Paper deep dive
Medical Imaging Fusing Vision Transformer: Laryngeal Cancer Screening with Explanation
Haiyang Wang, Luca Mainardi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/21/2026, 5:45:27 AM
Summary
The paper proposes a medical imaging framework for laryngeal cancer screening that fuses a Swin Transformer for classification with MedSAM for segmentation-based explanation. Using the CE-NBI dataset, the model achieves an F1-score of 82.72% and accuracy of 82.33%, providing clinicians with interpretable visual explanations of pathological regions to support decision-making.
Entities (7)
Relation Signals (5)
Swin Transformer → achievesmetric → F1 Score
confidence 95% · Results show it has good classification performance with F1 (82.72%)
Swin Transformer → usedforclassification → Laryngeal Cancer
confidence 95% · we proposed applying transformer and attention mechanism for analyzing the narrow band imaging and distinguish benign and malignant lesions.
MedSAM → usedforexplanation → Swin Transformer
confidence 92% · The explainability is utilizing the state of art segmentation method (MedSAM) to provide the useful pathological information area for clinicians.
CE-NBI Dataset → usedby → Swin Transformer
confidence 90% · We used the publicly available CE-NBI dataset... to develop our pipeline
NBI Endoscopy → inputfor → Swin Transformer
confidence 88% · applying transformer and attention mechanism for analyzing the narrow band imaging
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Early and timely screening of laryngeal cancer is crucial for improving clinical outcomes. In recent years, NBI endoscopy has become a standard diagnostic tool for the detection of laryngeal lesions. However, its effective use requires well-trained clinicians and the procedure is time-consuming and subject to interobserver variability. In this context, the application of artificial intelligence (AI) offers a promising solution to support clinical decision-making. In this work, we proposed applying transformer and attention mechanism for analyzing the narrow band imaging and distinguish benign and malignant lesions. Results show it has good classification performance with F1 (82.72%), accuracy(82.33%). In addition, the result of laryngeal cancer screening is explainable for clinicians. The explainability is utilizing the state of art segmentation method (MedSAM) to provide the useful pathological information area for clinicians. The proposed methodology fusing classification and segmentation provides a translating on laryngeal cancer screening.
Tags
Links
- Source: https://arxiv.org/abs/2607.17789v1
- Canonical: https://arxiv.org/abs/2607.17789v1
Trouble viewing inline? Open PDF directly →
Full Text
25,904 characters extracted from source content.
Expand or collapse full text
Medical Imaging Fusing Vision Transformer: Laryngeal Cancer Screening with Explanation 1st Haiyang WANG 2nd Luca Mainardi Abstract Early and timely screening of laryngeal cancer is crucial for improving clinical outcomes. In recent years, NBI endoscopy has become a standard diagnostic tool for the detection of laryngeal lesions. However, its effective use requires well-trained clinicians and the procedure is time-consuming and subject to interobserver variability. In this context, the application of artificial intelligence (AI) offers a promising solution to support clinical decision-making. In this work, we proposed applying transformer and attention mechanism for analyzing the narrow band imaging and distinguish benign and malignant lesions. Results show it has good classification performance with F1 (82.72%), accuracy (82.33%). In addition, the result of laryngeal cancer screening is explainable for clinicians. The explainability is utilizing the state of art segmentation method (MedSAM) to provide the useful pathological information area for clinicians. The proposed methodology fusing classification and segmentation provides a translating on laryngeal cancer screening. I Introduction Laryngeal cancer remains a significant global health concern, where early detection is paramount for effective treatment and improved survival rates[12] [10]. In clinical practice, endoscopic examination, particularly with Narrow Band Imaging (NBI), serves as a primary diagnostic tool by enhancing the visualization of vascular patterns and mucosal structures [31][13]. However, the interpretation of these images is inherently subjective, relying heavily on the clinician’s expertise, and can be a labor-intensive process [21] [14]. This subjectivity and inefficiency can lead to diagnostic delays and inconsistencies. Recent advances in Artificial Intelligence (AI), particularly in deep learning, present a compelling solution to these challenges by enabling automated, objective, and rapid medical image analysis [24] [2] [28]. Machine learning applied to laryngeal cancer diagnosed could date back to 2002 by Ritchings et. al. [25], using artificial neural networks for pathological voice quality assessment for laryngeal cancer. Then, as NBI endoscopy [3] became a more standard diagnostic tool for laryngeal cancer diagnosed, reserchers started working more on NBI. In 2017, [20] applied texture-based and SVM for laryngeal cancer diagnosis. In 2019, Hao Xiong et. al.[30] applied convolutional neural networks (CNNs) on laryngoscopic images for the first time. Following this work, several studies have employed CNN-based approaches across various types of laryngeal imaging data, including hyperspectral imaging[1]. Although CNNs have been extensively applied in medical image analysis, the recently developed transformer architecture, renowned for its self-attention mechanism, may offers an alternative for capturing long-range dependencies in image data. In applying AI model in medical imaging, an understandable explanation is always significant in medical translating [26]. In [9] , Gupta et. al tried to applied LIME (Local Interpretable Model-Agnostic Explanations) as explanation for skin cancer. LIME explanations, though simplified and accessible, tend to be inconsistent unless refined through several rounds of feedback[7]; Grad-CAM (Gradient- weighted Class Activation Mapping) is another common method as CNN explanation. Such as in [8] [18][27], Grad-CAM provides heatmaps that localize important medical image areas. But, it requires access to gradients and designed for CNN; [23] [6] [15] applying SHAP (SHapley Additive exPlanations) as an explanation, it provides more stable and reliable estimates of feature contributions. But it sufferes complex permutations. In [29] radiomics is combined with Global Interpretation Index as an explaination from feature perspective. It bridges the gap between high-dimensional handcrafted descriptors and model-level explanation. Many existing explanation methods focus heavily on technical details and overlook the medical imaging context, producing explanations intended primarily for engineers rather than clinicians. In our work, apart from applying AI to sort out laryngeal cancer patients, we provide an understandable explanation for clinicians. We harness the potential of the transformer model and its inherent attention mechanism for the classification of laryngeal cancer from NBI endoscopic images. Our proposed framework is designed to effectively identify and focus on critical pathological features. Meanwhile the state of art MedSAM [17] is applied to as the indicate potential suspicious pathological area as explanation. This could provide an understandable explanation for clinicians. I Methodology I-A Dataset We used the publicly available CE-NBI dataset hosted on the Zenodo Repository (https://zenodo.org/records/6674034) [5]. This dataset comprises images extracted from contact endoscopy with narrow-band imaging (CE-NBI) video recordings of adult patients with suspected benign, premalignant, and malignant vocal fold lesions. All patients were examined at the Department of Otorhinolaryngology, Head and Neck Surgery, Magdeburg University Hospital, Germany, between January 1, 2015, and December 31, 2021. In total, the dataset includes 11,144 images from 210 adult patients, encompassing both benign and malignant cases. Each image is annotated with classification labels: benign–malignant status [5]. Samples of malignant laryngeal cancer are shown in Figure 1. Images are provided in JPG format with variable resolutions (e.g., 1280×1008, 1736×1080, 1842×1080, 720×544, and 868×540 pixels). Further details on dataset composition and structure are described in [4]. Figure 1: Samples of Malignant Laryngeal Cancer. (A) Carcinoma in situ.