Artificial Intelligence‐Optimized Image Biomarker Identification for Sinonasal Inverted Papilloma Diagnosis Using Image‐Enhanced Versus White‐Light Endoscopy

ABSTRACT

Objective

This study aims to compare image feature identification and diagnostic performance between artificial intelligence‐aided image‐enhanced endoscopy (using the Storz Professional Image Enhancement System, SPIES) and conventional white‐light endoscopy (WLE) for sinonasal inverted papilloma (SIP).

Methods

A total of 163 patients were investigated using Storz Professional Image Enhancement System (SPIES) endoscopy (SPIES Spectral A[SA]/Spectral B[SB]) and WLE. From each patient, one or two images (only in the presence of bilateral lesions) were obtained, and a large number of potential features were extracted from WLE, SA, and SB. Independent t‐tests and LASSO methods were used to feature selection. Machine learning (ML) models were trained and evaluated for each channel of images.

Results

The Random Forest (RF) algorithm on SPIES SB mode achieved superior performance: accuracy = 0.86, sensitivity = 0.89, specificity = 0.83, AUC = 0.86, and Kappa = 0.71. SB significantly outperformed SA (accuracy = 0.77) and WLE (accuracy=0.75). ROC analysis confirmed SB's diagnostic superiority (area under curve [AUC]: SB = 0.86 vs. SA = 0.77, WLE = 0.75).

Conclusions

SPIES SB mode demonstrated the highest diagnostic accuracy and AUC for SIP identification, highlighting its potential as a superior imaging modality. SPIES endoscopy exhibits superior diagnostic accuracy and predictive performance compared to conventional WLE in the diagnosis of SIP. However, the single‐center design and limited sample size necessitate validation through multicenter studies with larger cohorts.

Key Points

• SPIES SB mode demonstrated the highest diagnostic accuracy and AUC for SIP identification, highlighting its potential as a superior imaging modality.

• SPIES endoscopy exhibits superior diagnostic accuracy and predictive performance compared to conventional WLE in the diagnosis of SIP.

1. Introduction

Sinonasal inverted papilloma (SIP) ranks among the most frequently encountered benign sinonasal tumors, characterized by a propensity for local recurrence (10%–20%) and a 2%–15% risk of malignant transformation to squamous cell carcinoma (SCC) [1, 2, 3, 4]. White light endoscopy (WLE) remains the first‐line diagnostic modality for suspected SIP. While magnetic resonance imaging (MRI) demonstrates high predictive value through its characteristic convoluted cerebriform pattern identification, and computed tomography (CT) provides superior anatomical localization of lesion origin, both imaging techniques exhibit limited sensitivity in detecting early‐stage SIP manifestations and subtle pathological changes [5].

The advent of image‐enhanced endoscopy (IEE) systems, such as narrow band imaging (NBI) and Storz Professional Image Enhancement System (SPIES). Notably, the SPIES has revolutionized early SIP detection [6, 7]. Kamphuis et al. compared to conventional WLE, these technologies significantly enhance visualization of mucosal microvascular networks (e.g., capillary loop patterns) and three‐dimensional architectural distortions through advanced optical filtering [8]. In our previous study, SPIES demonstrated significantly higher diagnostic accuracy than conventional WLE (sensitivity: 91.7% vs. 66.7%; specificity: 95.5% vs. 88.6%), with the small lobular sign serving as a critical discriminative feature [6]. However, not all small lobular signs can be readily identified on the surface of the tumor; sometimes it is necessary to move a portion of the mass to identify a lobulated surface. Nevertheless, quantitative analysis of endoscopic texture features imperceptible to human visual interpretation remains unexplored.

Recent advances in medical image analysis have established texture‐based radiomics as a powerful paradigm across multiple imaging modalities, including CT [9], MRI [10], ultrasound [11], and conventional endoscopy [12]. Particularly in the era of artificial intelligence, machine learning (ML)‐driven radiomic approaches enable systematic correlation of subvisual imaging biomarkers with clinicopathological characteristics. Therefore, this study aimed to extract texture features from SIP and nasal polyps (NP) images, combining them with ML techniques to establish and validate a SIP prediction model.

2. Materials and Methods

2.1. Subjects

The data were retrospectively collected from 2018 to 2025 at the Department of Otolaryngology‐Head and Neck Surgery, Beijing Tongren Hospital, Capital Medical University, China. This study was approved by the Ethics Committee of our center (approval number: TRECKY2018‐022).

