Introduction

Tumor is a serious threat to human life and health of major diseases. Malignant tumor (cancer) is one of the main causes of human death. According to an estimation from the world health organization (WHO) in 2019, cancer is the first or second leading cause of death before age 70 in 112 countries, ranking the third or fourth in the other 23 countries [1]. Because of the lack of cancer screening and cognitive deficiency for the strategy of “early detection, early diagnosis and early treatment”, the high or low 5 years of tumor patient survival rate becomes one of the key elements of life quality index.

At present, there are three main methods to detect tumors: tumor markers, imaging, and histopathology. The tumor marker detection is used for early tumor monitoring from the perspective of molecular biology, which is susceptible to individual differences and certain benign diseases [2,3,4]. Only several kinds of cancers have specific markers, such as alpha-fetoprotein, a tumor marker of liver cancer. Imaging diagnosis method is usually only used as an auxiliary means in clinical practice [5,6,7]. Because imaging diagnosis method only can initially identify of the shape and size of the tumor, can’t accurately identify the benign and malignant tumors qualitatively, with the high false positive ratio. Histopathology is the "gold standard" for the clinical application of tumor diagnosis [8,9,10], which can be traced back 100 years [11]. In this method, the benign and malignant, cell type and cancer subtype as well as the stage of the tumor can be identified by means of biopsy from the tissue and cell scale for further treatment of tumor. However, histopathology is a very complex subject with high professional and experience requiring for a long training cycle. To sum up, the existing methods of clinical cancer diagnosis have the problems of long diagnosis time, severe trauma, and high misjudgment rate, and relying heavily on the subjective experience of doctors.

Tissue and cell carcinogenesis are a very complex process, which will affect the content and structure of nucleic acids, proteins, and other biological macromolecules [12]. Therefore, there are differences in the morphology and composition of normal tissue/cell and cancerous tissue/cell at the molecular and/or cellular levels. Meanwhile, these differences can be detected optically, in the form of different intensity, peak location, peak width, etc. [13]. Then the position and/or intensity of these peaks can be analyzed corresponding to the transformation of cells and tissues, so as to realize the monitoring and diagnosis of tissue and cell cancerization [14].

Raman spectroscopy is a light scattering technique, whereby utilizes the differences between incident light wavelength (λincident) and scattering light wavelength (λscattering) for chemical analysis [15]. The elastic scattering is called Rayleigh scattering (λscattering = λincident), where the inelastic scattering is called Raman scattering which divided into anti-Stokes Raman scattering (λscattering < λincident) and Stokes Raman scattering (λscattering > λincident). Meanwhile, Raman spectroscopy, a label-free optical technology with the advantages of specificity, can also be used to analyze the biochemical characteristics of substances by such factors as the position and/or intensity of Raman peak [16]. Since the content and structure of biomacromolecules will change in the process of tissue and cell carcinogenesis, while Raman spectroscopy is mainly based on the interaction between light and molecules to obtain the information of molecular vibration and rotation. Therefore, we utilize the Raman spectroscopy to tumor detection. Spontaneous Raman scattering, as a rare phenomenon in comparison with its counterpart Rayleigh scattering, is typically identified as an insensitive technique [15]. However, with the technological advancement of highly efficient laser sources, low-noise detectors, valid filters and high-throughput optics, this applicability has significantly improved [17,18,19]. The commercial instrument “Confocal Raman Microscope”, which integrates the optical magnification power of laser and direct visualization of the sample, has widely applicated in tumor diagnosis.

Although the confocal Raman microscope can integrate the optical magnification power of laser and direct visualization of the sample and has widely applicated in tumor diagnosis, the intensity of spontaneous Raman scattering signal is weak and the anti-interference ability is poor. In practical applications, it usually takes a long integration time to obtain the spectrum with acceptable signal noise ratio (SNR), which is not conducive to spectral acquisition and fast image formation. Given these shortcomings, surface-enhanced Raman spectroscopy (SERS), a very excellent Raman enhancement technology, is based on the combination of Raman spectroscopy and nanomaterials, which can realize highly precise target detection [20,21,22]. The SERS phenomenon occurs on the local surface of some specially prepared metal or nonmetal conductor, where mainly caused by two mechanisms, called chemical enhancement mechanism and electromagnetic enhancement mechanism [20]. The former mainly emphasizes the adsorption between the metal substrate and adsorbed molecules, while the latter mainly for local field and dipole emission [23]. Thus, when adsorbed to a biomolecule, these nanoparticles result in greatly enhanced Raman spectrum signal by plasmon resonance. Recently, with high specificity, high sensitivity, rapidity and trace analysis, SERS is getting more and more frequently investigated in the field of tumor diagnosis.

Raman imaging technique is to study the molecular structure and dynamic characteristics by the interaction between light and matter [24, 25]. It can not only obtain the spectral information of light emission, absorption, and scattering, as well as the three-dimensional spatial imaging information, but also obtain the geometric shape, molecular structure, and dynamic characteristics of the research object [26, 27]. Hence, Raman imaging can be applied to the rapid imaging of tumor tissues to diagnose the subtypes and types of tumors according to the morphological and color changes of the imaging results [14, 28,29,30]. Especially the stimulated Raman histology (SRH), a label-free optical imaging method, provides rapid, label-free, sub-micrometer-resolution images of unprocessed biologic tissues, which has achieved rapid tumor diagnosis combining deep learning [31].

Based on widely applications of the Raman spectroscopy in the cancer diagnosis field, this article reviews the latest research and progress in the application of Raman spectroscopy in tumor diagnosis published between 2020 and 2022. It is mainly introduced from three aspects: Raman spectroscopy for tumor diagnosis, SERS for tumor diagnosis, Raman imaging for tumor diagnosis (Fig. 1). The conventional Raman spectra usually are utilized for tumor diagnosis by single spot acquisition, with simplicity but weak-signal highlighted. Therefore, the SERS are combined the Raman scatting and nanomaterials to enhance the Raman signal for tumor diagnosis, with high specificity, high sensitivity. However, these two techniques only leverage the Raman signal collected from the samples, but ignoring the morphological or positional information, especially in cells or tissues, which will affect the accuracy of diagnosis. Raman imaging technology can not only utilize the collected Raman signal, but also display the morphology or location information of the samples, which greatly improves the carrying capacity of the output information, but also increases the complexity of information processing. Whereas facing these complex Raman images, artificial intelligence (AI) has shown greater processing power for tumor diagnosis. Herein, the conventional procedure for tumor diagnosis based on three aspects are overviewed, with the introduction of the AI application in this data processing procedure. Through the introduction of the above three aspects, Raman spectroscopy will be a novel scientific approach for tumor diagnosis in the future. Especially, Raman imaging analysis combined with artificial intelligence may even replace the diagnosis process of pathologists, which will greatly promote the medical development of intelligent tumor diagnosis.

Fig. 1
figure 1

Schematic illustration of AI application of Raman spectroscopy and Raman figures for cancer diagnosis

Raman spectroscopy for tumor diagnosis

Raman spectroscopy can be easily collected from the commercial instrument “Confocal Raman Microscope”, which has widely applied in tumor diagnosis. Data processing is vital for tumor diagnosis based on Raman spectroscopy. Conventional spectrochemical analysis methods are to process the original Raman data through several simple algorithm and then output the results, manually identify the difference between the output data of normal samples and cancerous samples, and judge the belonging of unknown samples, such as such as principal components analysis (PCA), linear discriminate analysis (LDA), quadratic discriminant analysis (QDA), partial least squares (PLS), etc. [32]. These methods have great advantages in the processing of small sample data. However, with the development of technology, the amount of data obtained has increased significantly, making it difficult to calculate and extract subtle variations in complex hidden features from big data by conventional methods. Machine learning is a system that can acquire and integrate knowledge autonomously, find hidden features to significantly amplify the difference between normal and cancerous samples, and independently judge the affiliation of unknown samples, such as support-vector machines (SVM), random forest (RF), etc. Machine learning has been widely used in the field of biomedical photonics. In addition, AI is a kind of computational model abstracting the neural network of human brain from the perspective of information processing. It has the characteristics of nonlinear, unlimited, strong adaptability and fault tolerance. In contrast, non-AI methods mainly start from the data itself, extract features through matrix transformation and other methods, and finally conduct classification and regression through modeling. In this chapter, we first introduce the conventional procedure in tumor diagnosis based on Raman spectroscopy, and then focus on the AI application in tumor diagnosis.

