1. Introduction
Central nervous system (CNS) tumors present a considerable global health threat, with an annual report of more than 250,000 cases [1]. In recent years, there has been a remarkable proliferation of machine learning (ML) and artificial intelligence (AI) techniques across a diverse array of scientific disciplines, a development that is particularly salient in the context of the 2024 Nobel Prizes in Physics and Chemistry. These methods show great promise for transforming the field of neuro-oncology, with the potential to address critical challenges across various stages of clinical treatment. AI has been instrumental in empowering neuro-oncologists and radiologists by providing comprehensive diagnosis, offering a deeper understanding of the disease, facilitating accurate prognosis by predicting outcomes, and assisting in patient stratification to customize treatment plans to individual needs. In radiology in particular, AI applications are being used to build decision support systems using MRI images for automatic tumor detection and segmentation, identifying molecular subtypes of tumors, providing quantitative measurements, and differentiating between tumor and necrotic tissues [2,3,4]. Furthermore, neuronavigation systems have been developed, leveraging AI to identify the resected area during surgery and provide real-time diagnostic information, enhancing surgical precision and patient outcomes [5,6,7,8,9].
One of the broadest applications of AI in neuro-oncology concerns its use by neuropathologists to analyze histological images. This application encompasses the automated measurement of features, assistance in tumor classification and grading, enhancement of tumor detection, and comprehensive analysis of cellular and tissue structures through histomolecular classification [10]. Bioinformatics also represents a significant tool for brain tumor research [11,12]. By processing mutation data, individual cell information, methylation patterns, RNA sequencing, and more, AI empowers molecular pathologists by supporting biomarker identification, predicting responses to treatment, and optimizing the complex process of molecular analysis.
In summary, the main categories of data in neuro-oncology used to build AI-based decision support systems include imaging and genomic and clinical data. The analysis of both in vivo tissues and biopsy specimens of surgically removed tumors can be performed on a number of parameters, in addition to the standard histopathological examination. In particular, high speed and a large set of characteristic features are provided by optical spectroscopy methods [13]. These methods include steady-state or time-resolved fluorescence, which facilitates the estimation of the concentration of both endogenous fluorophores and fluorescent markers or photosensitizers administered prior to the procedure. Diffuse reflectance spectroscopy methods are also widely used to estimate the concentration of endogenous chromophores, such as hemoglobin, as well as the scattering properties of tissue, related to their structural integrity. Raman spectroscopy (spontaneous, stimulated, resonance, SERS, CARS) facilitates the evaluation of a wide range of molecular bonds and, as such, the molecular composition of tissues [14]. In recent years, IR spectroscopy, a method of molecular spectroscopy, has seen increased utilization in clinical settings [15].
2. General Description of Objects and Methods
2.1. Tumors of the Central Nervous System
A review of ML systems based on the spectroscopic analysis of tumor tissues in a clinical setting would be remiss without the consideration of the current classification of CNS tumors. Historically, the classification of CNS tumors was based on histological findings supported by ancillary tissue tests, such as immunohistochemical and ultrastructural analyses. However, in recent years, the importance of molecular biomarkers in providing both supportive and definitive diagnostic information has increased. Consequently, the WHO CNS5 classification system incorporates a comprehensive array of molecular alterations with significant clinicopathologic relevance, thereby facilitating the most precise classification of CNS neoplasms [16].
The new classification changes the approach to classifying gliomas, glioneuronal tumors, and neuronal tumors. These neoplasms are now categorized into the following six distinct families: diffuse adult-type gliomas; diffuse low-grade gliomas of childhood; diffuse high-grade gliomas of childhood; limited astrocytic gliomas; glioneuronal and neuronal tumors; and ependymomas. Isocitrate dehydrogenase (IDH) enzymes catalyze the oxidative decarboxylation of isocitrate and therefore play key roles in the Krebs cycle and cellular homoeostasis. Diffuse adult-type tumors encompass the majority of primary brain tumors encountered in adult neuro-oncology practice, including astrocytoma (IDH-mutant) and glioblastoma (IDH-wild type). Adult-type diffuse gliomas are stratified according to histological features (increased mitoses, necrosis and/or microvascular proliferation) and molecular alterations, including mutations in isocitrate dehydrogenase ½ (IDH1/2) genes and the whole-arm codeletion of chromosomes 1p and 19q [10]. From a diagnostic perspective, the intraoperative delineation of glial tumors poses a significant challenge, yet this can be addressed through the use of spectroscopic techniques. In the works mentioned in this review, groups of tumors such as meningiomas and CNS metastases (which are divided into metastases to the brain and spinal cord parenchyma and metastases to the cerebral membranes) are presented.
2.2. Light–Tissue Interactions Utilized in Optical Spectroscopy Methods in Neuro-Oncology
To understand the physical basis of the used optical-spectral approaches for building decision support systems in neuro-oncology, let us consider the main types of light interaction with matter at the molecular level. These techniques can be classified into the following four main categories: fluorescence spectroscopy, diffuse reflectance spectroscopy (which can be reduced, either instrumentally or algorithmically, to the study of absorption and elastic scattering), Raman (inelastic) scattering, and IR spectroscopy.
With rare exceptions, in its application to the analysis of biological media in a living organism, the geometry of measurements is realized where the source and receiver of light are on the same side of the object under study (Figure 1). The figure shows the simplified measurement schemes for the measurement of diffuse reflection signals, diffuse transmission signals, fluorescence, and Raman scattering. The diffuse reflection signal (Rd) represents the ratio of the light registered by the receiver as a result of multiple scattering and the absorption in tissues, relative to the incident light. The diffuse transmission signal (Td) represents the ratio of the light registered by the receiver that has transmitted through the sample, due to multiple scattering and the absorption in tissues, to the incident light.
This condition necessitates the consideration of the effect of multiple scattering on the resulting signal, which is a nontrivial task in general, since the information on the light path length in the sample becomes a value that is not uniquely determined. As shown in [17], when measuring diffuse reflectance spectra at a small distance between the illuminating and receiving fibers, the shape of the scattering phase function (PF) significantly impacts the signal intensity, along with the numerical aperture of the fibers. In Figure 1, the scattering phase function is conventionally depicted as an ellipse around the scattering particle.
When studying the interaction of light with biological tissues, it is necessary to take into account the complex composition of chromophores and fluorophores leading to light absorption and fluorescence excitation. These phenomena can be used for tissue characterization, but they can also serve as factors that distort the useful signal.
The main chromophores in biological tissues in the visible region of the spectra are water, lipids, and hemoproteins. A characteristic feature of the spectral dependence of the refractive index of water is a sharp decline in its imaginary part, responsible for absorption, at the short-wavelength boundary of the optical spectrum and a rise in the near-IR region. Absorption in the short-wavelength region of the spectrum is due to electronic transitions, and in the long-wavelength region due to vibrational-rotational transitions. In addition to water, biological tissues contain significant absorbers such as lipids, which contribute to their limited transparency. The cell membranes are lipid bilayer structures, which make this type of chromophores a significant contribution to the optical properties of biological tissues, in general, and the white matter of the brain and spinal cord, in particular, given that the latter consist of almost 50% myelin sheaths of nerve fibers.
One of the most functionally significant groups of biological chromophores are hemoproteins. Their structure is characterized by the presence of a protein and a non-protein (prosthetic) part. The group of hemoproteins includes hemoglobin and its derivatives, myoglobin and enzymes (the entire cytochrome system, catalase, and peroxidase). The prosthetic group of heme-containing proteins is structurally based on the porphyrin ring, the spectra of electron transitions of which are a subject of numerous works used in this review.
Porphyrins are composed of four modified pyrrole subunits interconnected at their α carbon atoms via methine bridges. Electron delocalization spreads throughout the macrocycle, which facilitates its easy excitation in the visible range of the spectrum.
The structure of the absorption spectra of various porphyrins is characterized by significant similarity. Hemoglobin is a hemoprotein binding and transporting molecular oxygen. These conformational changes affect the electronic absorption spectrum, and these differences allow for the spectral separation of the oxygenated and free forms of hemoglobin.
Protoporphyrin IX, a metabolic precursor of heme, lacks an iron ion in its coordination center, resulting in distinct differences in its electron transition spectrum compared to heme. The absorption spectrum of protoporphyrin IX contains a larger number of Q-band absorption lines. Additionally, protoporphyrin IX exhibits a relatively high quantum yield of fluorescence. However, when the heme molecule is excited, energy is transferred from the porphyrin ring to the iron ion, resulting in subsequent nonradiative deactivation. Protoporphyrin IX is frequently utilized as a fluorescent marker to detect metabolic changes occurring in cells with a deficiency of ferrochelatase, the enzyme that attaches an iron atom to protoporphyrin IX.
However, fluorescent markers are not the only means for differentiating between normal and tumor tissues. Spectroscopic measurements of autofluorescence conducted in vitro in the work [18] revealed significant differences between tumor tissue and the adjacent normal tissue of the human brain for all studied tumor types (gliomas, meningiomas, and schwannomas), with the exception of pituitary adenoma. From the results of this study, it was concluded that excitation wavelengths in the range of 410–470 nm are most suitable for detecting brain tumor tissues. The results of an ex vivo study that combined autofluorescence and diffuse reflectance for surgical navigation during the removal of brain tumors demonstrated the potential for distinguishing between normal and tumor tissues [19]. Studies have also been conducted on the combination of approaches such as ex vivo quantitative autofluorescence spectroscopy and OCT [20]. The molecular correlations of spectral differences between tumor and healthy tissues in the autofluorescence spectrum in the visible range are collagen, reduced nicotinamide adenine dinucleotide (NADH), flavin adenine dinucleotide (FAD), and porphyrins [21].
Raman spectroscopy (RS) is a vibrational type of spectroscopy that can detect the vibrations of various bonds in molecules. It is gaining increasing importance in neuro-oncology due to its ability to differentiate between tumor and normal tissues based on a combination of molecular features. The first results of a brain tumor analysis based on spontaneous Raman spectroscopy were obtained by Tashibu K [22]. In this work, the relative water concentration in normal and edematous rat brain tissues was investigated by analyzing the CH and OH groups in the high wave number region. In works [23,24], RS of brain tissue was performed for the qualitative and quantitative analysis of lipids. Higher lipid levels in normal tissues and a higher hemoglobin content but a lower lipid-to-protein ratio were reported in intracranial tumors. It was found that the ratio of phosphatidylcholine to cholesterol was increased in gliomas compared to healthy tissue [25].
2.3. Diffuse Reflectance and Fluorescence Spectroscopy Methods
The most common method used in clinical neurosurgery is fluorescence spectroscopy [13] because the analysis of fluorescence of markers injected into the organism and selectively accumulating in tumors does not require high costs for signal processing and interpretation compared to other methods presented here. Fluorescence is the emission of light by a molecule in an electronically excited state (e.g., due to light absorption). It is a form of radiative relaxation of a molecule from the excited state (Figure 2), in which an electron in the excited orbital is paired (with opposite spin) with a second electron in the ground state. Consequently, the return to the ground state is spin-resolved and occurs rapidly through the emission of a photon. The emission rate of fluorescence is typically 108 s−1, so a typical fluorescence lifetime is about 10 ns. Fluorescence lifetime (FL) spectroscopy holds significant potential for the study of biological media; however, it necessitates more cumbersome equipment and is less frequently employed in clinical settings. A second type of radiative relaxation from an excited state is phosphorescence, which is the emission of light from triplet excited states in which the electron in the excited orbital has the same spin orientation as the ground state electron. This transition is forbidden, resulting in phosphorescence lifetimes ranging from milliseconds to seconds. However, we will not encounter many decision support systems based on this process in this review.
Fluorescence spectroscopy has the highest sensitivity and interpretability, allowing it to be successfully extended to imaging systems. A prime example of this integration is the widespread use of additional modules in operating microscopes which are based on the phenomenon of fluorescence. The most widely used fluorescent marker in neurosurgery is 5-ALA-induced protoporphyrin IX, whose metabolism specificity leads to its selective accumulation in tumor cells [26,27,28]. Another fluorescent marker that has gained traction in the field of neurosurgery is chlorin e6, a porphyrin-based photosensitizer [29]. An important feature of these dyes is that they exhibit photosensitizing properties and allow for a combination of diagnostic and therapeutic functionalities [30]. Differences in the cellular accumulation of these dyes are also used for combined photodynamic therapy [31].
