In dental plaque, the cariogenic bacteria are wrapped in an organic matrix of polysaccharides, proteins, and DNA, which enhances resistance to host defense and antimicrobial agents (Selwitz et al., 2007). Endogenous cariogenic bacteria (mainly Streptococcus mutans, Streptococcus sobrinus, and Lactobacillus spp) ferment carbohydrates and produce organic acids, resulting in local pHs below the critical value and demineralization of teeth (Featherstone, 2004; Selwitz et al., 2007).

RS can detect metabolic differences to facilitate the differentiation between biofilms. Gieroba et al. collected and analyzed Raman spectra from single biofilms of S. mutans CAPM 6067, Streptococcus sanguis ATCC 10556, and several serotypes of S. sobrinus. The major differences were concentrated in the region representing lipids, amides, and carbohydrates, reflecting the corresponding biological characteristics. For example, the highest and lowest amide bands in several biofilms were different, indicating different protein compositions of the biofilms and adhesive and cariogenic characteristics (Gieroba et al., 2020). Daood et al. used the changes in Raman spectra to assist in evaluating the effect of quaternary ammonium on cariogenic biofilms. With exposure to QAS, the intensity of 484 cm^–1 in the Raman spectra, representing polysaccharides or carbohydrates, was significantly reduced, and the change was in concentration-dependent and time-dependent patterns (Daood et al., 2020a). Tao et al. focused on heavy water (D₂O) based on single-cell Raman microspectroscopy (D₂O-Raman) and analyzed the Raman spectral region from 2040 to 2300 cm^–1, representing the C–D vibration band, to evaluate the metabolic status of the S. mutans UA159, Streptococcus gordonii ATCC10558, S. sanguinis ATCC10556, and Lactobacillus fermentum ATCC9338 oral bacteria after drug exposure. They distinguished antibiotic-sensitive and -resistant S. mutans (Tao et al., 2017).

The DIAGNOdent pen (Germany) is a laser pen that can detect caries in vivo by collecting fluorescence. To evaluate the laser pen detection capability, RS has been applied to test changes in carious teeth, and Rodrigues et al. established an enamel demineralization model in vitro with cattle tooth blocks and chose phosphate apatite peaks at ~960 cm^–1 to estimate demineralization (Rodrigues et al., 2017). Carvalho et al. focused on the changes in fluoridated apatite, phosphate apatite, and organic matrix in carious teeth, which present in Raman spectra as ~575 cm^–1, ~960 cm^–1, and ~1450 cm^–1, respectively. As demineralization progressed, the intensities of ~575 cm^–1 and ~960 cm^–1 declined significantly and were negatively correlated with the fluorescence detected by the DIAGNOdent pen. This proved that fluoridated apatite and phosphate apatite decreased in caries and verified the caries detection accuracy of the DIAGNOdent pen (Carvalho et al., 2013). Point-scan and wide-field Raman imaging have also been investigated for caries detection as well as application in laser pens (Yang et al., 2014). RS combined with two-dimensional (2D) charge-coupled-device cameras can be assembled into wide-field Raman imaging, which is faster for diagnosing dental caries. Additionally, to compare the penetration depth of a photosensitizer (erythrosine) on intact dentin and decayed dentin in vitro, RS was applied to verify the intact and decayed dentin by providing organic and inorganic information (Occhi-Alexandre et al., 2018). However, the results were mainly obtained from tooth slices in vitro, and to determine whether saliva and bacteria interfere with the process in vivo requires further investigation. In addition, Almahdy et al. collected Raman and fluorescence spectra of carious tooth tissues (Almahdy et al., 2012). Combined with three fluorescence signals (porphyrin fluorescence, putative infected dentin signal, and affected dentin signal), two Raman signals, phosphate at 960 cm^–1 and protein at 1340 cm^–1, enabled differentiation between intact, infected, and affected dentin, indicating that vital transformations of phosphate and protein α-helices are presented. More information could be obtained by investigating organic-related bands more intensively.