Conventional procedure

The conventional procedure for tumor diagnosis based on Raman spectroscopy necessitates (i) sample preparation; (ii) spectral acquisition; (iii) data processing and analysis. Raman microscope mainly collect the Raman spectroscopy from the tissues, cells, body fluids, and other raw materials. After collecting spectra, Raman spectroscopy data are used multiple methods for computational processing and analysis. Conventional methods for diagnosing Raman spectroscopy data usually utilize Raman peak intensity comparison or various multivariate statistical analytical methods. Cancers in different parts of the body have been diagnosed among these conventional methods, such as: brain cancer [33,34,35,36,37,38,39,40], breast cancer [41,42,43,44], esophagus cancer [45], leukemia [46], liver cancer [47], lung cancer [48,49,50], oral cancer [51, 52], ovarian cancer [53], salivary gland neoplastic [54], rectal cancer [55], skin cancer [56, 57] (Table 1).

Table 1 Raman spectroscopy for tumor diagnosis

For sample preparation, tissues are the most directly and widely collected for tumor diagnosis. About the tissues, most researcher prefer to utilize the unprocessed surgical tissues, because the processed procedures of formalin fixation paraffin-embedding (FFPE) may influence Raman spectroscopy measurements. Indeed, Ning et al. [41] evaluated the diagnosis analysis capabilities of unprocessed surgical samples and dewaxed FFPE samples by collecting the Raman spectroscopy from different tissues (Fig. 2a). The results demonstrated that the dewaxing process significantly altered the biochemical composition of the tissues, particularly lipids, proteins, and carotenoids. Even though, the analytical result showed that PCA-LDA method and PLS-discriminate analysis (DA) method could distinguish the target tissue types effectively with satisfying overall accuracy in which the former of 88.3%, the matter of 93.0%. Although this study confirmed that FFPE sections had diagnostic potential with multivariate analytical model, biochemical changes still emerged, which may influence the accuracy. Therefore, it is preferable to use untreated samples.

Fig. 2
figure 2

Sample preparation for tumor diagnosis by Raman spectroscopy a White light micrographs of stained and unstained frozen healthy breast tissue section [41]; b Raman spectra of human body liquid (plasma, serum, urine, saliva) to identify cancer [45]; c Cryo-transmission electron microscope image of EVs [92]

In addition to tissues, body fluids are also easily accessible samples that are widely used to monitor various bodily functions and health conditions. Raman spectroscopy, a label-free analytical technique, has been proven useful in probing the blood components and the whole blood for over 40 years [93]. Blood is a vital bodily fluid responsible for numerous physiological functions, which contains plasma, erythrocytes (red blood cells), leucocytes (white blood cells) and platelets (thrombocytes). When an organ of the body becomes cancerous, the biochemical changes to the composition of blood follows. Therefore, researchers have utilized Raman spectroscopy to detect the changes for tumor diagnosis. For example, brain cancer [40], esophageal cancer [45], lung cancer [49]. In addition to blood as an important body fluid, saliva and urine are also readily available body fluids for tumor diagnosis. Raman spectra combined with saliva and/or urine have applied for tumor diagnosis, such as oral squamous cell carcinoma [52]. Particularly, Maitra et al. [45] collected Raman spectra from four kinds of human body liquid (plasma, serum, urine, saliva) to detect esophageal stages through to esophageal adenocarcinoma (Fig. 2b). For saliva and urine samples the analysis model achieved 100% classification for all classes, while for plasma and serum, the model achieved excellent accuracy in all esophageal stages (> 90%).

Cells are more delicate at the detection scale than tissues, so cell detection based on Raman spectroscopy can be used for cancer screening. Raman spectroscopy on live cells can classify among different disease stages, and play a significant role clinically as a diagnostic tool for cell phenotype. For instance, breast cancer cells [44, 58], colorectal cancer cells [58, 59], lymphoma cells [60]. Besides, extracellular vesicles (EVs) secreted by cancer cells provide a crucial insight into cancer biology and could be leveraged to enhance diagnostics and disease monitoring. Penders et al. [92] present a single particle automated Raman trapping analysis (SPARTA) system, a dedicated standalone device optimized for single particle analysis of EVs (Fig. 2c). They demonstrate that the dedicated SPARTA system can differentiate between cancer and noncancer EVs with a high degree of sensitivity and specificity (> 95% for both).

After sample preparation, it turns to collect Raman spectroscopy from sample. The commercial instrument “Confocal Raman Microscope”, which integrates the optical magnification power of laser and direct visualization of the sample, has widely applicated for Raman spectroscopy data acquisition. Different instruments equip different wavelength lasers, where different wavelengths of excitation light produce different sampling effects. The scattering intensity of visible light is higher than near infrared (NIR), while the NIR excitation was used to minimize tissue autofluorescence. Therefore, the appropriate laser wavelength should be selected according to the samples.

Data processing is vital for classification of different types of tumor tissues. The acquired spectra are firstly filtered, normalized, and corrected then classified by a variety of algorithms and methods. The advantages, limitations and corresponding suitable application of different data processing methods are as Table 2. For the conventional data processing methods, the Raman peak intensity comparison is the most straightforward method [53], which usually utilizes the intensity of several characteristic peaks or the ratio of a pair of characteristic peaks to distinguish. This method is suitable for spectral data with obvious characteristic peaks and large differences, and is more effective for small sample data. Once the sample data size increases, the accuracy is not high. Therefore, for vast amounts of data, abundant multivariate statistical analysis methods are often used to analyze spectral differences and distinguish the tissues. The multivariate statistical analysis methods contain PCA [37, 38, 44, 48, 54] (Fig. 3a), PCA-LDA [36, 40,41,42,43, 47, 50,51,52, 55] (Fig. 3b), PLS-DA [33, 35, 41, 49, 56, 57] (Fig. 3c), PLS-LDA [51] (Fig. 3d), PCA-QDA [39, 60] (Fig. 3e), K-means cluster analysis (KCA) [48], genetic algorithm (GA)-QDA [45] (Fig. 3f). PCA can effectively reduce the spectrum to a certain number of principal components (PCs) that account for significant spectral variance, thus retaining important spectral data while removing background noise [94]. PLS is also a data-reduction algorithms, which can be used to reduce individual spectra down to a few key factors [95]. PCA is the unsupervised method and ideal for exploratory studies but cannot distinguish well among samples that do not differ significantly between groups, while PLS is the supervised method where the characteristic variables of each group can be better selected to distinguish and the relationship between samples can be determined. PLS can be shown to be better than PCA to prepare for classification [96]. Although these two methods can also directly classify spectral data, the accuracy is relatively low, so classification methods need to be further used, which typically rely on the clustering technique and discriminate analysis methods. KCA is a popular option of the clustering technique, while LDA and QDA are the popular options of DA methods. For clustering technique, KCA method is an unsupervised clustering method of data exploration that helps to explore and discover data structures in a large amount of data with simple algorithm principle and fast processing speed [48]. However, KCA method utilizes the K given in advance, where the choice of K value is difficult to determine. And the result of KCA method is not necessarily global optimal, but only local optimal. Therefore, KCA method is suitable for samples with large differences between groups. For DA methods, LDA is a powerful supervised technique for achieving class classification, but it can overfit if the number of spectra is insufficient [95]. Therefore, a general guideline for applying a supervised technique is to have the number of spectra 5–10 times bigger than the number of variables [97], for example, using a PCA prior stage, hence PCA-LDA. PCA-LDA is the commonest method among the multivariate statistical analysis methods. QDA is the variant of LDA, where the both have similar algorithm characteristics. The difference is that LDA should be used when the covariance matrix of different classified samples is the same, while the QDA should be used when the covariance matrix of different classification samples is different [98]. In addition, the downside of QDA is that it cannot be used as a dimension reduction technique. Genetic algorithm is a branch of evolutionary computing, which simulates natural selection and genetic mechanisms to find optimal solutions. Meanwhile, GA is a general optimization technique and is applied for feature selection, where the feature selection is commonly applied as a stage prior to classification as a means to prevent overfitting and to circumvent the “curse of dimensionality” [99]. Therefore, GA method can be used for feature extraction of data and then QDA classification method is used for sample analysis.

Table 2 Data processing, advantages, limitations and suitable application of classification methods
Fig. 3
figure 3

Multivariate statistical analysis for tumor diagnosis by Raman spectroscopy a PCA analysis result of tumoral tissue and healthy margin [54]; b PCA-LDA analysis result of breast cancer [43]; c PLS-DA analysis result of different cancers [35]; d PLS-LDA analysis result of oral cancer [51]; e PCA-QDA analysis result of meningiomas [39]; f GA-QDA analysis result of esophageal cancer [45]

Although multivariate statistical analysis methods yielded high accuracy, these methods pose a limitation toward improving accuracy, especially facing large data sets. Therefore, it is critical to search more accurate methods for tumor diagnosis. With the development of computer science and technology, the machine-learning classification with neural network was applied for tissue diagnosis yielding higher sensitivity and specificity.

