1.

Introduction

Head and neck malignancies, including laryngeal and nasopharyngeal cancers, are diseases with high mortality rates.¹ In East Asia and Africa, the rates of incidence and mortality due to nasopharyngeal carcinomas (NPCs) and laryngeal cancer are remarkably higher than many parts of the world.² The five-year survival rate decreases significantly due to delayed diagnosis and NPC symptoms that are generally present at late tumor-node-metastasis. Early identification and adequate preoperative assessment of both NPCs and laryngeal cancers allow functional preserving therapy (e.g., radiation therapy, chemotherapy, surgery, etc.) and are also critical to reducing the mortality rates of the patients. Yet, early diagnosis of nasopharyngeal and laryngeal precancer and early cancer is clinically challenging, even for experienced clinicians with the aid of conventional white-light reflectance (WLR) endoscopy (e.g., microlaryngoscopy, transnasal esophagoscopy).³ Positive identification of these lesions relies heavily on the visualization of gross morphological manifestations, which can be very subjective. Therefore, it would be of imperative clinical value to develop a real-time, biomolecular-sensitive optical diagnostic technology (“optical biopsy”) that can assist in the early detection of nasopharyngeal and laryngeal dysplasia and neoplasia during transnasal endoscopic inspections.

In recent years, near-infrared (NIR) Raman spectroscopy has shown promising results for the pointwise diagnosis and characterization of disease progression in various organs (e.g., gastrointestinal tracts,^4–13 oral cavity,¹⁴ nasopharynx^15,16, larynx,^15,17–19 lung,²⁰ cervix,^21,22 bladder,²³ skin,^24,25 and breast²⁶) with high biomolecular specificity. To date, NIR Raman spectroscopic studies on the nasopharynx and larynx have been limited to in vitro tissue Raman measurements due to the lengthy data acquisition times, as well as technical challenges in making miniaturized flexible fiber-optic Raman probes with high collection efficiencies while also effectively eliminating interferences from fluorescence and silica Raman signals.²⁷ The diagnostic sensitivities and specificities of $\sim 70\%$ to 95% have been achieved in differentiating between different laryngeal pathologic types (e.g., normal, dysplasia and carcinoma) in vitro using NIR Raman spectroscopy.¹⁹ Teh et al. reported a diagnostic sensitivity of 88.0% and specificity of 91.4% for detecting laryngeal carcinoma in vitro when coupled with random recursive partitioning ensemble techniques.¹⁸ Lau et al. confirmed the feasibility of nasopharyngeal and laryngeal cancer diagnosis in vitro.^16,17 Very recently, we have initiated a targeted study aiming to translate the Raman spectroscopy technology from laboratory into routine, real-time, clinical endoscopic diagnostics, and have succesfully demonstrated the applications of Raman endoscopy for real-time in vivo diagnosis of upper gastrointestinal (i.e., esophageal and gastric) dysplasia and neoplasia.^11–13 With our successful development of a 1.8-mm Raman fiber probe that can pass down the instrument channel of medical endoscopes,²⁷ this allowed us to evaluate the clinical merit of the transnasal Raman endoscopy technique for in vivo tissue characterization in the head and neck. Since the compositional and morphological profiles of different organs (i.e., larynx and nasopharynx) in the head and neck are highly functionally specialized and exhibit significant variations in anatomical and morphological properties (e.g., lymphoid tissues, vascularity, secretion, cartilage, etc.), there is a fundamental ambiguity to which extent one may account for inter-anatomical variabililty in developing efficient algorithms for in vivo nasopharyngeal and laryngeal tissue Raman diagnostics. In this study, we report for the first time the implementation of transnasal image-guided Raman endoscopy [i.e., WLR and narrowband imaging (NBI)] to directly assess and characterize distinctive Raman spectral properties of nasopharyngeal and laryngeal tissue in vivo. Raman spectral differences reflecting the distinct composition and morphology among the nasopharynx and larynx are further evaluated using multivariate techniques [i.e., principal components analysis (PCA) and linear discriminant analysis (LDA)].

2.

Materials and Methods

3.