Enamel is the most superficial tissue of teeth and is covered by dental plaque, which consists mainly of bacteria. When sugar and other fermentable carbohydrates reach the bacteria and produce acids, teeth demineralization begins. Conversely, when sugar consumption ceases, saliva washes away the sugars and buffers the acids. Calcium and phosphates then enter the teeth again, resulting in remineralization. Thus, a cavity occurs if demineralization overtakes remineralization over time (Klokkevold, 2015). Kerr et al. evaluated the remineralization status of carious enamel treated with high-frequency microwave energy to sterilize and adjust pH by estimating the 960 cm^–1 intensity change (Kerr et al., 2016). For promoting enamel white spot lesion (WSL) remineralization, Raman spectra were also employed to assess the potential of bioactive glass (BAG) powder and BAG containing polyacrylic acid (PAA-BAG) (Milly et al., 2014). Four internal vibration modes of PO 4 3 − at 433 cm^–1, 579 cm^–1, 959 cm^–1, and 1043 cm^–1 were adopted to estimate enamel demineralization and remineralization. The intensity at 959 cm^–1 was the strongest among the four peaks.

Regarding remineralization in dentin, it is worth mentioning that Toledano et al. adopted many iconic bands to evaluate dentin changes from relative mineral concentration, crystallinity, and organic composition of dentin. Thus, the positive effects of zinc-containing amalgam mechanical loads, self-etching zinc-doped adhesives, and zinc-containing amalgam restoration on dentin remineralization before and after 24 h and after 3 weeks were clarified (Toledano et al., 2015a; Toledano et al., 2015b; Toledano et al., 2019).

Pioneering research began in 1999 when RS was used to detect the metabolites of periodontal bacteria. Evidence was found that an increase in spectral intensity at 1002 cm^–1 with time implies that the heme pigment is gradually accumulated on the cell surface when P. gingivalis is incubated on a blood plate. Moreover, P. gingivalis can synthesize the iron trivalent oxidation state Fe(III)PPIX with a band at 1373 cm^–1 on a horse blood plate while synthesizing the iron divalent oxidation state Fe(II)PPIX with a band at 1359 cm^–1 without horse blood. In 2003, researchers applied Mohs, Raman, and UV–vis spectrophotometry to characterize the heme pigment of Prevotella nigrescens ATCC 25261 and Prevotella intermedia ATCC 25611 (Smalley et al., 2003). They also explored the changes in heme pigment under different pH conditions. In 2016, in situ Raman microprobe spectroscopy was used to track the metabolic changes of P. gingivalis on the polished surfaces of bioceramics of the antibacterial substance silicon nitride (Si₃N₄), revealing the formation of peroxynitrite in P. gingivalis (Pezzotti et al., 2016).

In addition, Raman spectra can be used to distinguish different subgingival bacteria using data analysis. In 2020, Kriem et al. distinguished Fusobacterium nucleatum, S. mutans, Veillonella dispar, Actinomyces naeslundii, and Prevotella nigrescens with 100% accuracy in planktonic state. The accuracy of distinguishing S. mutans, V. dispar, and A. naeslundii single-species biofilms was 76%, and that for the others was 90% or higher (Kriem et al., 2020). Based on advanced technologies, in 2021, Witkowska et al. designed a standard Raman spectral detection process to differentiate different serotypes of P. gingivalis, Aggregatibacter actinomycetemcomitans, and Streptococcus spp.

The detection process applied microfluidics, Fe₂O₃@AgNPs combined with Ag/Si substrates, and successfully distinguished P. gingivalis from A. actinomycetemcomitans and Streptococcus spp. by principal component analysis (PCA) with an accuracy of 82–91%. They also confirmed the effectiveness of this detection system in a saliva environment (Witkowska et al., 2021). The detection process provides a feasible method for detecting periodontal pathogens in a clinical environment.

Regarding diagnosis, saliva and GCF have always been prominent as bodily fluids that can be obtained noninvasively and contain effective information as well as interfering noise. One important research approach for analyzing Raman spectra containing a large amount of information is to focus on peaks representing target substances, such as carotene, carotenoids, and sialic acid (SA).

