Leveraging Vision Transformer for Histological Grade Prediction in Laryngeal and Hypopharyngeal Squamous Cell Carcinoma: A Large-Scale Multicenter Study

Leveraging Vision Transformer for Histological Grade Prediction in Laryngeal and Hypopharyngeal Squamous Cell Carcinoma: A Large-Scale Multicenter Study | 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 Leveraging Vision Transformer for Histological Grade Prediction in Laryngeal and Hypopharyngeal Squamous Cell Carcinoma: A Large-Scale Multicenter Study Ran Guo, Xiaoxia Qu, Song Tian, Zheng Li, Xinyan Wang, Zhenchao Sun, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7249038/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 27 Dec, 2025 Read the published version in Neuroradiology → Version 1 posted You are reading this latest preprint version Abstract Background Pretreatment determination of histological differentiation grade is critical for prognostic evaluation in laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) .This study aimed to develop a contrast-enhanced CT (CECT)-based Vision Transformer (ViT) model for noninvasive evaluation of histological grades in LHSCC. Methods A total of 1,648 LHSCC patients who underwent CECT scans were enrolled from three hospitals in this study. Participants were divided into a training cohort (n=1,239) , an internal validation cohort (n=310) from one hospital, and an external validation cohort (n = 99) from the other two hospitals. The diagnostic model integrates a pre-trained ViT for CECT feature extraction and an XGBoost classifier for prediction. The model’s predictive performance was evaluated using the area under the curve (AUC), decision curve analysis (DCA), and calibration curve. Results The ViT model achieved AUCs of 0.887 (95%CI: 0.848-0.927) in internal validation and 0.796 (95%CI: 0.693-0.899) in external validation cohorts, significantly outperforming the conventional radiomics model (AUCs: 0.775, 95%CI: 0.714-0.837 and 0.544, 95%CI: 0.388-0.699; p< 0.001 and 0.002, respectively). Clinically, DCA demonstrated superior clinical utility, while calibration curves showed excellent prediction reliability. Gradient-weighted Class Activation Mapping visualization identified CT image regions most influential for the model's predictions, providing interpretability for clinical decision-making. Conclusion The ViT-based deep learning model developed in this study using CECT demonstrated excellent predictive performance for histological grading of LHSCC,with promising application for patient prognosis assessment. Deep learning Vision Transformer Histological grade Laryngeal and hypopharyngeal squamous cell carcinoma Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Head and neck squamous cell carcinoma (HNSCC) ranks as the sixth most prevalent cancer worldwide. Globally, approximately 890,000 new cases of head and neck cancer are diagnosed annually, of which laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) account for 189,000 and 86,000 new cases respectively, with corresponding mortality rates of 103,000 and 41,000 deaths per year [ 1 ]. Squamous cell carcinoma can be further classified into well-differentiated, moderately differentiated, and poorly differentiated types based on the degree of cell differentiation. Among these, well/moderately differentiated squamous cell carcinomas are the most common. Notably, although poorly differentiated tumors account for a smaller proportion, they exhibit more aggressive biological behavior: demonstrating significantly greater propensity for local invasion, distant metastasis [ 2 – 4 ], and higher sensitivity to chemoradiotherapy, yet generally poorer treatment outcomes and unfavorable prognosis [ 5 ]. Earlier study showed a substantial difference in 5-year overall survival rates between well/moderately-differentiated and poorly differentiated patients, ranging from 12–30% [ 6 , 7 ]. Currently, the assessment of histological differentiation grades primarily relies on biopsy and surgical resection specimens. Nevertheless, for LHSCC, invasive biopsies have limited predictive value and often fail to fully capture the tumor's comprehensive biological characteristics and heterogeneity. Therefore, developing a novel preoperative, non-invasive method to accurately predict tumor differentiation status is of significant clinical importance-this would optimize clinical decision-making