Results

High-quality in vivo Raman spectra can routinely be acquired in the nasopharynx and larynx in real time during transnasal image-guided [i.e., WLR and NBI] endoscopic inspections. Figure 1 shows an example of in vivo raw Raman spectrum (a weak Raman signal superimposed on a large-tissue autofluorescence background) acquired from the PN with an acquisition time of 0.1 s at endoscopy. The background-subtracted tissue Raman spectrum with a signal-to-noise ratio (SNR) of $>10$ (see the inset of Fig. 1) can be obtained and displayed online during clinical endoscopic measurements. Figure 2 depicts the inter-subject in vivo mean Raman spectra $\pm 1$ standard deviation (SD) of normal nasopharyngeal [PN ($n=521$) and FOR ($n=157$)] and laryngeal tissues [LVC ($n=196$)] when the Raman probe is gently put in contact with the tissue under WLR/NB imaging guidance. Also shown is WLR images obtained from the corresponding anatomical locations. Prominent Raman bands associated with proteins and lipids are identified as shown in Table 1 with tentative biomolecular assignments.^{5,11,13,20,30} Figure 3 shows the intra-subject mean spectra $\pm 1$ SD of a randomly chosen subject. The in vivo tissue Raman spectra were found to be reproducible with diminutive inter- and intra-subject variances ($<10\%$) in the nasopharynx and larynx. Further Raman endoscopic testings indicate that the variability between different tissue sites within the PN is very subtle (Raman intensity variations of $<5\%$; data not shown). We also calculated difference spectra $\pm 1$ SD between different tissue types (i.e., PN-LVC, LV-FOR, and PN-FOR) as shown in Fig. 4, resolving the distinctive compositional and morphological profiles of different anatomical tissue sites at the biomolecular level. ANOVA revealed 12 prominent and broad Raman spectral subregions that showed significant variability ($p<0.0001$) among the three anatomical tissue sites centered at 812, 875, 948, 986, 1,026, 1,112, 1,254, 1,340, 1,450, 1,558, 1,655, and $1,745{\mathrm{cm}}^{-1}$, reconfirming the importance of characterizing the Raman spectral properties of nasopharynx and larynx toward accurate in vivo tissue diagnostics.

Fig. 1

Representative in vivo raw Raman spectrum acquired from FOR with 0.1 s during clinical endoscopic examination. The inset is the processed tissue Raman spectrum after removing the intense autofluorescence background.

Fig. 2

In vivo (inter-subject) mean Raman spectra $\pm 1$ SD of PN ($n=521$), FOR ($n=157$) and LVCs ($n=196$). Note that the mean Raman spectra are vertically displaced for better visualization. In vivo, fiber-optic Raman endoscopic acquisitions from PN (upper), FOR (mid), and LVCs (lower) under WLR and NB imaging guidance are also shown.

Fig. 3

In vivo (intra-subject) mean Raman spectra $\pm 1$ SD of PN ($n=18$), FOR ($n=18$) and LVCs ($n=17$). Note that the mean Raman spectra are vertically displaced for better visualization.

Fig. 4

Comparison of difference spectra $\pm 1$ SD of different anatomical tissue types (inter- subject): (PN–LVCs); (PN–FOR) and (LVC–FOR).

Table 1

Tentative assignments5,11,13,20,30 of molecule vibrations and biochemicals involved in Raman scattering of nasopharyngeal and laryngeal tissue.

To investigate the significance of potential confounding factors during transnasal endoscopy, we also measured in vitro Raman spectra of blood, saliva, and nasal mucus obtained from healthy volunteers, as shown in Fig. 5. The most prominent Raman bands in saliva and nasal mucus are at $1638{\mathrm{cm}}^{-1}$ ($v_2$ bending mode of water), whereas blood exhibits porphyrin Raman bands near 1560 and $1620{\mathrm{cm}}^{-1}$. ³¹To assess further the spectral differences among different tissues in the head and neck, a five-component PCA model based on ANOVA and Student’s $t$-test ($p<0.05$) accounting for 57.41% of the total variance (PC1: 22.86%; PC2: 16.16%; PC3: 8.13%; PC4 6.22% PC5: 4.04%) was developed to resolve the significant peak variations of different anatomical locations. Figure 6 shows the PC loadings revealing the Raman bands associated with proteins (i.e., 853, 940, 1004, 1265, 1450, and $1660{\mathrm{cm}}^{-1}$) and lipids (i.e., 1078, 1302, 1450, 1655, and $1745{\mathrm{cm}}^{-1}$). Figure 7 displays box charts of PCA scores for the different tissue types (i.e., PN, FOR, and LVC). The line within each notch box represents the median, and the lower and upper boundaries of the box indicate first (25.0% percentile) and third (75.0% percentile) quartiles, respectively. Error bars (whiskers) represent the 1.5-fold interquartile range. The $p$-values are also represented among different tissue types. Dichotomous PCA algorithms integrated with LDA provided the sensitivities of 77.0% ($401/521$), 68.8% ($132/192$) and specificities of 89.2% ($140/157$) and 76.0% ($396/521$) for differentiation between PN and FOR and between LVC and PN, respectively, using the leave-one-subject-out (LOSO), cross-validation method. Overall, these results demonstrate that Raman spectra of nasopharynx and larynx in the head and neck can be measured in vivo with transnasal endoscopy, and the diagnostic algorithms development should be tissue site–specific to ensure minimum algorithm complexity.