In 2011, Gonchukov et al. collected saliva from 10 patients with periodontitis and 10 healthy subjects (Gonchukov et al., 2011). There were unique peaks in the periodontitis group at 1155 and 1525 cm^–1, representing C–C and C=C, respectively (Darvin et al., 2010), which indicates the existence of carotenoids (Darvin et al., 2009) and proves that as the severity of the periodontal disease increases, the total antioxidant level in saliva also rises (Kim et al., 2010). In 2014, Camerlingo et al. compared the Raman spectra of GCF collected from healthy and periodontitis patients and found that the band at 1537 cm^–1, which represents the isomerization product of the C=C group related to the degraded carotene, only appeared in the latter group (Camerlingo et al., 2014). The aforementoined studies confirm that carotenoid concentration may be beneficial in the diagnosis of periodontitis. Furthermore, SA is present in several proteins related to periodontitis (Ide et al., 2003). Hernandez-Cedillo et al. reported the potential of SA in the auxiliary diagnosis of periodontal diseases. In 2019, they collected saliva samples from patients with periodontitis or gingivitis as well as healthy controls. The peaks at 1002 cm^–1,1237 cm^–1, and 1391 cm^–1 or 910 cm^–1, 1171 cm^–1, and 1360 cm^–1 were compared with the standard SA peaks, indicating that the concentrations of SA in each group were significantly different (Hernández-Cedillo et al., 2019). Interleukin-1β (IL-1β) and tumor necrosis factor-α (TNF-α) are important cytokines in periodontitis, and their concentrations in GCF are higher with periodontitis (Klokkevold, 2015). Yang et al. labeled IL-1β and TNF-α with 5,5’-Dithiobis-(2-nitrobenzoic acid) (DTNB) and 4-mercaptobenzoic acid (4-MBA), respectively, and then collected the salivary surface-enhanced Raman spectra of patients with different periodontal conditions (Yang et al., 2020). The relative area values of the IL-1β and TNF-α peaks at 1335 cm^–1 and 1590 cm^–1, respectively, were calculated as the Raman intensity corresponding to the concentration of inflammatory factors, and significant differences were found between the three groups. The spectral characteristics of saliva reflect metabolic changes in periodontal tissues, and are of vital importance in the diagnosis of periodontitis. However, the composition of saliva is affected by many factors, such as diet and physical condition. Thus, a multi-center and large-sample study is necessary to acquire reliable spectral characteristics of different periodontal conditions.

Regarding Raman spectra of GCF, Jung et al. studied the changes in GCF composition during orthodontic tooth movement with RS. RS showed the degree of bone mineralization and accumulation of carbonate in the apatite lattice. During the alveolar bone remodeling, the mineral–matrix ratio decreased and the carbonate apatite–hydroxyapatite ratio increased. It is speculated that this results from insufficient mineralization during alveolar bone remodeling (Jung et al., 2014). Perillo et al. observed the changes in vibrational modes of proteins (amide I and amide III bands) and CH₂ and CH₃ modes in the periodontal ligament 2, 7, and 14 days after adding orthodontic force to obtain the molecular arrangements and conformational changes. The Raman spectra of the α-helix, 3₁₀-helix, β-turn, β-sheet and random coil in the amide I and amide III bands representing the secondary structure of the protein changed markedly with orthodontic tooth movement. The α-helical and the intensity of the entire amide I band were reduced compared with the control periodontal ligament sample. Compared to the α-helical mode, the remaining component mode of amide I became wider and stronger. The information from the Raman spectra provided quantitative insight into when and how the periodontal ligament molecular arrangement changed (Perillo et al., 2020).