AI application

With the advances in artificial intelligence, machine learning (ML) has been applied in tumor diagnosis based on Raman spectroscopy with higher accuracy. Herein, the conventional ML methods, SVM, RF, have harvested high accuracy for many years. Recently, deep learning, a branch of the machine learning, has achieved more excellent accuracy for tumor diagnosis. From the literature, these ML methods have applied in different parts of the body, such as in bladder cancer [81], bone tumors [68], brain cancer [61, 62, 86, 90, 91], breast cancer [63, 70, 82, 83], central nervous system tumor [84], cervical cancer [64], chondrogenic tumor [71], colon cancer [72, 87], kidney cancer [65, 66], laryngeal cancer [73], lung cancer [74,75,76], meningiomas [67], oral cancer [77, 78], osteosarcoma [69], ovarian cancer [88], pancreatic cancer [79, 85, 89], skin cancer [80, 100, 101] (Table 1).

For classical ML methods, the advantages, limitations and corresponding suitable application of different ML methods are as Table 2. SVM is a generalized classifier for binary classification, which seeks to determine the optimal hyperplane that maximizes the distance between the hyperplane and the nearest data sample in a high-dimensional space [102]. SVM can avoid local optimum and “curse of dimensionality”, and is less prone to overfitting. However, the prediction accuracy of SVM method highly depends on the kernel function, and SVM method has poor training efficiency when processing large-scale data. Therefore, SVM method is powerful in dealing with nonlinear, multi-dimensional problems, peculiarly those with limited samples [103]. Nowadays, SVM is the most widely used method for tumor diagnosis based on Raman spectroscopy, whatever the tissues as original sample preparation [61, 63, 65,66,67,68, 77], or cells as original sample preparation [69, 81, 82, 85, 100, 101], body fluids as original sample preparation [84, 86, 88, 89]. Other ML methods, such as boosted tree (BT) [62, 87], k-nearest neighbors (KNN) [64, 65, 68, 85], RF [62, 65, 83], also have applied for tumor diagnosis. BT method is an ensemble learning method, whose main idea is to assemble a weak classifier into a strong classifier. When using the boosted tree approach as learning algorithm, there is no need to do feature normalization/normalization for different types of data, and it is easy to balance runtime efficiency and accuracy. Whereas, BT method is sensitive to abnormal data and easy to overfit. Hence, BT method is suitable for low dimensional data, and the number of model layers should not be too high. RF method, as an ensemble learning method, can effectively reduce the risk of overfitting by leveraging the strategy of bagging and random feature selection to construct several decorrelated decision trees and output their average predictions [104, 105]. But the RF method has relatively lower learning speed, which is willing to dealing with those with limited samples. KNN is a nonparametric, supervised learning classifier that uses proximity of a single data point to classify or predict groupings [106]. KNN method can yield high precision and is insensitive to outliers, while with relatively large time complexity and space complexity. Therefore, KNN method is suitable for small-size samples and low-dimensional data. O’Dwyer et al. [81] utilized 11 classification method for different grade bladder cancer cells and normal cells, including SVM, RF, PLS, PCA-LDA, PCA-QDA, PCA-KNN, marginal relevance (MR)-LDA, MR-QDA, MR-KNN, MR-RF, MR-SVM method (Fig. 4a). And the results showed that cells can be distinguished by using a variety of approaches with accuracy, sensitivity, and specificity in excess of 95%, especially the SVM methods with the best performances of 0.996 accuracy, 0.996 sensitivity and 0.996 specificity.

Fig. 4
figure 4

AI method for tumor diagnosis by Raman spectroscopy a Different grade bladder cancer cell classification accuracy of the 11 classifiers [81]; b Three classic CNNs (AlexNet, ResNet, and GoogLeNet) model structure for diagnosis of glioma [90]; c RNN model structure for diagnosis of lung cancer and glioma [91]; d The conversion from the Raman spectral signal into 2D Raman spectrograms [74]; e Raman encoding figures [76]

For deep learning methods, convolutional neural network (CNN), one of the most popular basic deep leaning architectures, has been widely used tumor diagnosis based on Raman spectroscopy, where the original sample preparation can be obtained from tissues [68, 70, 73,74,75,76,77,78,79,80] and body fluids [90, 91]. CNN simulates the structure and function of biological neural networks in computers, which mainly contains three basic operations, i.e., the convolution, activation, and pooling. Convolution extracts the feature maps from the inputs with a kernel matrix, while the activation is to map its inputs into another space nonlinearly and is usually operated after the convolution, then pooling is a subsampling strategy, including max-pooling and average-pooling [32]. Indeed, a general CNN model is basically composed by several convolution, activation, and pooling layers and sometimes ties several fully connected layers. AlexNet, ResNet, and GoogLeNet are three kinds of classic CNNs, where the differences lie in the structure of the network. The initial AlexNet was proposed by Sutskever et al. [107] with 8 layers. It utilizes local response normalization to solve the overfitting problem and utilizes multiple GPUs to accelerate the performance of the model, shorten the training time, and balance the training speed and accuracy of the model [107]. The initial ResNet was proposed by He et al. [108] with 152 layers. It utilizes a residual algorithm to significantly improve the ability of a neural network to extract features. The use of residual blocks is to solve the vanishing and disappearing gradient problems with ordinary CNNs by not only deepening the depth of the network but also by improving the performance of the network [108]. The initial GoogLeNet was proposed by Szegedy et al. [109] with 22 layers. GoogLeNet introduces the “inception” new module, which connects filters with different sizes and dimensions into a new filter. Compared with other deep CNN models, GoogLeNet reduces the number of parameters and layers in the network and improves the utilization of computing resources inside the network [109]. After these excellent CNN models were proposed, researchers can further modify various models according to their own data, so as to establish required models and achieve high classification accuracy. In short, CNN can directly extract features from input data and classify the observed objects, but limited by the quality and features of the data. However, CNN is the first choice in most of the modeling tasks (classification and regression) because of the simple architecture and the ease of use. Tian et al. [90] utilized three classic CNNs, AlexNet, ResNet, and GoogLeNet, to build the classification model for diagnosis of glioma (Fig. 4b). The accuracy rates of the AlexNet, ResNet, and GoogLeNet models were 98.50%, 98.24%, and 99.50%, respectively, where GoogLeNet model yielded the best classification effect with the specificity and sensitivity of 98.98% and 98.48%, respectively. Meanwhile, other deep learning architectures are also used in tumor diagnosis and have achieved unprecedented success, such as recursive neural network (RNN) [91] (Fig. 4c). RNN is a type of neural networks with cyclic connections [110]. Compared with CNN, RNN has the characteristics of storing more long-term sequence information and mining temporal and semantic information in the data, which has a strong ability to learn the nonlinear data behavior of time sequence [91]. Considering the Raman spectroscopy can be regard as a special sequence, RNN may have more applications in Raman spectroscopy data classification.

Among these papers, most deep learning architectures utilized the Raman spectra as 1D data to fed into models for training and testing [68, 70, 77, 78, 80], however, 2D figures would be a better choice as inputs compared with 1D data for deep learning. Therefore, our group first proposed an unusual method wherein we considered the Raman spectral signal as a sequence and then converted it into a 2D Raman spectrogram by spectral short-time Fourier transform (SSTFT) [74, 75] (Fig. 4d). This novel method combined with deep learning yielded excellently accurate diagnosis of lung tissues. Subsequently, we extended the transformation approaches of converting the 1D Raman spectroscopy into 2D figures (Fig. 4e) and proposed a new concept called the Raman encoding figure, which can improve the accuracy [76]. Three new methods are proposed and implemented for the Raman spectrum conversion, i.e., spectral recurrence plot (SRP), spectral Gramian angular field (SGAF), and spectral Markov transition field (SMTF). For typical Raman spectrum, it contains two kinds of internal information, one for wavenumber position information, the other for intensity information. Particularly, SMTF is the conversion based on wavenumber position information, and SSTFT is the conversion based on wavenumber position information and intensity information. But for multiple spectra, they not only contain internal information, but also external information, such as shape information. Especially, SRP is the conversion based on structure of wavenumber series, while SGAF based on wavenumber series. The inclusion of different kinds of information in the conversion results in different performances. Furthermore, due to the more information involved in the transformation, SRP and SGAF methods are suitable for more complex original spectra, while SSTFT and SMTF are suitable for less varied spectra. These 2D-CNN methods all yielded more than 95% accuracy, 94% sensitivity, and 96% specificity when tested, where the SRP with best performances (98.9% accuracy, 99.5% sensitivity, 98.3% specificity), followed by SGAF, SSTFT, and SMTF. Meanwhile, we compared the diagnostic performances of the 2D-CNN method with that of the 1D-CNN method, which utilized the 1D Raman spectra as inputs and yielded a test accuracy of 94.1%, a test sensitivity of 91.8%, and a test specificity of 96.4%. In addition, Conforti et al. [71] proposed a chondrogenic tumor classification through wavelet transform of Raman spectra, which combines the hybrid 1D-2D deep learning classification process applied on Raman spectra both raw and after wavelet transform. The 1D deep learning classification makes it possible to distinguish between the tumor’s early stages and more advanced ones with great accuracy, while the 2D deep learning classification yield the high accuracy classification between the tumor’s malignant and non-cancerous stages, especially the 2D deep learning classification. Typical SSTFT method is limited by wavenumber (or frequency) resolution due to the application of a single wavenumber window. Here, Conforti et al.[71] proposed the continuous wavelet transform (CWT) method to solve this problem by using short windows at higher frequencies and long windows at lower frequencies, where the CWT method can be continuously changed during the procedure to still obtain a valid multi-resolution analysis. Finally, this method can classify Raman spectra obtained from bone tissues with high accuracy of 97% accuracy. This method extends the application of converting 1D Raman spectrum to 2D Raman encoding figure as deep learning input, and also shows that the conversion of 1D Raman spectrum to 2D Raman encoding figure has huge application space and can provide a reference method for tumor diagnosis based on Raman spectra. Overall, the 2D Raman encoding figure combined with deep learning shows great potential for diagnosing tissues and may provide a novel analysis method for other spectral techniques.