(B) High grade dysplasia. (C) Squamous cell carcinoma. Figure 2: Overview of the proposed framework combining Swin Transformer–based classification with MedSAM-based explanation. The input NBI is first processed by a hierarchical Swin Transformer architecture consisting of patch partitioning, linear embedding, and four sequential stages with patch merging and repeated Swin Transformer blocks, producing a final classification output (e.g. malignant). In parallel, the same input image is fed into MedSAM to generate a visual explanation in the form of a segmentation or highlighted region, providing explanation for the model’s prediction. I-B Vision Transformer The Vision Transformer (ViT) is a neural architecture that transposes the standard Transformer, based on self-attention mechanisms and initially designed for sequential data in NLP, to process visual information by treating images as sequences of patches[11][16]. ViT processes images by first partitioning the input into a sequence of non-overlapping patches. These patches are linearly projected into a lower-dimensional embedding space, forming the initial patch embeddings. To retain spatial information, learnable positional embeddings are added to these patch representations. This sequence, prepended with a special classification ([CLS]) token, is then fed into a standard Transformer encoder. The encoder comprises a stack of multi-headed self-attention layers, which enable global contextual integration by dynamically modeling dependencies between all patches, followed by position-wise feed-forward networks. For classification tasks, the final state of the [CLS] token serves as a comprehensive image representation, which is subsequently used for downstream tasks such as image classification. I-C Data Preparation To develop our pipeline, the dataset is divided into training, validation and testing subsets, comprising 67.2%67.2\%, 16.8%16.8\% and 16%16\% of the data, respectively. Splitting is performed at the patient level to ensure that images from the same patient were not shared between subsets, while maintaining a similar distribution of histopathology classes across both sets. The training set includes 7, 548 images, of which 5, 193 has benign lesions and 2, 355 has malignant lesions while the validaiton dataset is 1, 887 images with 1, 298 benign and 589 malignant. The testing set comprises 1, 709 images (Benign: 1, 166; Malignant: 543). Each image is resized and normalized using the pretrained Swin processor corresponding to the model Swin Transformer [19]. Here Swin-transformer is utizlied for its excellence in image classification than original Vision transformer. Figure 2 describled the overview of the proposed framework. I-D Patch Partition and Linear Embedding The input image ∈ℝH×W×3x ^H× W× 3, where H is the height and W is the width, is divided into non-overlapping 4×44× 4 patches. Each patch is flattened and linearly projected to a C-dimensional embedding using a learnable projection matrix: 0i=pi,∈ℝ(42⋅3)×C.z_0^i=x_p^iE, ^(4^2· 3)× C. (1) This patch embedding step is implemented internally in the Swin model as part of the patch embedding layer. I-E Hierarchical Feature Representation Unlike the original Vision Transformer, which maintains a fixed spatial resolution throughout the network, the Swin Transformer [16] constructs a hierarchical feature representation by progressively merging patches across stages. At each stage, groups of 2×22× 2 neighboring patches are concatenated and linearly projected, resulting in a doubling of the channel dimension: ^=Linear( Z=Linear( [(2i,2j),(2i+1,2j), [z(2i,2j),z(2i+1,2j), (2) (2i,2j+1),(2i+1,2j+1)]). (2i,2j+1),z(2i+1,2j+1)]). This operation reduces the spatial resolution by a factor of two while increasing the feature dimensionality, yielding a multi-scale representation analogous to that of convolutional neural networks. I-F Window-Based Self-Attention and Shifted Windows Each Swin Transformer block applies self-attention within non-overlapping local windows of size 7×77× 7. Given a feature map ∈ℝh×w×CX ^h× w× C, the feature map is partitioned into hw72 hw7^2 windows, and standard multi-head self-attention is performed independently within each window: h _h =whQ,h=whK,h=whV, =X_wW_h^Q, _h=X_wW_h^K, _h=X_wW_h^V, (3) headh _h =softmax(hh⊤dk+w)h =softmax\! ( Q_hK_h d_k+B_w )V_h (4) where wB_w denotes a learnable relative position bias specific to each window. To facilitate information exchange across windows, successive Swin Transformer blocks alternate between standard window partitioning and shifted window partitioning, in which windows are shifted by (M2,M2)( M2, M2). This shifted window mechanism enables cross-window interactions while maintaining computational efficiency. I-G Swin Transformer Block Figure 3: Architecture of a Swin Transformer block. The first block applies Window-based Multi-Head Self-Attention (W-MSA), while the second employs Shifted Window