In addition to saliva, GCF, and gingival tissue, periodontal hard tissue has also been considered for possible applications of RS. In 2019, Gatin et al. reported the application of RS to evaluate the bone differences between a periodontitis patient and a healthy patient before and after maxillary sinus lift surgery. Octacalcium phosphate (OCP) and amorphous HAP presented obvious peaks at 955–960 cm^–1, representing immature bone. Moreover, HAP crystals and biological HAP-based bone substitutes showed bands at 960–965 cm^–1, representing mature bone. Based on this analysis, t bone samples before surgery and 8 months after healing revealed that the two patients’ samples had peaks at 955–960 cm^–1 before surgery, and only peaks at 960–965 cm^–1 presented after surgery. The broad fluorescence peaks appearing at 800–900 cm^–1 represented collagen, and the changes in specific peaks could be used to quantify the healing process (Gatin et al., 2019). Raman spectra clearly depict the transformation from immature to mature bone.

In 2004, Krishna et al. first studied the potential of RS for detecting oral cancer (Krishna et al., 2004). They collected Raman spectra from healthy and malignant epithelial cells and differentiated them using PCA. Subsequently, many studies have validated this conclusion. In summary, Raman spectra have been studied to (1) clarify the tumor stage and histological classification of OSCCs, (2) distinguish OSCCs from precancerous lesions and other cancers, and (3) improve the accuracy of diagnostic models.

Clarifying the tumor stage and histological classification is critical for evaluating patients’ conditions and choosing the best treatment plan. Xue et al. established a diagnostic model based on the spectra of serum samples from 135 patients with OSCCs using PCA with linear discriminant analysis (PCA-LDA). The total accuracies of the diagnostic model in identifying tumors at different stages, distinguishing lymph node involvement, and distinguishing between different histological grades were 90.4%, 85.9%, and 90.4%, respectively (Xue et al., 2018). A novel SERS catheter (5–6 µm) helped to successfully obtain and differentiate healthy cells, moderate OSCCs, and severe OSCCs with an accuracy of 97.84% (Madathil et al., 2019). Surprisingly, even subtypes of head and neck cancer cells could be identified by analyzing the spectra from tissue engineering models. Mian et al. successfully identified subtypes of head and neck cancer cells (Mian et al., 2017). Significant differences in the spectra were observed in the lipid content (2881 cm^–1) and protein structure (amides I and III), the peaks of which are associated with several amino acids and nucleic acids (600 cm^–1 to 1003 cm^–1). Therefore, doctors can choose chemotherapy or radiotherapy based on the known subtypes of cancer. In addition, Yasser et al. successfully distinguished radioresistant cell sublines (70Gy-UPCI : SCC029B; 50Gy-UPCI : SCC029B) from parental oral cancer cell lines (UPCI : SCC029B) (Yasser et al., 2014). PCA presented a minor overlap between three clusters, indicating a large difference between three cell lines. Furthermore, Kumar et al. explored the change in differentiation efficacy during the cancer-inducing process, and the accuracy of cancer identification increased during the first 7 weeks, remained steady from 8 to 11 weeks, and exceeded 80% by the 14th week (Kumar et al., 2016).

Furthermore, RS has been found to be a powerful tool for distinguishing OSCCs from precancerous lesions. Krishna et al. studied the potential of Raman spectra obtained in vivo directly to differentiate malignant lesions (OSCCs, oral submucous fibrosis (OSMF) and leukoplakia (OLK)) in the oral cavity. The accuracy was 85% in HV, 89% for OSCCs, 85% for OSMF, and 82% for OLK. For spectra classified as normal and abnormal, the sensitivity and specificity were 94.2% and 94.4%, respectively (Krishna et al., 2014). However, another study reported that OSMF, OLK, and lichen planus were highly misclassified as OSCCs or habitues without lesions in the Raman spectra of sera (Dumal et al., 2020). OSCCs, verrucous carcinomas, and OLK were differentiated with 97.24% accuracy by taking thin cryosections of tissue specimens in a novel SERS catheter (5–6 µm) (Madathil et al., 2019). Moreover, RS combined with cytopathology can distinguish oral precancerous lesions (OPLs), healthy tobacco users (HT), and HV (Ghosh et al., 2019). The sensitivity of the OPL group was identified as approximately 77% when analyzed spectrally, which was higher than patient-wise, with a sensitivity of approximately 70%. The main changes in the spectra of OPLs were related to nucleic acids and proteins. This is consistent with changes in protein and DNA corresponding to cellular physiological changes during poor cell proliferation (Sahu et al., 2017). Connolly et al. obtained unlabeled spectra from saliva and exfoliated cells of HV and OSCC patients using SERS and established diagnostic models using PCA-LDA and PCA with logistic regression (PCA-LR) diagnostic algorithms. Consequently, it was concluded that the saliva and exfoliated cells could identify HV and patients with OSCCs, with sensitivities of 89% and 73% and overall accuracies of 68% and 60%, respectively (Connolly et al., 2016). In brief, the accuracy of Raman spectra from exfoliated cells in differentiating diagnoses must be improved. Misclassification of Raman exfoliative cytology also indicated field cancerization changes. The higher the misclassification rate between spectra of contralateral normal tissue and tumor tissue, the more similar the exfoliated cells are (Sahu et al., 2019).