The selection of a fluorescent marker is contingent upon the wavelength, which must coincide with one of the maxima of the absorption spectrum of the fluorophore. The parameters of the optical filters used to construct the filter cross-system are derived from the relationship between the source wavelengths and the fluorescence maximum. The spectrum range of fluorescence signal registration is known in advance, and the level of tissue malignancy correlates well with the fluorescence intensity. This method has been extensively utilized in the field of neuro-oncology for several decades, with minimal reliance on ML algorithms. Typically, the selection of a threshold value for a specific parameter is sufficient to determine the presence of tumor tissues in the designated area. However, in some cases, even highly malignant tumor tissues may not sufficiently accumulate the fluorescent marker, and in low-grade tumors, markers may accumulate less than in one third of cases. This necessitates the expansion of the space of diagnostic features and, consequently, the use of machine learning methods.
Since the vast majority of biological tissues belong to the class of highly scattering media, the next most obvious choice of spectroscopic method is diffuse reflectance (DR) spectroscopy [32,33]. Figure 3 shows the generalized schemes for the formation of the diffuse reflectance and fluorescence signal in the backward geometry. In both cases, we register the signal due to the multiple scattering of light. The diffuse reflectance signal is also affected by the absorption of part of the light by tissue chromophores. We register this signal, as a rule, in the entire spectral range of the spectrometer, using a broadband light source. The fluorescence signal in the backward geometry is also formed due to the multiple scattering of both the exciting and the fluorescence light itself. And both types of radiation are also affected by the absorption of light by tissue chromophores. However, fluorescence spectra are usually registered upon excitation by a monochromatic light source in the spectral range corresponding to Stokes fluorescence (as is the case in most clinical applications). The tissue probing depth in the backward geometry depends on the distance (r) between the source and the receiver (Figure 3). By altering this distance, the depth from which the light arrives at the detector can be modulated. The probing depth is also affected by tissue absorption, and we can increase it for the fluorescence signal by using a light source in the biological transparency window.
The mathematical approximation of the influence of scattering and absorption in diffuse reflectance spectra is essential for analysis. To this end, the theory of light transport in tissues and its different approximations are employed. Absorption spectra can be used to determine the concentration of hemoglobin in oxygenated and reduced forms. Tumors are generally characterized by higher hemoglobin concentrations and lower oxygenated levels due to the Warburg effect. However, these parameters can be less correlated with malignancy than fluorescent marker accumulation. Using light scattering, we can characterize the number of scatterers in tissues and their size distribution; changes in these parameters are usually more stable because they reflect the integrity of tissue structures [34].
2.4. Methods of Molecular Spectroscopy
Molecular spectroscopy methods that analyze the vibrational spectra of molecules have been shown to provide information about a large number of molecular bonds. These methods can work without the introduction of any additional substances, making their use extremely promising. We can distinguish among them the following absorption methods: infrared (IR) spectroscopy, and inelastic Raman scattering (RS) methods (Figure 4). However, when analyzing biological media, the complexity of the spectra necessitates the employment of machine learning algorithms for analysis. This analysis has led to the emergence of a specialized field within analytical chemistry known as chemometrics, which integrates methodologies of signal processing, statistics, and classification. The application of machine learning methods in the domain of spectroscopic neuro-oncology represents a significant extension of this research trajectory.
Among RS techniques (Figure 4), we can distinguish several of the most frequently methods used in biological and medical applications [14]. The basic variant is spontaneous RS. As a photon interacts with a molecule, it transitions to a higher energy virtual state from which it scatters the photon, in the vast majority of cases for Raman scattering with energy loss (this is the Stokes component). A variety of wavelength sources are suitable for the excitation of spontaneous Raman scattering, but it is most common in biological applications to use a wavelength of 785 nm to avoid background fluorescence. However, there are advantages in using shorter wavelengths of around 532 nm, since it is resonant for some important biomolecules, thus increasing the sensitivity of the method in diagnostics based on their detection. In resonant RS (RRS), the intensity of the Raman spectrum components increases up to six orders of magnitude if the pump beam energy matches the real electronic transition of the molecule.
As previously stated, the photon energy decreases (Stokes scattering) during the transition from the ground state to a higher vibrational state (through the virtual level). Conversely, in the case of a molecule initially occupying a higher vibrational state, the photon energy increases (anti-Stokes scattering). As the pump beam intensity increases, the scattered Stokes component intensity also increases. Under such conditions, the molecules simultaneously interact with two electromagnetic waves—the laser pump wave and the Stokes wave—which, under certain conditions, lead to an increase in the Stokes component. This effect is called stimulated Raman scattering. Anti-Stokes and stimulated scattering principles are used in coherent anti-Stokes Raman scattering spectroscopy (CARS). CARS uses an additional beam with a scanning frequency and a pump beam with a fixed frequency. The pump beam is used both for the initial transfer of the molecule through the virtual level to a higher vibrational level of the ground state (Stokes component), which determines the scanning beam, and for re-excitation from this higher vibrational level, after which we obtain the anti-Stokes photon.
Another important extension of Raman scattering is surface-enhanced Raman scattering (SERS). Positioning the analyzed molecules near the surface of suitable nanostructured electrically conductive surfaces typically amplifies Raman signals by 104 − 106, and in some special cases, the amplification factor can be as high as 101 − 1012. The overall enhancement results from a combination of electromagnetic enhancement (based on the excitation of surface plasmons) and a non-electromagnetic contribution called chemical enhancement (due to the interaction of the analyte with the surface and thus modification of the molecular polarizability). A review on the application of Raman techniques in neurosurgery [15] underscores the necessity for close collaboration between clinicians, engineers, and data scientists to translate Raman spectroscopy technologies into clinical practice.
In [35], the authors identified the following two key characteristics of machine learning methods applied to spectroscopy data: high dimensionality and a relatively small sample size. For each sample, a feature vector is typically generated, which generally corresponds to the number of points in the spectrum or the number of selected maxima. In some cases, these vectors can have hundreds or thousands of values per sample, respectively. The process of collecting and processing samples in clinical practice is demanding in terms of cost, time, and registration conditions. The number of samples is often relatively small, particularly compared to the big data industry. This limits the applicability of some advanced cluster analysis techniques, such as clustering with deep neural networks (deep clustering), which require a larger sample size for their effective implementation.
2.5. Data Processing and Machine Learning Methods in Optical Spectroscopy
The generalized scheme of the data analysis and machine learning methods used in optical spectroscopy in neuro-oncology is presented on Figure 5.
One of the key stages in the application of machine learning and artificial intelligence methods is the data preprocessing stage, which is also true for the application of these methods to optical-spectral data. For the various types of spectral dependencies discussed in this article, both general steps are applied, including background subtraction and smoothing, and specific ones, due to the physical effects that underlie the spectra under study.
The accurate interpretation of fluorescence spectra is complicated by the distorting effects of tissue scattering and absorption, necessitating the use of attenuation correction techniques during the pre-processing of these spectra. These techniques can be divided into three broad categories as follows: empirical techniques, measurement method-based techniques, and theory-based techniques [36]. We will omit measurement-based methods here because they are hardware-based rather than spectra preprocessing-based. The most common empirical techniques are based on a ratio of fluorescence to reflectance. However, significantly different optical properties at the excitation and emission wavelengths can lead to inaccuracies, for example. For a small Stokes shift this difference can be neglected. Theory-based techniques utilize different approximations of the radiative transfer equation. The modified Beer–Lambert law, the Kubelka–Munk theory, and diffusion theory, have been employed as a basis for proposing a number of correction techniques [37,38,39]. The combined approaches also show good results [40].
For diffuse reflectance spectra, a crucial preliminary step, dictated by the inherent physical phenomenon, entails the calculation of the ratio of reflected to incident radiation. This is followed by the logarithmic reduction in this ratio to facilitate further analysis of the attenuation coefficient [41]. The obtained attenuation coefficient can be decomposed into absorption and scattering components through various methodologies. Given that most human tissues are turbid media, i.e., heterogeneous structures with spatial fluctuations of optical properties, it is necessary to take into account the effects of multiple scattering when describing the propagation of light in them, which complicates the direct application of electromagnetic theory [42]. Therefore, a radiative transfer theory (RTT) model is used, which ignores the wave properties of light, such as interference and polarization, and considers solely the transport of light energy within the medium. However, the model does implicitly account for the wave properties of light due to the tissue parameters included in the basic equation. These parameters are a generalization of the optical parameters derived by solving Maxwell’s equations for ensembles of particles. A significant simplification of the RTT model is the equation used for the steady state, in which the energy loss and gain are balanced so that the intensity at any point in the medium does not vary with time. The analytical solution of this equation is conveniently sought for the limiting cases where either absorption or scattering predominates. The second variant is also known as the diffusion approximation. The diffusion approximation is based on the diffusion equation widely used for various physical applications of RTT [43]. This model has a number of limitations as follows: first, on the isotropy of scattering in the medium, which limits its applicability near the light source; and second, on the ratio of scattering and absorption indices. In certain instances, alternative models based on empirical results may be more suitable. For example, the modified Beer–Lambert law describing the attenuation of radiation in highly scattering absorbing media and taking into account the increase in the photon path length in the tissue due to the effect of multiple scattering [44]. The model of laser radiation propagation in highly scattering media is currently often based on numerical modeling [45]. It is possible to model both the random trajectories of photons in the medium and random fields [46].
For spectra in the near-infrared range, pre-processing methods include not only the methods for correcting the influence of light scattering described above, but also methods based on taking derivatives [47]. For biological tissue, undesired systematic variations are primarily caused by multiple light scattering and differences in the effective path length, which can be observed in IR spectra as shifts in baseline (multiplicative effects) and other phenomena called nonlinearities. Scatter-correction methods can be carried out with one of the following three pre-processing concepts: Multiplicative Scatter (or, in general, Signal) Correction (MSC); Standard Normal Variate (SNV); and normalization. Derivatives have been used in analytical spectroscopy for decades. And smoothing of the spectrum is a common prior step to the derivative calculating in order to decrease the detrimental effect of this manipulation on the signal-to-noise ratio. First, the derivative has the capability to remove the baseline. The second derivative removes both the baseline and linear trend, and also resolves overlapped peaks. A review of the preprocessing strategy in the ATR-FTIR spectrum [48] also mentions scatter correction (SC), baseline correction (BC), and derivative (Drv) algorithms.
We can also distinguish such important stages of Raman spectrum preprocessing to improve its decoding properties such as smoothing, base signal correction, normalization, and in some cases, its decomposition into spectral components, peaks, or reference spectra [49,50]. It is also possible to use neural networks in the preprocessing stage of spectra [51]; a convolutional neural network (CNN) is trained on synthetic spectra to implement the three main preprocessing steps (cosmic ray removal, smoothing, and baseline subtraction) in a single operation. The generation of synthetic spectra entails the random addition of peaks, baselines, background noise, and cosmic rays. The CNN is then trained on the synthetic spectra and known peaks.
Figure 6 shows the data preprocessing methods most frequently used in various spectroscopy methods.
The subsequent phase in the processing of spectral data is the extraction of features, a process that is critical for the subsequent classification results. These methodologies can be categorized into two distinct groups as follows: theory-based techniques using the knowledge about the formation of the spectrum and methods (when “data knows best”) that automatically extract features based on the specified requirements for the representation of the obtained data.
Basic algorithms use approaches such as the fitting of individual spectral peaks (e.g., by Gaussian or Lorentzian approximation) or on the decomposition into biochemical components (by linear combination approximation of spectral standards).