and facilitate personalized treatment strategies. Current studies demonstrate the promising applications of radiomics in oncology [ 8 , 9 ], with preoperative features capable of predicting pathological differentiation grades in various tumors including hepatocellular carcinoma, pancreatic neuroendocrine tumors, bladder cancer, and HNSCC [ 10 – 13 ]. Additionally, histogram and texture analysis of Fs-T2WI enables noninvasive prediction of differentiation grades in HNSCC [ 14 ]. As a cornerstone of computer vision, convolutional neural networks (CNNs) can automatically extract hierarchical image features through multilayer convolution operations and have been effectively applied to assess HNSCC differentiation [ 8 ]. Zheng et al. developed an integrated model combining radiomics and deep learning features, achieving an AUC of 0.822 in the test cohort [ 13 ]. However, current research has several limitations: 1) Significant heterogeneity in study populations, often mixing distinct anatomical subtypes of HNSCC (e.g., oropharyngeal, laryngeal, and hypopharyngeal SCC) with divergent treatment approaches and prognoses [ 15 , 16 ]. 2) Insufficient sample sizes hindering the model's ability to learn complex feature-differentiation relationships; 3) Lack of well-established pretrained weights for 3D medical imaging of LHSC. In recent years, the Vision Transformer (ViT), a Transformer-based deep learning model, has shown strong knowledge transfer capabilities [ 17 ]. Originally successful in natural language processing, ViT has since been applied to medical image analysis with promising results. Studies have shown that ViT outperforms conventional CNNs in various medical imaging tasks [ 8 , 18 – 19 ], mainly due to its self-attention mechanism, which captures global context and long-range dependencies. In contrast, CNNs are limited by their local receptive fields, restricting their ability to model global features [ 20 ]. The Segment Anything Model for 3D medical images (SAM-Med3D) is a foundation model for automatic segmentation across different imaging modalities such as CT and MRI [ 21 ]. Notably, its encoder shares a similar architecture with ViT. This study explored an early and innovative application of ViT models fine-tuned with pre-trained weights from SAM-Med3D for predicting histological differentiation grades in LHSCC. Leveraging large-scale, multicenter contrast-enhanced CT (CECT) datasets, we developed a deep learning framework and conducted comparative analyses with conventional radiomics approach. Materials and methods Patients’ selection We retrospectively collected patients with LHSCC confirmed by postoperative pathology between March 2009 and May 2024 from three medical institutions. The inclusion criteria were as follows: (1) LHSCC confirmed by pathology; (2) Histopathological examination performed within 2 weeks post-scanning; (3) No prior tumor-related treatment before surgery. Exclusion criteria included: (1) Recurrent patient; (2) Patient was diagnosed with a second primary tumor; (3) Lession maximum diameter less than 5 mm; (4) Inadequate CT imaging quality, which could not be used for the imaging analysis, or the loss of images. The patient recruitment pathway is shown in Fig. 1 . Ultimately, 1,648 eligible LHSCC cases were included in this study from the three hospitals. Next, they were randomly divided into a training cohort (n = 1,239) and an internal validation cohort (n = 310) according to the 8:2 ratio from Hospital 1, with the external validation cohort consisted of 99 patients from Hospital 2 and 3. CECT image acquisition, segmentation CECT images were obtained using five different scanners: SOMATOM Definition Flash and SOMATOM Force CT (Siemens Healthineers, Germany), the Brilliance 64 and iCT 256 CT (Philips Medical Systems, Nederland), and Revolution CT (GE Healthcare, USA). CECT scans were performed in the supine position. Patients were instructed to remain motionless and avoid swallowing during image acquisition, with both slice thickness and interval set at 1 mm. The region of interest delineation was performed using ITK-SNAP (version 3.8.0, http://www.itksnap.org/ ) software by a radiologist with 8 years of clinical experience, employing manual slice-by-slice annotation to comprehensively encompass both cystic and necrotic components of the lesions. All delineations were subsequently reviewed