In conclusion, the key of tumor diagnosis based on Raman spectroscopy is to find the difference between normal and cancerous samples. It is relatively unremarkable to look for such differences from the Raman characteristic peaks. Therefore, a variety of Raman spectral data methods have been developed to improve the diagnosis and recognition rate, from multivariate statistical methods, to classical machine learning methods, and finally to deep learning. Especially facing a large amount of data, deep learning can yield a better diagnosis and recognition rate. In the future, we should start from two aspects: data volume and further classification to achieve accurate diagnosis in view of the complexity of tumor diagnosis. On the one hand, the magnitude of sample data should be improved to explore the classification effect of various diagnostic methods in the case of large data volume and find better solutions. On the other hand, appropriate data processing methods should be selected for further typing of cancerous samples to achieve accurate diagnosis of cancerous subtypes.

SERS for tumor diagnosis

SERS is a very excellent Raman enhancement technology, which can realize highly precise target detection. SERS substrate preparation is the key of tumor diagnosis based on SERS, because the appropriate substrate can greatly improve the intensity of Raman signal and is more conducive to the diagnosis of tumors. Therefore, SERS based tumor diagnosis is more suitable for detection of samples with low concentration, such as the detection of tumor biomarkers. Since the SERS substrates greatly improve the intensity of Raman signal, the data processing of SERS spectrum was relatively simple. The conventional methods mainly utilize peak intensity comparison and multivariate statistical analysis. Meanwhile, artificial intelligence methods are also being used to improve diagnosis accuracy. In this chapter, we first introduce the conventional procedure in tumor diagnosis based on SERS, and then introduce on the AI application in tumor diagnosis.

Conventional procedure

The conventional procedure for tumor diagnosis based on SERS necessitates (i) sample preparation and nanoparticles (NPs) preparation; (ii) spectral acquisition; (iii) data processing and analysis. The first step of the preparation process is mainly divided into sample preparation and nanoparticle preparation. After preparation, the Raman microscope would collect the Raman spectroscopy from the preprocessing samples. Then, theses Raman spectra would be processing and analysis for tumor diagnosis, for example, bladder cancer [111,112,113], brain cancer [114,115,116], breast cancer [117,118,119,120,121,122,123,124,125,126], cervical cancer [127], gastric cancer [128], liver cancer [129,130,131,132], lung cancer [133, 134], melanoma [135, 136], nasopharyngeal cancer [137], osteosarcoma [138], pancreatic cancer [139], prostate cancer [140,141,142], salivary gland neoplastic [143], thyroid cancer [144] (Table 3).

Table 3 SERS for tumor diagnosis

For samples preparation, tissue [114] or tissue homogenates [115, 143] have been relatively less studied as samples unlike conventional Raman spectroscopy diagnostic methods. Besides, it is difficult to gain the tumor tissues via traditional biopsies repeatedly for analysis. Therefore, body fluids, easily accessible samples, are widely used to diagnose tumor, such as tears [122], serum [123, 128, 129, 140,141,142], plasma [112, 144], urine [113]. Since SERS can significantly improve the Raman signal intensity, researchers are more inclined to detect samples with finer scale. Cells are more delicate at the detection scale than tissues, so cell detection based on SERS can be used for cancer diagnosis, such as breast cancer cell [118, 145], hepatoma cell [145], esophageal cancer cell [145], bladder cancer cell [111].

In addition, tumor biomarkers, a non-invasive and rapid analysis, has recently demonstrated the potential to solve the limitations of conventional biopsy. Because of the low abundance and chemical complexity of these materials, SERS, a very excellent Raman enhancement technology, is appropriate for detection and analysis tumor biomarkers. Recently, several researches have conducted about tumor biomarkers by SERS, such as aldehydes in exhaled breath [133], carcinoembryonic antigen (CEA) [148], circulating tumor cells [120, 126, 131, 132, 136], epithelial cell adhesion molecule (EpCAM) [148], epidermal growth factor receptors (EGFR) [124], exosomes [125, 134, 138], matrix metalloproteinases (MMPs) [137], micro RNAs (miRNAs) [117, 119, 121], mRNA [111], p16/Ki-67 [127], prostate specific antigen [142], protein [130, 141]. In addition to tumor biomarkers, other biomolecules related to tumor growth and dissemination are also detected by SRES technology to monitor or diagnose tumor changes, for example, cancer stem cells [146], cell lysates [118], proteases [137], telomerase [147], vascular endothelial growth factor [135].

For nanoparticles preparation, noble metal NPs are the mostly applied for tumor diagnosis with the advantages of low cost and easy synthesis, such as Au [111, 119,120,121,122,123, 125, 131, 134,135,136,137, 141, 142, 145, 147], Ag [112,113,114,115, 117, 118, 127,128,129, 132, 133, 140, 143, 144, 148, 155], where the single metal as the original nanoparticle (Fig. 5a, b). However, the single SERS nanoparticle has its own shortcomings, the bimetallic nanoparticle substrate such as Au@Ag [130] combines the advantages of two metals (Fig. 5c), which maybe have better SERS enhancement effect. Except these noble metals (Au/Ag), other transition metals such as Ni [146] also produce an enhanced effect and have been applied for tumor diagnosis (Fig. 5d). Oxide semiconductors such as TiO2 [124, 126] also have a SERS enhanced effect and have been utilized for detecting the cancer biomarkers (Fig. 5e).

Fig. 5
figure 5

SERS substrate a Au substrate [122]; b Ag substrate [118]; c Au@Ag substrate [130]; d Ni substrate [146]; e TiO2 substrate [126]

After sample preparation and nanoparticle preparation, the Raman spectroscopy acquisition will be the next step. Compared with the spontaneous Raman signal intensity, SERS greatly improved the Raman signal intensity, so commercial Raman spectrometers could be directly used in the process of Raman spectroscopy data acquisition. At the same time, some miniaturized Raman spectrometers, such as hand-held Raman spectrometers [148] or portable Raman spectrometers [122, 128, 133, 147], can be used for signal acquisition because the Raman signal has been greatly improved, which can better achieve a fast, flexible, and convenient acquisition process.

Data processing and analysis is the final and vital step for tumor diagnosis by SERS, which can be divided into two types: detection of biomarkers or biomolecules and diagnosis classification. The limit of detection (LOD) is used to evaluate the detection ability of the former, while the accuracy, sensitivity and specificity to demonstrate the classification capacity of the latter. At the same time, the limit of quantification (LOQ) can also be used for evaluate the detection ability of the detection of biomarkers or biomolecules [118]. Moreover, several researches can simultaneously achieve the detection of biomarkers and diagnosis of cancer [118].

For detecting the biomarkers biomolecules, the LOD of circulating tumor cells (CTCs) in the peripheral blood of cancer patients could reach 1 cell per mL [132], 2 cells/mL [126], up to 5 cells per mL [120]. The LOD of two miRNA markers for breast cancer (miR-21 and miR-155) was 451 zmol and 1.65 amol respectively [117],while the LOD of exosomal miRNA reaching as low as 1 pmol/L [121]. The LOD of mRNA for detection of bladder tumor could extend 3.4 pM [111]. The LOD of exosomes can achieve the detection limit of 2.4 × 103 particles/mL [134], 5.3 × 103 particles/mL [125]. Furthermore, Kim et al. [148] successfully detected the cancer biomarkers (EpCAM, CEA) in tiny amount of sample solutions (~ 2 μl) with as low as 0.2 pmol of the protein biomarkers. Er et al. [130] developed a SERS immunosensor and exhibited a wide linear detection range (1 pg/mL to 10 ng/mL) with a LOD as low as 0.03 pg/mL toward α-fetoprotein with good reproducibility and stability. Lin et al. [137] have designed the metal carbonyl nanobarcodes for detecting MMPs with strong sensitivity (LOD 0.07 ng/mL). Turan et al. [142] introduced a novel designed sensor for detection of prostate specific antigen with the LOD and LOQ of 0.9 pg/mL and 3.2 pg/mL, respectively. Keshavarz et al. [124] achieved semiconductor materials (TiOX) as a SERS template to diagnose breast cancer with the LOD of 1 nM. Huang et al. [133] successfully achieved the ultrasensitive detection of aldehydes in exhaled breath of lung cancer patients with a LOD 1.35 nM. Liu et al. [118] yielded a high sensitivity for the detection of creatinine with a low LOD of 5.3 μM and a LOQ of 17.68 μM. Huang et al. [135] traced vascular endothelial growth factor (VEGF) in cell lysis samples with an excellent limit of detection of 2.3 pg/mL under the optimum conditions.