Another important tool of feature extraction in machine learning for spectral data is dimensionality reduction techniques. This step may seem redundant for data obtained from fluorescence or diffuse reflectance spectra, when you have already applied your a priori theoretical information for spectral deconvolution, but it becomes necessary when analyzing molecular spectra with a large number of characteristic peaks. The RS or IR spectra of biological tissues, in general, and intracranial tumors, in particular, are characterized by a well-developed structure, a large number of characteristic peaks, and their overlap. Furthermore, numerous peaks are shared between healthy and tumor tissues, thus, they do not contribute to their differentiation. This poses a challenge for researchers in selecting a subset of characteristic peaks (features) that would provide the optimal classification of the studied samples by tissue type. Consequently, it is imperative to devise an approach for the identification of a subset of characteristics that are pertinent to the specific task at hand, which should be the focal point of attention, while disregarding the rest.
In the fields of statistics, machine learning, and information theory, dimensionality reduction is a data transformation method that reduces the number of variables by obtaining the underlying variables. This transformation can be approached in the following two ways (Figure 7): feature selection [52] and feature projection [53]. The feature selection method is designed to identify a subset of the original features, such as peaks or attributes. Feature selection can be implemented using wrapper methods, filtering algorithms, and embedding methods. In the context of machine learning, wrapper methods refer to a specific algorithmic approach wherein a feature subset selection algorithm is integrated as a “wrapper” around a pre-existing classification algorithm. This integration serves to utilize the classification result as an estimate of the efficacy with which the feature extraction process was executed. Conversely, embedding methods constitute a more generalized group of algorithms that are designed to perform feature selection as an integral component of the machine learning model building process. Filters typically require fewer complex computations, but they are not tuned for a particular type of predictive model. To evaluate the subsets of the features, filter methods turn to a third-party metric rather than the classification algorithm’s error metric to simplify computation. These metrics include correlation coefficients, the results of statistical significance tests of differences for each combination of classes/attributes, or the distance between/within classes.
Feature projection-based methods transform data from multidimensional space to low-dimensional space and are divided into linear and nonlinear methods [54]. Linear methods include principal component analysis (PCA) and linear discriminant analysis (LDA). PCA works on unlabeled data and rotates the (feature) axes so that the direction of the principal axis coincides with the direction of maximum variance of the data. In contrast, LDA, a supervised learning method, utilizes class/diagnosis data to identify the directions of maximum separation between classes. As LDA operates on labeled data, it is more adept at discerning features that accentuate the distinction between classes. Furthermore, LDA exhibits enhanced robustness in the analysis of highly correlated features, a prevalent challenge in spectroscopy. This allows LDA to be more effective in feature extraction and dimensionality reduction for classification tasks. Nonlinear methods include Multidimensional Scaling (MDS), Isometric Object Mapping (Isomap), Local Linear Embedding (LLE), Local Linear Hessian Embedding (HLLE), Spectral Embedding (Laplacian Eigenmaps, SE), and t-Distributed Stochastic Neighbor Embedding (t-SNE) [55].
In the field of biomedicine, the presence of missing data is a common occurrence, often attributable to the unavailability of the complete feature vector for all patients. This presents a significant challenge, and the process of data imputation is crucial in addressing this issue. There are various approaches to data imputation, ranging from straightforward methods (by filling the missing value with boundary values, the mean of the whole sample, or a random variable within a couple of standard deviations from the mean) to more sophisticated data-oriented techniques. One such approach is the use of clustering as a foundation for analyzing missing data.
Cluster analysis methods are very important in machine learning, including spectroscopy data [35]. Since these are unsupervised classification methods, they allow us to explore natural structures in data not labeled by tissue type or diagnosis, highlighting the most important features. In addition, they also can be used to improve the quality of experimental sampling.
Cluster analysis methods can be classified into two main categories as follows: hierarchical and sectional [35]. They both aim at minimizing distances within a cluster or maximizing distances between clusters. Hierarchical algorithms are characterized by the way links between data points and hence clusters are formed as follows: single linkage, full linkage, average linkage, and Ward’s method, shown to be the most effective for molecular optical spectroscopy data.
The k-Nearest Neighbors (KNN) algorithm is a widely used and relatively simple method. In addition to the data, this method requires additional input parameters such as the number of clusters (K) and the initial cluster center. DBSCAN (density-based spatial clustering of applications with noise) is a widely used algorithm for selecting closely spaced but sufficiently dense clusters of a complex shape.
In fuzzy clustering methods, each observation is eventually assigned to more than one cluster with a given probability. One of the widely used soft clustering algorithms is the Fuzzy C Means algorithm, and it has been shown to outperform k-means clustering in scenarios involving overlapping datasets.
The methods used by the authors of the papers discussed in this review to classify the labeled data are extremely broad. Linear discriminant analysis, previously mentioned as a supervised dimensionality reduction method, is also employed for data classification and is one of the most frequently utilized methods in the articles reviewed here. Another prevalent method is the support vector machine (SVM), which belongs to the family of linear classifiers. A linear classifier does not always yield an optimal partitioning, so the kernel function can be used to create the additional feature to transfer data to a higher dimensional space where linear classification can be performed [56]. Figure 8 shows diagrams illustrating the linear method as well as its nonlinear variant using the kernel function.
Ensemble methods integrate the predictions of the multiple baseline estimators built using basic learning algorithms to improve generalizability/robustness over a single estimator. Two very well-known examples of ensemble methods are gradient-boosted trees (GBDT) and random forests (RF). In gradient boosting, an ensemble of weak training elements is used to improve the performance of the machine learning model. The weak training elements are usually decision trees. Collectively, their results lead to an improvement in the model. The process includes the following: an initial model with data, initial predictions on the entire dataset, error calculation using predictions and actual values, assigning more weight to incorrect predictions, another model to correct the errors of the last model, new predictions with it, and so on. The final model is obtained by weighting the average of all models.
Artificial neural networks are currently not widely used for tissue classification in neuro-oncology in relation to optical-spectral methods, unlike, for example, methods related to image analysis. There are several reasons for this, and one of the most important is the insufficiently large volume of optical-spectral databases. Obviously, this limitation will be lifted over time, so we can expect a significant expansion of the applicability of neural network algorithms in optical spectroscopy in neuro-oncology. However, neural network algorithms are already finding interesting applications in related areas, such as the classification of cognitive events recorded using functional near-infrared spectroscopy (fNIRS) [57], including a combination of spectral analysis with electroencephalography [58] for developing brain–computer interface systems [59].
If we go beyond the applications of neural networks in neuroscience, we see interesting approaches to neural networks in theoretical spectroscopy. For example, in paper [60], the authors presented the application of the following three neural network architectures: the multilayer perceptron (MLP), which is one of the simplest architectures and takes vectors as input; the convolutional neural network (CNN), which takes tensors as input; and the deep tensor neural network (DTNN), developed specifically for molecular data. In this work, the authors considered molecules from an atomistic point of view, assuming that the coordinates of all atoms composing the molecule are precisely known, which is natural for theoretical spectroscopy since spectral properties can be directly calculated from approximations to the Hamiltonian of each molecule. MLP was chosen because of its architectural simplicity. In contrast, CNN is a neural network used for image recognition. Similarly to an image which is a matrix (or tensor) representation of a real-world object, a Coulomb matrix is a matrix representation of a real molecule containing spatially repeating patterns. The CNN transforms the input data through a sequence of intermediate representations by convolving the input data with one or more learned filter matrices in the convolutional layer and passing the output through a nonlinear union operation in the unifying layer. In DTNN, atoms are embedded in each molecule like words in text. The interaction between atoms and their surroundings is represented by an interaction tensor. The gradual development of methods for recording spectra of diagnostically significant biological molecules and the accumulation of corresponding databases might make it possible to move from theoretical constructions to practical applications in this field in the future.
For example, paper [61] analyzed bladder cancer grades using RS. The authors used a custom-built CNN, inspired by the ResNet architecture, and then trained it using two completely different datasets. The primary dataset consists of previously acquired Raman spectra of various human bladder tissues, with a total of 2592 spectra obtained using a portable, low-resolution Raman spectrometer with the following three categories: healthy tissue and tumors of either low or high malignancy. This raises hope for the expansion of the use of neural network algorithms in neuro-oncology with the gradual expansion of the spectra base.
3. A Review of the Application of Machine Learning Algorithms to Various Optical-Spectral Methods
3.1. Fluorescence and Diffuse Reflectance Spectroscopy Methods
As previously stated, with regard to fluorescence spectroscopy, machine learning techniques are often superfluous, given that empirical data regarding the threshold value in a specific location is sufficient for malignancy analysis based on the fluorescence of a single fluorophore. However, there are instances where fluorescence neuronavigation systems are implemented based on the analysis of multiple fluorophores, as illustrated in [62], where excitation at three wavelengths is performed. Their study focused on intracranial glial tumors, based on 10 patients and 50 samples. It shows promising results with the best performance showing 77% accuracy in healthy tissue prediction from margin tissue. The methodology of all the works analyzed in Section 3 is described in the corresponding tables. Thus, the methodology of using ML for fluorescence and DR spectra is presented in Table 1. In their previous work [63], the authors used only a 405 nm source and analyzed two protoporphyrin IX fluorescence peaks, but in this work, they expanded the feature vector by adding two light sources for fluorescence excitation. Their original feature space is composed of a raw fluorescence spectrum with 900 emission wavelengths (from 435 to 840 nm), taken at three distinct excitation wavelengths including 2700 features. The PCA shows that a minimum of five components is required to describe the data variance, which is compatible with the expert-based model (six components). However, these methods are often combined with diffuse reflectance spectroscopy, which increases the feature vector [7]. Authors of this work used theoretical explanations and mathematical models of light tissue interaction to fit the measured spectra with the main biochemical components while taking into account the scattering trend. The developed diagnostic algorithm for intra-operative tissue delineation achieved accuracies of up to 94% (specificity = 94%, sensitivity = 94%) for tumor tissue delineation. Another application of ML methods can be the combined analysis of fluorescence spectra and fluorescence lifetime, as in [64]. In the pre-processing stage, they used deconvolution of FL data with the Laguerre expansion technique, the steady-state fluorescence spectrum restoration, and attenuation correction. For dimensionality reduction, they used the results of Fisher test with a 5% significance level and adopted a classification algorithm based on the binary separation of tissue types using linear discriminant function analysis (LDFA). There was also an application of ML methods to diffuse reflectance spectra in a wide enough range, enabling the investigation of not only chromophore composition but also the optical properties of tissues [32,33]. In both papers, they used a theory-based model of light–tissue interaction to generate a feature vector. In first case [32], diffusion theory was used, and classification using random forest yielded a sensitivity and specificity for the detection of low-grade gliomas of 82.0% and 82.7%, respectively; the area under curve (AUC) was 0.91. In a separate study [33], the Monte Carlo simulation was employed to analyze the spectra, resulting in a sensitivity of 93% and a specificity of 95% for distinguishing between high-grade gliomas and normal white matter.
In sum, the reviewed works demonstrate that for the accurate analysis of fluorescence and diffuse reflectance spectra, projection methods for reducing dimensionality are unnecessary. Instead, the expert method, which involves selecting the significant features of the spectra based on theoretical models and the a priori knowledge of the biochemical composition of accessible tissues, proves to be an effective approach.
An analysis of the results of clinical studies presented in Table 1 allows us to conclude that the approaches used are approximately equally successful. The lowest results presented in Table 1 were obtained in the work that used only violet and blue light for fluorescence excitation; therefore, the obtained result can be explained by the low depth of light penetration into the tissue. The approximation of spectral data in accordance with one or another model of light transport in tissues was also omitted in this work, which indicates the need for a primary interpretation of the spectra from the point of view of the physical model as the first step in feature extraction. At the same time, the most optimal outcomes were attained through the implementation of diffuse reflectance spectroscopy, thereby invalidating the hypothesis that fluorescence analysis is adequate for tumor delineation in neurosurgical contexts.
3.2. Methods of Machine Learning in Molecular Spectroscopy
The analysis of publications demonstrating various approaches to the application of machine learning and artificial intelligence methods to brain tumor classification based on molecular spectroscopy techniques, such as Raman spectroscopy and IR spectroscopy, has shown that the use of ML methods in this application has already began in the last decade, with a notable surge in recent years. This is due to the greater availability of compact molecular spectroscopy systems applicable for the classification of brain tumors, as well as their use for the classification of brain tumors.