and validated by a senior radiologist with 20 years of professional experience. Notably, both radiologists were blinded to the patients' definitive pathological diagnoses throughout the entire annotation process. Data Preprocessing Figure 2 shows the workflow of our study. All scans were preprocessed using a window level of 40 HU and a window width of 350 HU. Before being fed into the model, each scan underwent Z-score normalization to ensure consistency in intensity distribution. Radiomics and deep learning model development We first trained a radiomics-based model as a baseline, following the Image Biomarker Standardization Initiative (IBSI) guidelines. The extracted radiomics features included first-order statistics, Gray Level Co-occurrence Matrix, Gray Level Dependence Matrix, Gray Level Run Length Matrix, Gray Level Size Zone Matrix, Neighbouring Gray Tone Difference Matrix, and shape features. Additionally, wavelet transformation was applied, resulting in a total of 1,454 radiomics features per sample. For feature selection, we used Pearson correlation (threshold = 0.8) and Least absolute shrinkage and selection operator (LASSO) regression. Ultimately, nine radiomics features were retained, consisting of five texture features, three shape features, and one first-order feature. For deep learning, we used a large-scale 3D ViT model with 90 million parameters to extract high-order texture features of the tumor. The CT images were first processed through a 3D embedding layer to convert them into a sequence of flattened patch embeddings (a transformer-compatible format). These transformed features were then passed through ViT, yielding a final set of 384 extracted features. To leverage pre-trained knowledge, we incorporated SAM-Med3D as the initial pre-trained weights for 3D ViT encoder. Although SAM-Med3D is originally designed for instance segmentation, its encoder shares the architecture and has generalizable representations of 3D anatomical structures and textures from a massive and diverse medical image dataset. Therefore, we used its encoder weights as pre-trained encoder in our model. Finally, we used EXtreme Gradient Boosting (XGBoost) as the classifier to train models based on both radiomics and ViT-extracted features. We defined a hyperparameter search space with a maximum tree depth of 2 to 3 and estimators ranging from 50 to 200. A five-fold cross-validation strategy was applied to the training set to optimize and evaluate the final model. Statistical analysis Statistical analysis was performed using SPSS software (version 25.0, IBM) and Python software (version 3.5.6; http://www.python.org ). Continuous variables were assessed for normality using the Shapiro-Wilk test. Normally distributed data are presented as mean ± standard deviation (x̄±s) and compared using independent samples t-tests. Categorical variables are reported as frequencies and compared using chi-square tests. Model predictive performance was evaluated using receiver operating characteristic curve analysis with area under the curve (AUC), and compared by DeLong's test. Calibration was assessed via calibration curves and Hosmer-Lemeshow test. Clinical utility was evaluated using decision curve analysis (DCA). A p -value < 0.05 was considered statistically significant. Results Characteristic of patients This study ultimately enrolled 1,648 patients. Table 1 lists the details of the clinical data of all patients. The training cohort included 1,239 patients from Hospital 1 (1,190 males and 49 females; mean age 61.64 ± 8.82 years), consisting of 946 well/moderately differentiated cases (313 well differentiated and 633 moderately differentiated) and 293 poorly differentiated cases. The internal validation cohort included 310 patients from Hospital 1 (294 males and 16 females; mean age 62.63 ± 9.24 years; 264 well/moderately differentiated including 80 well and 184 moderately differentiated, 46 poorly differentiated). The external validation cohorts comprised patients from two centers: 80 from Hospital 2 (78 males; mean age 65.63 ± 8.90 years; 69 well/moderately differentiated including 26 well and 43 moderately differentiated, 11 poorly differentiated), 19 from Hospital 3 (all males; mean age 64.26 ± 10.36 years; 15 well/moderately differentiated including 6 well and 9 moderately differentiated, 4 poorly differentiated), Compared to the training and internal