. H. G. Schulze and R. F. Turner, “A two-dimensionally coincident second difference cosmic ray spike removal method for the fully automated processing of Raman spectra (in eng),” Appl. Spectrosc. 68(2), 185–191 (2014). [CrossRef]
30. H. Martens, J. P. Nielsen, and S. B. Engelsen, “Light scattering and light absorbance separated by extended multiplicative signal correction. application to near-infrared transmission analysis of powder mixtures,” Anal. Chem. 75(3), 394–404 (2003). [CrossRef]
31. P. H. C. Eilers, “A perfect smoother,” Anal. Chem. 75(14), 3631–3636 (2003). [CrossRef]
32. C. A. Lieber and A. Mahadevan-Jansen, “Automated method for subtraction of fluorescence from biological Raman spectra,” Appl. Spectrosc. 57(11), 1363–1367 (2003). [CrossRef]
33. J. Zhao, H. Lui, D. I. McLean, and H. Zeng, “Automated autofluorescence background subtraction algorithm for biomedical Raman spectroscopy,” Appl. Spectrosc. 61(11), 1225–1232 (2007). [CrossRef]
34. R. Perez-Pueyo, M. J. Soneira, and S. Ruiz-Moreno, “Morphology-based automated baseline removal for Raman spectra of artistic pigments,” Appl. Spectrosc. 64(6), 595–600 (2010). [CrossRef]
35. J. Wahl, E. Klint, M. Hallbeck, J. Hillman, K. Wardell, and K. Ramser, “Artificial neural network for preprocessing of Raman spectra,” version 3, figshare (2022), https://doi.org/10.6084/m9.figshare.21195454.
36. S. Lloyd, “Least squares quantization in PCM,” IEEE Trans. Inf. Theory 28(2), 129–137 (1982). [CrossRef]
37. R. Tibshirani, G. Walther, and T. Hastie, “Estimating the number of clusters in a data set via the gap statistic,” J. Royal Stat. Soc.: Ser. B 63(2), 411–423 (2001). [CrossRef]
38. S. Elumalai, S. Managó, and A. C. De Luca, “Raman microscopy: progress in research on cancer cell sensing,” Sensors 20(19), 5525 (2020). [CrossRef]
39. C. G. Atkins, K. Buckley, M. W. Blades, and R. F. B. Turner, “Raman spectroscopy of blood and blood components,” Appl. Spectrosc. 71(5), 767–793 (2017). [CrossRef]
40. K. Ramser, E. Malinina, and S. Candefjord, “Resonance micro-Raman investigations of the rat medial preoptic nucleus: effects of a low-iron diet on the neuroglobin content (in eng),” Appl. Spectrosc. 66(12), 1454–1460 (2012). [CrossRef]
41. J. C. Merlin, “Resonance Raman spectroscopy of carotenoids and carotenoid-containing systems,” Pure Appl. Chem. 57(5), 785–792 (1985). [CrossRef]
42. M. Kirsch, G. Schackert, R. Salzer, and C. Krafft, “Raman spectroscopic imaging for in vivo detection of cerebral brain metastases,” Anal. Bioanal. Chem. 398(4), 1707–1713 (2010). [CrossRef]
43. R. Wolthuis, M. van Aken, K. Fountas, Robinson, H. A. Bruining, and G. J. Puppels, “Determination of water concentration in brain tissue by Raman spectroscopy,” Anal. Chem. 73(16), 3915–3920 (2001). [CrossRef]
44. J. Alleon, G. Montagnac, B. Reynard, T. Brulé, M. Thoury, and P. Gueriau, “Pushing Raman spectroscopy over the edge: purported signatures of organic molecules in fossil animals are instrumental artefacts,” BioEssays 43(4), 295 (2021). [CrossRef]
45. J. Liu, M. Osadchy, L. Ashton, M. Foster, C. J. Solomon, and S. J. Gibson, “Deep convolutional neural networks for Raman spectrum recognition: a unified solution,” Analyst 142(21), 4067–4074 (2017). [CrossRef]
46. C. Beleites, U. Neugebauer, T. Bocklitz, C. Krafft, and J. Popp, “Sample size planning for classification models,” Anal. Chim. Acta 760, 25–33 (2013). [CrossRef]