For tumor diagnosis classification based on SERS, statistical analysis methods are usually used to analyze spectral differences and achieve the tumor diagnosis, which is the same as the conventional method for tumor diagnosis classification based on Raman spectroscopy. The multivariate statistical analysis methods contain PCA [114, 115, 138], PCA-LDA [112, 113, 122, 129, 140, 144, 145], PLS [116], PLS-DA [143], PCA-QDA [128]. Among using PCA method, Kowalska et al. [115] proved that tumor brain samples can be discriminated well from the healthy tissues by using only three main principal components with 96% of accuracy, while Li et al. [114] reported that healthy brain tissue and Grade II gliomas as low grade gliomas as well as Grade III and Grade IV as high-grade gliomas can be clearly distinguished by three-dimensional PCA. Moreover, Han et al. [138] identified the osteosarcoma by combining SERS and matrix-assisted laser desorption/ionization time-of-flight mass spectrometry profiling of exosomes with all 100% sensitivity, specificity, and accuracy. Among PCA-LDA method, Hu et al. [113] utilized SERS for detecting urine metabolites of bladder cancer with the total diagnostic sensitivity and specificity 100% and 98.85%, respectively, while Qian et al. [112] utilized SERS of pretreated plasma samples to predict disease recurrence in muscle-invasive bladder cancer patients undergoing neoadjuvant chemotherapy and radical cystectomy revealing a high accuracy of 85.2% in prediction of disease recurrence. Especially for determination of cancer stages, Gao et al. [129] screened liver cancer (LC) at different tumor (T) stages by serum albumin, in conjunction with SERS and finally classified the early T (T1) stage LC vs. normal group and advanced T (T2–T4) stage LC vs. normal group, yielding high diagnostic accuracies of 90.00% and 96.55%, respectively by PCA-LDA method. For PLS/PLS-DA method, Kowalska et al. [116] reported that the differentiation between primary and secondary brain tumors SERS data was completed PLS method with a very high 85% of accuracy, while Czaplicka et al. [143] analyzed salivary glands carcinoma, tumor and healthy tissues and their homogenates by SERS and PLS-DA method, showing correlation accuracy as 0.98 with sensitivity 0.97, and specificity 0.89. In addition, Avram et al. [128] utilized PCA-QDA method for classifying gastrointestinal tumors by different testing models, where the results showed that the classification accuracy yielded by combining SERS analysis of serum with C-reactive protein levels, neutrophil counts, platelet counts and hemoglobin levels was superior (accuracy 83.33%) to the classification accuracy yielded by SERS profiling alone (accuracy 76.92%) and to the one yielded by blood tests (accuracy 73.08%). In a word, multivariable statistical analysis methods are commonly used with spectral data processing, no matter the tumor diagnosis method based on Raman spectroscopy or the tumor diagnosis method based on SERS. Similarly, AI methods, as a complement to conventional methods, will serve as a superior method to improve the accuracy of tumor diagnosis.

AI application

Since SERS can greatly improve the detection signal of samples, the tumor diagnosis methods based on SERS are mostly used for the detection of small molecules such as cells and biomarkers, resulting in the collected Raman signal processing often used for the calibration of detection limits. Of course, some Raman signals are used for classification diagnosis, where researchers usually utilized multivariate statistical methods for classification. Even so, artificial intelligence methods also provide unique methods for classification diagnosis to improve accuracy, such as breast cancer [123, 149], liver cancer [152, 154], lung cancer [150], ovarian cancer [151], pancreatic cancer [153] (Table 3).

For typical ML methods, SVM is the most widely used method for classification. Dawuti et al. [154] utilized urine SERS combined with SVM method to identify liver cirrhosis (sensitivity 88.9%, specificity 83.3%, and accuracy 85.9%) and hepatocellular carcinoma (sensitivity 85.5%, specificity 84.0%, and accuracy 84.8%), while Fang et al. [150] also implemented SERS technique combined with SVM to identify and distinguish non-small-cell lung cancer (NSCLC) and small cell lung cancer (SCLC) cells from normal cells including blood cells and immortalized lung cells, where achieving the classification accuracy of 98.8% between NSCLC cells and normal cells and reaching the accuracy 100% in the classification of SCLC cells and normal cells, as well as SCLC cells and NSCLC cells (Fig. 6a). Except SVM method, Banaei et al. [153] applied the classification tree method to the analysis of the expression level of EVs biomarkers in pancreatic cancer, chronic pancreatitis, and normal controls individuals, measuring the sensitivity and specificity as 0.95 and 0.96, respectively (Fig. 6b). Moreover, Ćulum et al. [151] discriminated the EVs isolated from different ovarian cancer cell lines by a logistic regression-based machine learning method with ∼99% accuracy, sensitivity, and specificity.

Fig. 6
figure 6

AI method for tumor diagnosis by SERS a 3D map of classification accuracy for the SVM model [150]; b The classification tree analysis of the expression level of EVs biomarkers [153]; c Residual network classification results of various tumor cells and blood cells [149]; d CNN classifier results of the independent test dataset for recognizing serum SERS spectra [152]

Deep learning is an advanced machine learning method that can be used to discriminate various data accurately. Furthermore, deep learning has been applied for tumor diagnosis based on SERS. Fang et al. [149] utilized the feature peak ratio method, PCA combined with KNN, and residual network to classify the SERS spectra from blood cells and tumor cells. The results show that the ratio method and PCA combined with the KNN could only identify some tumor cells from blood cells, but residual network method could quickly distinguish various tumor cells and blood cells with an accuracy of 100% (Fig. 6c). This indicates that deep learning has great potential in tumor classification and diagnosis. In addition, Cheng et al. [152] constructed a convolutional neural network classifier for recognizing serum SERS spectra as deep learning inputs. The 1D-CNN method achieved a prediction accuracy of 97.78% on an independent test dataset randomly sampled from normal controls, hepatocellular carcinoma cases, and hepatitis B patients (Fig. 6d).

In conclusion, the key of tumor diagnosis based on SERS is to find suitable enhanced nanoparticle substrates so as to greatly increase the original Raman signal. Since the original Raman signal is weak, the differences between normal and cancerous samples will become larger after enhancement. Therefore, the exploration of detection limit is a very important aspect in tumor diagnosis based on SERS. Otherwise, due to the signal enhancement brought by SERS technology, the data processing of tumor classification and diagnosis based on SERS is easier, and multivariate statistical method is often used to distinguish. However, despite the multivariate statistical methods have achieved a high classification accuracy, but from the data processing process of tumor diagnosis methods based on SERS, the classification accuracy still can be improved, especially when artificial intelligence method is used for data processing and classification, which has been proved by some literatures. Whereas, there is still relatively little literature on these diagnostic methods using artificial intelligence [70, 76]. In the future, tumor diagnosis based on SERS should start from two aspects: signal enhancement and classification diagnosis. On the one side, more suitable SERS substrates with greater signal intensity increased should be sought from the perspectives of biocompatibility and spatial distribution. On the other side, for tumor classification, new data processing methods such as artificial intelligence can be selected to provide the accuracy of classification and diagnosis.

Raman imaging for tumor diagnosis

Raman imaging can obtain more information, so it has broad application potential in tumor diagnosis. The processing of Raman figure is the most important part of this method. Compared with spectral data, Raman images carry more information. Therefore, in the processing of Raman images, the conventional method is to directly carry out comparative analysis of images, or adopt multivariate statistical analysis method, which is feasible for small-batch images. But when faced with a large number of images, this method is time consuming. As an automatic method of image analysis, processing and mining, artificial intelligence has important application value in tumor diagnosis based on Raman imaging. In this chapter, we first introduce the conventional procedure in tumor diagnosis based on Raman imaging, and then focus on the AI application in tumor diagnosis.