Multi-Head Self-Attention (SW-MSA) to enable cross-window information exchange. In both blocks, Layer Normalization (LN) precedes the attention and MLP modules, and residual connections are applied after each sub-layer. Each Swin Transformer block is illustrated in Figure 3, it is defined as: ^l Z_l =W-MSA(LN(l−1))+l−1, =W-MSA(LN(Z_l-1))+Z_l-1, (5) l _l =MLP(LN(^l))+^l, =MLP(LN( Z_l))+ Z_l, (6) ^l+1 Z_l+1 =SW-MSA(LN(l))+l, =SW-MSA(LN(Z_l))+Z_l, (7) l+1 _l+1 =MLP(LN(^l+1))+^l+1 =MLP(LN( Z_l+1))+ Z_l+1 (8) where W-MSA for window-based attention and SW-MSA for shifted-window attention. Layer Normalization (LN) is applied before each sub-layer (pre-norm structure), and MLP denotes a two-layer feed-forward network with GELU activation: MLP()=GELU(1+1)2+2MLP(X)=GELU(XW_1+b_1)W_2+b_2 (9) I-H Model Fine-Tuning and Optimization To accelerate training, the Swin backbone parameters θswin _swin were optionally frozen, ∂ℒ∂θswin=0 ∂ _swin=0 and only the classifier head parameters θcls _cls were optimized. The classifier head is composed by a linear layer mapping the final pooled feature vector avgz_avg to two output classes: =softmax(avgcls+cls)y=softmax(z_avgW_cls+b_cls) (10) where cls∈ℝD×2W_cls ^D× 2 , clsb_cls is the bias term. I-I Training and Evaluation During training, standard data augmentation techniques including random horizontal flipping, rotation, and color jittering are applied to improve generalization. Weighted random sampling is applied to address class imbalance. Training was performed for 20 epochs using the AdamW optimizer with a learning rate of 1×10−41× 10^-4 and weight decay of 0.010.01. The model was evaluated using validation loss, with early stopping (patience = 3) and best-model checkpointing. After training, the best-performing checkpoint was evaluated on the held-out test set (1,709 images). Performance was reported using overall accuracy, precision, recall, F1-score, and the confusion matrix. I-J Explanation MedSAM is a foundational model designed to bridge the gap toward universal medical image segmentation [17]. The model is trained on a comprehensive dataset comprising 1.57 million paired medical images and corresponding segmentation masks, including endoscopy images. The dataset spans 10 distinct imaging modalities and covers more than 30 types of cancer. Extensive experiments across 146 tasks (86 internal and 60 external), encompassing diverse anatomies, pathologies, and imaging modalities, demonstrate that MedSAM consistently outperforms state-of-the-art segmentation foundation models [22]. In clinical applications of artificial intelligence, explainability is critical. However, providing meaningful explanations for AI predictions remains a significant challenge. To address this issue, we propose a novel approach that integrates segmentation as a form of explanation alongside classification tasks. Specifically, segmentation can visually highlight anatomical structures and pathological regions that are most relevant to the model’s decision, closely aligning with the regions clinicians routinely examine during diagnosis and treatment planning. In this work, we employ MedSAM as the explanatory mechanism. MedSAM is based on a transformer backbone. Using segmentation outputs from the same transformer backbone provides a natural and consistent explanation for the classification decisions, ensuring architectural coherence between prediction and explanation. I Results and discussion The performance of the proposed transformer-based model was quantitatively evaluated on a test set of 1,709 samples, with the detailed classification report presented in Table I. The results demonstrate a robust overall capability of the model in the defined classification task, while also revealing distinct inter-class performance variations that warrant further discussion. Figure 4 shows the confusion matrix of the proposed Transformer model on the test set. The model achieved 956 true negatives, 451 true positives, 92 false negatives, and 210 false positives, yielding an overall accuracy of 82.33%. This is further corroborated by the weighted average F1-score of 82.72% in Table. I, which provides a more reliable measure of overall performance given the class imbalance in the dataset (Support: Class B=1, 166, Class M=543). It demonstrated strong precision for malignant cases (91.22%)—clinically valuable for reducing unnecessary biopsies. The close alignment between accuracy and the weighted average F1-score indicates that the model’s performance is consistent