2.2. Inclusion and Exclusion Criteria

Inclusion criteria for the study cohort comprised: (1) Age greater than or equal to 18 years; (2) patients suspected of having SIP and NP. Exclusion criteria included: (1) Severe, active hemorrhagic disorders; (2) fungal rhinosinusitis; (3) sinonasal malignancies; (4) suboptimal endoscopic image quality; (5) secretions obscuring the mass surface.

Informed consent was waived by the ethics committee due to the retrospective nature of this study.

This study utilized a dataset comprising images from 163 patients (86 with SIP and 77 with NP). To minimize data correlation from individual patients, the image acquisition was limited to one or two per patient (the latter only in the presence of bilateral lesions). The collected images were processed into three independent datasets based on their color channels: white light endoscopy (WLE, n = 189), narrow‐band imaging A (SA, n = 194), and narrow‐band imaging B (SB, n = 192). Separate models were subsequently developed and evaluated for each channel‐specific dataset.

The same splitting strategy was applied independently to all three datasets: each was divided by random stratified sampling (60% training, 40% test), with a fixed random seed to guarantee reproducible and consistent partitions.

2.3. Instruments

The SPIES endoscopic system (commercially designated as IMAGE 1S) comprises 4‐mm rigid nasal endoscopes with angulation options of 0°, 30°, and 70°, integrated with the SPIES Connect (TC200) image processing unit and SPIES H3‐Link (TC300) camera control module (Karl Storz SE & Co. KG, Tuttlingen, Germany).

SPIES endoscopy is an advanced digital imaging platform employing full high‐definition (HD) technology with enhanced spectral processing capabilities. This system optimizes visible light spectrum acquisition through its specialized HD camera system, enabling precise color rendering for improved tissue differentiation. By selecting the optimal wavelength of light for imaging, those using SPIES endoscopy can enhance the appearance of mucosal and submucosal structures and associated vasculature.

Besides standard WLE mode, the system incorporates five distinct IEE modalities: Clara (tissue contrast enhancement), Chroma (chromatic amplification), Chroma+Clara (CC) composite mode, Spectra A (SA), and Spectra B (SB). These proprietary algorithms operate through spectral separation and recombination processes, generating enhanced color contrast that can reveal subtle pathological features imperceptible in conventional white light endoscopic imaging. SA and SB are two predefined modalities for virtual chromoendoscopy, specifically engineered to amplify hemoglobin absorption contrast through targeted wavelength optimization (400–570 nm for SA, 550‐630 nm for SB), enabling high‐definition mapping of microvascular architecture within sinonasal mucosal surfaces. SPIES SA is based mainly on the green and blue light spectral signals. SPIES SB mode is based on a color tone shift algorithm to reduce the dominant diffuse red spectral reflection, while still keeping the whole spectral information contained in the original color bands. This real‐time angioarchitectural visualization significantly enhances the detection of inflammatory hypervascularization and neovascular patterns characteristic of early‐stage lesions [8].

2.4. Endoscopic Examination

All endoscopic examination were performed by two senior surgeons using a 4 mm rigid endoscope (SPIES/IMAGE 1S) in the outpatient clinic or the operating room. Decongestant was not used to contract the nasal mucosa prior to examination. The endoscopic examination was performed initially using WLE, followed by SA and SB models. All images were digitally stored.

2.5. Pathological Diagnosis

Following SPIES endoscopic examination and image acquisition, all enrolled patients underwent standardized targeted biopsy procedures in either outpatient or operative settings. The study cohort was subsequently stratified into two distinct diagnostic groups: SIP and chronic rhinosinusitis with nasal polyps (CRSwNP).

2.6. Image Data Processing

The endoscopic images (WLE, SA, and SB) were decomposed into three different color spaces: HSV, RGB, and LAB matrices by using IMAGE J (Softonic International, Barcelona, Spain) software assisted by FIJI [13]. The region of interest (ROI) was outlined using freehand selection of the IMAGE J (Figure 1) by our attending physician (ZF Li) and confirmed and revised by the senior author (ZX Huang). The region of interest (ROI) of SIP and NP groups was manually delineated on endoscopic images by two otolaryngologists. All images (WLE: 189 images, SA: 194 images, SB: 192 images) were converted to nine color channels: a_star, b_star, blue, green, hue, L_star, red, sat and val (Figure 2).

