A Deep Learning-Based Ni Classification System for Laryngeal NBI Images: A Multicenter Diagnostic Study Running title: DL for Ni Classification in Larynx

A Deep Learning-Based Ni Classification System for Laryngeal NBI Images: A Multicenter Diagnostic Study Running title: DL for Ni Classification in Larynx | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Deep Learning-Based Ni Classification System for Laryngeal NBI Images: A Multicenter Diagnostic Study Running title: DL for Ni Classification in Larynx Jie-Lin Huang, Li-Juan Li, Ji-Qing Zhu, Li-Zhou Dou, Xue Zhang, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8157582/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Background Laryngeal squamous cell carcinoma accounts for more than 95% of laryngeal tumors. Early diagnosis is crucial for function preservation and prognosis improvement. Narrow Band Imaging (NBI) technology and Ni classification system provide an important basis for early diagnosis; however, poor interobserver agreement limits its standardized clinical application. To develop and validate the first deep learning system (DL-Ni) for Ni classification of laryngeal NBI images, and to evaluate its effectiveness in improving diagnostic agreement among physicians with varying levels of experience. Methods This multicenter diagnostic study retrospectively collected 3,023 high-quality laryngeal NBI images to construct the dataset. A dual-branch collaborative learning architecture was developed, comprising a UNet + + semantic segmentation branch and an improved ResNet classification branch. A randomized controlled crossover experiment was conducted to evaluate the improvement in diagnostic agreement among 12 physicians with different levels of experience under AI assistance. Results The developed DL-Ni system demonstrated robust performance in both internal and external validations, with accuracy of 0.858 (95% CI: 0.821–0.895) and 0.827 (95% CI: 0.813–0.841), respectively. AI assistance significantly improved interobserver diagnostic agreement: the Fleiss’ κ value increased from 0.488 to 0.685 (P < 0.05) in the junior physician group, and from 0.621 to 0.791 (P < 0.05) in the expert group. Conclusion This study is the first to develop and validate an automated deep learning system for Ni classification of laryngeal NBI images. The system significantly improved interobserver diagnostic consistency, offering an effective tool and solution for the standardized clinical application of NBI technology. Narrow Band Imaging Ni classification deep learning laryngeal cancer interobserver agreement Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Laryngeal squamous cell carcinoma (LSCC) accounts for over 95% of malignant laryngeal tumors, with 5-year survival rates ranging from 80–90% for early-stage disease (T1-T2) to approximately 60% for advanced stage( 1 , 2 ). However, superficial lesions often present with non-specific changes under conventional white light imaging (WLI), such as mucosal roughness or erythema, which frequently lead to missed diagnoses. Narrow Band Imaging (NBI) enhances visualization of mucosal microvascular patterns using specific wavelengths (415nm and 540nm), demonstrating superior sensitivity and specificity in detecting early laryngeal malignancies( 3 – 7 ). The Ni classification system, established in 2011, standardizes NBI interpretation by categorizing intrapapillary capillary loop (IPCL) morphology into types I through V, providing objective criteria for diagnosing malignant and precancerous laryngeal lesions( 8 – 11 ). Despite clear morphological criteria, clinical adoption faces a critical challenge: significant inter-observer variability. Meta-analyses report NBI sensitivity ranging from 81% to 94% and specificity between 85% and 96%, with substantial heterogeneity across studies( 12 ). Multiple independent studies demonstrate modest interobserver agreement (κ values 0.40–0.58), far below the clinical application threshold (κ > 0.8)( 13 – 15 ). This inconsistency compromises diagnostic accuracy, potentially causing missed early malignancies or overtreatment of benign lesions. Improving NBI classification consistency has traditionally relied on standardized training and experience accumulation. However, even after systematic training courses, interobserver κ value improvements rarely exceed 0.1–0.15( 16 ). Moreover, high-quality educational resources remain concentrated in tertiary hospitals, limiting widespread NBI adoption in primary care settings( 17 ). Artificial intelligence (AI), particularly deep learning (DL), offers transformative potential in medical image analysis by identifying complex patterns and providing objective, reproducible interpretations. This technology presents a promising solution to inter-observer variability: DL models trained on expert-annotated NBI images can learn characteristic IPCL features, enabling standardized, automated classification. While existing studies demonstrate feasibility of DL in laryngeal image analysis, no research has systematically validated AI efficacy in resolving NBI interpretative discrepancies among physicians( 18 – 20 ). This study aims to develop a deep learning system for automatic Ni classification of laryngeal NBI images (DL-Ni system) and systematically evaluate its effectiveness as an assistive tool through randomized controlled human-AI collaborative experiments. By assessing whether AI assistance improves diagnostic accuracy and consistency across physicians with varying experience levels, we seek to standardize NBI interpretation, reduce diagnostic variability, and optimize early precision management of laryngeal cancer and precancerous lesions. Method This study was approved by the Institutional Review Board of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (Approval No. 22/454–3656), and written informed consent was obtained from all participants. The study comprised two components: a retrospective phase for model development and internal validation, and a prospective phase for human-AI collaborative evaluation. Data Collection and Study Population Laryngeal NBI images were retrospectively collected from the Cancer Hospital and Institute of the Chinese Academy of Medical Sciences (CHCAMS) between January 2008 and March 2022 for model development and internal validation. External validation data were obtained from the Shenzhen Hospital of the Chinese Academy of Medical Sciences (CHSZCAMS) and the Shanxi Provincial Cancer Hospital (SXPCH). All images and videos were captured using Olympus Medical Systems devices (CV-170/ENF-VT2, CV-290/BF-H290, and CV-260/BF-260). The inclusion criteria for images were as follows: 1) clear visualization of IPCL structures with good vascular contrast, and absence of motion blur or equipment artifacts; 2) uniform and moderate illumination without overexposure or underexposure affecting the observation of vascular patterns; 3) a lesion area occupying ≥ 30% of the image to ensure sufficient diagnostic information; and 4) availability of a histopathological gold standard for diagnostic confirmation. Exclusion criteria included: 1) technical issues such as blurring, inaccurate focus, or artifacts caused by device malfunction; 2) biological interference including hemorrhage covering ≥ 30% of the field of view, or secretions and necrotic tissue covering ≥ 20% of the field; 3) imaging parameter issues such as excessive or insufficient illumination leading to inadequate IPCL contrast; and 4) lack of a definitive pathological diagnosis or unclear pathological findings. Following stringent quality control screening, 3,308 high-quality NBI images from 1,421 patients were included after exclusion of 1,805 images (37.4%) from an initial pool of 4,828 candidate images. The primary reasons for exclusion were image artifacts (n = 699, 14.5%), biological interference (n = 808, 16.7%), and insufficient IPCL contrast (n = 298, 6.2%). Image annotation was performed independently by three senior NBI specialists, each with over 10 years of NBI experience and an annual examination volume exceeding 1,500 cases. Using the Ni classification criteria, the experts precisely delineated lesion regions and assigned classifications via the LabelMe software. The initial interobserver agreement among the experts yielded a Fleiss' κ of 0.780 (95% CI: 0.756–0.804). Discrepant cases (n = 156, 5.1%) were resolved through panel discussion and re-evaluation to reach a final consensus. The histopathological diagnosis served as the gold standard. Dataset Partition The internal dataset (sourced from CHCAMS) consisted of 3,023 images, while the external validation set (from CHSZCAMS and SXPCH) contained 285 images. The internal dataset was randomly partitioned at the patient level into training (2,144 images), validation (578 images), and internal testing (301 images) subsets in a ratio of 7:2:1. Stratified sampling was employed to ensure a balanced distribution of Ni classification types across all subsets. The distribution of the dataset is shown in Fig. 1 . Deep Learning Model Development We proposed a dual-branch architecture that jointly performs semantic segmentation and image classification, designed with two dedicated sub-networks: one for pixel-wise lesion area segmentation and the other for image-level lesion type classification. Each network was optimized to enhance feature representation capability for its respective task. For the semantic segmentation branch, we constructed a segmentation network based on the UNet + + framework, which utilizes nested skip connections to effectively integrate features across different semantic levels. To improve the model’s focus on critical regions, we incorporated the Convolutional Block Attention Module (CBAM), which adaptively emphasizes lesion-relevant features through channel and spatial attention while suppressing background interference. Deep supervision was introduced to enhance the discriminative power of intermediate features and facilitate gradient propagation. The loss function combines weighted Binary Cross-Entropy (BCE), Dice loss, and Focal loss, jointly optimizing pixel accuracy, region overlap, and hard example learning. For the image classification task, we designed a dual-path network comprising a main branch and a lightweight branch. The main branch employs a 17-layer ResNet to extract high-level semantic and fine-grained structural features. The lightweight branch consists of a 19-layer convolutional network integrated with a CBAM attention module, guiding the model to focus on globally salient regions while significantly reducing parameter count and improving inference efficiency. Features from both branches are independently extracted and subsequently integrated within a feature fusion module to produce the final classification output. The classification network receives input as cropped RGB images of size 224×224. During training, data augmentation techniques including normalization, rotation, and color jittering were applied. To address class imbalance, a class-weighted cross-entropy loss was adopted. A phased training strategy was employed: the first 65 epochs froze the main branch and updated only the lightweight branch and fusion module, while the subsequent 65 epochs unfroze the entire network for end-to-end fine-tuning, totaling 130 epochs. The AdamW optimizer was used with an initial learning rate of 1×10⁻³, coupled with a dynamic decay scheduling strategy to enhance convergence in later stages. The architectural framework of the model is depicted in Fig. 2 . Guman-AI Collaboration Experiment A prospective, fully crossed, multi-reader study was conducted to quantify the improvement in diagnostic consistency among physicians with the assistance of the AI system. A total of 12 