Conventional procedure

The conventional procedure for tumor diagnosis based on Raman imaging contains (i) sample preparation; (ii) image acquisition; (iii) image processing and analysis. Tissues are the most applied materials for Raman imaging, as well as cells. Due to weak scattering effect, measurements of spontaneous Raman spectra with a competent signal to noise ratio can be time-consuming [156]. Therefore, conventional Raman spectroscopy measurements are limited to smaller sample areas. For large scale sample, nonlinear optical technologies (coherent anti-Stokes Raman scattering; stimulated Raman scattering) have shown success in tissue and cell imaging for tumor diagnosis [157,158,159]. After collecting the Raman imaging, it is vital for image processing with suitable methods. Comparative analysis of Raman images is the most direct method, but it is for small-scale data volume. When the amount of data is huge, artificial intelligence is often used for analysis and diagnosis, which has achieved a very high diagnostic performance. At present, multiple types of tumor samples have been accurately diagnosed based on Raman imaging, such as benign cementoma [160], bladder cancer [161, 162], brain cancer [163,164,165], breast cancer [163, 166,167,168], chondrogenic tumor [169], colorectal cancer [59, 170], glioma [171], meningiomas [172], prostate cancer [156, 173, 174], spine tumors [164], skin cancer [175], skull base tumors [176, 177] (Table 4). In addition to diagnosing tumors directly, Raman imaging can also capture the components or metabolites, such as glycogen [178], lipid droplets [179], in cells to analyze the differences and changes between normal cells and cancer cells, thus playing a role in monitoring the cancer progression and tumorigenesis.

Table 4 Raman imaging for tumor diagnosis

For sample preparation, tissue is the most common source of samples, especially fresh, unprocessed biological tissues. Cordero et al. [161] utilized the entire extracted biopsy without thin-sectioning to characterize the tumor grading ex vivo, using a compact fiber probe-based Raman imaging system. Liao et al. [167] used the adipose tissue in breast resections to assess the surgical margins by high wavenumber Raman imaging and fingerprint Raman spectroscopy, which can reduce the overall tissue analysis time and maintain high diagnostic accuracy. Boitor et al. [180] presented a prototype device based on integrated auto-fluorescence imaging and Raman spectroscopy for intraoperative assessment of surgical margins during Mohs micrographic surgery of basal cell carcinoma (BCC), and the results showed that typically more than 95% of the resection area is analyzed by the Fast Raman device, which includes both the epidermal and deep margin, without requiring tissue trimming (Fig. 7a). In addition to imaging from fresh, untreated biological tissue, it is also able to image unstained tissue after sectioning. D’Acunto et al. [169] utilized the 5 μm unstained tissue specimens for Raman imaging to the diagnosis and grading of chondrogenic tumors, including enchondroma and chondrosarcomas of increasing histologic grades. Brozek-Pluska et al. [170] applied 16 μm sections from the nonfixed, fresh samples to differentiate noncancerous and cancerous human sigmoid colon mucosa based on Raman spectroscopy and imaging (Fig. 7b). Feng et al. [175] adopted the skin sections of 20 μm thickness for rapid discrimination of basal cell carcinoma tumor by a superpixel acquisition approach, which expedited acquisition with two to five orders of magnitude faster than conventional point-by-point scanning by trading off spatial resolution.

Fig. 7
figure 7

Conventional method for tumor diagnosis by Raman imaging a Integration of the Fast Raman device in the pathway of Mohs surgery [180]; b The microscopy image and Raman image of noncancerous sigmoid colon mucosa tissue [170]; c FTIR and RS images of cells [162]; d AF mapping images and Raman mapping images of live adipocytes at the time of differentiation [204]

Cells, as a much minute scale of observation than tissue, show more subtle signs of changes in the tumor. Therefore, Raman imaging of the cells can more clearly find the changes on the cellular scale of the tumor. Kujdowicz et al. [162] employed Fourier transform infrared (FTIR) and Raman spectroscopic (RS) imaging to investigate bladder cancer cell lines of various invasiveness, and demonstrated that FTIR and Raman spectroscopy can be employed to distinguish between different bladder cancer cells of various malignancy (Fig. 7c). Beton et al. [59] evaluated the biochemical and structural features of human colon cell lines based on Raman spectroscopy and imaging, and shown that normal reactive oxygen species-injured and cancerous human colon cells could be distinguished based on their unique vibrational properties. Paidi et al. [168] utilized 3D optical diffraction tomography and Raman spectroscopy for optical phenotyping of cancer cells at single-cell resolution, and demonstrated that coarse Raman microscopy is capable of rapidly mapping a sufficient number of cells for training a classifier that can accurately predict the metastatic potential of cells at a single-cell level.

Cell imaging based on Raman spectroscopy, not only can be used to image the cellular morphology, but also can be used to image the cell cytosolic microstructures, to achieve subcellular level analysis of the components of cancer cells. Roman et al. [174] leveraged the Raman mapping technique to investigate lipid droplets (LDs) composition in untreated and irradiated with X-ray beams prostate cancer cells, proved lipids accumulation in PC-3 cells by Raman mapping technique, and revealed the heterogeneous composition of LDs. Suhito et al. [204] reported a novel optical method called “autofluorescence-Raman mapping integration (ARMI)”, which used cell autofluorescence to reveal cellular morphology and cytosolic microstructures, while Raman mapping allowed site-specific intensive analysis of target molecules, which enables ultra-fast identification of cell types. The novel technique has rapidly and precisely analyzed the adipogenesis (Fig. 8d). Abramczyk et al. [186] utilized the Raman imaging to detect molecular processes that occur in normal and cancer brain cells due to retinol transport in human cancers at the level of isolated organelles and found that aberrant expression of retinoids and retinol binding proteins in human tumors could be localized in lipid droplets, and mitochondria. Uematsu et al. [184] presented a new method for simultaneously visualizing up to five atomically labeled intracellular fatty acid species by Raman imaging and revealed that fatty acids with more double bonds tend to concentrate more efficiently at lipid droplets. Radwan et al. [185] considered the astaxanthin as a new Raman probe for the detection of lipids in the endothelial cells of various vascular beds, where the astaxanthin colocalized with lipids in cells could enable Raman imaging of lipid-rich cellular components with lower laser power. Horgan et al. [183] presented a new strategy for simultaneous quantitative in vitro imaging and molecular characterization of EVs in 2D and 3D based on Raman spectroscopy and metabolic labelling, and showed that metabolic deuterium incorporation demonstrated no apparent adverse effects on EV secretion, marker expression, morphology, or global composition.

Fig. 8
figure 8

SRS/SRH/SERS method for tumor diagnosis by Raman imaging a SRS imaging of different components in live and fixed HeLa cells after culturing in d7-glucose medium [178]; b Workflow of SRH imaging in neurosurgery [165]; c Raman imaging of MCF-7, HeLa and NHDF cells treated with Au NPs-H2 and end assembly [191]; d Comparison of cellular features available with SRH versus conventional H&E stained slides in representative case of Meningioma [177]

For the Raman imaging, spontaneous Raman spectroscopy gives an opportunity to investigate biochemical changes in biological samples [59, 161, 162, 167,168,169,170, 174, 175, 183,184,185,186, 204]. However, due to weak scattering effect and low resolution, spontaneous Raman imaging is limited to small area samples [156]. Therefore, Raman enhancement techniques are needed to provide signal strength to achieve higher resolution and imaging effects [205]. Stimulated Raman scattering (SRS) and surface enhanced Raman scattering have been applied for Raman signal amplifications to achieve tumor diagnosis.

SRS amplifies the weak spontaneous Raman signal via stimulated emission by orders of magnitude to enable fast imaging with molecular specificity inherited from spontaneous Raman spectroscopy [206]. Lee et al. [178] leveraged the stimulated Raman scattering microscopy with metabolic incorporation of deuterium-labeled glucose to visualize glycogen in live cancer cells (Fig. 8a), and characterized different glycogen metabolic phenotypes in a series of mutant melanoma cell lines by this method. Tipping et al. [187] demonstrated the multi-wavelength SRS imaging together with spectral phasor analysis to characterize a panel of breast cancer cell lines treated with two clinically relevant statins, and revealed the lipid droplet distribution throughout populations of live breast cancer cells by SRS imaging within the high wavenumber. Du et al. [179] utilized SRS of intracellular lipid droplets to identify a previously unknown susceptibility of lipid mono-unsaturation within dedifferentiated mesenchymal cells. Bae et al. [188] reported a unique spatial light-modulated stimulated Raman scattering microscopy to monitor real-time cancer treatment effects, and showed immediate apoptotic response when monitoring the therapeutic effect of mild alkaline solution on cancer cells. Sepp et al. [189] utilized the SRS to image label-free ponatinib in live human chronic myeloid leukemia cell lines with high sensitivity and specificity. Lin et al. [190] developed a deformable mirror-based remote-focusing SRS microscope, and performed volumetric chemical imaging of living cells.