Figure 1.

Representative endoscopic images of ROI. (A–C) Sinonasal inverted papilloma group, (A) is white light endoscopy, (B) is SPIES SA imaging, (C) is SPIES SB imaging; (D–F) Nasal polyp group, (D) is white light endoscopy, (E) is SPIES SA imaging, (F) is SPIES SB imaging (ROI, regions of interest; SA, spectral A; SB, spectral B; SPIES, Storz Professional Image Enhancement System).

Figure 2.

Processing of the images in the sinonasal inverted group. (A) ROI confirmation on WLE; (B) ROI confirmation on SPIES SA; (C) ROI confirmation on SPIES SB; (D) a_star; (E) b_star; (F) Blue; (G) Green; (H) Hue; (I) L_star; (J) Red; (K) Sat; (L) Val; (M) ROI Mask (ROI, regions of interest; SA, spectral A; SB, spectral B; SPIES, Storz Professional Image Enhancement System; WLE, white light endoscopy).

2.7. Feature Extraction

The images from nine color channels and masks (Figure 2) were input into PyCharm 2022.2.5 (JetBrains, Nusle, Czech Republic) for texture feature extraction. All the extracted features are saved in “csv” file format for further machine learning and feature analysis.

The dataset was divided into SIP and NP groups. Differential image features were identified using WLE, SA, and SB methods, with independent samples t‐tests for statistical analysis (p < 0.05).

2.8. Features Selection

Further feature selection was conducted via a least absolute shrinkage and selection operator (LASSO) regression algorithm [14] using Python code, processing data from two cohorts with WLE and two SPIES modes: SA and SB. The optimal λ value was selected using 10‐fold cross‐validation based on the minimum mean squared error (MSE). Subsequently, a groupwise analysis of the characteristics between the SIP group and the NP group was conducted. The selected features were used for the establishment of the next model.

2.9. Establishment of Models and Analysis of Feature Importance

Primary model training: Initial algorithm screening was conducted through fivefold cross‐validation on the preprocessed dataset, utilizing Python 3.9 (Python Software Foundation, Oregon, USA) and the scikit‐learn framework for standardized implementation.

Interpretability enhancement: The SHAP (SHapley Additive exPlanations) framework was systematically applied to quantify feature importance, identifying the most predictive biomarkers while improving model transparency.

Model performance was evaluated using accuracy, sensitivity, specificity, and receiver operating characteristic (ROC) curves, with the area under the curve (AUC) quantified for further analysis. The complete methodological workflow is illustrated in Figure 3.

Figure 3.

The research flowchart of the study. (AUC, area under the receiver operating characteristic curve; NP, nasal polyps; ROI, region of interest; SPIES, Storz Professional Image Enhancement System; SA, spectral A; SB, spectral B; SIP, sinonasal inverted papilloma; WLE, white light endoscopy).

2.10. Statistical Analysis

Statistical analyses were performed using SPSS 20.0 (IBM, USA). Intergroup differences were assessed using independent samples t‐tests, with two‐tailed p‐value < 0.05 considered as statistically significant. Age was included as a covariate in the model.

3. Results

3.1. Patient Cohort

A total of 163 patients (SIP: 86 and NP: 77) with 575 endoscopic images (WLE: 189, SA: 194 and SB: 192) were included for the study (Table 1). The mean age of the SIP and NP groups was (61.83 ± 8.21) years and (37.83 ± 8.09) years, respectively. In the SIP group, there were 51 males and 35 females, while the NP group comprised 41 males and 36 females.

Table 1.

The models of performance of random forest and comparison of the other machine learning algorithms in different digital images enhance system channel.