physicians from three centers were recruited in a stratified manner: 1) the expert group (n = 6), comprising physicians with over 10 years of NBI experience and an annual operation volume exceeding 500 cases; and 2) the junior group (n = 6), consisting of physicians with 1–3 years of NBI experience. Each physician evaluated 200 representative images randomly selected from the external validation set, covering all Ni classification types. The evaluation was performed in two independent sessions: during the baseline phase, only the NBI images and basic clinical information were provided; in the AI-assisted phase, real-time outputs from the DL-Ni system—including probability distribution of classification, confidence score, and Grad-CAM-generated heatmaps highlighting IPCL key regions—were added. A washout period of at least 4 weeks was implemented between the two sessions to mitigate memory effects. All assessments were conducted under standardized lighting conditions using calibrated medical displays. The order of image presentation was fully randomized across sessions to eliminate sequence effects. The primary endpoint was the inter-observer agreement (Fleiss’ κ) in Ni classification with and without AI assistance. Secondary endpoints included diagnostic accuracy, sensitivity, and specificity within each physician group. Statistical Analysis Interobserver agreement was quantified using Fleiss’ κ with 95% confidence intervals. Model performance was evaluated in terms of sensitivity, specificity, accuracy, and receiver operating characteristic (ROC) curves, all reported with 95% CIs. The improvement in performance metrics before and after AI assistance was compared using paired t-tests or Wilcoxon signed-rank tests, as appropriate. All statistical analyses were performed using R version 4.3.3 and Python 3.9. A two-sided p-value < 0.05 was considered statistically significant. Results Data Distribution and Baseline Characteristics Following rigorous quality control screening, a total of 3,308 high-quality NBI images were included in this study. The distribution across Ni classification types was as follows: Type I, 333 images (10.1%); Type II, 231 images (7.0%); Type III, 1,158 images (35.0%); Type IV, 135 images (4.1%); Type Va, 745 images (22.5%); Type Vb, 338 images (10.2%); and Type Vc, 368 images (11.1%). From CHCAMS, 1,270 patients were enrolled, constituting an internal dataset of 3,023 images. An external validation set of 285 images was formed from 151 patients recruited from CHSZCAMS and SXPCH. Detailed clinical characteristics of each dataset are summarized in Table 1 . Table 1 Summary of Demographic and Clinical Characteristics Across Study Cohorts CHCAMS Cohorts (Internal) External Cohorts Characteristic Train (n = 873) Validation (n = 269) Test (n = 128) SXPCH (n = 77) CHSZCAMS (n = 74) Age, Mean (± SD), years 46.6 ± 12.3 46.1 ± 12.4 47.1 ± 13.2 47.4 ± 13.9 47.8 ± 12.9 Female, n (%) 501 (41.7%) 165 (40.2%) 70 (42.3%) 51 (44.1%) 44 (45.3%) Male, n (%) 372 (53.2%) 104 (55.5%) 49 (54.8%) 26 (53.9%) 30 (51.5%) NBI category, n (%) All images 2144 578 301 139 146 I 211 (9.84%) 57 (9.86%) 30 (9.97%) 12 (8.63%) 23 (15.75%) II 154 (7.18%) 41 (7.09%) 18 (5.98%) 10 (7.19%) 8 (5.48%) III 753 (35.12%) 205 (35.47%) 102 (33.89%) 47 (33.81%) 51 (34.93%) IV 99 (4.62%) 22 (3.81%) 10 (3.32%) 3 (2.16%) 1 (0.68%) Va 464 (21.64%) 133 (23.01%) 80 (26.58%) 30 (21.58%) 38 (26.03%) Vb 222 (10.35%) 61 (10.55%) 26 (8.64%) 17 (12.23%) 12 (8.22%) Vc 241 (11.24%) 59 (10.21%) 35 (11.63%) 20 (14.39%) 13 (8.90%) Model Performance Evaluation On the internal test dataset, the model demonstrated the following performance: overall accuracy 0.858 (95% CI: 0.821–0.895), sensitivity 0.809 (0.799–0.819), specificity 0.974 (0.965–0.983), AUC 0.967 (0.960–0.974) (Fig. 3 ), and mIoU 0.731 (0.711–0.751). For Type Va lesions, the accuracy was 0.940 (0.929–0.951) and sensitivity 0.873 (0.866–0.890). The model exhibited strong generalization capability on the external validation set, achieving an overall accuracy of 0.827 (0.813–0.841), specificity of 0.770 (0.765–0.775), sensitivity of 0.969 (0.958–0.980), AUC of 0.961 (0.955–0.967), and mIoU of 0.719 (0.704–0.734). For Type Va lesions in the external set, accuracy reached 0.918 (0.911–0.925) with a sensitivity of 0.836 (0.830–0.842). The classification performance across different center-specific datasets is detailed in Table 2 . Figure 4 depicts the segmentation and prediction results of the DL-Ni system, along with corresponding heatmap visualizations. Table 2 Evaluation of DL-Ni's Diagnostic and Segmentation Performance in Independent Cohorts Validation Set Category Accuracy (95% CI) Sensitivity (95% CI) Specificity (95% CI) AUC (95% CI) IoU (95% CI) Internal validation All category 0.858 (0.821–0.895) 0.809 (0.799–0.819) 0.974 (0.965–0.983) 0.967 (0.960–0.974) 0.731 (0.711–0.751) Va 0.940 (0.929–0.951) 0.873 (0.866–0.890) 0.959 (0.955–0.964) 0.957 (0.951–0.963) 0.726 (0.708–0.742) External validation All category 0.827 (0.813–0.841) 0.770 (0.765–0.775) 0.969 (0.958–0.980) 0.961 (0.955–0.967) 0.719 (0.704–0.734) Va 0.918 (0.911–0.925) 0.836 (0.830–0.842) 0.943 (0.937–0.949) 0.963 (0.959–0.967) 0.721 (0.710–0.732) Real-time validation All category 0.827 (0.811–0.843) 0.751 (0.742–0.760) 0.967 (0.953–0.981) 0.971 (0.961–0.981) 0.713 (0.698–0.728) Va 0.915 (0.899–0.931) 0.830 (0.822–0.838) 0.947 (0.933–0.961) 0.968 (0.955–0.981) 0.714 (0.695–0.733) To assess the model’s potential for real-time application, performance was evaluated on a test set containing 200 video clips. The model achieved an overall accuracy of 0.827 (0.811–0.843), sensitivity of 0.751 (0.742–0.760), specificity of 0.967 (0.953–0.981), AUC of 0.971 (0.961–0.981), and mIoU of 0.713 (0.698–0.728). For Type Va lesions, accuracy was 0.915 (0.899–0.931) and sensitivity 0.830 (0.822–0.838), confirming the model’s diagnostic capability in dynamic imaging scenarios. Demonstrations of real-time detection and Ni-type probability prediction during laryngoscopy are provided in Supplementary Videos S1 and S2. Impact of AI Assistance on Physician Diagnostic Agreement Results from the human-AI collaboration experiment demonstrated that AI assistance significantly improved the diagnostic performance of both physician groups. In the junior group, the diagnostic accuracy increased from 0.661 to 0.801 with AI support, and interobserver agreement (κ value) improved from 0.488 (moderate agreement) to 0.685 (substantial agreement) (P < 0.05). Among experts, diagnostic accuracy rose from 0.787 to 0.885, while the κ value increased from 0.621 (substantial agreement) to 0.791 (substantial agreement) (P < 0.05). For the clinically critical Type Va lesions, AI assistance also led to a marked improvement in recognition accuracy: the junior group improved from 0.652 to 0.834 (P < 0.05), and the expert group increased from 0.759 to 0.891 (P < 0.05). Detailed comparative data are presented in Table 3 . Table 3 Comparison of Sensitivity, Specificity, Accuracy, and Kappa for Junior and Senior Endoscopists With Versus Without AI Assistance Group Condition Accuracy (95% CI) Sensitivity (95% CI) Specificity (95% CI) Kappa P value (vs Without AI) Junior Without AI 0.661 (0.630–0.692) 0.658 (0.625–0.691) 0.943 (0.939–0.949) 0.488 < 0.05 With AI 0.801 (0.765–0.837) 0.801 (0.762–0.840) 0.967 (0.961–0.972) 0.685 Senior Without AI 0.787 (0.767–0.806) 0.784 (0.758–0.810) 0.964 (0.961–0.968) 0.621 < 0.05 With AI 0.885 (0.850–0.921) 0.882 (0.842–0.922) 0.981 (0.975–0.987) 0.791 Discussion This study successfully developed and validated a deep learning system (DL-Ni) for the automated Ni classification of laryngeal NBI images. The system demonstrated excellent diagnostic performance in both internal and external validation (accuracy: 0.858 and 0.827; AUC: 0.967 and 0.961), with particularly high sensitivity in identifying Type Va lesions (0.873 and 0.836). Through a rigorously designed human-AI collaboration experiment, we confirmed that the DL-Ni system significantly improves interobserver agreement in Ni classification among practitioners with varying experience levels, with κ values increasing by 0.197 and 0.170 in the junior and expert groups, respectively, reaching moderate to substantial agreement. This effectively addresses a core bottleneck hindering the standardized clinical application of NBI technology. The high subjectivity in NBI interpretation—specifically, the variability among physicians in judging the morphological features of intrapapillary capillary loops (IPCLs)—has remained a major challenge limiting its widespread clinical adoption( 21 – 24 ). This interobserver variability directly leads to inconsistent diagnostic results, ultimately affecting the accuracy of clinical decision-making. The DL-Ni system developed in this study represents a fundamental shift from the traditional subjective interpretation model toward a standardized and objective analytical paradigm. Through the automatic extraction and quantitative analysis of IPCL morphological features by deep convolutional networks, it transforms traditional qualitative descriptions into quantitative parameters. The system's outputs—including probability distribution of classification, confidence scores, and Grad-CAM-generated heatmaps—provide physicians with a transparent navigation tool for diagnostic decision-making, converting a process traditionally dependent on personal experience into one based on objective data and visual evidence( 25 ). Compared to previous AI studies on laryngeal NBI, this work offers multiple innovations: it is the first to systematically develop and validate a deep learning system for automatic Ni classification of laryngeal NBI images; through a rigorous human-AI collaboration experiment, we quantitatively evaluated for the first time the improvement in inter-physician diagnostic consistency with AI assistance, demonstrating that its effect significantly surpasses traditional training methods; and the visual explanations and quantitative confidence levels enhance the transparency and interpretability of the diagnostic process, which is crucial for clinical acceptance and adoption of AI-assisted diagnosis( 26 – 28 ). This study has several limitations that should be addressed in future work. First, although multi-center external validation was performed, the training data were primarily acquired from Olympus endoscopy systems. Differences in spectral characteristics and image quality among devices from different manufacturers may exist; this limitation could be mitigated through cross-device generalizability studies utilizing techniques such as domain adaptation to improve model performance across devices with varying spectral and image characteristics. Second, although model performance was tested on a video dataset, its ability to capture subtle mucosal movements in real-time during dynamic imaging remains to be enhanced; this could be addressed through optimization for real-time video stream analysis, developing algorithms that can capture IPCL morphological changes and hemodynamic features in real-time under conditions of subtle mucosal movement. Third, the histopathological gold standard used in this study may be affected by biopsy sampling errors, especially for heterogeneously distributed lesions; this challenge could be addressed through multimodal integrated diagnosis, exploring comprehensive models that combine NBI with other imaging modalities to compensate for the limitations of NBI in detecting deep tissue infiltration. Additionally, long-term, multi-center prospective cohort studies are needed to systematically evaluate the impact of AI-assisted NBI classification on clinical treatment decisions and patient outcomes. In conclusion, this study demonstrates that