Researchers have differentiated OSCCs, other oral cancers (such as verrucous carcinoma, mucoepidermoid carcinoma, parotid pleomorphic adenoma, and Warthin’s tumor) (Yan et al., 2011; Yan et al., 2015; Tan et al., 2017; Madathil et al., 2019), and cancers in other regions (such as breast, colorectal, lung, and ovarian) (Moisoiu et al., 2019). The difference in the SERS spectra of sera between OSCCs, mucoepidermoid carcinomas, and healthy humans is mainly represented by nucleic acids and proteins. The spectral differences are mainly distributed in the spectral bands represented by the specific molecular structures of carotenoids and lipids. OSCCs were successfully distinguished from a control group with a sensitivity of 80.7% and specificity of 84.1% (Tan et al., 2017). Another study has also attempted to distinguish OSCCs from parotid pleomorphic adenomas, Warthin’s tumors, and mucoepidermoid carcinomas by support vector machine (SVM) according to the SERS information of sera. The results showed that the SVM had a favorable effect on SERS spectral classification, with an accuracy of 84.1–88.3%, sensitivity of 82.2–97.4%, and specificity of 73.7–86.7%. Although this method can easily differentiate mucoepidermoid carcinoma from the other two benign tumors, it is difficult to distinguish between the two benign tumors themselves (Yan et al., 2011; Yan et al., 2015). Moisoiu et al. successfully discriminated several cancers with an accuracy of 88% for oral cancer, 76% for breast cancer, 86% for colorectal cancer, 59% for lung cancer, and 80% for ovarian cancer (Moisoiu et al., 2019). AgNP substrate enhanced the signal in serum samples and PCA-LDA differentiated spectra in different groups.

Obtaining more important information can increase the differentiation efficacy. Ghosh et al. combined Fourier-transform infrared spectroscopy (FTIR) and RS to increase the classification accuracy from 85% (FTIR) and 82% (RS) to 98% (Ghosh et al., 2019). Spectra of DNA from dehydrated cancer cells and high-wavenumber regions of spectra have attracted the attention of researchers. Panikkanvalappil et al. provided a new method for improving the accuracy of the diagnostic model. Their study found that the conformational induction of DNA from dehydrated cancer cells presents a series of unique Raman-labeled bands. According to these bands, cancer and healthy cell DNA can be distinguished. It is speculated that nucleobase damage in tumor cell DNA and subsequent changes in the electron cloud during the dehydration-driven conformational change results in a Raman spectral change (Panikkanvalappil et al., 2013). The high-wavenumber region contains more distinct information for identifying subcellular structure, i.e., the nucleus and cytoplasm, than the fingerprint region. Carvalho et al. analyzed the spectra of the nucleolus, nucleus, and cytoplasm of oral epithelial carcinomas (SCC-4), dysplasia (DOK) cell lines, and normal oral epithelial cells. The O–H bond from cell-membrane-bound water or intracellular fluid provided key information for distinguishing cell lines. The sensitivity and specificity of cytoplasm recognition were up to ~100% and 97%, respectively. Regarding nuclear recognition, the specificity was 99%. Compared to other methods, high-wavenumber regions provide more information about subcellular structures to distinguish between normal, premalignant, and malignant tissues (Carvalho et al., 2015; Carvalho et al., 2017).