Fig. 5

In vitro Raman spectra of possible confounding factors from human body fluids (nasal mucus, saliva, and blood).

Fig. 6

PC loadings resolving the biomolecular variations among different tissues in the head and neck, representing a total of 57.41% (PC1: 22.86%; PC2: 16.16%; PC3: 8.13%; PC4 6.22% PC5: 4.04%) of the spectral variance.

Fig. 7

Box charts of the 5 PCA scores for the different tissue types (i.e., PN, FOR, and LVC). The line within each notch box represents the median, but the lower and upper boundaries of the box indicate the first (25.0% percentile) and third (75.0% percentile) quartiles, respectively. Error bars (whiskers) represent the 1.5-fold interquartile range. The $p$-values are also given among different tissue types.

4.

Discussion

Minimally invasive technologies, such as Raman spectroscopy, can greatly benefit in transnasal inspections of the larynx and nasopharynx in clinical endoscopy. In vitro studies have thoroughly demonstrated that NIR Raman spectroscopy is sensitive to carcinogenesis (i.e., precancer and cancer) in the head and neck, including the nasopharynx and larynx.^16–19 Direct translation of this technology into the clinic environment, however, remains very challenging, as efficient fiber probes, short measurement times, and online data processing and diagnosis are needed.²⁷ This study demonstrates for the first time the feasibility of Raman spectroscopy in transnasal endoscopic applications, providing the foundation for large-scale clinical studies in the head and neck. Our unique image-guided Raman endoscopy platform, integrated with a miniaturized fiber Raman probe, provides a rapid and minimally invasive assessment of endogenous tissue constituents of the head and neck at the molecular level during clinical endoscopic examination. This greatly facilitates clinicians to obtain detailed biomolecular fingerprints of tissue in the head and neck, reflecting the genuine compositional and morphological signatures without introducing the artifacts caused by vascular puncturing or tissue dehydration, morphological and anatomical effects, etc. Distinct Raman bands near $936{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}\text{-}\mathrm{C})$ proteins], $1004{\mathrm{cm}}^{-1}$ [${\nu }_s(\mathrm{C}\text{-}\mathrm{C})$ ring breathing of phenylalanine], $1078{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}\text{-}\mathrm{C})$ of lipids], $1265{\mathrm{cm}}^{-1}$ [amide III $v(\mathrm{C}-\mathrm{N})$ and $\delta (\mathrm{N}\text{-}\mathrm{H})$ of proteins], $1302{\mathrm{cm}}^{-1}$ [${\mathrm{CH}}_3{\mathrm{CH}}_2$ twisting and wagging of proteins], $1450{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$ deformation of proteins and lipids], $1618{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}=\mathrm{C})$ of porphyrins], $1655{\mathrm{cm}}^{-1}$ [amide I $v(\mathrm{C}=\mathrm{O})$ of proteins] are consistently observed in different anatomical sites of the nasopharynx and larynx (Figs. 1Fig. 2 to 3). The difference spectra $\pm 1$ SD (Fig. 4) reveal that the Raman-active tissue constituents are comparable among different anatomical sites, but the subtle (while highly molecular-specific) inter-anatomical variations are observed [e.g., relative tissue Raman (spectral shape, bandwidth, peak position and intensity) differences]. With the 785-nm laser light penetration depth in the vicinity of $\sim 800\mathrm{\mu m}$ in epithelial tissue,¹² it is plausible that Raman spectra of nasopharyngeal tissue reflect the lymphoid-rich mucosa and epithelia type (i.e., mostly stratified squamous epithelium). On the other hand, the distinct morphology in the FOR (i.e., cartilage) likely explains the spectral appearance associated with the pharyngeal recess. To investigate further the properties of inter-anatomical variability in the head and neck, PCA was employed to resolve the spectral variability. ANOVA and noise level was used to select the PCs in the model. The PCA modeling captured a total variation of 57.41% after mean-centering of the dataset (PC1: 22.86%; PC2: 16.16%; PC3: 8.13%; PC4 6.22% PC5: 4.05%) that were found to have significant different means [Fig. 7(a) to 7(e)]. Indeed, this suggests that the majority of the spectral variation is related to inter-anatomical variability. We anticipate that the remaining variance not accounted for is due to probe handling variations associated with in vivo Raman endoscopic trials. The loadings on PC1 and PC2 (Fig. 6) are generally associated with lipid signals {i.e., $1302{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_2$ twisting and wagging), $1440\mathrm{cm}-1$ [$\delta ({\mathrm{CH}}_2)$], $1655{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}=\mathrm{C})$] and $1745{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}=\mathrm{O})$]} suggesting that FOR exhibit fewer signals from lipids (Fig. 7). As both PC1 and PC2 components largely reflect variations in lipids, indicating that distinct lipid types might be associated with the FOR tissues. In contrast, PC3, PC4, and PC5 represent complex signals related to proteins {i.e., $853{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}\text{-}\mathrm{C})$], $936{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}\text{-}\mathrm{C})$], $1004{\mathrm{cm}}^{-1}$ [$v(\mathrm{C}\text{-}\mathrm{C})$], $1450{\mathrm{cm}}^{-1}$ [$\delta ({\mathrm{CH}}_2)$], and $1660{\mathrm{cm}}^{-1}$ [Amide I $v(\mathrm{C}=\mathrm{O})$]}, signifying that distinct anatomical locations are present with a highly specific compositional and morphological signature (Figs. 6 and 7). The two binary PCA-LDA classification algorithms with the LOSO method of cross validation provide the sensitivities of 77.0%, 67.3%, and specificities of 89.2% and 76.0% for differentiation between PN and FOR and between LVC and PN, respectively, reconfirming that the larynx and nasopharynx are unique organs. Therefore, the distinct Raman active biomolecules as well as morphology (e.g., mucosa thickness, cartilage, blood vessels, etc.) and optical properties together contribute to the complex spectral differences observed between different tissue sites in the nasopharynx and larynx. Overall, the subtle spectral differences of distinct anatomical sites observed would inevitably induce additional model complexity in developing efficient Raman diagnostic algorithms [e.g., PCA-LDA, PLS-DA, classification and regression trees (CART), etc.] ^5,16–19 for precancer and cancer diagnosis in the head and neck. To investigate further the influence of potential confounding factors at in vivo transnasal applications, the Raman spectra of body fluids (e.g., blood, saliva, and nasal mucus) in the head and neck were also measured in vitro (Fig. 5). Saliva and nasal mucus are mostly associated with broad Raman peak at $1638{\mathrm{cm}}^{-1}$ ($v_2$ bending mode of water), whereas the blood exhibits signals of porphyrins near 1560 and $1620{\mathrm{cm}}^{-1}$. Comparisons with the nasopharyngeal and laryngeal tissue Raman spectra acquired (Fig. 2) reveal that those biochemicals in the body fluids do not contribute significantly to the in vivo tissue Raman spectra in transnasal endoscopy.