the DL-Ni system not only achieves high diagnostic accuracy in Ni classification of laryngeal NBI images but, more importantly, significantly improves interobserver agreement among practitioners with varying experience levels. This breakthrough provides a novel technological pathway for the early and precise diagnosis of laryngeal cancer and precancerous lesions and is expected to promote the standardized use of NBI in clinical practice, ultimately contributing to organ-preserving treatment and improving long-term patient outcomes. Conclusion This study successfully developed and validated the first deep learning-based automated classification system (DL-Ni) for laryngeal NBI images and demonstrated its significant effectiveness in improving diagnostic agreement among physicians with varying levels of experience. The system achieved high-precision automatic classification of NBI images (accuracy exceeding 85%), and more importantly, markedly enhanced inter-observer diagnostic agreement, with κ values increasing by 0.17–0.20. These results provide an effective technical solution to overcome the key bottleneck in the standardized clinical application of NBI technology. Our work offers a new technological pathway for the early and precise diagnosis of laryngeal cancer and precancerous lesions, with strong potential to promote the standardized adoption of NBI in clinical practice. Ultimately, this approach is expected to improve patient outcomes and healthcare quality. With continued technical optimization and further clinical validation, AI-assisted Ni classification could become a standardized tool in the diagnosis of laryngeal lesions. Declarations Ethics approval and consent to participate This study was approved by the Institutional Review Board of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (Approval No. 22/454–3656), and written informed consent was obtained from all participants. Consent for publication Not applicable. Competing interests The authors declare that they have no competing interests. Funding This work was supported by National High Level Hospital Clinical Research Funding (grant number LC2024A04) and CAMS Innovation Fund for Medical Sciences (CIFMS) (grant number 2022-I2M-C&T-B-059). Author Contribution Jie-Lin Huang and Li-Juan Li contributed equally to this work. Conceptualization: Xiao-Guang Ni, Jian-Hui Wang. Methodology: Jie-Lin Huang, Li-Juan Li, Ji-Qing Zhu. Software: Jie-Lin Huang, Li-Zhou Dou. Validation: Xue Zhang, Yu-Meng Liu, Yan Ke. Formal analysis: Yu-Da Zhao, Mei-Ling Wang. Investigation: All authors. Resources: Xiao-Guang Ni, Quan-Mao Zhang, Jian-Hui Wang. Data Curation: Li-Juan Li, Ji-Qing Zhu. Writing – Original Draft: Jie-Lin Huang, Li-Juan Li. Writing – Review & Editing: All authors. Visualization: Li-Zhou Dou, Xue Zhang. Supervision: Xiao-Guang Ni. Project administration: Xiao-Guang Ni. Funding acquisition: Xiao-Guang Ni. All authors read and approved the final manuscript. Acknowledgements Not applicable. Data Availability The datasets used and analyzed during the current study are available from the corresponding author on reasonable request. The source code of the deep learning model is not publicly available at this time to protect intellectual property prior to patent filing but may be made available under a material transfer agreement. References Chu EA, Kim YJ. Laryngeal cancer: diagnosis and preoperative work-up. Otolaryngol Clin North Am. 2008;41(4):673–95, v. Marioni G, Marchese-Ragona R, Cartei G, Marchese F, Staffieri A. Current opinion in diagnosis and treatment of laryngeal carcinoma. Cancer Treat Rev. 2006;32(7):504–15. Bertino G, Cacciola S, Fernandes WB, Jr., Fernandes CM, Occhini A, Tinelli C, et al. Effectiveness of narrow band imaging in the detection of premalignant and malignant lesions of the larynx: validation of a new endoscopic clinical classification. Head Neck. 2015;37(2):215–22. Staníková L, Šatanková J, Kučová H, Walderová R, Zeleník K, Komínek P. The role of narrow-band imaging (NBI) endoscopy in optical biopsy of vocal cord leukoplakia. Eur Arch Otorhinolaryngol. 2017;274(1):355–9. Westra JM, Zwakenberg MA, Halmos GB, van der Laan B, van der Vegt B, Plaat BEC. Narrow band imaging reveals field cancerisation undetected by conventional white light: Optical diagnosis versus histopathology. Clin Otolaryngol. 2024;49(4):429–35. Galli J, Settimi S, Mele DA, Salvati A, Schiavi E, Parrilla C, et al. Role of Narrow Band Imaging Technology in the Diagnosis and Follow up of Laryngeal Lesions: Assessment of Diagnostic Accuracy and Reliability in a Large Patient Cohort. J Clin Med. 2021;10(6). Ni XG, Wang GQ. The Role of Narrow Band Imaging in Head and Neck Cancers. Curr Oncol Rep. 2016;18(2):10. Ni XG, He S, Xu ZG, Gao L, Lu N, Yuan Z, et al. Endoscopic diagnosis of laryngeal cancer and precancerous lesions by narrow band imaging. J Laryngol Otol. 2011;125(3):288–96. Staníková L, Kántor P, Fedorová K, Zeleník K, Komínek P. Clinical significance of type IV vascularization of laryngeal lesions according to the Ni classification. Front Oncol. 2024;14:1222827. Nerurkar NK, Sarkar A. Correlation of narrow-band imaging findings using the Ni and European Laryngeal Society classification systems during in-office flexible laryngoscopy with histopathology. J Laryngol Otol. 2024;138(2):203–7. Fang Y, Yang Y, Chen M, Chen J, He P, Cheng L, et al. Correlating intraepithelial papillary capillary loops of vocal cord leukoplakia with histopathology. Acta Otolaryngol. 2022;142(1):106–11. Sanda IA, Hainarosie R, Ionita IG, Voiosu C, Ristea MR, Zamfir Chiru Anton A. A Systematic Review Evaluating the Diagnostic Efficacy of Narrow-Band Imaging for Laryngeal Cancer Detection. Medicina (Kaunas). 2024;60(8). Scholman C, Zwakenberg MA, Wedman J, Wachters JE, Halmos GB, van der Laan B, et al. The influence of clinical experience on reliable evaluation of pharyngeal and laryngeal lesions: comparison of high-definition laryngoscopy using narrow band imaging with fibre-optic laryngoscopy. J Laryngol Otol. 2024;138(4):425–30. Kraus F, Gehrke S, Ehrmann-Müller D, Hofer F, Shehata-Dieler W, Hagen R, et al. Comparison of three different image enhancement systems for detection of laryngeal lesions. J Laryngol Otol. 2024;138(1):105–11. Pietruszewska W, Morawska J, Rosiak O, Leduchowska A, Klimza H, Wierzbicka M. Vocal Fold Leukoplakia: Which of the Classifications of White Light and Narrow Band Imaging Most Accurately Predicts Laryngeal Cancer Transformation? Proposition for a Diagnostic Algorithm. Cancers (Basel). 2021;13(13). Ni XG, Wang GQ, Hu FY, Xu XM, Xu L, Liu XQ, et al. Clinical utility and effectiveness of a training programme in the application of a new classification of narrow-band imaging for vocal cord leukoplakia: A multicentre study. Clin Otolaryngol. 2019;44(5):729–35. Żurek M, Rzepakowska A, Osuch-Wójcikiewicz E, Niemczyk K. Learning curve for endoscopic evaluation of vocal folds lesions with narrow band imaging. Braz J Otorhinolaryngol. 2019;85(6):753–9. Esmaeili N, Sharaf E, Gomes Ataide EJ, Illanes A, Boese A, Davaris N, et al. Deep Convolution Neural Network for Laryngeal Cancer Classification on Contact Endoscopy-Narrow Band Imaging. Sensors (Basel). 2021;21(23). Xiong H, Lin P, Yu JG, Ye J, Xiao L, Tao Y, et al. Computer-aided diagnosis of laryngeal cancer via deep learning based on laryngoscopic images. EBioMedicine. 2019;48:92–9. Azam MA, Sampieri C, Ioppi A, Africano S, Vallin A, Mocellin D, et al. Deep Learning Applied to White Light and Narrow Band Imaging Videolaryngoscopy: Toward Real-Time Laryngeal Cancer Detection. Laryngoscope. 2022;132(9):1798–806. Arthur C, Huangfu H, Li M, Dong Z, Asamoah E, Shaibu Z, et al. The Effectiveness of White Light Endoscopy Combined With Narrow Band Imaging Technique Using Ni Classification in Detecting Early Laryngeal Carcinoma in 114 Patients: Our Clinical Experience. J Voice. 2023. Chabrillac E, Dupret-Bories A, Vairel B, Woisard V, De Bonnecaze G, Vergez S. Narrow-Band Imaging in oncologic otorhinolaryngology: State of the art. Eur Ann Otorhinolaryngol Head Neck Dis. 2021;138(6):451–8. Kumazawa Y, Ikenoyama Y, Takamatsu M, Kido K, Namikawa K, Tokai Y, et al. Differences in Clinical Characteristics Between Missed and Detected Laryngopharyngeal Cancers. J Gastroenterol Hepatol. 2025;40(8):2037–45. Nogués-Sabaté A, Aviles-Jurado FX, Ruiz-Sevilla L, Lehrer E, Santamaría-Gadea A, Valls-Mateus M, et al. Intra and interobserver agreement of narrow band imaging for the detection of head and neck tumors. Eur Arch Otorhinolaryngol. 2018;275(9):2349–54. Jahmunah V, Ng EYK, Tan RS, Oh SL, Acharya UR. Explainable detection of myocardial infarction using deep learning models with Grad-CAM technique on ECG signals. Comput Biol Med. 2022;146:105550. Inaba A, Hori K, Yoda Y, Ikematsu H, Takano H, Matsuzaki H, et al. Artificial intelligence system for detecting superficial laryngopharyngeal cancer with high efficiency of deep learning. Head Neck. 2020;42(9):2581–92. Li Y, Gu W, Yue H, Lei G, Guo W, Wen Y, et al. Real-time detection of laryngopharyngeal cancer using an artificial intelligence-assisted system with multimodal data. J Transl Med. 2023;21(1):698. Yumii K, Ueda T, Kawahara D, Chikuie N, Taruya T, Hamamoto T, et al. Artificial intelligence-based diagnosis of the depth of laryngopharyngeal cancer. Auris Nasus Larynx. 2024;51(2):417–24. Additional Declarations No competing interests reported. Supplementary Files SupplementaryVideos1.mp4 SupplementaryVideos2.mp4 Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 16 Feb, 2026 Reviews received at journal 18 Jan, 2026 Reviews received at journal 15 Jan, 2026 Reviewers agreed at journal 21 Dec, 2025 Reviewers agreed at journal 19 Dec, 2025 Reviewers agreed at journal 18 Dec, 2025 Reviewers agreed at journal 18 Dec, 2025 Reviewers invited by journal 16 Dec, 2025 Editor assigned by journal 24 Nov, 2025 Submission checks completed at journal 24 Nov, 2025 First submitted to journal 19 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8157582","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":561181841,"identity":"d061d77e-4ad4-4095-9ff9-93f29a375a92","order_by":0,"name":"Jie-Lin Huang","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Jie-Lin","middleName":"","lastName":"Huang","suffix":""},{"id":561181842,"identity":"72911663-cf2d-4992-a827-6bc9a2c43a17","order_by":1,"name":"Li-Juan Li","email":"","orcid":"","institution":"The People's Hospital of Wenshan Prefecture","correspondingAuthor":false,"prefix":"","firstName":"Li-Juan","middleName":"","lastName":"Li","suffix":""},{"id":561181843,"identity":"21451dd9-f866-4e3e-a0c5-cce69317c74e","order_by":2,"name":"Ji-Qing Zhu","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Ji-Qing","middleName":"","lastName":"Zhu","suffix":""},{"id":561181844,"identity":"93b7e512-fa53-437b-a8f8-abd480dd73bf","order_by":3,"name":"Li-Zhou Dou","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Li-Zhou","middleName":"","lastName":"Dou","suffix":""},{"id":561181845,"identity":"e45efde0-1abe-4b7a-a7c7-53b46b2cb958","order_by":4,"name":"Xue Zhang","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Xue","middleName":"","lastName":"Zhang","suffix":""},{"id":561181846,"identity":"5b293f62-e37e-42ea-ad24-57e8512b1412","order_by":5,"name":"Yu-Meng Liu","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Yu-Meng","middleName":"","lastName":"Liu","suffix":""},{"id":561181847,"identity":"f7547af2-7c8e-4c0d-93bd-0dc16b21e458","order_by":6,"name":"Yan Ke","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Ke","suffix":""},{"id":561181848,"identity":"c9ca01da-0624-4540-92fd-010fad00b10a","order_by":7,"name":"Yu-Da Zhao","email":"","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Yu-Da","middleName":"","lastName":"Zhao","suffix":""},{"id":561181849,"identity":"336411c4-8332-4fec-b401-4d848d9a48b9","order_by":8,"name":"Mei-Ling