across the distribution of the data. TABLE I: Transformer Classification Report Class Precision Recall F1-Score Support B (Benign) 0.9122 0.8199 0.8636 1166 M (Malignant) 0.6823 0.8306 0.7492 543 Overall Accuracy 0.8233 1709 Macro Avg 0.7973 0.8252 0.8064 1709 Weighted Avg 0.8392 0.8233 0.8272 1709 Figure 4: Confusion Matrix of the Transformer Model Figure 5: Malignant Predication with Segmentation as Explanation Fig. 5. illustrates a representative image of a histopathologically confirmed malignant lesion alongside its corresponding segmentation area, which serves as a visual explanation of the model’s decision-making process. The transparent green could cover most of the area of the pathological region. This proves MedSAM works with this laryngeal cancer dataset. There is a certain improvement space for the model. But it still indicates a relatively good explanation for AI model. Compared to conventional black-box classification approaches, the integration of segmentation as an intrinsic explanation mechanism enhances model transparency and trustworthiness — critical requirements for computer-aided diagnosis systems in laryngeal medical imaging diagnosis. Experimental results confirm that our method achieves superior classification performance with understandable explanation. It provides a significant step towards translating AI-driven laryngeal cancer diagnostic aids from research into clinical practice. IV Conclusion In our study, this paper has made a contribution in the field of laryngeal cancer diagnosis with vision transformer. Transformer architecture, particularly with selective fine-tuning—markedly improves classification performance in laryngeal cancer detection on endoscopic images. (F1-score: 82.72%; Accuracy: 82.33%). Meanwhile, using MedSAM to achieve the area of the pathological region as the explantion to provide more instructive information. Here, we emphasized the importance of fusing classification and segmentation for AI in medical imaging. The findings underscore the transformative role of transformer in medical image analysis and suggest promising directions for real-world deployment in laryngoscope. With continued refinement, such AI-assisted in laryngeal cancer could become invaluable. References [1] M. Bengs, S. Westermann, N. Gessert, D. Eggert, A. O. Gerstner, N. A. Mueller, C. Betz, W. Laffers, and A. Schlaefer (2020) Spatio-spectral deep learning methods for in-vivo hyperspectral laryngeal cancer detection. In Medical Imaging 2020: Computer-Aided Diagnosis, Vol. 11314, p. 369–374. Cited by: §I. [2] W. L. Bi, A. Hosny, M. B. Schabath, M. L. Giger, N. J. Birkbak, A. Mehrtash, T. Allison, O. Arnaout, C. Abbosh, I. F. Dunn, et al. (2019) Artificial intelligence in cancer imaging: clinical challenges and applications. CA: a cancer journal for clinicians 69 (2), p. 127–157. Cited by: §I. [3] A. De Vito, G. Meccariello, and C. Vicini (2017) Narrow band imaging as screening test for early detection of laryngeal cancer: a prospective study. Clinical otolaryngology 42 (2), p. 347–353. Cited by: §I. [4] N. Esmaeili, N. Davaris, A. Boese, A. Illanes, N. Navab, M. Friebe, and C. Arens (2023) Contact endoscopy–narrow band imaging (ce-nbi) data set for laryngeal lesion assessment. Scientific Data 10 (1), p. 733. Cited by: §I-A. [5] N. Esmaeili (2022) Contact endoscopy – narrow band imaging (ce-nbi) data set. Note: https://zenodo.org/record/6674034#.ZAIYx3bML52. Cited by: §I-A, §I-A. [6] Y. Feng, Y. Xu, J. Wang, Z. Cao, B. Liu, Z. Du, L. Zhou, H. Hua, W. Wang, J. Mei, et al. (2025) Prediction of early recurrence after bronchial arterial chemoembolization in non-small cell lung cancer patients using dual-energy ct: an interpretable model based on shap methodology. Academic Radiology. Cited by: §I. [7] D. Garreau and U. Luxburg (2020) Explaining the explainer: a first theoretical analysis of lime. In International conference on artificial intelligence and statistics, p. 1287–1296. Cited by: §I. [8] S. Guluwadi et al. (2024) Enhancing brain tumor detection in mri images through explainable ai using grad-cam with resnet 50. BMC medical imaging 24 (1), p. 1–19. Cited by: §I. [9] R. K. Gupta (2025) Interpretable ai-enabled model for skin cancer diagnosis using lime. Procedia Computer Science 260, p. 3–11. Cited by: §I. [10] D. Han, H. L. Lee, Q. W. Oung, and C. H. Lee (2025) Global, regional, and national mortality of larynx cancer from 1990 to 2021: results from the global burden of disease study. World Journal of Surgical Oncology 23 (1), p. 