Since chemometrics has historically evolved to analyze this type of spectral signal, there is already something of a standard algorithm for analyzing them [65]. Raman spectra preprocessing methods include background subtraction (BS), instrument response correction (IRC), fluorescence correction (FC), and normalization. For smoothing, the most commonly used is the piecewise polynomial approximation using the Savitsky–Golay (S-G) method. The most commonly used for normalization are Standard Normal Variate (SNV), Root Mean Square (RMS), and vector normalization. The open-source code of standardized Raman spectrum processing was developed and presented in [66]. The most common variant of dimensionality reduction is PCA. In this section, it was used in 18 out of the 22 papers reviewed (Table 2). Genetic algorithm (GA) and t-SNE were used as alternatives. Also, the SVM and LDA methods, discussed in the second paragraph of this review, were the most commonly used methods used for classification. The classic combination of PCA and SVM is presented in [67], where the resonance Raman (RR) spectra of six types of human brain tissues were examined in vitro, yielding a diagnostic sensitivity of 90.9% and a specificity of 100% for meningeal tissues with two principal components.
In [68,69,70], we also see the use of Quadratic Discriminant Analysis (QDA). QDA is a variant of LDA that allows the nonlinear partitioning of data. Regularized Discriminant Analysis (RDA) is a compromise between LDA and QDA. In QDA, an individual covariance matrix is estimated for each class of observations. The disadvantage of QDA is that it cannot be used as a dimensionality reduction method. The results of [68] show the best classification performance for the combination of PCA-QDA and the successive projection algorithm (SPA)-QDA, resulting in accuracies of 96.2%, sensitivities of 85.7%, and specificities of 100% using both methods. In [69], both Raman and Fourier-transform infrared (FTIR) spectroscopy were used to differentiate normal brain tissue, glioma, or meningioma. The combination of these two types of molecular spectroscopy is very promising. Both IR and Raman spectroscopy describe the interaction of incident radiation with molecular vibrations. However, IR spectra are observed as a result of absorption, whereas CR is a scattering technique in which the incident radiation causes the polarization of the molecule. The different mechanisms underlying IR and CR spectroscopy lead to the complementarity of the two techniques. For a fluctuation to be active in IR spectroscopy, a dipole change is required, whereas for it to be active in CR spectroscopy, polarizability is required. Fluctuations of asymmetric polar bonds show up well in IR spectra, while the Raman scattering of light is particularly suitable as a probe of symmetric nonpolar groups. This study demonstrates the ability of both Raman and ATR-FTIR spectroscopy to discriminate tumor from non-tumor fresh frozen brain tissue with 94% and 97.2% of cases being correctly classified, with sensitivities of 98.8% and 100%, respectively.
Among the works discussed, we also see the use of ensemble methods, gradient boosting (GB), and random forest (RF) [71,72]. In work [71], Raman spectroscopy with a 785 nm light source was used to analyze 63 fresh tissue biopsies within a few minutes after resection (a total of 3450 spectra were collected, of which 1377 were labeled as Healthy and 2073 as Tumor). As a result of the optimization, the optimal global parameters for the classifiers of RF were as follows: numbers of trees, 150; maximum depth, 5. Meanwhile for GB, the parameters were as follows: numbers of estimators, 200; max depth, 5; learning rate, 0.1. The GB method displayed better performance in distinguishing between neoplastic and normal tissue, with an accuracy of 83%. The RF method displayed an accuracy of 80%. In article [72], formalin-fixed, paraffin-embedded (FFPE) tissue slides of malignant glioma were analyzed with a Raman microscope (532 nm). They developed APOLLO (rAman-based PathOLogy of maLignant gliOma)—a computational workflow that predicts different subtypes of glioma from spontaneous Raman spectra of FFPE tissue slides. They used a clustering algorithm to automatically separate tumor from non-tumor tissue in each sample. To this end, they trained several classifiers for the IDH wild-type versus mutant characterization and for the G-CIMP-high versus G-CIMP-low characterization.
Regarding the type of bio-objects that are analyzed by molecular spectroscopy techniques, this is much less frequent in in vivo studies due to the lower cross-sectional area of the process leading to the requirement of longer exposure times compared to fluorescence spectroscopy and diffuse reflectance techniques. The majority of research efforts are focused on the analysis of fresh tissue samples obtained in saline solutions directly from the operating room for urgent optical biopsy, or paraffin sections, allowing the a posteriori analysis of the material stored in biobanks and its comparison with new samples.
When using (SERS), the tissue sample was examined as a homogenate (frozen material from brain tumors and normal control were placed in RIPA lysis buffer, incubated on ice for 15 min and centrifuged) [73]. In another work with SERS [74], the ex vivo specimens were analyzed. The classical combination of PCA and LDA showed an ability to discriminate primary from metastatic tumors (especially for normal and low-grade lesions), with a sensitivity and specificity of 94.4% and 99.5% for normal brain tissue; 92.2% and 97.0% for low-grade glial tumors; 89.7% and 90.6% for high-grade glial tumors; 63.9% and 97.1% for meningiomas; 79.2% and 98.8% for metastases; and 88.9% and 100% for lymphoma, based on smear samples (κ = 0.87). Due to the higher sensitivity of SERS, this method may be more widely used in the future.
In summary, the results of this approach demonstrate the methodology’s potential as a leading optical biopsy instrument in the domain of neurosurgery. The methodology employs a range of sample preparation options and machine learning algorithms. The subsequent stage of development for the method should involve its translation into real-time in vivo systems.
Table 2Methodology for the use of ML techniques in IR and Raman spectroscopy for neuro-oncology.
| Article | Diagnosis | Sampling (Number of Patients p, Number of Spectra N) | Spectroscopy | Data Processing | Dimensionality Reduction | ML Algorithm |
|---|---|---|---|---|---|---|
| Zhou, 2012 [67] | Meningioma, glioblastoma, acoustic neuroma, pituitary adenoma | Ex vivo (p = 7, N = 43) | Resonance Raman spectroscopy (532 nm) | PCA | SVM (90.9%, 100%) | |
| Jermyn, 2015 [75] | Gliomas (Grade II–IV) | In vivo (p = 17, N = 161) | Raman spectroscopy (785 nm) with handheld probe | Boosted trees (89–94%, 91%) | ||
| Jermyn, 2016 [76] | Glioma | In vivo (p = 17, N = 177) | Raman spectroscopy (785 nm) | BS, IRC, AF removal, SNV normalization | Boosted trees (93%, 91%), ANN (94%, 89%) | |
| Bury, 2019 [74] | Brain tumors | Fresh ex vivo (p = 27, N = 290) | SERS (785 nm) | Polynomial BC, vector normalization | PCA | LDA (64–94%, 91–100%) |
| Kowalska, 2019 [73] | Brain tumor | Homogenates (p = 5, N = 40) | SERS | PCA | ||
| Livermore, 2019 [77] | Astrocytoma, IDH-wild-type, IDH-mutant; oligodendroglioma | Ex vivo (p = 62, N > 6200) | Raman spectroscopy | Normalization, EMSC, averaging, the second derivative | PCA | LDA (79–94%, 90–100%) |
| Morais, 2019 [68] | Meningioma brain tumor | FFPE (p = 90) | Raman microspectroscopy imaging (785 nm) | CR removal, S-G smoothing, asymmetric LS BC | PCA, SPA, GA | LDA, QDA, SVM (85.7%, 100% with QDA) |
| Zhou, 2019 [78] | Gliomas | Brain tumor specimens (p = 121, N = 510) | VRR spectroscopy (532 nm) | Smoothing, BC | PCA | SVM (100%, 71%) |
| Bury, 2020 [69] | Glioma, meningioma | Frozen brain tissues (p = 96, N = 1911) | Raman and ATR FTIR spectroscopy | S-G smoothing, MSC, BC, vector normalization | PCA, GA | LDA, QDA (98.8%, 41.7% for PCA + QDA) |
| Depciuch, 2020 [79] | Glioblastoma | FFPE (p = 8) | ATR FTIR and Raman spectroscopy | Kramers–Kronig transformation for FTIR, fitting with Lorenz function | PCA | |
| Klamminger, 2021 [80] | Glioblastoma | FFPE (p = 53, N = 471) | Raman spectroscopy (785 nm) | SVM (59% and 94% for necrosis, 54% and 94% for perifocal area, 79% and 58% for vital zone) | ||
| Riva, 2021 [71] | Astrocytoma, oligodendroglioma, glioblastoma | Ex vivo (p = 63, N = 3450) | Raman spectroscopy (785 nm) | BC, CR, outlier removal, RMS normalization | RF, GB (80% and 83% accuracy) | |
| Zhang, 2022 [81] | Meningioma | Fresh tissue samples (p = 33, N = 1180) | VRR (532 nm) | Normalization | Peak ratios, PCA | SVM with a Gaussian kernel (100% and 70% for peak ratios, 100% and 100% for PCA) |
| Romanishkin, 2022 [50] | Intracranial tumors | Ex vivo (p = 8, N = 40) | Raman spectroscopy (785 nm) | Smoothing, BC, FC (Morlet wavelets) | Feature selection + PCA | SVM (96% and 44% for PCA, 92% and 56% for feature filtration, 77% and 92% for combination) |
| Steiner, 2022 [82] | Glioblastoma (I), other types of primary brain tumors (II), metastases (III) | Fresh samples (p = 790, N = 842) | ATR FTIR | AC, a quadratic BC, vector normalization | PCA | LDA (86% and 100% (I), 92% and 100% (II), 92% and 100% (III)) |
| Wahl, 2022 [49] | Low- or high-grade tumor located in the cerebrum or pituitary | Ex vivo (p = 34, N > 900) | Raman spectroscopy (532 nm) | EMSC, smoothing, BC, normalization by L2-norm, CNN | PCA, tSNE | k-means cluster analysis |
| Quesnel, 2023 [83] | Gliomas | Serum, FFPE (p = 30, N = 750) | Raman spectroscopy (785 nm) | Min-max normalization | PCA | Accuracy for SVM (63.3%), LDA (60%), cos k-NN (53.3%), bilayed NN (36.7%), narrow NN (30%) |
| Tołpa, 2023 [84] | Glioblastoma | FFPE (p = 9) | ATR FTIR | Vector normalization, BC, 2nd derivative, smoothing | Curve fitting, PCA | RF (89% and 78%), k-NN (100% and 100%), SVM (100% and 100%), XGBoost (78% and 89%) |
| Ember, 2024 [85] | Glioblastoma, brain metastases, meningioma | In vivo (p = 67, N ~ 976) | Raman spectroscopy (785 nm) with handheld probe | BC, CR correction, IRC, averaging, BS, S-G smoothing, SNV normalization, Gaussian fitting | Feature filtering with RF | SVM (91% and 91% for glioblastoma, 98% and 96% for metastases, 96% and 96% for meningioma) |
| Galli, 2024 [86] | Meningioma | Cryosections (p = 47, N = 35,643) | Infrared transmission spectroscopy | AC, binning, BC | PCA, binning, feature selection | LDA (AUC 0.89), fully connected neural network (AUC 0.91) |
| Lita, 2024 [72] | Malignant glioma | FFPE (p = 46, N = 300 506) | Raman microscope (532 nm) | BC with airPLS algorithm, L2-normalization | PCA, tSNE | Clustering (k-means, DBSCAN) to find subregions, SVM, RF |
| Liu, 2024 [70] | Glioma, pituitary adenoma, meningioma, acoustic neuroma | Fresh tissue samples (p = 40, N = 120) | Raman spectroscopy | Removing of CR, BC, reducing of external noises | PCA | Accuracy for LDA (95.07%), QDA (99.47%) |
| Tołpa, 2024 [87] | Glioblastoma, meningioma | (p = 33) | FT-Raman spectroscopy (1064 nm) | PCA | SVM (75% and 95%), RF (75% and 90%) |
The analysis of the results shown in Table 2 allows us to conclude that the combination of the standard algorithm for spectra processing, including background correction, smoothing, and normalization, followed by dimensionality reduction is quite effective. We see the least successful classification for the cases where these steps were not mentioned in the paper, i.e., they may not have been used. Thus, we can consider them necessary for success. We can also infer from Table 2 the superiority of analyzing fresh samples over paraffin-embedded samples. However, there are exceptions to this observation [68] due to either the object of study (meningiomas) or the method (Raman microspectroscopy imaging). The use of a handheld probe for Raman spectroscopy [76,85] has also shown very encouraging results.