Wang","email":"","orcid":"","institution":"National Cancer Center, National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Mei-Ling","middleName":"","lastName":"Wang","suffix":""},{"id":561181851,"identity":"1cb2a6b0-1235-4d06-b71c-97b5955edfdd","order_by":9,"name":"Jian-Hui Wang","email":"","orcid":"","institution":"Shanxi Province Cancer Hospital, Chinese Academy of Medical Sciences, Shanxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jian-Hui","middleName":"","lastName":"Wang","suffix":""},{"id":561181854,"identity":"7c330697-1363-44ff-8166-ca90cb856d4a","order_by":10,"name":"Quan-Mao Zhang","email":"","orcid":"","institution":"Shanxi Province Cancer Hospital, Chinese Academy of Medical Sciences, Shanxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Quan-Mao","middleName":"","lastName":"Zhang","suffix":""},{"id":561181855,"identity":"04f20984-bfc9-4ac9-8963-4679bb479aca","order_by":11,"name":"Xiao-Guang Ni","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA30lEQVRIiWNgGAWjYDACZgaGA1Am44OEChvStDAbPDiTRpqFbJIP2w4RVmbOzmN44EdFbWL/7PZrFQlsBxj427sT8GqxbGZLONhz5njijDtnym4k8NxhkDhzdgNeLQaHmQ8c4G07lthwIyftRoLEMwYDiVxCWhgbDv4FapkP1FKQYHCYGC3MBw7zttUkbriRfowhIYEoLWwJh2XOHDDeeCOHWSLhQBoPYb+cP2P88U1Fney8G+kPP/78ZyPH396LXwsUHAZiHgMQi4cY5SBQB8TsD4hVPQpGwSgYBSMMAADpKlPI1hjEawAAAABJRU5ErkJggg==","orcid":"","institution":"National Cancer Center, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":true,"prefix":"","firstName":"Xiao-Guang","middleName":"","lastName":"Ni","suffix":""}],"badges":[],"createdAt":"2025-11-19 17:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8157582/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8157582/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98759439,"identity":"f479bcce-b79e-4535-aecb-f74ed555e7b3","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"tif","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1613700,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/a312f6afe21574f91febad44.tif"},{"id":98759443,"identity":"24753674-829a-4a21-b044-00d1a46fe59c","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1318905,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/f844aac2add10056cd085649.docx"},{"id":98780357,"identity":"0e884f3b-b772-4f86-bd90-202dc967380d","added_by":"auto","created_at":"2025-12-22 12:31:14","extension":"tif","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2435524,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/513fd3a8f503d587b19e5b34.tif"},{"id":98759446,"identity":"ae5e8a09-df30-4f76-a93a-d08b65e8fb51","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"tif","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1536812,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/70808453509cbfa1146cb5eb.tif"},{"id":98778879,"identity":"bfa2fdd3-5123-4559-8e1a-ab72291dc2c1","added_by":"auto","created_at":"2025-12-22 12:29:46","extension":"tif","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7187896,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/cb2534640fa1781f1a758f58.tif"},{"id":98779928,"identity":"d1fdb1f7-42d3-4584-ac21-5ce4756fc212","added_by":"auto","created_at":"2025-12-22 12:30:56","extension":"json","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12680,"visible":true,"origin":"","legend":"","description":"","filename":"4e1ec93ab82c4850b4650a90bb594f6e.json","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/055940593bbd750934115167.json"},{"id":98759463,"identity":"f3944214-f00d-46b4-bdf2-e66d40edd034","added_by":"auto","created_at":"2025-12-22 09:48:49","extension":"mp4","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":53472704,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryVideos1.mp4","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/64f7a425e39c1df770b9ac0a.mp4"},{"id":98759464,"identity":"0e3da7a6-db7d-472d-9be0-10a8b690f936","added_by":"auto","created_at":"2025-12-22 09:48:50","extension":"mp4","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119051462,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryVideos2.mp4","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/efc667af1f2c48d566ac7207.mp4"},{"id":98759447,"identity":"6b7db748-b45d-4ae7-b41a-6616f5e966f8","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101248,"visible":true,"origin":"","legend":"","description":"","filename":"4e1ec93ab82c4850b4650a90bb594f6e1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/be5dc43b6ace858b8498f1e5.xml"},{"id":98778129,"identity":"5796b5ff-f47e-46f2-bbed-c74be0ce0d40","added_by":"auto","created_at":"2025-12-22 12:28:55","extension":"tif","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1613700,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/4dfd2bcd5b669188b7e30e4d.tif"},{"id":98780664,"identity":"1cc5a867-7011-47cf-87d2-b08e9ea43b36","added_by":"auto","created_at":"2025-12-22 12:31:33","extension":"tif","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2435524,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/96f8f2b717cf53c6c95b0bb0.tif"},{"id":98759455,"identity":"ef304f82-d75b-4a9a-8876-806292c5150e","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"tif","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1536812,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/9bed7e0d2ea33a62c1847a40.tif"},{"id":98759459,"identity":"fddeb56f-6ba0-4b3b-85ae-6003cb8228d6","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"tif","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7187896,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/3b02c9db2aacce8662522229.tif"},{"id":98759448,"identity":"c60f6d0f-9d3e-410b-a117-a1f93a94314b","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115992,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/fa0f093735668fb774b391af.jpeg"},{"id":98759449,"identity":"54b1df41-629f-4fa5-b95c-c26db6bee3b2","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":188468,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/85751496b233cc75d160aba5.jpeg"},{"id":98778721,"identity":"a4fc8c74-ed99-4342-9640-55448dfac828","added_by":"auto","created_at":"2025-12-22 12:29:33","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69298,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/9bf1293a09a75b3ba083f631.jpeg"},{"id":98759462,"identity":"70603d6f-9d39-4c12-9ce0-845a87994146","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"jpeg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":861664,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/cd3ef71cce3b634aa6e31e28.jpeg"},{"id":98778324,"identity":"991e2b49-0ae3-4341-82f0-584e1b9928de","added_by":"auto","created_at":"2025-12-22 12:29:10","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32443,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/5a57ca86b873d88ef4fb34e9.png"},{"id":98779076,"identity":"713c8ab2-83cb-4338-b315-a58237f9b269","added_by":"auto","created_at":"2025-12-22 12:29:56","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35914,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/90361f6ba9bd8bcd5117cf46.png"},{"id":98759460,"identity":"b33c8799-bb35-41eb-9490-46b35acee028","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15429,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/14c7650dc8da8d0774ca97c2.png"},{"id":98778068,"identity":"74c55605-21a1-487f-b575-c5c66bf34e1d","added_by":"auto","created_at":"2025-12-22 12:28:52","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":500778,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/05901ece409db130aabbe161.png"},{"id":98779041,"identity":"6a8fcffe-6cda-4790-99a1-ca14447f35a0","added_by":"auto","created_at":"2025-12-22 12:29:53","extension":"xml","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101286,"visible":true,"origin":"","legend":"","description":"","filename":"4e1ec93ab82c4850b4650a90bb594f6e1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/66c6ab00d8dc28e07a751e68.xml"},{"id":98778295,"identity":"e16d2602-b92e-46e9-b18c-69cd3b655e79","added_by":"auto","created_at":"2025-12-22 12:29:09","extension":"html","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111617,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/f906f7a352d633dd4a6acb3f.html"},{"id":98759438,"identity":"b0eb7d7a-8f07-452e-99cf-2f5cda56e16b","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":450933,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart illustrating the development and validation process of the DL-Ni system.\u003c/p\u003e","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/b4e037548fd7b6ea083c06f0.png"},{"id":98759440,"identity":"da79c19d-c994-4628-b29f-227c0450a192","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":408808,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the proposed DL-Ni system for Ni classification. (A) Overall network architecture. (B) The Convolutional Block Attention Module (CBAM) is integrated after each convolutional layer within the residual blocks. It consists of two sub-modules: the Channel Attention Module (CAM) and the Spatial Attention Module (SAM). The CAM utilizes both global average-pooled and max-pooled features, computes channel weights through a shared multi-layer perceptron (MLP), and thereby emphasizes informative channels. Meanwhile, the SAM aggregates channel information via average and max pooling operations to generate a spatial attention map, which is then processed by a convolutional layer to produce spatial weights, enabling the network to focus on critical regions.\u003c/p\u003e","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/7cafd764aa4d3179bca4da54.png"},{"id":98780494,"identity":"1d164348-8d4f-46dc-9186-d0f3f7027dff","added_by":"auto","created_at":"2025-12-22 12:31:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":234248,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic (ROC) curves demonstrating the diagnostic performance of the DL-Ni model across different datasets. (A) ROC curve for the internal test set from CHCAMS. (B) ROC curve for the external validation cohort from CHSZCAMS and SXPCH. (C) ROC curve for the video validation cohort, reflecting the model's performance in dynamic real-time scenarios.\u003c/p\u003e","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/25b9725f2c634a3b350e13e2.png"},{"id":98759445,"identity":"2d2afeac-6300-40ea-9b28-93c532fc9912","added_by":"auto","created_at":"2025-12-22 09:48:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":770427,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of DL-Ni predictions, segmentation results, and heatmaps. This figure illustrates the model's ability to identify Ni classification types across various NBI lesion images, along with corresponding heatmaps indicating lesion probability. (A) Original NBI image. (B) Lesion boundaries annotated by laryngoscopy experts. (C) Lesion boundaries predicted by DL-Ni segmentation. (D) Heatmap prediction of the lesion region generated by DL-Ni. (E) Overlay visualization combining the predicted lesion boundaries and heatmap from DL-Ni.