76. Cited by: §I. [11] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, et al. (2022) A survey on vision transformer. IEEE transactions on pattern analysis and machine intelligence 45 (1), p. 87–110. Cited by: §I-B. [12] A. Hut, E. R. Boia, D. Para, G. Iovanescu, D. Horhat, L. Mikša, M. Chiriac, R. Galant, A. C. Motofelea, and N. C. Balica (2025) Laryngeal cancer in the modern era: evolving trends in diagnosis, treatment, and survival outcomes. Journal of Clinical Medicine 14 (10), p. 3367. Cited by: §I. [13] J. Jang (2015) The past, present, and future of image-enhanced endoscopy. Clinical endoscopy 48 (6), p. 466–475. Cited by: §I. [14] S. W. S. Lam, M. H. Lee, M. Dorosan, S. Altonji, H. K. Tan, and W. T. Lee (2025) Use of a preliminary artificial intelligence-based laryngeal cancer screening framework for low-resource settings: development and validation study. JMIR Formative Research 9 (1), p. e66110. Cited by: §I. [15] W. Liu, Z. Cai, Y. Chen, X. Guan, J. Feng, H. Chen, B. Guo, F. OuYang, C. Luo, R. Zhang, et al. (2025) Gadoxetic acid-enhanced mri for identifying cholangiocyte phenotype hepatocellular carcinoma by interpretable machine learning: individual application of shap. BMC cancer 25 (1), p. 1–13. Cited by: §I. [16] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo (2021) Swin transformer: hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, p. 10012–10022. Cited by: §I-B, §I-E. [17] J. Ma, Y. He, F. Li, L. Han, C. You, and B. Wang (2024) Segment anything in medical images. Nature Communications 15 (1), p. 654. Cited by: §I, §I-J. [18] J. A. Marmolejo-Saucedo and U. Kose (2024) Numerical grad-cam based explainable convolutional neural network for brain tumor diagnosis. Mobile Networks and Applications 29 (1), p. 109–118. Cited by: §I. [19] Microsoft (2021) Swin model. Note: https://huggingface.co/microsoft/swin-tiny-patch4-window7-224 Cited by: §I-C. [20] S. Moccia, E. De Momi, M. Guarnaschelli, M. Savazzi, A. Laborai, L. Guastini, G. Peretti, and L. S. Mattos (2017) Confident texture-based laryngeal tissue classification for early stage diagnosis support. Journal of Medical Imaging 4 (3), p. 034502–034502. Cited by: §I. [21] R. Nocini, G. Molteni, C. Mattiuzzi, and G. Lippi (2020) Updates on larynx cancer epidemiology. Chinese Journal of Cancer Research 32 (1), p. 18. Cited by: §I. [22] M. Nouman, G. Khoriba, and E. A. Rashed (2024) Rethinking medsam: performance discrepancies in clinical applications. In 2024 IEEE International Conference on Future Machine Learning and Data Science (FMLDS), p. 301–307. Cited by: §I-J. [23] A. Rahman, M. Hayat, N. Iqbal, F. K. Alarfaj, S. Alkhalaf, and F. Alturise (2025) Enhanced mri brain tumor detection using deep learning in conjunction with explainable ai shap based diverse and multi feature analysis. Scientific Reports 15 (1), p. 29411. Cited by: §I. [24] N. L. Rane, M. Paramesha, S. P. Choudhary, and J. Rane (2024) Artificial intelligence, machine learning, and deep learning for advanced business strategies: a review. Partners Universal International Innovation Journal 2 (3), p. 147–171. Cited by: §I. [25] R. Ritchings, M. McGillion, and C. J. Moore (2002) Pathological voice quality assessment using artificial neural networks. Medical engineering & physics 24 (7-8), p. 561–564. Cited by: §I. [26] S. N. Saw, Y. Y. Yan, and K. H. Ng (2025) Current status and future directions of explainable artificial intelligence in medical imaging. European journal of radiology 183, p. 111884. Cited by: §I. [27] P. K. Seerala and S. Krishnan (2020) Grad-cam-based classification of chest x-ray images of pneumonia patients. In International Symposium on Signal Processing and Intelligent Recognition Systems, p. 161–174. Cited by: §I. [28] M. Tsuneki (2022) Deep learning models in medical image analysis. Journal of Oral Biosciences 64 (3), p. 312–320. Cited by: §I. [29] H. Wang, K. Zhang, and L. Mainardi (2024) An interpretable framework utilizing radiomics for laryngeal cancer classification through narrow band imaging. IEEE Access 12, p. 129804–129812. Cited by: §I. [30] H. Xiong, P. Lin, J. Yu, J. Ye, L. Xiao, Y. Tao, Z. Jiang, W. Lin, M. Liu, J. Xu, et al. (2019) Computer-aided diagnosis of laryngeal cancer via deep learning based on laryngoscopic images. EBioMedicine 48, p. 92–99. Cited by: §I. [31] Q. Yang, Z. Liu, H. Sun, F. Jiao, B. Zhang, and J. Chen (2023) A narrative review: narrow-band imaging endoscopic classifications. Quantitative Imaging in Medicine and Surgery 13 (2), p. 1138. Cited by: §I.