In addition, improving the efficiency of the classifier can increase the distinguishing accuracy. Jeng et al. gathered the spectra from 36 normal and 44 OSCC tissues and analyzed them with PCA, followed by LDA or quantitative discrimination analysis (QDA). The accuracies of the PCA-LDA and PCA with QDA (PCA-QDA) classifiers were 81.25% (sensitivity: 77.27%, specificity: 86.11%) and 87.5% (sensitivity: 90.90%, specificity: 83.33%), respectively. PCA-QDA showed better classification efficiency than PCA-LDA (Jeng et al., 2019). Moreover, Cals et al. adopted PCA with (hierarchical) LDA (PCA-(h)LDA) (Cals et al., 2016). Because there is a large difference between lipid, nerve, and tumor tissue, researchers first differentiated lipids and nerves from other tissues and then distinguished OSCCs from the remaining components (squamous epithelium and connective tissue, muscle, and glands). Compared to the one-step PCA-LDA model, the two-step PCA-(h)LDA model presented higher accuracy (91% accuracy; sensitivity: 100%, specificity: 78%). However, the PCA-(h)LDA model also increased the data analysis workload.

Furthermore, focusing on biomarkers is also a way to distinguish between normal and abnormal tissues. Chen et al. compared Raman spectra decomposed by multivariate curve resolution with alternating least squares (MCR-ALS) with the standard Raman spectra of keratin, a well-known molecular marker of OSCC. However, some spectra were neither divided into the OSCC group nor the normal group, and as a result, with different classification methods of suspicious samples, the sensitivity differed widely (77% or 92%) (Chen et al., 2016). Recently, Fălămaş et al. attempted to differentiate six OSCC patients from five HV by salivary Raman spectra (Fălămaş et al., 2020). They noticed that thiocyanate is not only an indicator of smokers (Tsuge et al., 2000) but also related to cancer (Shiue, 2015). The characteristic band of thiocyanate is 2126 cm^–1, the intensity of which was higher in the cancer group than in the healthy group. The peak at 738 cm^–1 is another characteristic band of thiocyanate, contributing to the differentiation of the two groups by PCA. Fălămaş et al. also found that the peaks at 752, 884, 928, 989, and 1047 cm^–1, representing tryptophan, collagen, proline, and glycogen, respectively, contributed to finding the biggest difference between OSCC patients and HV in salivary spectra (Falamas et al.). Daniel et al. also concentrated on the presence of thiocyanate at 2108 cm^–1 in the SERS spectra of saliva among smokers (Daniel et al., 2020). However, due to the small sample size, whether thiocyanate can be a potent biomarker for recognizing cancer patients requires further exploration. S100 calcium-binding protein P (S100P) mRNA has been reported as a valid salivary biomarker for oral cancer detection without periodontitis interference (Y.-S. et al., 2017). Han et al. designed a sandwich assay format consisting of oligonucleotides, AuNPs as the SERS substrate, and malachite green isothiocyanate as a reporter molecule to quantify S100P mRNA in saliva (Han et al., 2019). The concentration of S100P mRNA was three times higher in the oral cancer group than in the healthy group.

The extent to which the cancerous tissue can be removed by a clinician (i.e., whether it can be completely removed or not) significantly affects the prognosis. It has been reported that the 5-year disease-free survival rate significantly declines in patients with inadequate surgical margins. Even after surgical treatment and radiotherapy or chemotherapy, malignant lesions may still recur, which directly affects the prognosis and patients’ quality of life. Thus, assessing surgical margins and the potential for oral cancer recurrence are of major importance. RS shows the application potential and unique advantages of these two aspects.