\u003c/p\u003e","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/20a470b6151f279f2ef1e3ca.png"},{"id":98786083,"identity":"57d0cf43-091d-4eb7-844e-4a217d04c5e4","added_by":"auto","created_at":"2025-12-22 12:43:27","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3721822,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/49c2b0b7-a795-4dc9-8bda-8cde3fb4f0f5.pdf"},{"id":98778510,"identity":"5592cc48-6fee-4c14-bbf3-32c43f414fa3","added_by":"auto","created_at":"2025-12-22 12:29:23","extension":"mp4","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":53472704,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryVideos1.mp4","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/e1e5d8df67e873c1640cbe0c.mp4"},{"id":98759466,"identity":"c69c10a3-8313-4072-9e7e-fd448e53661f","added_by":"auto","created_at":"2025-12-22 09:48:52","extension":"mp4","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":119051462,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryVideos2.mp4","url":"https://assets-eu.researchsquare.com/files/rs-8157582/v1/db3e8bad8ce61f045e6811a5.mp4"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Deep Learning-Based Ni Classification System for Laryngeal NBI Images: A Multicenter Diagnostic Study Running title: DL for Ni Classification in Larynx","fulltext":[{"header":"Background","content":"\u003cp\u003eLaryngeal squamous cell carcinoma (LSCC) accounts for over 95% of malignant laryngeal tumors, with 5-year survival rates ranging from 80\u0026ndash;90% for early-stage disease (T1-T2) to approximately 60% for advanced stage(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). However, superficial lesions often present with non-specific changes under conventional white light imaging (WLI), such as mucosal roughness or erythema, which frequently lead to missed diagnoses.\u003c/p\u003e \u003cp\u003eNarrow Band Imaging (NBI) enhances visualization of mucosal microvascular patterns using specific wavelengths (415nm and 540nm), demonstrating superior sensitivity and specificity in detecting early laryngeal malignancies(\u003cspan additionalcitationids=\"CR4 CR5 CR6\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). The Ni classification system, established in 2011, standardizes NBI interpretation by categorizing intrapapillary capillary loop (IPCL) morphology into types I through V, providing objective criteria for diagnosing malignant and precancerous laryngeal lesions(\u003cspan additionalcitationids=\"CR9 CR10\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Despite clear morphological criteria, clinical adoption faces a critical challenge: significant inter-observer variability. Meta-analyses report NBI sensitivity ranging from 81% to 94% and specificity between 85% and 96%, with substantial heterogeneity across studies(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Multiple independent studies demonstrate modest interobserver agreement (κ values 0.40\u0026ndash;0.58), far below the clinical application threshold (κ\u0026thinsp;\u0026gt;\u0026thinsp;0.8)(\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). This inconsistency compromises diagnostic accuracy, potentially causing missed early malignancies or overtreatment of benign lesions.\u003c/p\u003e \u003cp\u003eImproving NBI classification consistency has traditionally relied on standardized training and experience accumulation. However, even after systematic training courses, interobserver κ value improvements rarely exceed 0.1\u0026ndash;0.15(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). Moreover, high-quality educational resources remain concentrated in tertiary hospitals, limiting widespread NBI adoption in primary care settings(\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI), particularly deep learning (DL), offers transformative potential in medical image analysis by identifying complex patterns and providing objective, reproducible interpretations. This technology presents a promising solution to inter-observer variability: DL models trained on expert-annotated NBI images can learn characteristic IPCL features, enabling standardized, automated classification. While existing studies demonstrate feasibility of DL in laryngeal image analysis, no research has systematically validated AI efficacy in resolving NBI interpretative discrepancies among physicians(\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study aims to develop a deep learning system for automatic Ni classification of laryngeal NBI images (DL-Ni system) and systematically evaluate its effectiveness as an assistive tool through randomized controlled human-AI collaborative experiments. By assessing whether AI assistance improves diagnostic accuracy and consistency across physicians with varying experience levels, we seek to standardize NBI interpretation, reduce diagnostic variability, and optimize early precision management of laryngeal cancer and precancerous lesions.\u003c/p\u003e"},{"header":"Method","content":"\u003cp\u003e This study was approved by the Institutional Review Board of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (Approval No. 22/454\u0026ndash;3656), and written informed consent was obtained from all participants. The study comprised two components: a retrospective phase for model development and internal validation, and a prospective phase for human-AI collaborative evaluation.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData Collection and Study Population\u003c/h2\u003e \u003cp\u003eLaryngeal NBI images were retrospectively collected from the Cancer Hospital and Institute of the Chinese Academy of Medical Sciences (CHCAMS) between January 2008 and March 2022 for model development and internal validation. External validation data were obtained from the Shenzhen Hospital of the Chinese Academy of Medical Sciences (CHSZCAMS) and the Shanxi Provincial Cancer Hospital (SXPCH). All images and videos were captured using Olympus Medical Systems devices (CV-170/ENF-VT2, CV-290/BF-H290, and CV-260/BF-260).\u003c/p\u003e \u003cp\u003eThe inclusion criteria for images were as follows: 1) clear visualization of IPCL structures with good vascular contrast, and absence of motion blur or equipment artifacts; 2) uniform and moderate illumination without overexposure or underexposure affecting the observation of vascular patterns; 3) a lesion area occupying\u0026thinsp;\u0026ge;\u0026thinsp;30% of the image to ensure sufficient diagnostic information; and 4) availability of a histopathological gold standard for diagnostic confirmation. Exclusion criteria included: 1) technical issues such as blurring, inaccurate focus, or artifacts caused by device malfunction; 2) biological interference including hemorrhage covering\u0026thinsp;\u0026ge;\u0026thinsp;30% of the field of view, or secretions and necrotic tissue covering\u0026thinsp;\u0026ge;\u0026thinsp;20% of the field; 3) imaging parameter issues such as excessive or insufficient illumination leading to inadequate IPCL contrast; and 4) lack of a definitive pathological diagnosis or unclear pathological findings.\u003c/p\u003e \u003cp\u003eFollowing stringent quality control screening, 3,308 high-quality NBI images from 1,421 patients were included after exclusion of 1,805 images (37.4%) from an initial pool of 4,828 candidate images. The primary reasons for exclusion were image artifacts (n\u0026thinsp;=\u0026thinsp;699, 14.5%), biological interference (n\u0026thinsp;=\u0026thinsp;808, 16.7%), and insufficient IPCL contrast (n\u0026thinsp;=\u0026thinsp;298, 6.2%).\u003c/p\u003e \u003cp\u003eImage annotation was performed independently by three senior NBI specialists, each with over 10 years of NBI experience and an annual examination volume exceeding 1,500 cases. Using the Ni classification criteria, the experts precisely delineated lesion regions and assigned classifications via the LabelMe software. The initial interobserver agreement among the experts yielded a Fleiss' κ of 0.780 (95% CI: 0.756\u0026ndash;0.804). Discrepant cases (n\u0026thinsp;=\u0026thinsp;156, 5.1%) were resolved through panel discussion and re-evaluation to reach a final consensus. The histopathological diagnosis served as the gold standard.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eDataset Partition\u003c/h3\u003e\n\u003cp\u003eThe internal dataset (sourced from CHCAMS) consisted of 3,023 images, while the external validation set (from CHSZCAMS and SXPCH) contained 285 images. The internal dataset was randomly partitioned at the patient level into training (2,144 images), validation (578 images), and internal testing (301 images) subsets in a ratio of 7:2:1. Stratified sampling was employed to ensure a balanced distribution of Ni classification types across all subsets. The distribution of the dataset is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eDeep Learning Model Development\u003c/h3\u003e\n\u003cp\u003eWe proposed a dual-branch architecture that jointly performs semantic segmentation and image classification, designed with two dedicated sub-networks: one for pixel-wise lesion area segmentation and the other for image-level lesion type classification. Each network was optimized to enhance feature representation capability for its respective task.\u003c/p\u003e \u003cp\u003eFor the semantic segmentation branch, we constructed a segmentation network based on the UNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;framework, which utilizes nested skip connections to effectively integrate features across different semantic levels. To improve the model\u0026rsquo;s focus on critical regions, we incorporated the Convolutional Block Attention Module (CBAM), which adaptively emphasizes lesion-relevant features through channel and spatial attention while suppressing background interference. Deep supervision was introduced to enhance the discriminative power of intermediate features and facilitate gradient propagation. The loss function combines weighted Binary Cross-Entropy (BCE), Dice loss, and Focal loss, jointly optimizing pixel accuracy, region overlap, and hard example learning.\u003c/p\u003e \u003cp\u003eFor the image classification task, we designed a dual-path network comprising a main branch and a lightweight branch. The main branch employs a 17-layer ResNet to extract high-level semantic and fine-grained structural features. The lightweight branch consists of a 19-layer convolutional network integrated with a CBAM attention module, guiding the model to focus on globally salient regions while significantly reducing parameter count and improving inference efficiency. Features from both branches are independently extracted and subsequently integrated within a feature fusion module to produce the final classification output. The classification network receives input as cropped RGB images of size 224\u0026times;224. During training, data augmentation techniques including normalization, rotation, and color jittering were applied. To address class imbalance, a class-weighted cross-entropy loss was adopted. A phased training strategy was employed: the first 65 epochs froze the main branch and updated only the lightweight branch and fusion module, while the subsequent 65 epochs unfroze the entire network for end-to-end fine-tuning, totaling 130 epochs. The AdamW optimizer was used with an initial learning rate of 1\u0026times;10⁻\u0026sup3;, coupled with a dynamic decay scheduling strategy to enhance convergence in later stages. The architectural framework of the model is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eGuman-AI Collaboration Experiment\u003c/h3\u003e\n\u003cp\u003eA prospective, fully crossed, multi-reader study was conducted to quantify the improvement in diagnostic consistency among physicians with the assistance of the AI system. A total of 12 physicians from three centers were recruited in a stratified manner: 1) the expert group (n\u0026thinsp;=\u0026thinsp;6), comprising physicians with over 10 years of NBI experience and an annual operation volume exceeding 500 cases; and 2) the junior group (n\u0026thinsp;=\u0026thinsp;6), consisting of physicians with 1\u0026ndash;3 years of NBI experience. Each physician evaluated 200 representative images randomly selected from the external validation set, covering all Ni classification types. The evaluation was performed in two independent sessions: during the baseline phase, only the NBI images and basic clinical information were provided; in the AI-assisted phase, real-time outputs from the DL-Ni system\u0026mdash;including probability distribution of classification, confidence score, and Grad-CAM-generated heatmaps highlighting IPCL key regions\u0026mdash;were added. A washout period of at least 4 weeks was implemented between the two sessions to mitigate memory effects. All assessments were conducted under standardized lighting conditions using calibrated medical displays. The order of image presentation was fully randomized across sessions to eliminate sequence effects. The primary endpoint was the inter-observer agreement (Fleiss\u0026rsquo; κ) in Ni classification with and without AI assistance. Secondary endpoints included diagnostic accuracy, sensitivity, and specificity within each physician group.