Stimulated Raman histology, developed from SRS, utilizes the intrinsic vibrational properties of lipids, proteins and nucleic acids to generate image contrast, revealing diagnostic microscopic features and histologic findings poorly visualized with hematoxylin and eosin (H&E)-stained images, such as axons and lipid droplets [207], while eliminating the artifacts inherent in frozen or smear tissue preparations [208]. The SRH has shown great potential for rapid and accurate tumor diagnosis [201,202,203], demonstrating diagnosis in near-perfect agreement with conventional H&E [31, 164, 171, 172, 176, 177, 181]. Neidert et al. [165] established a dedicated workflow (Fig. 8b) for SRH serving as an intraoperative diagnostic, research, and quality control tool in the neurosurgical operating room, and suggested to optimize the process regarding tissue collection, preparation, and imaging during using this novel imaging modality for intraoperative diagnostic.

SERS enhances the Raman spectrum signal by plasmon resonance [20], and can also achieve Raman imaging for tumor diagnosis rather than acquiring the Raman spectra. Liu et al. [191] demonstrated a target-triggered regioselective assembly strategy of plasmonic nanoprobes for dual Raman imaging of intracellular cancer biomarkers, and successfully performed Raman imaging of MCF-7, HeLa and normal human dermal fibroblasts (NHDF) cells with this strategy(Fig. 8c). Chen et al. [192] designed an Ag SERS substrate to track of the intracellular distribution of exosomes and the concurrent quantitative sensing of environmental pH, and demonstrated that exosomes first attached with the tumor cell surfaces, and then entered into the cells and accumulated in lysosomes as time prolonged. Yuan et al. [194] reported a special nanoparticle which can be detected with high specificity in furin-overexpressing tumor cells, and applied in high-resolution image-guided surgery to precisely delineate tumor margins during and after resection in real-time. Burgio et al. [196] achieved the stable and specifically targeting SERS tags for visualization of the exact tumor borders and infiltrating foci of glioblastoma through application of the appropriate gold nanoparticles surface chemistry and by the correct balance of inert and active targeting functionalities.

For image processing and analysis, picture comparison analysis and Raman peak intensity comparison is the intuitive and simple processing method. Abramczyk et al. [163] used Raman spectroscopy and Raman imaging to monitor changes in the redox state of the mitochondrial cytochromes in ex vivo human brain and breast tissues, and found that the concentration of reduced cytochrome c becomes abnormally high in human brain tumors and breast cancers and correlates with the grade of cancer. Furthermore, multivariate statistical methods can also be used for image processing and analysis. Marro et al. [166] reported the 3D biochemical analysis of breast cancer microcalcifications, combining 3D Raman spectroscopy imaging and advanced multivariate analysis for investigating the molecular composition of HAp calcifications found in breast cancer tissue biopsies.

In particular, SRH is a novel technology that leverages laser spectroscopy and color-matching algorithms to create images similar to the formalin-fixed paraffin-embedded section [176]. The key to SRH analysis and diagnosis lies in the consistency between SRH images and H&E sections. Fitzgerald et al. [176] assessed the time taken for results and diagnostic concordance between SRH images and FFPE section from the patients undergoing sinonasal and skull base surgery, and the results showed that the sensitivity, specificity, precision, and overall accuracy of SRH were 93.3%, 94.1%, 93.8%, and 93.3%, respectively, and near-perfect concordance was seen between SRH and frozen section with Cohen's kappa of 0.89. Pekmezci et al. [171] acquired glioma margin specimens for SRH, histology, and tumor specific tissue characterization, and the results yielded that consistency between immunohistochemistry (IHC) and SRH was near perfect with Cohen's kappa of 0.84 while the substantial agreement between IHC and H&E with Cohen's kappa of 0.67 and between SRH and H&E with Cohen's kappa of 0.72. Shin et al. [177] evaluated the skull base tumor diagnostic accuracy beyond cancer/non-cancer determination and neuropathologist confidence for SRH images contrasted to H&E-stained frozen and FFPE tissue sections (Fig. 8d), and the results revealed that SRH was effective for establishing a diagnosis using fresh tissue in most cases with 87% accuracy relative to H&E-stained FFPE sections. Straehle et al. [164] found a substantial diagnostic correlation between SRH-based neuropathological diagnosis and H&E-stained frozen sections (κ = 0.8), and the results showed that when diagnosing the brain and spine tumors, the accuracy of neuropathological diagnosis based on SRH images was 87.7% and was non-inferior to the current standard of fast frozen H&E-stained Sects. (87.3 vs. 88.9%). Di et al. [172] compared the diagnostic time and accuracy of SRH images with the gold standard (frozen section), and results revealed that the mean time-to-diagnosis was significantly shorter for SRH-mediated diagnosis compared with frozen Sect. (9.2 vs. 35.8 min), and the diagnostic accuracy was not significantly different between methods. Einstein et al. [181] explored the non-inferiority of SRH as compared to frozen section on the same piece of tissue in neurosurgical patients, and the results showed that SRH was sufficient for diagnosis in 78% of specimens as compared to 94% of specimens by frozen section of the same specimen. In a word, SRH images shows a high degree of consistency with H&E staining sections, which provides an important prerequisite for rapid diagnosis using SRH. However, manual diagnosis of SRH images would undoubtedly increase the diagnosis time. Therefore, artificial intelligence methods should be sought to diagnose SRH images to reduce the diagnosis time, so as to truly apply SRH diagnosis to intraoperative diagnosis.

AI application

Images, as two-dimensional data, are more suitable as the input of artificial intelligence models. Therefore, artificial intelligence has great significance application in tumor diagnosis based on Raman imaging, whether using classical artificial intelligence method or the deep learning method. At present, Raman imaging based on artificial intelligence has been applied to a variety of tumor types, such as brain cancer [31, 197, 199, 202], breast cancer [198], gastric cancer [201], laryngeal cancer [203], prostate cancer [209] (Table 4).

For typical ML methods, SVM is a commonly used method for tumor diagnosis. Bae et al. [197] applied the hyperspectral SRS microscopy combined with SVM method for assessment of glioblastoma intertumoral heterogeneity (Fig. 9a), and found that the predominant proportion of glioblastoma tissue was consistent with the diagnosis from genomic analysis, but a significant portion of the remaining SRS image blocks in the specimens belonged to other molecular subtypes, implying a large degree of heterogeneity in glioblastoma. Yang et al. [198] leveraged hyperspectral SRS microscopy to evaluate the breast tumor malignancy based on tissue calcifications, and reached a precision of 98.21% and recall of 100.00% for classifying benign and malignant cases by using SVM method, significantly improving from the pure spectroscopy or imaging based methods. Doherty et al. [209] utilized the multimodal approach of Raman chemical imaging (RCI) and digital histopathology to image the prostate cancer tissues, then the multimodal approach achieved a sensitivity of 73.8% and specificity of 88.1% for Gleason grade 3/4 classification by SVM method.

Fig. 9
figure 9

AI method for tumor diagnosis by Raman imaging a Schematic illustration of the SRS imaging diagnostic platform for rapid glioblastoma subtyping by SVM method [197]; b Diagnostic results of 33 independent cases using residual network [203]; c Confusion matrix of the three diagnostic subtypes [201]

Deep learning, as a machine learning method for large-scale image processing, has great application value in tumor diagnosis based on Raman imaging, especially the SRH combined with deep learning for tumor diagnosis. Group Orringer is the pioneer and leader of this diagnosis pattern. They reported a parallel workflow that combines SRH and deep convolutional neural networks to predict brain tumor diagnosis in near real-time in an automated fashion, where after training on over 2.5 million SRH images in their CNN, the workflow could predict brain tumor diagnosis in the operating room in under 150 s, an order of magnitude faster than conventional techniques (for example, 20 ~ 30 min) [31]. This amazing work shows the great potential for near real-time intraoperative diagnosis, which streamlines the workflow of intraoperative cancer diagnosis and creates a complementary pathway for tissue diagnosis that is independent of a traditional pathology laboratory. Moreover, they also utilized the SRH and deep neural networks to improve the intraoperative detection of glioma recurrence, and achieved a diagnostic accuracy of 95.8% when facing the external SRH validation dataset [202]. Except the wonderful work of Group Orringer, the Group Ji also recently yielded the excellent advances in SRH combined with deep learning for tumor diagnosis. They applied this important pattern to diagnose the laryngeal squamous cell carcinoma on fresh, unprocessed surgical specimens, where they demonstrated near-perfect diagnostic concordance (κ > 0.90) between SRS and standard histology, and yielded the 100% accuracy of classifying 33 independent surgical specimens using deep-learning based SRS [203] (Fig. 9b). Furthermore, they recently leveraged the single-shot femtosecond SRS to reach the maximum speed and sensitivity with preserved chemical resolution, achieved < 60 s imaging the Fresh gastroscopic biopsy, and harvested the accuracy > 96% of predicting gastric cancer by CNN [201] (Fig. 9c).