Algorithms	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC (%)	Kappa (%)
White light (SIP vs. NP)
Random forest	0.75	0.83	0.68	0.75	0.50
Logistic regression	0.72	0.77	0.68	0.73	0.45
K‐nearest neighbors	0.59	0.56	0.63	0.59	0.59
Linear discriminant analysis	0.64	0.55	0.71	0.63	0.26
Decision tree	0.66	0.63	0.68	0.66	0.32
Adaptive boosting	0.71	0.76	0.67	0.72	0.42
Multi‐layer perception	0.63	0.68	0.60	0.65	0.27
eXtreme gradient boosting	0.70	0.84	0.55	0.70	0.39
Light gradient boosting machine	0.72	0.79	0.66	0.72	0.45
Spectra A (SIP vs. NP)
Random forest	0.77	0.74	0.79	0.77	0.54
Logistic regression	0.73	0.68	0.79	0.73	0.46
K‐nearest neighbors	0.54	0.53	0.55	0.55	0.08
Linear discriminant analysis	0.65	0.64	0.68	0.66	0.31
Decision tree	0.68	0.77	0.60	0.69	0.37
Adaptive boosting	0.73	0.73	0.70	0.73	0.46
Multi‐layer perception	0.68	0.86	0.44	0.65	0.32
eXtreme gradient boosting	0.68	0.61	0.74	0.67	0.35
Light gradient boosting machine	0.72	0.68	0.76	0.72	0.43
Spectra B (SIP vs. NP)
Random forest	0.86	0.89	0.83	0.87	0.71
Logistic regression	0.70	0.56	0.83	0.69	0.39
K‐nearest neighbors	0.61	0.63	0.59	0.61	0.22
Linear discriminant analysis	0.70	0.63	0.74	0.69	0.38
Decision tree	0.79	0.68	0.90	0.79	0.58
Adaptive boosting	0.78	0.77	0.77	0.77	0.52
Multi‐layer perception	0.53	0.91	0.23	0.57	0.13
eXtreme gradient boosting	0.78	0.73	0.81	0.77	0.54
Light gradient boosting machine	0.75	0.74	0.74	0.75	0.55

Abbreviations: AUC, area under the curve; NP, nasal polyps; SIP, sinonasal inverted papilloma.

3.2. Features Selection

Preliminary screening identified 386 (WLE), 458 (SA), and 474 (SB) features with statistically significant differences (p < 0.05) between SIP and NP images. Respectively, the top 23, 25, and 28 most relevant features were selected by Lasso (Figure 4) to construct the machine learning models.

Figure 4.

The lambda figures (coefficients and MSE) of the Lasso feature selection procedure. (A, B) MSE and coefficients lambda figure of WLE; (C, D) MSE and coefficients lambda figure of SA; (E, F) MSE and coefficients lambda figure of SB. (MSE, minimum mean squared error; SA, spectral A; SB, spectral B; WLE, white light endoscopy).

3.3. Machine Learn Model and Features Identification

The performance metrics of the best machine learning models for WLE, SA, and SB are summarized in Table 1. Among these models across all three SPIES channels, the random forest (RF) algorithm demonstrated the best test performance. Notably, the SB channel achieved the highest predictive accuracy using the RF algorithm, with values of 0.86 (accuracy), 0.89 (sensitivity), 0.83 (specificity), 0.86 (AUC), and 0.71 (kappa). The ROC curve and confusion matrix of the best‐performing model (Random Forest, RF) were evaluated on the independent test set (Figure 5). Complete performance metrics for all 10 machine learning models in WLE, SA, and SB are presented in Table 1.

Figure 5.

The ROC curve (A, C, E) and confusion matrix (B, D, F) of the model tested using the independent test set. (ROC, receiver operating characteristic; SA, spectral A; SB, spectral B; WLE, white light endoscopy).

3.4. SHAP Analysis Results

The best model with the highest scores in SB was selected for the SHAP analysis. The “Sat (Saturation)_glrlm (gray level run length matrix)_Gray Level Non Uniformity Normalized” contributed the most to the model, followed by, “Red_glcm (gray‐level co‐occurrence matrix) _Inverse Variance”, “Hue_gldm (gray level dependence matrix)_DV (dependence variance)”, “Blue_glrlm (gray level run length matrix)_Gray Level Non Uniformity”, “b_star_Root Mean Squared”, “Green_glrlm_Run Entropy”, “Val_glrlm_Gray Level Non Uniformity Normalized”, “L_star_Robust Mean Absolute Deviation”, “Sat_glcm_Difference Average”, “Blue_gldm_LDHGLE (Large Dependence High Gray Level Emphasis)”, “L_star_gldm_DE (Dependence Entropy)”, “Sat_glszm_Zone Entropy”, “L_star_glcm_Sum Squares”, “Sat_glszm_Gray Level Variance”, “Val_gldm_LDHGLE”, “Val_Mean Absolute Deviation”, “Sat_glcm_idmn (Inverse Difference Moment Normalized)”, “Blue_glcm_Joint Entropy”, “Red_gldm_DV”, and “Hue_gldm_DNN(Deep Neural Network)” (Figure 6).