validation cohorts, the external validation cohort has a slightly older age distribution ( p 0.05). Table 1 Clinical characteristics of patients with LHSCC in the training, internal validation, and external validation cohorts Characteristics Training cohort (n = 1,239) Internal validation cohort (n = 310) External validation cohort (n = 99) p value Age (years) 61.64 ± 8.82 62.63 ± 9.24 65.36 ± 9.16 < 0.001* Sex /n(%) 0.358 Male 1,190(96.0) 294 (94.8) 97(98.0) Female 49 (4.0) 16 (5.2) 2(2.0) location/n(%) 0.112 Larynx 1,102 (88.9) 283(91.3) 83(83.8) Hypopharynx 137 (11.1) 27 (8.7) 16 (16.2) T stage/n(%) < 0.001* T1 412 (33.3) 137(44.2) 19 (19.2) T2 400 (32.3) 94 (30.3) 41 (41.4) T3 307 (24.8) 54 (17.4) 28 (28.3) T4 120(9.7) 25 (8.1) 11 (11.1) LNM/n(%) 0.962 Absent 489 (61.2) 102 (60.7) 55 (59.8) Present 310 (38.8) 66 (39.3) 37 (40.2) Unknown 440 142 7 *Represents p < 0.05; LNM: lymph node metastasis; Data of age are shown as mean ± standard deviation; other data are number of patients with percentage in parentheses.Unknown: No lymphadenectomy was done Comparison of the ViT and radiomics models The ViT model demonstrated robust performance in discriminating well/moderately differentiated from poorly differentiated grades across validation cohorts, achieving an AUC of 0.887 [95% (confidence interval,CI): 0.848–0.927]in the internal validation cohort and 0.796 (95%CI: 0.693–0.899) in the external validation cohort. In contrast, the radiomics model showed significantly lower performance, with AUCs of 0.775 (95%CI: 0.714–0.837) and 0.544 (95%CI: 0.388–0.699) in the internal and external validation cohorts, respectively (Fig. 3 a and d). The results of the ROC analysis of each model were shown in Table 2 . The DeLong test confirmed the superior predictive efficacy of the ViT model over the radiomics model in both the internal ( p < 1x10 − 7 ) and external validation cohorts ( p = 0.002). DCA (Fig. 3 b and e) revealed that the ViT model had a higher clinical net benefit in predicting the histological differentiation grades of LHSCC. The calibration curves (Fig. 3 c and f) further evidenced its excellent calibration performance indicating strong concordance between predicted and observed risks. Figure 4 presents the Gradient-weighted Class Activation Mapping (Grad-CAM), which is used to visualize salient regions in CT images that contribute most to the model’s predictions. Table 2 Diagnostic performance of the ViT and radiomics models in the validation cohorts model AUC (95%CI) Sensitivity (95%CI) Specificity (95%CI) Accuracy (95%CI) F1-score (95%CI) Internal validation cohort ViT model 0.887 (0.848–0.927) 0.784 (0.733–0.832) 0.825 (0.714–0.926) 0.805 (0.743–0.859) 0.864 (0.829–0.896) Radiomics model 0.775 (0.714–0.837) 0.731 (0.676–0.783) 0.673 (0.532–0.809) 0.702 (0.619–0.722) 0.817 (0.780–0.854) External validation cohort ViT model 0.796 (0.693–0.899) 0.774 (0.682–0.860) 0.602 (0.333–0.857) 0.688 (0.549–0.824) 0.838 (0.772–0.897) Radiomics model 0.544 (0.388–0.699) 1.000 (1.000–1.000) 0.000 (0.000–0.000) 0.500 (0.500–0.500) 0.919 (0.875–0.958) ViT: Vision Transformer; CI: confidence interval Discussion The ViT model developed in our study demonstrated excellent performance in predicting histological grades of LHSCC, with an AUC of 0.887 and 0.796 in the internal and external validation cohort, respectively. To the best of our knowledge, this is the largest LHSCC cohort analyzed by radiomics and deep learning. Previous studies have established the significant correlation between histological differentiation and disease prognosis in LHSCC. Daneshi et al. [ 22 ] demonstrated a 59% increased mortality risk in advanced-stage poorly differentiated LSCC compared to well-differentiated tumors. Zhu et al. [ 23 ] identified tumor differentiation as an independent prognostic factor for 5-year survival in LSCC patients. Wang et al. [ 24 ] reported significantly higher rates of lymph node metastasis in early-stage poorly differentiated LSCC versus moderately/well-differentiated cases. Poorly differentiated histology correlates with extranodal extension, an established predictor of poor survival in LHSCC [ 25 ]. These findings highlight the critical importance of accurately assessing tumor differentiation before treatment. Although advanced imaging techniques, such as the 18 F-FDG PET/MRI model proposed by Meng et al. [ 26 ] achieved excellent