The in vivo nasopharyngeal and laryngeal tissue Raman results presented in this report is of immense importance as it transfers the Raman technology from laboratory into in vivo, real-time, transnasal Raman endoscopy, paving the way for realizing early diagnosis and detection of the head and neck cancer and precancer, as well as post-therapeutic surveillance of head and neck cancer during clinical examination.^16–19 One notes that the pharyngeal recess (FOR) could be too deep to be evaluated visually under conventional wide-field imaging modalities (Fig. 2), resulting in blind/random biopsies with poor diagnostic accuracy.³ Our in vivo, fiber-optic Raman endoscopy technique can be used for multiple but instant biochemical assessments of the tissue in situ (Fig. 2) that could increase the diagnostic yield and ultimately improve the effectiveness of surveillance of epithelial lesions in the head and neck. Overall, the in vivo Raman spectra of the head and neck tissue are very similar, but with subtle spectral differences. This observation correlates well with our previous study defining the Raman spectral properties of the upper gastrointestinal tract (i.e., esophagus and gastric).¹¹ We have also established the Raman spectral profiles of dysplastic and neoplastic tissue in the lung, esophagus, stomach, and colon, which in fact exhibit comparable difference spectra (e.g., cancer tissue shows upregulated DNA and protein, but a relative reduction in lipid content) among different organs associated with neoplastic tissue transformation.^{11–13,20,32,33} We anticipate that similar spectral difference profiles could be found in different pathologic tissues of the head and neck, but a large Raman dataset is needed to evaluate the specific impact of cancerous transformation within the distinct anatomical sites in the nasopharynx and larynx. As such, in vivo, transnasal Raman endoscopic measurements on a larger number of patients are currently in progress to assess the specific biochemical foundations of the Raman spectral differences between different pathologic types of nasopharynx and larynx tissue for developing robust Raman diagnostic algorithms for in vivo tissue diagnostics and characterization in the head and neck.

In summary, this work demonstrates for the first time that transnasal, image-guided Raman endoscopy can be used to acquire in vivo Raman spectra from nasopharyngeal and laryngeal tissue in real time. Significant Raman spectral differences are identified, reflecting that the distinct compositions and morphology in the nasopharynx and larynx should be considered to be important parameters in the interpretation and rendering of diagnostic decision algorithms for in vivo tissue diagnosis and characterization in the head and neck.