A number of works using ML methods are dedicated to the analysis of microscopic images obtained with spectral resolution. For example, a series of works [88,89,90,91] presents the results of the development and application of a Stimulated Raman Histology (SRH) technique for the analysis of images of ex vivo brain tissue, structurally similar to the hematoxylin and eosin (H&E) staining, but obtained using stimulated Raman scattering with a dual-wavelength fiber laser with a fixed pump wavelength of 790 nm and Stokes wavelengths in the range from 1015 nm to 1050 nm. For the SRH study, samples were sequentially scanned at two Raman shifts as follows: 2850 cm−1 and 2950 cm−1. Lipid-rich brain regions show a high signal at 2845 cm−1 and cellular regions show high intensity at 2930 cm−1 and a high S2930/S2845 ratio, indicating high protein and DNA content. The resulting maps of the intensity ratios at these two peaks were converted to pseudocolor images using a H&E staining scheme. CNNs were used to classify the resulting images. This is a very promising method for using the whole array of H&E histological sections, accumulated so far as a training sample. However, these works are somewhat removed from the subject of this review, since image analysis, including spectrally resolved images, is a separate actively developing field that requires the careful consideration of algorithms and approaches.
3.3. Combined Techniques
In recent years, there has been a notable increase in the number of papers devoted to the combined analysis of optical spectra of various kinds (Table 3). This increased interest has necessitated the combination of different types of spectrometers and the reduction in features obtained from spectra of different origins to a common form. For such approaches, the stage of feature extraction from the data is of particular importance.
In work [92], Raman spectroscopy and imaging were combined with atomic force microscopy (AFM) topography and mechanical indentation by AFM for the analysis of medulloblastoma (grade IV WHO) (n = 5) and the tissue from the negative margins was used as normal controls. The AFM modality provided new information about the conformational correlations of molecules, which we can detect with Raman spectroscopy. A continuation of this study may contribute to a better understanding of tumorigenesis. The authors of this paper confirm our conclusion from Table 2. They used fresh samples because they have shown that the lipid phenotype of cancer is disturbed in FFPP samples due to tissue processing (de-waxing, washing, and clearing with solvents, e.g., xylene).
There are more clinically oriented works in the literature [54,93,94]. In all cases, the combination of fluorescence, Raman, and diffuse reflectance spectroscopies was employed, which appears to be a promising approach. In these works, freshly excised tissue specimens in saline were analyzed ex vivo. In [54], cluster analysis of spontaneous Raman spectra of intracranial tumors was carried out. The results showed natural clusters, grouping meningeal and glial tumors. The results of [94] demonstrated brain, glioma, and necrosis detection with an accuracy of 94.7%, 84.2%, and 89.5% using linear discriminant analysis and a naive Bayesian classifier. The work of [93] showed good results with discriminating tumor/dysplastic tissues against control tissue with 91%/86% sensitivity and 100%/100% specificity, respectively, whereas tumor from dysplastic tissues was discriminated with 89% sensitivity and 86% specificity.
4. Conclusions
In this review, we have considered the methodology of applying ML algorithms in neurosurgery to fluorescence spectroscopy, both steady-state and time-resolved; diffuse reflectance; Raman spectroscopy in various technical implementations; and IR. Fluorescence and diffuse reflectance spectroscopy are the most widely used tools in neurosurgery and were used in vivo in all cases reviewed by us for intraoperative navigation during the removal of diffuse glial tumors. For these methods, the optimal way to extract features is by using methods based on theoretical models of the light–tissue interaction. Molecular spectroscopy methods such as Raman and IR spectroscopy produce much more loaded spectra, from which the extraction of significant features can be a problem. Although one of the basic chemometric approaches to such spectra is the use of the principal component analysis for dimensionality reduction, in this review, we have considered other options for using both feature projection and feature filtering techniques. Many works devoted to molecular spectroscopy demonstrate the importance of the spectral preprocessing and feature extraction stages for the effective sample classification. In recent years, there has been a surge of interest in leveraging machine learning methodologies in the field of spectroscopy, despite the challenges associated with acquiring substantial datasets for training purposes. Technological advancements, such as the integration of diverse spectroscopic modalities or the transition to spectral mapping, are enabling the continuous expansion of the scope of optical biopsy methods in the context of neuro-oncology.
Conceptualization, T.S., I.R., G.P. and S.G.; methodology, T.S.; investigation, T.S., I.R., A.O.; writing—original draft preparation, T.S.; writing—review and editing, T.S. and I.R.; visualization, T.S. and I.R.; supervision, V.L.; project administration, I.P.; funding acquisition, I.P. All authors have read and agreed to the published version of the manuscript.
The raw data supporting the conclusions of this article can be made available by the authors without undue reservation.
The authors declare no conflicts of interest.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Enlarge this image.Figure 1. The most common types of light–tissue interactions used in optical spectroscopy analysis in neurosurgery (green arrows for incident and elastically scattered light, red arrows for fluorescence, orange arrows for inelastically scattered light).
Enlarge this image.Figure 2. Energy-level diagram of different types of light interactions with molecules used in absorption, fluorescence, and elastic and inelastic scattering spectroscopy.
Enlarge this image.Figure 3. Diffuse reflectance (left) and fluorescence (right) spectroscopy registration schemes (r—distance between illuminating and receiving fibers; NAS—numerical aperture of source; NAR—numerical aperture of receiving fiber). The crescent-shaped region between the illumination and receiving fibers marks the region of the most probable path of the detected photons (the dotted lines represent an example of such a path).
Enlarge this image.Figure 5. A general diagram of the most commonly used steps of spectral data processing and classification methods in neuro-oncology. Methods used in the reviewed works are highlighted in bold.
Enlarge this image.Figure 6. A general diagram of the most commonly used spectral data preprocessing methods.
Enlarge this image.Figure 7. Approaches to dimensionality reduction in optical spectroscopy. The blue and orange circles belong to two different classes.
Enlarge this image.Figure 8. One-dimensional illustration of the support vector machine principle (left) and transition to higher dimensional space with kernel function (right). The blue and orange circles belong to two different classes.
Methodology for the use of ML techniques in fluorescence and diffuse reflectance spectroscopy for neuro-oncology.
| Article | Diagnosis | Sampling (Number of Patients p, Number of Spectra N) | Spectroscopy | Data Processing | Dimensionality Reduction | ML Algorithm (Sensitivity, Specificity) |
|---|---|---|---|---|---|---|
| Butte, 2011 [ | Glioma | In vivo (p = 42, N = 186) | FL spectroscopy (337 nm) | LE deconvolution, attenuation correction | Feature filtering with F-test | LDFA (100%, 98% for LGG, 47%, 95% for HGG) |
| Valdés, 2011 [ | Glioma | In vivo (p > 70) | Fluorescence (405 nm ex) and DR (450 to 720 nm) | Light transport modeling | Fitting to biochemical components spectra | SVM (94%, 94%) |
| Leclerc, 2020 [ | Glioma | In vivo (p = 10, N = 50) | Fluorescence spectroscopy (385 nm, 405 nm, and 420 nm ex) | Normalization by global energy, feature extraction | PCA, t-SNE | K-Means, the Gaussian mixture model (GMM) (77% accuracy) |
| Skyrman, 2022 [ | Glioma | In vivo, ex vivo (p = 30) | DR (450 to 1100 nm and 900 to 1600 nm) | Fitting with a diffusion theory model | RF (82%, 82.7%) | |
| Li, 2023 [ | Glioma | Ex vivo (p = 19, N = 55) | DR | Fitting with Monte Carlo simulation-based model | GBDT (93%, 95%) |
Methodology for the use of ML techniques for combination of different modalities of spectroscopic analysis.
| Article | Diagnosis | Sampling (Number of Patients p, Number of Spectra N) | Spectroscopy | Data Processing | Dimensionality Reduction | ML Algorithm |
|---|---|---|---|---|---|---|
| Abramczyk, 2018 [ | Medulloblastoma | Frozen sections (p = 5) | Raman spectroscopy and imaging combined with AFM | PCA | PLSDA (98.5%, 96%) | |
| Baria, 2023 [ | Brain tumor and cortical dysplasia | Fresh samples in saline (p = 51, N = 163) | Fluorescence, Raman, and diffuse reflectance spectroscopies | Smoothing, BC, IRC, and normalization (fluorescence on maximum, Raman on intensity) | PCA | LDA-based algorithm (91%, 100%) |
| Ospanov, 2023 [ | Intracranial tumors | Ex vivo (p = 54, N = 93) | Raman spectroscopy (785 nm), fluorescence (632.8 nm), and diffuse reflectance (400–800 nm) | Smoothing, BC, FC, normalization, ratiometry for fluorescence, fitting with diffuse theory-based model for diffuse reflectance | PCA, LDA, Isomap, SE, LLE, t-SNE | Clustering with k-means (72% and 82%), agglomerative hierarchical clustering (70% and 78%), DBSCAN, fuzzy C-means |
| Romanishkin, 2023 [ | Intracranial tumors | Ex vivo (p = 150, N = 195) | Raman spectroscopy (785 nm), fluorescence (405 nm and 632.8 nm), and diffuse reflectance (400–800 nm) | Smoothing, BC, FC, normalization, ratiometry for fluorescence, fitting with diffuse theory-based model for diffuse reflectance | Feature selection + PCA/LDA/SE/t-SNE | SVM, logistic regression, naïve Bayes (81.25% and 100%) |
References
1. Khalighi, S.; Reddy, K.; Midya, A.; Pandav, K.B.; Madabhushi, A.; Abedalthagafi, M. Artificial Intelligence in Neuro-Oncology: Advances and Challenges in Brain Tumor Diagnosis, Prognosis, and Precision Treatment. NPJ Precis. Oncol.; 2024; 8, 80. [DOI: https://dx.doi.org/10.1038/s41698-024-00575-0] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38553633]
2. Weinberg, B.; Ramesh, K.; Gurbani, S.; Schreibmann, E.; Kleinberg, L.; Shu, H.-K.; Shim, H. NIMG-23. Brain tumor reporting and data system (bt-rads) and quantitative tools to guide its implementation. Neuro-Oncol.; 2019; 21, vi166. [DOI: https://dx.doi.org/10.1093/neuonc/noz175.695]
3. Monsour, R. Neuroimaging in the Era of Artificial Intelligence: Current Applications. Fed. Pract.; 2022; 39, pp. S14-S20. [DOI: https://dx.doi.org/10.12788/fp.0231]