In a word, Raman imaging has a wide application space in tumor diagnosis, because imaging can not only display the morphology of cells, but also show the distribution of cell components by using the vibration of special bonds. However, tumor diagnosis based on Raman imaging still has some problems such as imaging speed and resolution.

Therefore, tumor diagnosis based on Raman imaging can be improved from the following two aspects: One for Raman imaging technology. New Raman imaging techniques can be explored to improve the resolution and speed of imaging. At the same time, new Raman enhancement techniques can also be explored to enhance weak signals so as to achieve more detailed imaging of cells or tissues. The other for imaging processing method. In terms of diagnosis, more excellent image processing methods can be applied, such as artificial intelligence methods, especially when SRH technology is used, which will greatly improve the efficiency of diagnosis and have high accuracy.

Conclusion, challenges and outlook

Tumor ranks as a leading cause of death and an important barrier to increasing life expectancy in each country of the world. In view of the problems of the current tumor detection methods, it is urgent to develop tumor diagnosis technology with intelligent attributes. According to recent studies, it is shown that Raman spectroscopy, as a label-free optical technique, has attracted more and more attention in tumor diagnosis, especially the Raman spectroscopy technique combined with AI. Here in this section, we compare and analyze firstly the three methods from sample acquisition, data collection and data processing, and give the clinical application direction of the three methods (Table 5). Then, we briefly discuss the challenges in in tumor diagnosis based on Raman spectroscopy, SERS, and Raman imaging, and the outlook of the future of AI for spectroscopic application.

Table 5 The comparison of three methods

For the acquisition of samples, tissues, cells, body fluids, and other kinds of specimens can be used to collect the Raman signals for tumor diagnosis. Conventional Raman spectroscopy and imaging methods can directly collect signals from samples, so as to achieve nondestructive testing. Compared with the other two methods, SERS need add the substrate as one more step, where the SERS substrate can greatly enhance the Raman signal of the detecting substance. However, the added SERS substrate would affect the natural characteristics of samples, which could not really achieve nondestructive testing, compared with the other two methods. Therefore, SERS is used for trace sample detection, such as tumor markers. At the same time, biocompatible SERS substrate should be developed to achieve better diagnostic effect.

For data collection, confocal Raman instrument is mostly applied if only spectral information is collected. However, due to the weak spontaneous Raman intensity of the sample, SERS technology can be used to increase the signal. In this case, confocal Raman instrument can also be used for collection. However, in order to avoid the influence of SERS substrate on the sample's own signal, some other Raman enhancement techniques, such as SRS, could be adopted to improve the signal. For Raman imaging of samples, SRS technology is often used to enhance the signal to achieve higher resolution imaging due to the weak Raman signal of samples themselves. However, the implementation of these technologies requires more expensive and larger footprint equipment. Therefore, it is supposed to design a cheaper Raman equipment and miniaturize the equipment to finally achieve efficient signal acquisition of samples.

For data processing, the data collected based on Raman spectroscopy and SERS are basically one-dimensional spectral data, mainly including peak intensity information and peak position information, while those based on Raman imaging are usually two-dimensional picture information, including both peak intensity information and peak position information, as well as the information of the sample itself, such as morphology. Therefore, different methods are applied to process the collected data. The former two methods mainly use peak intensity contrast or multivariate statistical methods, where these methods are more accurate when dealing with small sample data, however, there is a bottleneck when dealing with large sample data. At this time, artificial intelligence can be used for analysis, especially for sample classification. The one-dimensional spectral data can be directly input into the artificial intelligence model, or the one-dimensional spectral data can be converted to the two-dimensional Raman encoding figure before the training and validation of the artificial intelligence model. Relatively, artificial intelligence is a better choice when processing Raman images because Raman imaging contains more information. It can not only process a large amount of data analysis, but also be more suitable for processing two-dimensional images. Meanwhile, some new artificial intelligence models, such as convolutional neural network and residual network, can better analyze and classify samples.

For clinical application, Raman spectroscopy can be easily obtained by confocal Raman instruments. Therefore, due to the easy availability of data, Raman spectroscopy is more suitable for the diagnosis of easily available samples, such as normal and cancerous diagnosis of body fluids. In addition, the difference of Raman spectra of different samples mainly lies in the difference of Raman peaks and positions, which are relatively not obvious, so the diagnosis of Raman spectra is more suitable for simple diagnostic applications, such as tissue diagnosis of normal and cancerous and cell detection for cancer screening. However, due to the weak spontaneous Raman intensity of the sample, it usually takes a long integration time to obtain the spectrum with acceptable signal noise ratio. SERS is a very excellent Raman enhancement technology and can realize highly precise target detection. Therefore, SERS is very suitable for the detection of low concentration biological molecules, such as tumor biomarkers, and biomolecules related to tumor growth and dissemination. Raman spectroscopy and SERS methods mainly utilize the spectra to diagnosis, which contains less biological information. Raman imaging can not only obtain the spectral information, as well as the three-dimensional spatial imaging information, but also obtain the geometric shape, molecular structure, and dynamic characteristics of the research object. Hence, Raman imaging is more suitable for imaging tissue samples to analyze the benign and malignant tissues and subtypes. Further, Raman imaging can also be used to analyze cellular biochemical and structural characteristics, and tumor biomarker diagnosis. The appropriate method should be selected based on the samples and the equipment.

In a word, the diagnosis methods of tumor are constantly developing. How to detect, diagnose and treat tumors as early as possible is particularly crucial for human life and health. For early tumor diagnosis, some tumor markers can play a reference significance. However, how to ensure the accuracy of these tumor markers is the most important thing to realize early tumor diagnosis. For these tumor markers with small content, the SERS method could enhance their signal, thus achieving accurate diagnosis of early tumors. However, when the tumor develops to the stage of excision, imaging is usually only used as an auxiliary means, because it can only identify the initial shape of the tumor and cannot accurately distinguish the benign and malignant tumors. Therefore, endoscope/probe-based Raman diagnostic method can be applied to achieve tumor diagnosis and guided resection. For histopathology, the gold standard of tumor diagnosis, necessitates multiple processing steps, and interpretation by a pathologist, which is time, resource, and labor intensive. Here, the diagnostic method based on Raman imaging can streamlined the tumor diagnosis, creating a complementary pathway for tissue diagnosis that is independent of a traditional pathology. In particular, the near real-time SRH + AI diagnostic method can deliver rapid intraoperative pathological diagnosis, greatly minish the waiting time for pathological diagnosis results with compressing the operation time and reducing the damage to patients.

With the development of computer science, applying AI algorithm to classify and diagnose Raman spectra from tumors is the key to greatly improve the accuracy of tumor diagnosis. Two main challenges of using AI to realize tumor diagnosis and cancer detection remain. The first challenge is data set acquisition. At present, the data set acquisition method is still manual. However, in the actual process of data acquisition, the specific parameters or processes used in sample preparation, instrument operation, data labeling and other steps vary from person to person, which will cause certain differences in the data collected by different groups of research, so that the model established according to the data cannot be universal. Secondly, most data sets are obtained from single-center sources. Although single-center studies are favored by researchers due to fewer restrictions on conditions and convenient cooperation, multi-center studies are more suitable for the application of AI for diagnosis. On the one hand, multi-center studies can obtain more data; on the other hand, multi-center studies can achieve data consistency and comparability. Thereby gaining experience in establishing standardized processes. The second challenge is the transparency and interpretability of the models. At present, AI research is largely results-oriented, but the black-box nature of the training process makes the models much less interpretable, and thus less acceptable to clinicians when it comes to clinical applications. Therefore, it is necessary to establish a transparent model to realize the interpretability of its process, so as to dispel the doubt of clinicians and build trust in it. Still, the research on the interpretability and visualization of deep learning model is still in the initial stage, and more researchers are expected to explore it.

Future clinical applications can be explored and studied from the following two aspects. The first is to do as many multi-center studies as possible. Since most current studies are single-center, there is poor data consistency among various studies, and in-depth comparison cannot be made. Therefore, conducting multi-center studies can not only improve the consistency and comparability of data, but also establish normative procedures and eventually form consensus standards, laying a foundation for actual clinical application. The second aspect is the equipment miniaturization. The Raman equipment used by most of research covers a large area and is fixed in one location. However, stationary and large footprint devices are a major obstacle to clinical application and, in particular, to near-real-time intraoperative diagnosis. Therefore, it is necessary to explore and study the miniaturization of Raman equipment, which can be explored from two directions: handheld probe + Raman spectrometer or miniaturized portable Raman spectrometer.

Anyway, the tumor diagnosis method based on Raman spectroscopy has a broad application prospect. Especially the artificial intelligence fusion with Raman spectroscopy can provide an effective supplement to the existing tumor diagnosis methods, and even replace the existing diagnostic methods in the near future.