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eInterobserver agreement was quantified using Fleiss\u0026rsquo; κ with 95% confidence intervals. Model performance was evaluated in terms of sensitivity, specificity, accuracy, and receiver operating characteristic (ROC) curves, all reported with 95% CIs. The improvement in performance metrics before and after AI assistance was compared using paired t-tests or Wilcoxon signed-rank tests, as appropriate. All statistical analyses were performed using R version 4.3.3 and Python 3.9. A two-sided p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eData Distribution and Baseline Characteristics\u003c/h2\u003e \u003cp\u003eFollowing rigorous quality control screening, a total of 3,308 high-quality NBI images were included in this study. The distribution across Ni classification types was as follows: Type I, 333 images (10.1%); Type II, 231 images (7.0%); Type III, 1,158 images (35.0%); Type IV, 135 images (4.1%); Type Va, 745 images (22.5%); Type Vb, 338 images (10.2%); and Type Vc, 368 images (11.1%). From CHCAMS, 1,270 patients were enrolled, constituting an internal dataset of 3,023 images. An external validation set of 285 images was formed from 151 patients recruited from CHSZCAMS and SXPCH. Detailed clinical characteristics of each dataset are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary of Demographic and Clinical Characteristics Across Study Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eCHCAMS Cohorts (Internal)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eExternal Cohorts\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTrain (n\u0026thinsp;=\u0026thinsp;873)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation (n\u0026thinsp;=\u0026thinsp;269)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest (n\u0026thinsp;=\u0026thinsp;128)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSXPCH (n\u0026thinsp;=\u0026thinsp;77)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCHSZCAMS (n\u0026thinsp;=\u0026thinsp;74)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, Mean (\u0026plusmn;\u0026thinsp;SD), years\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e46.6\u0026thinsp;\u0026plusmn;\u0026thinsp;12.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46.1\u0026thinsp;\u0026plusmn;\u0026thinsp;12.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e47.1\u0026thinsp;\u0026plusmn;\u0026thinsp;13.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e47.4\u0026thinsp;\u0026plusmn;\u0026thinsp;13.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e47.8\u0026thinsp;\u0026plusmn;\u0026thinsp;12.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e501 (41.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e165 (40.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e70 (42.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e51 (44.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e44 (45.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e372 (53.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e104 (55.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e49 (54.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e26 (53.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e30 (51.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNBI category, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAll images\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2144\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e578\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e301\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e139\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e146\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e211 (9.84%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e57 (9.86%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30 (9.97%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e12 (8.63%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e23 (15.75%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e154 (7.18%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41 (7.09%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (5.98%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e10 (7.19%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8 (5.48%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e753 (35.12%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e205 (35.47%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e102 (33.89%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e47 (33.81%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e51 (34.93%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e99 (4.62%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22 (3.81%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10 (3.32%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3 (2.16%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1 (0.68%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e464 (21.64%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e133 (23.01%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e80 (26.58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e30 (21.58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e38 (26.03%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVb\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e222 (10.35%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e61 (10.55%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e26 (8.64%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e17 (12.23%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12 (8.22%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVc\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e241 (11.24%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59 (10.21%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e35 (11.63%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e20 (14.39%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e13 (8.90%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eModel Performance Evaluation\u003c/h3\u003e\n\u003cp\u003eOn the internal test dataset, the model demonstrated the following performance: overall accuracy 0.858 (95% CI: 0.821\u0026ndash;0.895), sensitivity 0.809 (0.799\u0026ndash;0.819), specificity 0.974 (0.965\u0026ndash;0.983), AUC 0.967 (0.960\u0026ndash;0.974) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), and mIoU 0.731 (0.711\u0026ndash;0.751). For Type Va lesions, the accuracy was 0.940 (0.929\u0026ndash;0.951) and sensitivity 0.873 (0.866\u0026ndash;0.890). The model exhibited strong generalization capability on the external validation set, achieving an overall accuracy of 0.827 (0.813\u0026ndash;0.841), specificity of 0.770 (0.765\u0026ndash;0.775), sensitivity of 0.969 (0.958\u0026ndash;0.980), AUC of 0.961 (0.955\u0026ndash;0.967), and mIoU of 0.719 (0.704\u0026ndash;0.734). For Type Va lesions in the external set, accuracy reached 0.918 (0.911\u0026ndash;0.925) with a sensitivity of 0.836 (0.830\u0026ndash;0.842). The classification performance across different center-specific datasets is detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e depicts the segmentation and prediction results of the DL-Ni system, along with corresponding heatmap visualizations.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation of DL-Ni's Diagnostic and Segmentation Performance in Independent Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation Set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eIoU (95% CI)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eInternal validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAll category\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.858 (0.821\u0026ndash;0.895)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.809 (0.799\u0026ndash;0.819)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.974 (0.965\u0026ndash;0.983)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.967 (0.960\u0026ndash;0.974)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.731 (0.711\u0026ndash;0.751)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.940 (0.929\u0026ndash;0.951)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.873 (0.866\u0026ndash;0.890)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.959 (0.955\u0026ndash;0.964)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.957 (0.951\u0026ndash;0.963)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.726 (0.708\u0026ndash;0.742)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eExternal validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAll category\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.827 (0.813\u0026ndash;0.841)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.770 (0.765\u0026ndash;0.775)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.969 (0.958\u0026ndash;0.980)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.961 (0.955\u0026ndash;0.967)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.719 (0.704\u0026ndash;0.734)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.918 (0.911\u0026ndash;0.925)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.836 (0.830\u0026ndash;0.842)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.943 (0.937\u0026ndash;0.949)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.963 (0.959\u0026ndash;0.967)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.721 (0.710\u0026ndash;0.732)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eReal-time validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAll category\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.827 (0.811\u0026ndash;0.843)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.751 (0.742\u0026ndash;0.760)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.967 (0.953\u0026ndash;0.981)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.971 (0.961\u0026ndash;0.981)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.713 (0.698\u0026ndash;0.728)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVa\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.915 (0.899\u0026ndash;0.931)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.830 (0.822\u0026ndash;0.838)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.947 (0.933\u0026ndash;0.961)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.968 (0.955\u0026ndash;0.981)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.714 (0.695\u0026ndash;0.733)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo assess the model\u0026rsquo;s potential for real-time application, performance was evaluated on a test set containing 200 video clips. The model achieved an overall accuracy of 0.827 (0.811\u0026ndash;0.843), sensitivity of 0.751 (0.742\u0026ndash;0.760), specificity of 0.967 (0.953\u0026ndash;0.981), AUC of 0.971 (0.961\u0026ndash;0.981), and mIoU of 0.713 (0.698\u0026ndash;0.728). For Type Va lesions, accuracy was 0.915 (0.899\u0026ndash;0.931) and sensitivity 0.830 (0.822\u0026ndash;0.838), confirming the model\u0026rsquo;s diagnostic capability in dynamic imaging scenarios. Demonstrations of real-time detection and Ni-type probability prediction during laryngoscopy are provided in Supplementary Videos S1 and S2.