Figure 6.

Results of SHapley Additive exPlanations (SHAP) feature importance analysis. SHAP value (A) and mean SHAP value (B) from the random forest (RF) algorithm for SB.

4. Discussion

Recent developments in endoscopic techniques aim to improve our capability to detect pathological abnormalities in an optimal way. NBI uses a narrow‐band light to highlight contrast between different vascularized mucosal areas. SPIES SA and SB modalities modulate the spectral response characteristics within the imaging system, resulting in enhanced color contrast visualization [8]. This study pioneers systematic identification of key endoscopic texture features and establishes a preliminary SIP predictive model using three digital enhanced imaging modalities: WLE, SPIES SA, and SPIES SB modes. Machine learning evaluation revealed SPIES SB mode with optimized random forest achieved superior performance (accuracy: 85.71%, sensitivity: 89.19%, specificity: 82.50%, AUC: 0.86, κ = 0.715), significantly outperforming WLE (75.00%, 83.33%, 67.50%, 0.754, κ = 0.50) and SA (76.92%, 74.35%, 79.49%, 0.77, κ = 0.54). SB demonstrated substantial histopathological agreement versus moderate agreement in WLE/SA, highlighting its enhanced detection of microvascular irregularities and mucosal microstructure for SIP diagnosis.

It is hypothesized that the difference between SA and SB arises from specific color renderings used to pronounce spectral differentiation, as light penetration in tissue is wavelength‐dependent. SB reduces red reflection while preserving full spectral information, which may improve visualization of submucosal microvasculature. A previous study showed greater light absorbance at 100–200 μm tissue thicknesses, which could potentially explain why SB contrast appeared higher compared with SA [8]. We therefore speculate, pending further validation, that within an optimal range, SPIES SB mode enhances color contrast more effectively, enabling the capture of more discriminative image features and might ultimately yield superior predictive performance.

AI and machine learning techniques have been described for assistance of SIP diagnosis. Yui et al. [15] developed a diagnostic system using AI to identify SIP based on white light endoscopic videos from 53 patients with good accuracy (AUC of 0.87, accuracy of 0.84, false positive rate of 0.12, and false negative rate of 0.19). However, to our knowledge, there are no reports on the use of texture features from endoscopic images to predict SIP. In our study, we showed that the pathological findings can be distinguished by textures extracted from the nine color channels through decomposition of the white light and enhanced images.

There were 20 most important texture features found in the optimized SPIES SB mode in this study for predicting SIP, including Sat_glrlm_Gray Level Non Uniformity Normalized, Red_glcm_Inverse Variance, Hue_gldm_DV, Blue_glrlm_Gray Level Non Uniformity, b_star_Root Mean Squared. Here, Sat and glrlm mean saturation and gray level run length matrix. Sat_glrlm_Gray Level Non Uniformity Normalized means that tumors with irregular surfaces (e.g., lobulated or papillary morphology) may be captured by imaging, manifested as an elevated value of this feature. Red_glcm_Inverse Variance is an important parameter for quantifying the texture homogeneity of the red channel in tumor regions, and changes in its value can indirectly reflect the structural heterogeneity of tumors [16, 17]. Hue_gldm_DV is a radiomic feature that quantifies textural heterogeneity in medical imaging by combining color space analysis (Hue channel) and texture metrics derived from the Gray Level Dependence Matrix (GLDM) [18]. This feature means “coarse” or “clustered” pattern and homogeneity of the tumor.

The Blue_gldm_Gray Level Non Uniformity is a radiomic feature that measures nuclear or tissue heterogeneity in the blue channel of medical images using GLDM. b_star_Root Mean Squared quantifies intensity variability in the blue‐yellow channel of images, offering insights into tissue heterogeneity [19]. Blue_gldm_GLNU and b_star_RMS co‐elevation collectively indicate significant intratumoral nuclear heterogeneity and chromatic/architectural complexity.