4. Philip, A.; Samuel, B.; Bhatia, S.; Khalifa, S.; El-Seedi, H. Artificial Intelligence and Precision Medicine: A New Frontier for the Treatment of Brain Tumors. Life; 2022; 13, 24. [DOI: https://dx.doi.org/10.3390/life13010024] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36675973]
5. Schonfeld, E.; Mordekai, N.; Berg, A.; Johnstone, T.; Shah, A.; Shah, V.; Haider, G.; Marianayagam, N.J.; Veeravagu, A. Machine Learning in Neurosurgery: Toward Complex Inputs, Actionable Predictions, and Generalizable Translations. Cureus; 2024; 16, e51963. [DOI: https://dx.doi.org/10.7759/cureus.51963] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38333513]
6. Puustinen, S.; Vrzáková, H.; Hyttinen, J.; Rauramaa, T.; Fält, P.; Hauta-Kasari, M.; Bednarik, R.; Koivisto, T.; Rantala, S.; Von Und Zu Fraunberg, M. et al. Hyperspectral Imaging in Brain Tumor Surgery—Evidence of Machine Learning-Based Performance. World Neurosurg.; 2023; 175, pp. e614-e635. [DOI: https://dx.doi.org/10.1016/j.wneu.2023.03.149]
7. Valdés, P.A.; Kim, A.; Leblond, F.; Conde, O.M.; Harris, B.T.; Paulsen, K.D.; Wilson, B.C.; Roberts, D.W. Combined Fluorescence and Reflectance Spectroscopy for in Vivo Quantification of Cancer Biomarkers in Low- and High-Grade Glioma Surgery. J. Biomed. Opt.; 2011; 16, 116007. [DOI: https://dx.doi.org/10.1117/1.3646916] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/22112112]
8. Goryaynov, S.A.; Okhlopkov, V.A.; Golbin, D.A.; Chernyshov, K.A.; Svistov, D.V.; Martynov, B.V.; Kim, A.V.; Byvaltsev, V.A.; Pavlova, G.V.; Batalov, A. et al. Fluorescence Diagnosis in Neurooncology: Retrospective Analysis of 653 Cases. Front. Oncol.; 2019; 9, 830. [DOI: https://dx.doi.org/10.3389/fonc.2019.00830] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31552168]
9. Rynda, A.Y.; Olyushin, V.E.; Rostovtsev, D.M.; Zabrodskaya, Y.M.; Papayan, G.V. Results of Microsurgical Resection of Glioblastomas under Endoscopic and Fluorescent Control. Biomed. Photonics; 2024; 13, pp. 20-30. [DOI: https://dx.doi.org/10.24931/2413-9432-2024-13-3-20-30]
10. Redlich, J.-P.; Feuerhake, F.; Weis, J.; Schaadt, N.S.; Teuber-Hanselmann, S.; Buck, C.; Luttmann, S.; Eberle, A.; Nikolin, S.; Appenzeller, A. et al. Applications of Artificial Intelligence in the Analysis of Histopathology Images of Gliomas: A Review. NPJ Imaging; 2024; 2, 16. [DOI: https://dx.doi.org/10.1038/s44303-024-00020-8]
11. Lah Turnšek, T.; Jiao, X.; Novak, M.; Jammula, S.; Cicero, G.; Ashton, A.W.; Joyce, D.; Pestell, R.G. An Update on Glioblastoma Biology, Genetics, and Current Therapies: Novel Inhibitors of the G Protein-Coupled Receptor CCR5. Ijms; 2021; 22, 4464. [DOI: https://dx.doi.org/10.3390/ijms22094464] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33923334]
12. Zeller, S.L.; Spirollari, E.; Chandy, A.M.; Hanft, S.J.; Gandhi, C.D.; Jhanwar-Uniyal, M. Understanding the Genomic Landscape of Glioblastoma: Opportunities for Targeted Therapies. Anticancer. Res.; 2024; 44, pp. 4677-4690. [DOI: https://dx.doi.org/10.21873/anticanres.17295]
13. Valdés, P.A.; Roberts, D.W.; Lu, F.-K.; Golby, A. Optical Technologies for Intraoperative Neurosurgical Guidance. FOC; 2016; 40, E8. [DOI: https://dx.doi.org/10.3171/2015.12.FOCUS15550] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26926066]
14. Vlasov, A.V.; Maliar, N.L.; Bazhenov, S.V.; Nikelshparg, E.I.; Brazhe, N.A.; Vlasova, A.D.; Osipov, S.D.; Sudarev, V.V.; Ryzhykau, Y.L.; Bogorodskiy, A.O. et al. Raman Scattering: From Structural Biology to Medical Applications. Crystals; 2020; 10, 38. [DOI: https://dx.doi.org/10.3390/cryst10010038]
15. DePaoli, D.; Lemoine, É.; Ember, K.; Parent, M.; Prud’homme, M.; Cantin, L.; Petrecca, K.; Leblond, F.; Côté, D.C. Rise of Raman Spectroscopy in Neurosurgery: A Review. J. Biomed. Opt.; 2020; 25, 1. [DOI: https://dx.doi.org/10.1117/1.JBO.25.5.050901] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32358930]
16. Louis, D.N.; Perry, A.; Wesseling, P.; Brat, D.J.; Cree, I.A.; Figarella-Branger, D.; Hawkins, C.; Ng, H.K.; Pfister, S.M.; Reifenberger, G. et al. The 2021 WHO Classification of Tumors of the Central Nervous System: A Summary. Neuro-Oncol.; 2021; 23, pp. 1231-1251. [DOI: https://dx.doi.org/10.1093/neuonc/noab106] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34185076]
17. Mourant, J.R.; Boyer, J.; Hielscher, A.H.; Bigio, I.J. Influence of the Scattering Phase Function on Light Transport Measurements in Turbid Media Performed with Small Source–Detector Separations. Opt. Lett.; 1996; 21, 546. [DOI: https://dx.doi.org/10.1364/OL.21.000546] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19865467]
18. Saraswathy, A.; Jayasree, R.S.; Baiju, K.V.; Gupta, A.K.; Pillai, V.P.M. Optimum Wavelength for the Differentiation of Brain Tumor Tissue Using Autofluorescence Spectroscopy. Photomed. Laser Surg.; 2009; 27, pp. 425-433. [DOI: https://dx.doi.org/10.1089/pho.2008.2316]
19. Lu, H.; Grygoryev, K.; Bermingham, N.; Jansen, M.; O’Sullivan, M.; Nunan, G.; Buckley, K.; Manley, K.; Burke, R.; Andersson-Engels, S. Combined Autofluorescence and Diffuse Reflectance for Brain Tumour Surgical Guidance: Initial Ex Vivo Study Results. Biomed. Opt. Express; 2021; 12, 2432. [DOI: https://dx.doi.org/10.1364/BOE.420292]
20. Zhu, M.; Chang, W.; Jing, L.; Fan, Y.; Liang, P.; Zhang, X.; Wang, G.; Liao, H. Dual-Modality Optical Diagnosis for Precise in Vivo Identification of Tumors in Neurosurgery. Theranostics; 2019; 9, pp. 2827-2842. [DOI: https://dx.doi.org/10.7150/thno.33823] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31244926]
21. Zuluaga, A.F.; Utzinger, U.; Durkin, A.; Fuchs, H.; Gillenwater, A.; Jacob, R.; Kemp, B.; Fan, J.; Richards-Kortum, R. Fluorescence Excitation Emission Matrices of Human Tissue: A System for in Vivo Measurement and Method of Data Analysis. Appl. Spectrosc.; 1999; 53, pp. 302-311. [DOI: https://dx.doi.org/10.1366/0003702991946695]
22. Tashibu, K. Analysis of Water Content in Rat Brain Using Raman Spectroscopy. No Shinkei = Brain Nerve; 1990; 42, pp. 999-1004. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/2288780]
23. Krafft, C.; Neudert, L.; Simat, T.; Salzer, R. Near Infrared Raman Spectra of Human Brain Lipids. Spectrochim. Acta Part A Mol. Biomol. Spectrosc.; 2005; 61, pp. 1529-1535. [DOI: https://dx.doi.org/10.1016/j.saa.2004.11.017] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15820887]
24. Krafft, C.; Sobottka, S.B.; Schackert, G.; Salzer, R. Near Infrared Raman Spectroscopic Mapping of Native Brain Tissue and Intracranial Tumors. Analyst; 2005; 130, pp. 1070-1077. [DOI: https://dx.doi.org/10.1039/b419232j] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15965532]
25. Köhler, M.; Machill, S.; Salzer, R.; Krafft, C. Characterization of Lipid Extracts from Brain Tissue and Tumors Using Raman Spectroscopy and Mass Spectrometry. Anal. Bioanal. Chem.; 2009; 393, pp. 1513-1520. [DOI: https://dx.doi.org/10.1007/s00216-008-2592-9]
26. Traylor, J.I.; Pernik, M.N.; Sternisha, A.C.; McBrayer, S.K.; Abdullah, K.G. Molecular and Metabolic Mechanisms Underlying Selective 5-Aminolevulinic Acid-Induced Fluorescence in Gliomas. Cancers; 2021; 13, 580. [DOI: https://dx.doi.org/10.3390/cancers13030580] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33540759]
27. Ivanova-Radkevich, V.I.; Kuznetsova, O.M.; Filonenko, E.V. The Role of Membrane Transport Proteins in 5-ALA-Induced Accumulation of Protoporphyrin iX in Tumor Cells. Biomed. Photonics; 2024; 13, pp. 43-48. [DOI: https://dx.doi.org/10.24931/2413-9432-2024-13-2-43-48]
28. Nasir-Moin, M.; Wadiura, L.I.; Sacalean, V.; Juros, D.; Movahed-Ezazi, M.; Lock, E.K.; Smith, A.; Lee, M.; Weiss, H.; Müther, M. et al. Localization of Protoporphyrin IX during Glioma-Resection Surgery via Paired Stimulated Raman Histology and Fluorescence Microscopy. Nat. Biomed. Eng.; 2024; 8, pp. 672-688. [DOI: https://dx.doi.org/10.1038/s41551-024-01217-3]
29. Matsumura, H.; Akimoto, J.; Haraoka, J.; Aizawa, K. Uptake and Retention of the Photosensitizer Mono-l-Asparthyl Chlorine E6 in Experimental Malignant Glioma. Lasers Med. Sci.; 2008; 23, pp. 237-245. [DOI: https://dx.doi.org/10.1007/s10103-007-0469-3]
30. Olyushin, V.E.; Kukanov, K.K.; Nechaeva, A.S.; Sklyar, S.S.; Vershinin, A.E.; Dikonenko, M.V.; Golikova, A.S.; Mansurov, A.S.; Safarov, B.I.; Rynda, A.Y. et al. Photodynamic Therapy in Neurooncology. Biomed. Photonics; 2023; 12, pp. 25-35. [DOI: https://dx.doi.org/10.24931/2413-9432-2023-12-3-25-35]
31. Kozlikina, E.I.; Trifonov, I.S.; Sinkin, M.V.; Krylov, V.V.; Loschenov, V.B. The Combined Use of 5-ALA and Chlorin E6 Photosensitizers for Fluorescence-Guided Resection and Photodynamic Therapy under Neurophysiological Control for Recurrent Glioblastoma in the Functional Motor Area after Ineffective Use of 5-ALA: Preliminary Results. Bioengineering; 2022; 9, 104. [DOI: https://dx.doi.org/10.3390/bioengineering9030104]
32. Skyrman, S.; Burström, G.; Lai, M.; Manni, F.; Hendriks, B.; Frostell, A.; Edström, E.; Persson, O.; Elmi-Terander, A. Diffuse Reflectance Spectroscopy Sensor to Differentiate between Glial Tumor and Healthy Brain Tissue: A Proof-of-Concept Study. Biomed. Opt. Express; 2022; 13, 6470. [DOI: https://dx.doi.org/10.1364/BOE.474344]
33. Li, K.; Wu, Q.; Feng, S.; Zhao, H.; Jin, W.; Qiu, H.; Gu, Y.; Chen, D. In Situ Detection of Human Glioma Based on Tissue Optical Properties Using Diffuse Reflectance Spectroscopy. J. Biophotonics; 2023; 16, e202300195. [DOI: https://dx.doi.org/10.1002/jbio.202300195] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37589177]
34. Savelieva, T.A.; Kalyagina, N.A.; Kholodtsova, M.N.; Loschenov, V.B.; Goryainov, S.A.; Potapov, A.A. Numerical Modelling and in Vivo Analysis of Fluorescent and Laser Light Backscattered from Glial Brain Tumors; Wax, A.P.; Backman, V. SPIE: San Francisco, CA, USA, 2012; 82300L.
35. Crase, S.; Hall, B.N.; Thennadil, S. Cluster Analysis for IR and NIR Spectroscopy: Current Practices to Future Perspectives. Comput. Mater. Contin.; 2021; 69, pp. 1945-1965. [DOI: https://dx.doi.org/10.32604/cmc.2021.018517]
36. Bradley, R.S.; Thorniley, M.S. A Review of Attenuation Correction Techniques for Tissue Fluorescence. J. R. Soc. Interface.; 2006; 3, pp. 1-13. [DOI: https://dx.doi.org/10.1098/rsif.2005.0066] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/16849213]
37. Richards-Kortum, R.; Rava, R.P.; Fitzmaurice, M.; Tong, L.L.; Ratliff, N.B.; Kramer, J.R.; Feld, M.S. A One-Layer Model of Laser-Induced Fluorescence for Diagnosis of Disease in Human Tissue: Applications to Atherosclerosis. IEEE Trans. Biomed. Eng.; 1989; 36, pp. 1222-1232. [DOI: https://dx.doi.org/10.1109/10.42117] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/2606498]
38. Durkin, A.J.; Richards-Kortum, R. Comparison of Methods to Determine Chromophore Concentrations from the Fluorescence Spectra of Turbid Samples. Proceedings of the Conference on Lasers and Electro-Optics (1995); Baltimore, MY, USA, 21–26 May 1995; Paper CFE5 Optica Publishing Group: Washington, DC, USA, 1995; CFE5.