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eImpact of AI Assistance on Physician Diagnostic Agreement\u003c/h2\u003e \u003cp\u003eResults from the human-AI collaboration experiment demonstrated that AI assistance significantly improved the diagnostic performance of both physician groups. In the junior group, the diagnostic accuracy increased from 0.661 to 0.801 with AI support, and interobserver agreement (κ value) improved from 0.488 (moderate agreement) to 0.685 (substantial agreement) (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Among experts, diagnostic accuracy rose from 0.787 to 0.885, while the κ value increased from 0.621 (substantial agreement) to 0.791 (substantial agreement) (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). For the clinically critical Type Va lesions, AI assistance also led to a marked improvement in recognition accuracy: the junior group improved from 0.652 to 0.834 (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05), and the expert group increased from 0.759 to 0.891 (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Detailed comparative data are presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of Sensitivity, Specificity, Accuracy, and Kappa for Junior and Senior Endoscopists With Versus Without AI Assistance\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCondition\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eKappa\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eP value (vs Without AI)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eJunior\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWithout AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.661 (0.630\u0026ndash;0.692)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.658 (0.625\u0026ndash;0.691)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.943 (0.939\u0026ndash;0.949)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.488\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWith AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.801 (0.765\u0026ndash;0.837)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.801 (0.762\u0026ndash;0.840)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.967 (0.961\u0026ndash;0.972)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.685\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eSenior\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWithout AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.787 (0.767\u0026ndash;0.806)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.784 (0.758\u0026ndash;0.810)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.964 (0.961\u0026ndash;0.968)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWith AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.885 (0.850\u0026ndash;0.921)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.882 (0.842\u0026ndash;0.922)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.981 (0.975\u0026ndash;0.987)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.791\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study successfully developed and validated a deep learning system (DL-Ni) for the automated Ni classification of laryngeal NBI images. The system demonstrated excellent diagnostic performance in both internal and external validation (accuracy: 0.858 and 0.827; AUC: 0.967 and 0.961), with particularly high sensitivity in identifying Type Va lesions (0.873 and 0.836). Through a rigorously designed human-AI collaboration experiment, we confirmed that the DL-Ni system significantly improves interobserver agreement in Ni classification among practitioners with varying experience levels, with κ values increasing by 0.197 and 0.170 in the junior and expert groups, respectively, reaching moderate to substantial agreement. This effectively addresses a core bottleneck hindering the standardized clinical application of NBI technology.\u003c/p\u003e \u003cp\u003eThe high subjectivity in NBI interpretation\u0026mdash;specifically, the variability among physicians in judging the morphological features of intrapapillary capillary loops (IPCLs)\u0026mdash;has remained a major challenge limiting its widespread clinical adoption(\u003cspan additionalcitationids=\"CR22 CR23\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). This interobserver variability directly leads to inconsistent diagnostic results, ultimately affecting the accuracy of clinical decision-making.\u003c/p\u003e \u003cp\u003eThe DL-Ni system developed in this study represents a fundamental shift from the traditional subjective interpretation model toward a standardized and objective analytical paradigm. Through the automatic extraction and quantitative analysis of IPCL morphological features by deep convolutional networks, it transforms traditional qualitative descriptions into quantitative parameters. The system's outputs\u0026mdash;including probability distribution of classification, confidence scores, and Grad-CAM-generated heatmaps\u0026mdash;provide physicians with a transparent navigation tool for diagnostic decision-making, converting a process traditionally dependent on personal experience into one based on objective data and visual evidence(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). Compared to previous AI studies on laryngeal NBI, this work offers multiple innovations: it is the first to systematically develop and validate a deep learning system for automatic Ni classification of laryngeal NBI images; through a rigorous human-AI collaboration experiment, we quantitatively evaluated for the first time the improvement in inter-physician diagnostic consistency with AI assistance, demonstrating that its effect significantly surpasses traditional training methods; and the visual explanations and quantitative confidence levels enhance the transparency and interpretability of the diagnostic process, which is crucial for clinical acceptance and adoption of AI-assisted diagnosis(\u003cspan additionalcitationids=\"CR27\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study has several limitations that should be addressed in future work. First, although multi-center external validation was performed, the training data were primarily acquired from Olympus endoscopy systems. Differences in spectral characteristics and image quality among devices from different manufacturers may exist; this limitation could be mitigated through cross-device generalizability studies utilizing techniques such as domain adaptation to improve model performance across devices with varying spectral and image characteristics. Second, although model performance was tested on a video dataset, its ability to capture subtle mucosal movements in real-time during dynamic imaging remains to be enhanced; this could be addressed through optimization for real-time video stream analysis, developing algorithms that can capture IPCL morphological changes and hemodynamic features in real-time under conditions of subtle mucosal movement. Third, the histopathological gold standard used in this study may be affected by biopsy sampling errors, especially for heterogeneously distributed lesions; this challenge could be addressed through multimodal integrated diagnosis, exploring comprehensive models that combine NBI with other imaging modalities to compensate for the limitations of NBI in detecting deep tissue infiltration. Additionally, long-term, multi-center prospective cohort studies are needed to systematically evaluate the impact of AI-assisted NBI classification on clinical treatment decisions and patient outcomes.\u003c/p\u003e \u003cp\u003eIn conclusion, this study demonstrates that the DL-Ni system not only achieves high diagnostic accuracy in Ni classification of laryngeal NBI images but, more importantly, significantly improves interobserver agreement among practitioners with varying experience levels. This breakthrough provides a novel technological pathway for the early and precise diagnosis of laryngeal cancer and precancerous lesions and is expected to promote the standardized use of NBI in clinical practice, ultimately contributing to organ-preserving treatment and improving long-term patient outcomes.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study successfully developed and validated the first deep learning-based automated classification system (DL-Ni) for laryngeal NBI images and demonstrated its significant effectiveness in improving diagnostic agreement among physicians with varying levels of experience. The system achieved high-precision automatic classification of NBI images (accuracy exceeding 85%), and more importantly, markedly enhanced inter-observer diagnostic agreement, with κ values increasing by 0.17\u0026ndash;0.20. These results provide an effective technical solution to overcome the key bottleneck in the standardized clinical application of NBI technology.\u003c/p\u003e \u003cp\u003eOur work offers a new technological pathway for the early and precise diagnosis of laryngeal cancer and precancerous lesions, with strong potential to promote the standardized adoption of NBI in clinical practice. Ultimately, this approach is expected to improve patient outcomes and healthcare quality. With continued technical optimization and further clinical validation, AI-assisted Ni classification could become a standardized tool in the diagnosis of laryngeal lesions.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003eThis study was approved by the Institutional Review Board of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (Approval No. 22/454\u0026ndash;3656), and written informed consent was obtained from all participants.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis work was supported by National High Level Hospital Clinical Research Funding (grant number LC2024A04) and CAMS Innovation Fund for Medical Sciences (CIFMS) (grant number 2022-I2M-C\u0026amp;T-B-059).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJie-Lin Huang and Li-Juan Li contributed equally to this work. Conceptualization: Xiao-Guang Ni, Jian-Hui Wang. Methodology: Jie-Lin Huang, Li-Juan Li, Ji-Qing Zhu. Software: Jie-Lin Huang, Li-Zhou Dou. Validation: Xue Zhang, Yu-Meng Liu, Yan Ke. Formal analysis: Yu-Da Zhao, Mei-Ling Wang. Investigation: All authors. Resources: Xiao-Guang Ni, Quan-Mao Zhang, Jian-Hui Wang. Data Curation: Li-Juan Li, Ji-Qing Zhu. Writing \u0026ndash; Original Draft: Jie-Lin Huang, Li-Juan Li. Writing \u0026ndash; Review \u0026amp; Editing: All authors. Visualization: Li-Zhou Dou, Xue Zhang. Supervision: Xiao-Guang Ni. Project administration: Xiao-Guang Ni. Funding acquisition: Xiao-Guang Ni. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and analyzed during the current study are available from the corresponding author on reasonable request. The source code of the deep learning model is not publicly available at this time to protect intellectual property prior to patent filing but may be made available under a material transfer agreement.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eChu EA, Kim YJ. Laryngeal cancer: diagnosis and preoperative work-up. Otolaryngol Clin North Am. 2008;41(4):673\u0026ndash;95, v.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarioni G, Marchese-Ragona R, Cartei G, Marchese F, Staffieri A. Current opinion in diagnosis and treatment of laryngeal carcinoma. Cancer Treat Rev. 2006;32(7):504\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBertino G, Cacciola S, Fernandes WB, Jr., Fernandes CM, Occhini A, Tinelli C, et al. Effectiveness of narrow band imaging in the detection of premalignant and malignant lesions of the larynx: validation of a new endoscopic clinical classification. Head Neck. 2015;37(2):215\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStan\u0026iacute;kov\u0026aacute; L, Šatankov\u0026aacute; J, Kučov\u0026aacute; H, Walderov\u0026aacute; R, Zelen\u0026iacute;k K, Kom\u0026iacute;nek P. The role of narrow-band imaging (NBI) endoscopy in optical biopsy of vocal cord leukoplakia. Eur Arch Otorhinolaryngol. 