For instance, frozen section is an intraoperative choice to assess whether the surgical margin is adequate. This technique works well for soft tissues, but it is difficult to use it to assess bone edges (Nieberler et al., 2016). RS has the potential to distinguish between malignant and normal tissues. The Raman map composed of the spectra of all sites of a specimen is a candidate method for determining the surgical margins more accurately during operation. Barroso et al. found that the main factor in distinguishing a tumor from surrounding healthy tissue is water concentration (Barroso et al., 2015). Based on this, they constructed a 2D Raman map to observe the change in water concentration from tumor tissue to the surrounding healthy tissue. In 2016, Barroso et al. (Barroso et al., 2016) applied the ratio of 3390 cm^–1 (O–H stretching band of water) to 2935 cm^–1 (C–H stretching band of lipids and proteins) as an indicator of water concentration and then assigned different colors to various water concentrations for 2D Raman map construction. In the 4–6 mm transition region between the tumor and surrounding normal tissue, the water content in the tissue ranged from 76% ± 8% in the tumor tissue to 54% ± 24% in the surrounding healthy tissue. The 2D Raman map presented the water concentration transformation from the tumor tissue to the surrounding healthy tissue directly, providing valid information for clinicians to determine the surgical edge more accurately. Two years later, Barroso et al. further focused on the potential of RS to evaluate the margin of bone resection during OSCC mandibular resection (Barroso et al., 2018). They also assessed the water concentrations of healthy and tumor bone tissues (more than 3 mm from the tumor boundary), and a Manne–Whitney U-test revealed significant differences in water concentration between the two groups. Furthermore, they built a PCA-LDA model to distinguish two types of samples based on the intensity of 2800–3050 cm^–1 C–H stretching. For suspicious spectra proven by water concentration, the PCA-LDA model presented 95% accuracy (sensitivity: 95%, specificity: 87%). By assigning different colors to the tumor bone tissue and normal tissue, they completed a 2D Raman map. Therefore, Barroso et al. believed that RS is a candidate method for intraoperative surgical edge assessment. Compared to the reports from Barroso et al., Cals et al. adopted two-step PCA-(h)LDA to achieve higher accuracy (accuracy: 91%, sensitivity: 100%, specificity: 78%) (Cals et al., 2016). A 2D Raman map can also be constructed using the tumor posterior probability of each sub-site spectrum to present visual information.

RS can also sensitively detect minor changes in a sample, presenting the potential to predict the recurrence rate of oral cancer patients (Sahu et al., 2015; Malik et al., 2017). Carcinoembryonic antigen (CEA) has the potential to predict tumor recurrence, which has been confirmed in other types of cancers (Lumachi et al., 2008). The Raman spectra of sera can detect cancer-related proteins and DNA (Harris et al., 2010). In 2015, Sahu et al. reported that the Raman spectra of sera are related to OSCC recurrence (Sahu et al., 2015). They collected serum samples immediately before and 1 week after surgery. The PCA-LDA classifier could differentiate patients with or without tumor recurrence within 2 years, and the classification effectiveness was approximately 78%. It is worth mentioning that only postoperative serum spectra are related to tumor recurrence. However, serum spectra changes cannot identify the specific location of a tumor, but as a preliminary test to screen high-risk people, they can be effective. Malik et al. applied oral mucosa to predict tumor recurrence. They analyzed the Raman spectra of malignant and contralateral normal mucosa in OSCC patients using PCA-LDA and leave-one-out cross-validation (LOOCV). They focused on misclassification, and found a relationship with recurrence 2 years post-surgery (Malik et al., 2017). They included 57 patients with OSCCs (including tongue, cheek, mandible, molar posterior pad, hard plate, and mouth cancers). Eight of 41 patients with misclassified spectra and 2 of 16 with correct spectra classification had second primary cancer or recurrence 2 years after surgery. It can be concluded that the misclassified group presented 1.5 times greater recurrence risk than the correctly classified group. Tissues have changed at the molecular level before the visible malignant change appears, and Raman spectra can detect this minor transformation. The sensitivity and specificity of the classifier were 80% and 29.7%, respectively. Although the specificity was low, it is acceptable as a preliminary screening tool, and following studies may increase the specificity by increasing the sample size.