39. Zhadin, N.N. Correction of the Internal Absorption Effect in Fluorescence Emission and Excitation Spectra from Absorbing and Highly Scattering Media: Theory and Experiment. J. Biomed. Opt.; 1998; 3, 171. [DOI: https://dx.doi.org/10.1117/1.429874] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23015054]
40. Savelieva, T.A.; Kuryanova, M.N.; Akhlyustina, E.V.; Linkov, K.G.; Meerovich, G.A.; Loschenov, V.B. Attenuation Correction Technique for Fluorescence Analysis of Biological Tissues with Significantly Different Optical Properties. Front. Optoelectron.; 2020; 13, pp. 360-370. [DOI: https://dx.doi.org/10.1007/s12200-020-1094-z]
41. Stratonnikov, A.A.; Loschenov, V.B. Evaluation of Blood Oxygen Saturation in Vivo from Diffuse Reflectance Spectra. J. Biomed. Opt.; 2001; 6, 457. [DOI: https://dx.doi.org/10.1117/1.1411979]
42. Ishimaru, A. Wave Propagation and Scattering in Random Media; IEEE/OUP Series on Electromagnetic Wave Theory; IEEE Press: New York, NY, USA, 1997; ISBN 978-0-470-54704-5
43. Martelli, F.; Del Bianco, S.; Ismaelli, A.; Zaccanti, G. Light Propagation Through Biological Tissue and Other Diffusive Media: Theory, Solutions, and Software; SPIE: New York, NY, USA, 2009; ISBN 978-0-8194-8183-2
44. Delpy, D.T.; Cope, M.; Zee, P.V.D.; Arridge, S.; Wray, S.; Wyatt, J. Estimation of Optical Pathlength Through Tissue from Direct Time of Flight Measurement. Phys. Med. Biol.; 1988; 33, pp. 1433-1442. [DOI: https://dx.doi.org/10.1088/0031-9155/33/12/008]
45. Bahl, A.; Segaud, S.; Xie, Y.; Shapey, J.; Bergholt, M.S.; Vercauteren, T. A Comparative Study of Analytical Models of Diffuse Reflectance in Homogeneous Biological Tissues: Gelatin-based Phantoms and Monte Carlo Experiments. J. Biophotonics; 2024; 17, e202300536. [DOI: https://dx.doi.org/10.1002/jbio.202300536]
46. Kandidov, V.P.; Militsin, V.O.; Bykov, A.V.; Priezzhev, A.V. Application of Corpuscular and Wave Monte-Carlo Methods in Optics of Dispersive Media. Quantum Electron.; 2006; 36, pp. 1003-1008. [DOI: https://dx.doi.org/10.1070/QE2006v036n11ABEH013436]
47. Rinnan, Å.; Berg, F.V.D.; Engelsen, S.B. Review of the Most Common Pre-Processing Techniques for near-Infrared Spectra. TrAC Trends Anal. Chem.; 2009; 28, pp. 1201-1222. [DOI: https://dx.doi.org/10.1016/j.trac.2009.07.007]
48. Lee, L.C.; Liong, C.-Y.; Jemain, A.A. A Contemporary Review on Data Preprocessing (DP) Practice Strategy in ATR-FTIR Spectrum. Chemom. Intell. Lab. Syst.; 2017; 163, pp. 64-75. [DOI: https://dx.doi.org/10.1016/j.chemolab.2017.02.008]
49. Wahl, J.; Klint, E.; Hallbeck, M.; Hillman, J.; Wårdell, K.; Ramser, K. Impact of Preprocessing Methods on the Raman Spectra of Brain Tissue. Biomed. Opt. Express; 2022; 13, 6763. [DOI: https://dx.doi.org/10.1364/BOE.476507]
50. Romanishkin, I.; Savelieva, T.; Kosyrkova, A.; Okhlopkov, V.; Shugai, S.; Orlov, A.; Kravchuk, A.; Goryaynov, S.; Golbin, D.; Pavlova, G. et al. Differentiation of Glioblastoma Tissues Using Spontaneous Raman Scattering with Dimensionality Reduction and Data Classification. Front. Oncol.; 2022; 12, 944210. [DOI: https://dx.doi.org/10.3389/fonc.2022.944210] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36185245]
51. Wahl, J.; Sjödahl, M.; Ramser, K. Single-Step Preprocessing of Raman Spectra Using Convolutional Neural Networks. Appl. Spectrosc.; 2020; 74, pp. 427-438. [DOI: https://dx.doi.org/10.1177/0003702819888949] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31961223]
52. Crase, S.; Hall, B.N.; Thennadil, S. Feature Selection for Cluster Analysis in Spectroscopy. Comput. Mater. Contin.; 2022; 71, pp. 2435-2458. [DOI: https://dx.doi.org/10.32604/cmc.2022.022414]
53. Zhao, H.; Lai, Z.; Leung, H.; Zhang, X. Feature Learning and Understanding: Algorithms and Applications; Information Fusion and Data Science Springer International Publishing: Cham, Switzerland, 2020; ISBN 978-3-030-40793-3
54. Ospanov, A.; Romanishkin, I.; Savelieva, T.; Kosyrkova, A.; Shugai, S.; Goryaynov, S.; Pavlova, G.; Pronin, I.; Loschenov, V. Optical Differentiation of Brain Tumors Based on Raman Spectroscopy and Cluster Analysis Methods. Ijms; 2023; 24, 14432. [DOI: https://dx.doi.org/10.3390/ijms241914432] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37833879]
55. Van Der Maaten, L.; Postma, E.; Van den Herik, J. Dimensionality Reduction: A Comparative Review. J. Mach. Learn. Res.; 2009; 10, pp. 66-71.
56. Boser, B.E.; Guyon, I.M.; Vapnik, V.N. A Training Algorithm for Optimal Margin Classifiers. Proceedings of the Fifth Annual Workshop on Computational Learning Theory—COLT ’92; Pittsburgh, PA, USA, 27–29 July 1992; ACM Press: Cambridge, MA, USA, 1992; pp. 144-152.
57. Kumar, A.; Karmakar, S.; Agarwal, I.; Pal, T. Mental Workload Classification with One-Dimensional CNN Using fNIRS Signal. Pattern Recognition and Machine Intelligence; Maji, P.; Huang, T.; Pal, N.R.; Chaudhury, S.; De, R.K. Lecture Notes in Computer Science Springer Nature Switzerland: Cham, Switzerland, 2023; Volume 14301, pp. 746-755. ISBN 978-3-031-45169-0
58. Saadati, M.; Nelson, J.; Curtin, A.; Wang, L.; Ayaz, H. Application of Recurrent Convolutional Neural Networks for Mental Workload Assessment Using Functional Near-Infrared Spectroscopy. Advances in Neuroergonomics and Cognitive Engineering; Ayaz, H.; Asgher, U.; Paletta, L. Lecture Notes in Networks and Systems Springer International Publishing: Cham, Switzerland, 2021; Volume 259, pp. 106-113. ISBN 978-3-030-80284-4
59. Sharma, V.; Ahirwal, M.K. An End-to-End Brain Computer Interface System for Mental Workload Estimation through Hybrid Deep Learning Model. Hum-Cent. Intell. Syst.; 2024; [DOI: https://dx.doi.org/10.1007/s44230-024-00086-y]
60. Ghosh, K.; Stuke, A.; Todorović, M.; Jørgensen, P.B.; Schmidt, M.N.; Vehtari, A.; Rinke, P. Deep Learning Spectroscopy: Neural Networks for Molecular Excitation Spectra. Adv. Sci.; 2019; 6, 1801367. [DOI: https://dx.doi.org/10.1002/advs.201801367]
61. Kazemzadeh, M.; Hisey, C.L.; Zargar-Shoshtari, K.; Xu, W.; Broderick, N.G.R. Deep Convolutional Neural Networks as a Unified Solution for Raman Spectroscopy-Based Classification in Biomedical Applications. Opt. Commun.; 2022; 510, 127977. [DOI: https://dx.doi.org/10.1016/j.optcom.2022.127977]
62. Leclerc, P.; Ray, C.; Mahieu-Williame, L.; Alston, L.; Frindel, C.; Brevet, P.-F.; Meyronet, D.; Guyotat, J.; Montcel, B.; Rousseau, D. Machine Learning-Based Prediction of Glioma Margin from 5-ALA Induced PpIX Fluorescence Spectroscopy. Sci. Rep.; 2020; 10, 1462. [DOI: https://dx.doi.org/10.1038/s41598-020-58299-7] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31996727]
63. Alston, L.; Mahieu-Williame, L.; Hebert, M.; Kantapareddy, P.; Meyronet, D.; Rousseau, D.; Guyotat, J.; Montcel, B. Spectral Complexity of 5-ALA Induced PpIX Fluorescence in Guided Surgery: A Clinical Study towards the Discrimination of Healthy Tissue and Margin Boundaries in High and Low Grade Gliomas. Biomed. Opt. Express; 2019; 10, 2478. [DOI: https://dx.doi.org/10.1364/BOE.10.002478] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31149380]
64. Butte, P.V.; Mamelak, A.N.; Nuno, M.; Bannykh, S.I.; Black, K.L.; Marcu, L. Fluorescence Lifetime Spectroscopy for Guided Therapy of Brain Tumors. NeuroImage; 2011; 54, pp. S125-S135. [DOI: https://dx.doi.org/10.1016/j.neuroimage.2010.11.001]
65. Chu, X.; Huang, Y.; Yun, Y.-H.; Bian, X. Chemometric Methods in Analytical Spectroscopy Technology; Springer Nature Singapore: Singapore, 2022; ISBN 978-981-19162-4-3
66. Sheehy, G.; Picot, F.; Dallaire, F.; Ember, K.; Nguyen, T.; Petrecca, K.; Trudel, D.; Leblond, F. Open-Sourced Raman Spectroscopy Data Processing Package Implementing a Baseline Removal Algorithm Validated from Multiple Datasets Acquired in Human Tissue and Biofluids. J. Biomed. Opt.; 2023; 28, 025002. [DOI: https://dx.doi.org/10.1117/1.JBO.28.2.025002] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36825245]
67. Zhou, Y.; Liu, C.-H.; Sun, Y.; Pu, Y.; Boydston-White, S.; Liu, Y.; Alfano, R.R. Human Brain Cancer Studied by Resonance Raman Spectroscopy. J. Biomed. Opt.; 2012; 17, 116021. [DOI: https://dx.doi.org/10.1117/1.JBO.17.11.116021] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23154776]
68. Morais, C.L.M.; Lilo, T.; Ashton, K.M.; Davis, C.; Dawson, T.P.; Gurusinghe, N.; Martin, F.L. Determination of Meningioma Brain Tumour Grades Using Raman Microspectroscopy Imaging. Analyst; 2019; 144, pp. 7024-7031. [DOI: https://dx.doi.org/10.1039/C9AN01551E] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31650137]
69. Bury, D.; Morais, C.L.M.; Martin, F.L.; Lima, K.M.G.; Ashton, K.M.; Baker, M.J.; Dawson, T.P. Discrimination of Fresh Frozen Non-Tumour and Tumour Brain Tissue Using Spectrochemical Analyses and a Classification Model. Br. J. Neurosurg.; 2020; 34, pp. 40-45. [DOI: https://dx.doi.org/10.1080/02688697.2019.1679352]
70. Liu, J.; Wang, P.; Zhang, H.; Wu, N. Distinguishing Brain Tumors by Label-Free Confocal Micro-Raman Spectroscopy. Photodiagnosis Photodyn. Ther.; 2024; 45, 104010. [DOI: https://dx.doi.org/10.1016/j.pdpdt.2024.104010]