2017;274(1):355\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWestra JM, Zwakenberg MA, Halmos GB, van der Laan B, van der Vegt B, Plaat BEC. Narrow band imaging reveals field cancerisation undetected by conventional white light: Optical diagnosis versus histopathology. Clin Otolaryngol. 2024;49(4):429\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGalli J, Settimi S, Mele DA, Salvati A, Schiavi E, Parrilla C, et al. Role of Narrow Band Imaging Technology in the Diagnosis and Follow up of Laryngeal Lesions: Assessment of Diagnostic Accuracy and Reliability in a Large Patient Cohort. J Clin Med. 2021;10(6).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNi XG, Wang GQ. The Role of Narrow Band Imaging in Head and Neck Cancers. Curr Oncol Rep. 2016;18(2):10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNi XG, He S, Xu ZG, Gao L, Lu N, Yuan Z, et al. Endoscopic diagnosis of laryngeal cancer and precancerous lesions by narrow band imaging. J Laryngol Otol. 2011;125(3):288\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStan\u0026iacute;kov\u0026aacute; L, K\u0026aacute;ntor P, Fedorov\u0026aacute; K, Zelen\u0026iacute;k K, Kom\u0026iacute;nek P. Clinical significance of type IV vascularization of laryngeal lesions according to the Ni classification. Front Oncol. 2024;14:1222827.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNerurkar NK, Sarkar A. Correlation of narrow-band imaging findings using the Ni and European Laryngeal Society classification systems during in-office flexible laryngoscopy with histopathology. J Laryngol Otol. 2024;138(2):203\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFang Y, Yang Y, Chen M, Chen J, He P, Cheng L, et al. Correlating intraepithelial papillary capillary loops of vocal cord leukoplakia with histopathology. Acta Otolaryngol. 2022;142(1):106\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSanda IA, Hainarosie R, Ionita IG, Voiosu C, Ristea MR, Zamfir Chiru Anton A. A Systematic Review Evaluating the Diagnostic Efficacy of Narrow-Band Imaging for Laryngeal Cancer Detection. Medicina (Kaunas). 2024;60(8).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eScholman C, Zwakenberg MA, Wedman J, Wachters JE, Halmos GB, van der Laan B, et al. The influence of clinical experience on reliable evaluation of pharyngeal and laryngeal lesions: comparison of high-definition laryngoscopy using narrow band imaging with fibre-optic laryngoscopy. J Laryngol Otol. 2024;138(4):425\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKraus F, Gehrke S, Ehrmann-M\u0026uuml;ller D, Hofer F, Shehata-Dieler W, Hagen R, et al. Comparison of three different image enhancement systems for detection of laryngeal lesions. J Laryngol Otol. 2024;138(1):105\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePietruszewska W, Morawska J, Rosiak O, Leduchowska A, Klimza H, Wierzbicka M. Vocal Fold Leukoplakia: Which of the Classifications of White Light and Narrow Band Imaging Most Accurately Predicts Laryngeal Cancer Transformation? Proposition for a Diagnostic Algorithm. Cancers (Basel). 2021;13(13).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNi XG, Wang GQ, Hu FY, Xu XM, Xu L, Liu XQ, et al. Clinical utility and effectiveness of a training programme in the application of a new classification of narrow-band imaging for vocal cord leukoplakia: A multicentre study. Clin Otolaryngol. 2019;44(5):729\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eŻurek M, Rzepakowska A, Osuch-W\u0026oacute;jcikiewicz E, Niemczyk K. Learning curve for endoscopic evaluation of vocal folds lesions with narrow band imaging. Braz J Otorhinolaryngol. 2019;85(6):753\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsmaeili N, Sharaf E, Gomes Ataide EJ, Illanes A, Boese A, Davaris N, et al. Deep Convolution Neural Network for Laryngeal Cancer Classification on Contact Endoscopy-Narrow Band Imaging. Sensors (Basel). 2021;21(23).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiong H, Lin P, Yu JG, Ye J, Xiao L, Tao Y, et al. Computer-aided diagnosis of laryngeal cancer via deep learning based on laryngoscopic images. EBioMedicine. 2019;48:92\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAzam MA, Sampieri C, Ioppi A, Africano S, Vallin A, Mocellin D, et al. Deep Learning Applied to White Light and Narrow Band Imaging Videolaryngoscopy: Toward Real-Time Laryngeal Cancer Detection. Laryngoscope. 2022;132(9):1798\u0026ndash;806.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArthur C, Huangfu H, Li M, Dong Z, Asamoah E, Shaibu Z, et al. The Effectiveness of White Light Endoscopy Combined With Narrow Band Imaging Technique Using Ni Classification in Detecting Early Laryngeal Carcinoma in 114 Patients: Our Clinical Experience. J Voice. 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChabrillac E, Dupret-Bories A, Vairel B, Woisard V, De Bonnecaze G, Vergez S. Narrow-Band Imaging in oncologic otorhinolaryngology: State of the art. Eur Ann Otorhinolaryngol Head Neck Dis. 2021;138(6):451\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumazawa Y, Ikenoyama Y, Takamatsu M, Kido K, Namikawa K, Tokai Y, et al. Differences in Clinical Characteristics Between Missed and Detected Laryngopharyngeal Cancers. J Gastroenterol Hepatol. 2025;40(8):2037\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNogu\u0026eacute;s-Sabat\u0026eacute; A, Aviles-Jurado FX, Ruiz-Sevilla L, Lehrer E, Santamar\u0026iacute;a-Gadea A, Valls-Mateus M, et al. Intra and interobserver agreement of narrow band imaging for the detection of head and neck tumors. Eur Arch Otorhinolaryngol. 2018;275(9):2349\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJahmunah V, Ng EYK, Tan RS, Oh SL, Acharya UR. Explainable detection of myocardial infarction using deep learning models with Grad-CAM technique on ECG signals. Comput Biol Med. 2022;146:105550.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInaba A, Hori K, Yoda Y, Ikematsu H, Takano H, Matsuzaki H, et al. Artificial intelligence system for detecting superficial laryngopharyngeal cancer with high efficiency of deep learning. Head Neck. 2020;42(9):2581\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Y, Gu W, Yue H, Lei G, Guo W, Wen Y, et al. Real-time detection of laryngopharyngeal cancer using an artificial intelligence-assisted system with multimodal data. J Transl Med. 2023;21(1):698.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYumii K, Ueda T, Kawahara D, Chikuie N, Taruya T, Hamamoto T, et al. Artificial intelligence-based diagnosis of the depth of laryngopharyngeal cancer. Auris Nasus Larynx. 2024;51(2):417\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-medical-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejmr","sideBox":"Learn more about [European Journal of Medical Research](http://eurjmedres.biomedcentral.com)","snPcode":"40001","submissionUrl":"https://submission.nature.com/new-submission/40001/3","title":"European Journal of Medical Research","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Narrow Band Imaging, Ni classification, deep learning, laryngeal cancer, interobserver agreement","lastPublishedDoi":"10.21203/rs.3.rs-8157582/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8157582/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eLaryngeal squamous cell carcinoma accounts for more than 95% of laryngeal tumors. Early diagnosis is crucial for function preservation and prognosis improvement. Narrow Band Imaging (NBI) technology and Ni classification system provide an important basis for early diagnosis; however, poor interobserver agreement limits its standardized clinical application. To develop and validate the first deep learning system (DL-Ni) for Ni classification of laryngeal NBI images, and to evaluate its effectiveness in improving diagnostic agreement among physicians with varying levels of experience.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis multicenter diagnostic study retrospectively collected 3,023 high-quality laryngeal NBI images to construct the dataset. A dual-branch collaborative learning architecture was developed, comprising a UNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;semantic segmentation branch and an improved ResNet classification branch. A randomized controlled crossover experiment was conducted to evaluate the improvement in diagnostic agreement among 12 physicians with different levels of experience under AI assistance.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe developed DL-Ni system demonstrated robust performance in both internal and external validations, with accuracy of 0.858 (95% CI: 0.821\u0026ndash;0.895) and 0.827 (95% CI: 0.813\u0026ndash;0.841), respectively. AI assistance significantly improved interobserver diagnostic agreement: the Fleiss\u0026rsquo; κ value increased from 0.488 to 0.685 (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the junior physician group, and from 0.621 to 0.791 (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the expert group.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis study is the first to develop and validate an automated deep learning system for Ni classification of laryngeal NBI images. The system significantly improved interobserver diagnostic consistency, offering an effective tool and solution for the standardized clinical application of NBI technology.\u003c/p\u003e","manuscriptTitle":"A Deep Learning-Based Ni Classification System for Laryngeal NBI Images: A Multicenter Diagnostic Study Running title: DL for Ni Classification in Larynx","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 09:48:43","doi":"10.21203/rs.3.rs-8157582/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-16T05:46:35+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-18T09:26:57+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-15T14:39:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"53066804394986821119711890132639753726","date":"2025-12-21T07:51:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"297020849256955622791066580121070134929","date":"2025-12-19T09:17:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"189797462567420352286607936893273077583","date":"2025-12-18T12:54:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"32347648985516005250300983609179314141","date":"2025-12-18T12:07:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-16T12:45:39+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-24T06:21:33+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-24T06:19:54+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Medical Research","date":"2025-11-19T17:01:52+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-medical-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejmr","sideBox":"Learn more about [European Journal of Medical Research](http://eurjmedres.biomedcentral.com)","snPcode":"40001","submissionUrl":"https://submission.nature.com/new-submission/40001/3","title":"European Journal of Medical Research","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0765c93c-5f51-459c-aea7-c0aa78405eba","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T09:10:30+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-22 09:48:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8157582","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8157582","identity":"rs-8157582","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}