Among the top five SHAP features, Red_gldm_InvSeraVariance was lower in inverted papilloma than in NP, reflecting loss of local texture homogeneity due to endophytic growth and cellular pleomorphism. Blue_gldm_DNV and Green_gldm_RunEntropy were elevated in papilloma, indicating irregular nuclear dependence patterns and complex vascular or inflammatory architecture. Val_gldm_GrayLevelNonUniformityNormalized captured extreme gray‐level skewness from intratumoral calcification or cystic changes, while L_star_RobustMeanAbsoluteDeviation highlighted marked luminance variability caused by interdigitating epithelial nests and stroma. Collectively, these features provide a radiomic signature of the chaotic, heterogeneous microstructure of inverted papilloma versus the relatively ordered, edematous background of NP.

The advantages of this study are threefold: First, we systematically compared WLE, SPIES SA, and SPIES SB modes for diagnosing SIP. This represents the first study to employ texture feature extraction, decomposing WLE, SPIES SA, and SPIES SB images into nine channels and extracting 76 features imperceptible to the human eye. Second, after feature selection using artificial intelligence algorithms, we developed diagnostic models for all three endoscopic modalities. Finally, we identified the optimal prediction model based on accuracy and ROC performance metrics.

This study utilized a hand‐crafted radiomics approach with classical machine learning. Given our modest cohort size, this decision was appropriate because deep learning methods typically require substantially larger datasets to achieve robust performance without overfitting. We prioritized model robustness and interpretability. Our objective was not only to differentiate SIP from NP but also to identify specific, interpretable texture features that may offer clinical insights. We acknowledge that the proposed mechanism for SB's superiority, while plausible and consistent with optical principles, remains speculative at this stage. Therefore, our conclusions are presented with appropriate caution, and the identified feature set should be viewed as a preliminary benchmark for future deep learning studies with expanded, multi‐center datasets.

This study has several limitations. First, its retrospective and single‐center design may introduce selection bias and limit the generalizability of our findings. Second, while the cohort size is modest, it is comparable to other preliminary studies in this niche domain. In addition, the significant age difference between the SIP and NP groups (61.8 vs. 37.8 years) may have confounded the textural features, as age itself can affect tissue properties. Future studies with age‐matched cohorts or statistical adjustments are needed to confirm our findings. Finally, the significant age difference between the SIP and NP groups, while reflective of the true epidemiology of these conditions, could be a potential confounder. In future studies, we plan to expand our dataset by including more samples from multiple centers to further optimize the model. Promising future directions include multimodal information fusion, deep learning‑based end‑to‑end learning, and a real‑time diagnostic assistant system.

The clinical translation pathway remains underdeveloped at this early stage. To move toward clinical applicability, future work should integrate the proposed texture feature pipeline into existing endoscopic workstations, ideally as a real‐time computer‐aided diagnosis module that highlights suspicious regions during live examination. This would require addressing computational latency, user interface design, and prospective validation in routine endoscopic practice. While our current findings are foundational, we recognize that workflow integration—for example, automatic image capture, feature extraction, and output of a probability score—will be essential before any clinical deployment.

5. Conclusions

In conclusion, we have developed a proof‐of‐concept AI model that shows promising discriminatory power for differentiating SIP from NP. SPIES SB mode demonstrated the highest diagnostic accuracy specifically for SIP identification, outperforming both SPIES SA and WLE. Nevertheless, given the single‐center, retrospective design and limited sample size, these findings are preliminary. Rigorous prospective external validation is therefore essential before the model can be considered for clinical application.

Author Contributions

Zhen‐Xiao Huang: design, data collection, surgery and preparation of manuscript; Cheng Li: design, data collection and surgery. Min Li: design, follow‐up data collection and surgery. Ying‐Hong Zhang: data collection and surgery. Gui‐Min Zhang: data collection and surgery. Qian Huang: data collection and surgery. Zu‐Fei Li: data analysis, preparation of manuscript and revision of manuscript. Michael T. Chang: design and preparation of manuscript. Bing Zhou: study design, data collection, review and revision of manuscript.

Ethics Statement

This study was approved by the Ethics Committee of Beijing Tongren Hospital (approval number: TRECKY2018‐022).

Conflicts of Interest

Professor Bing Zhou is a member of World Journal of Otorhinolaryngology – Head & Neck Surgery (WJOHNS) editorial board and is not involved in the peer review process of this article. The other authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.