71. Riva, M.; Sciortino, T.; Secoli, R.; D’Amico, E.; Moccia, S.; Fernandes, B.; Conti Nibali, M.; Gay, L.; Rossi, M.; De Momi, E. et al. Glioma Biopsies Classification Using Raman Spectroscopy and Machine Learning Models on Fresh Tissue Samples. Cancers; 2021; 13, 1073. [DOI: https://dx.doi.org/10.3390/cancers13051073] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33802369]
72. Lita, A.; Sjöberg, J.; Păcioianu, D.; Siminea, N.; Celiku, O.; Dowdy, T.; Păun, A.; Gilbert, M.R.; Noushmehr, H.; Petre, I. et al. Raman-Based Machine-Learning Platform Reveals Unique Metabolic Differences between IDHmut and IDHwt Glioma. Neuro-Oncol.; 2024; 26, pp. 1994-2009. [DOI: https://dx.doi.org/10.1093/neuonc/noae101] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38828478]
73. Kowalska, A.A.; Berus, S.; Szleszkowski, Ł.; Kamińska, A.; Kmiecik, A.; Ratajczak-Wielgomas, K.; Jurek, T.; Zadka, Ł. Brain Tumour Homogenates Analysed by Surface-Enhanced Raman Spectroscopy: Discrimination among Healthy and Cancer Cells. Spectrochim. Acta Part A Mol. Biomol. Spectrosc.; 2020; 231, 117769. [DOI: https://dx.doi.org/10.1016/j.saa.2019.117769] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31787534]
74. Bury, D.; Morais, C.L.M.; Ashton, K.M.; Dawson, T.P.; Martin, F.L. Ex Vivo Raman Spectrochemical Analysis Using a Handheld Probe Demonstrates High Predictive Capability of Brain Tumour Status. Biosensors; 2019; 9, 49. [DOI: https://dx.doi.org/10.3390/bios9020049]
75. Jermyn, M.; Mok, K.; Mercier, J.; Desroches, J.; Pichette, J.; Saint-Arnaud, K.; Bernstein, L.; Guiot, M.-C.; Petrecca, K.; Leblond, F. Intraoperative Brain Cancer Detection with Raman Spectroscopy in Humans. Sci. Transl. Med.; 2015; 7, 274ra219. [DOI: https://dx.doi.org/10.1126/scitranslmed.aaa2384]
76. Jermyn, M.; Desroches, J.; Mercier, J.; St-Arnaud, K.; Guiot, M.-C.; Leblond, F.; Petrecca, K. Raman Spectroscopy Detects Distant Invasive Brain Cancer Cells Centimeters beyond MRI Capability in Humans. Biomed. Opt. Express; 2016; 7, 5129. [DOI: https://dx.doi.org/10.1364/BOE.7.005129] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28018730]
77. Livermore, L.J.; Isabelle, M.; Bell, I.M.; Scott, C.; Walsby-Tickle, J.; Gannon, J.; Plaha, P.; Vallance, C.; Ansorge, O. Rapid Intraoperative Molecular Genetic Classification of Gliomas Using Raman Spectroscopy. Neuro-Oncol. Adv.; 2019; 1, vdz008. [DOI: https://dx.doi.org/10.1093/noajnl/vdz008] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31608327]
78. Zhou, Y.; Liu, C.-H.; Wu, B.; Yu, X.; Cheng, G.; Zhu, K.; Wang, K.; Zhang, C.; Zhao, M.; Zong, R. et al. Optical Biopsy Identification and Grading of Gliomas Using Label-Free Visible Resonance Raman Spectroscopy. J. Biomed. Opt.; 2019; 24, 1. [DOI: https://dx.doi.org/10.1117/1.JBO.24.9.095001]
79. Depciuch, J.; Tołpa, B.; Witek, P.; Szmuc, K.; Kaznowska, E.; Osuchowski, M.; Król, P.; Cebulski, J. Raman and FTIR Spectroscopy in Determining the Chemical Changes in Healthy Brain Tissues and Glioblastoma Tumor Tissues. Spectrochim. Acta Part A Mol. Biomol. Spectrosc.; 2020; 225, 117526. [DOI: https://dx.doi.org/10.1016/j.saa.2019.117526]
80. Klamminger, G.G.; Gérardy, J.-J.; Jelke, F.; Mirizzi, G.; Slimani, R.; Klein, K.; Husch, A.; Hertel, F.; Mittelbronn, M.; Kleine-Borgmann, F.B. Application of Raman Spectroscopy for Detection of Histologically Distinct Areas in Formalin-Fixed Paraffin-Embedded Glioblastoma. Neuro-Oncol. Adv.; 2021; 3, vdab077. [DOI: https://dx.doi.org/10.1093/noajnl/vdab077] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34355170]
81. Zhang, L.; Shen, L.; Wu, D. Clinical Significance of Cancer Stem Cell Markers in Lung Carcinoma. Acta Biochim. Pol.; 2021; 68, pp. 187-191. [DOI: https://dx.doi.org/10.18388/abp.2020_5498] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33877783]
82. Steiner, G.; Galli, R.; Preusse, G.; Michen, S.; Meinhardt, M.; Temme, A.; Sobottka, S.B.; Juratli, T.A.; Koch, E.; Schackert, G. et al. A New Approach for Clinical Translation of Infrared Spectroscopy: Exploitation of the Signature of Glioblastoma for General Brain Tumor Recognition. J. Neurooncol; 2023; 161, pp. 57-66. [DOI: https://dx.doi.org/10.1007/s11060-022-04204-3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36509907]
83. Quesnel, A.; Coles, N.; Angione, C.; Dey, P.; Polvikoski, T.M.; Outeiro, T.F.; Islam, M.; Khundakar, A.A.; Filippou, P.S. Glycosylation Spectral Signatures for Glioma Grade Discrimination Using Raman Spectroscopy. BMC Cancer; 2023; 23, 174. [DOI: https://dx.doi.org/10.1186/s12885-023-10588-w]
84. Tołpa, B.; Depciuch, J.; Jakubczyk, P.; Paja, W.; Pancerz, K.; Wosiak, A.; Kaznowska, E.; Gala-Błądzińska, A.; Cebulski, J. Fourier Transform Infrared Spectroscopic Marker of Glioblastoma Obtained from Machine Learning and Changes in the Spectra. Photodiagnosis Photodyn. Ther.; 2023; 42, 103550. [DOI: https://dx.doi.org/10.1016/j.pdpdt.2023.103550] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37024000]
85. Ember, K.; Dallaire, F.; Plante, A.; Sheehy, G.; Guiot, M.-C.; Agarwal, R.; Yadav, R.; Douet, A.; Selb, J.; Tremblay, J.P. et al. In Situ Brain Tumor Detection Using a Raman Spectroscopy System—Results of a Multicenter Study. Sci. Rep.; 2024; 14, 13309. [DOI: https://dx.doi.org/10.1038/s41598-024-62543-9] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38858389]
86. Galli, R.; Lehner, F.; Richter, S.; Kirsche, K.; Meinhardt, M.; Juratli, T.A.; Temme, A.; Kirsch, M.; Warta, R.; Herold-Mende, C. et al. Prediction of WHO Grade and Methylation Class of Aggressive Meningiomas: Extraction of Diagnostic Information from Infrared Spectroscopic Data. Neuro-Oncol. Adv.; 2024; 6, vdae082. [DOI: https://dx.doi.org/10.1093/noajnl/vdae082] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39006162]
87. Tołpa, B.; Paja, W.; Trojnar, E.; Łach, K.; Gala-Błądzińska, A.; Kowal, A.; Gumbarewicz, E.; Frączek, P.; Cebulski, J.; Depciuch, J. FT-Raman Spectra in Combination with Machine Learning and Multivariate Analyses as a Diagnostic Tool in Brain Tumors. Nanomed. Nanotechnol. Biol. Med.; 2024; 57, 102737. [DOI: https://dx.doi.org/10.1016/j.nano.2024.102737] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38341010]
88. Orringer, D.A.; Pandian, B.; Niknafs, Y.S.; Hollon, T.C.; Boyle, J.; Lewis, S.; Garrard, M.; Hervey-Jumper, S.L.; Garton, H.J.L.; Maher, C.O. et al. Rapid Intraoperative Histology of Unprocessed Surgical Specimens via Fibre-Laser-Based Stimulated Raman Scattering Microscopy. Nat. Biomed. Eng.; 2017; 1, 0027. [DOI: https://dx.doi.org/10.1038/s41551-016-0027]
89. Hollon, T.C.; Pandian, B.; Adapa, A.R.; Urias, E.; Save, A.V.; Khalsa, S.S.S.; Eichberg, D.G.; D’Amico, R.S.; Farooq, Z.U.; Lewis, S. et al. Near Real-Time Intraoperative Brain Tumor Diagnosis Using Stimulated Raman Histology and Deep Neural Networks. Nat. Med.; 2020; 26, pp. 52-58. [DOI: https://dx.doi.org/10.1038/s41591-019-0715-9] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31907460]
90. Hollon, T.C.; Pandian, B.; Urias, E.; Save, A.V.; Adapa, A.R.; Srinivasan, S.; Jairath, N.K.; Farooq, Z.; Marie, T.; Al-Holou, W.N. et al. Rapid, Label-Free Detection of Diffuse Glioma Recurrence Using Intraoperative Stimulated Raman Histology and Deep Neural Networks. Neuro Oncol.; 2021; 23, pp. 144-155. [DOI: https://dx.doi.org/10.1093/neuonc/noaa162] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32672793]
91. Hollon, T.; Jiang, C.; Chowdury, A.; Nasir-Moin, M.; Kondepudi, A.; Aabedi, A.; Adapa, A.; Al-Holou, W.; Heth, J.; Sagher, O. et al. Artificial-Intelligence-Based Molecular Classification of Diffuse Gliomas Using Rapid, Label-Free Optical Imaging. Nat. Med.; 2023; 29, pp. 828-832. [DOI: https://dx.doi.org/10.1038/s41591-023-02252-4] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36959422]
92. Abramczyk, H.; Imiela, A. The Biochemical, Nanomechanical and Chemometric Signatures of Brain Cancer. Spectrochim. Acta Part A Mol. Biomol. Spectrosc.; 2018; 188, pp. 8-19. [DOI: https://dx.doi.org/10.1016/j.saa.2017.06.037] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28688336]
93. Baria, E.; Giordano, F.; Guerrini, R.; Caporalini, C.; Buccoliero, A.M.; Cicchi, R.; Pavone, F.S. Dysplasia and Tumor Discrimination in Brain Tissues by Combined Fluorescence, Raman, and Diffuse Reflectance Spectroscopies. Biomed. Opt. Express; 2023; 14, 1256. [DOI: https://dx.doi.org/10.1364/BOE.477035]
94. Romanishkin, I.D.; Savelieva, T.A.; Ospanov, A.; Linkov, K.G.; Shugai, S.V.; Goryajnov, S.A.; Pavlova, G.V.; Pronin, I.N.; Loschenov, V.B. Classification of Intracranial Tumors Based on Optical-Spectral Analysis. Biomed. Photonics; 2023; 12, pp. 4-10. [DOI: https://dx.doi.org/10.24931/2413-9432-2023-12-3-4-10]
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Romanishkin, Igor 2
; Ospanov, Anuar 3
; Goryaynov, Sergey 4 ; Pavlova, Galina 5 ; Pronin, Igor 4 ; Loschenov, Victor 1
1 Prokhorov General Physics Institute of the Russian Academy of Sciences, 119991 Moscow, Russia;
2 Prokhorov General Physics Institute of the Russian Academy of Sciences, 119991 Moscow, Russia;
3 Engineering Physics Institute of Biomedicine, National Research Nuclear University MEPhI (Moscow Engineering Physics Institute), 115409 Moscow, Russia;
4 N.N. Burdenko National Medical Research Center of Neurosurgery, 125047 Moscow, Russia;
5 N.N. Burdenko National Medical Research Center of Neurosurgery, 125047 Moscow, Russia;