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.

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.

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.

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.

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.

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.

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.

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.

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.