preoperative prediction of LHSCC differentiation (AUC = 0.936), and dual-energy CT has also demonstrated promising diagnostic performance [ 27 ], these high-end imaging modalities remain largely inaccessible in primary care settings. Our study developed a more universally applicable differentiation prediction model through advanced analytical methods based on routinely performed neck CECT (the standard imaging modality for LHSCC staging). This approach aims to provide a widely accessible and precise diagnostic tool for institutions with varying levels of healthcare resources. Multiple studies have demonstrated that radiomics can be used to assess tumor histological differentiation grades. Arthur et al reported that a CT-based radiomics model can predict the histological grades of retroperitoneal sarcomas with excellent performance (AUC of 0.88 on an external test set) [ 28 ]. Liu et al. successfully constructed an MRI-based radiomics model that integrates intratumoral and peritumoral features for distinguishing the pathological differentiation of hepatocellular carcinoma, achieving an AUC of 0.86 on the validation cohorts [ 12 ]. It is noteworthy that this approach also demonstrates potential utility in predicting the differentiation degree of HNSCC. Li et al. [ 29 ] (n = 178) and Wu et al. [ 30 ] (n = 206) reported that CECT-based radiomics features demonstrated excellent performance in predicting the histological differentiation grades of HNSCC. In the current study, the performance of our radiomics model was not as good as that reported in previous studies, with the AUC of our radiomics model being only 0.776 and 0.544 for the validation cohort. This difference may be related to different sample sizes and disease subtypes. Deep learning has also demonstrated significant value in assessing tumor histological differentiation grades. Zhao et al. developed a deep learning model based on conventional MRI that enables noninvasive preoperative differentiation of histological differentiation grades in early-stage renal cell carcinoma [ 31 ]. Notably, ViT, originally developed for natural language processing, has emerged as a foundational architecture for models like GPT. In computer vision, ViT have demonstrated effectiveness for non-medical image classification [ 32 ], with growing applications in medical imaging that show promising performance [ 33 ], Currently, ViT represent the state-of-the-art in both image and language processing tasks [ 34 ]. Yang et al. used Transformer-based model to classify the grade of clear cell renal cell carcinoma with high performance[ 35 ], and Yu et al. proposed a novel 3D block aggregation Transformer for efficient segmentation of kidney substructures with small datasets[ 36 ]. In ViT, images are divided into fixed-size patches that are sequentially fed into the Transformer as input tokens. This architecture leverages self-attention mechanisms to capture global relationships among image patches, enabling comprehensive image modeling-a particularly valuable feature for processing high-dimensional, high-resolution medical images requiring anatomical context [ 37 , 38 ]. However, it should be noted that Transformers typically demand larger training datasets compared to conventional CNNs, presenting challenges in medical imaging where annotated data is often scarce. To address this, we used SAM-Med3D as a pre-trained model, which was extensively trained on millions of high-quality medical image–mask pairs, demonstrating strong generalizability and accurate segmentation across diverse anatomical structures and multimodal images. By applying transfer learning with SAM-Med3D encoder weights to initialize our 3D ViT architecture, we enhanced CT image feature representation. Unlike models pre-trained on natural images (e.g., ImageNet), SAM-Med3D’s weights are tailored for medical imaging, mitigating domain shift and improving performance. Additionally, the use of an XGBoost classifier helped reduce overfitting in predicting histological differentiation of LHSCC. Our full 3D feature extraction preserved the spatial and biological characteristics of tumors more comprehensively than traditional 2D approaches, enabling richer tumor characterization. In this study, we developed a ViT-based deep learning model specifically for LHSCC, utilizing the largest multicenter cohort to date in this field (n = 1,648) while meticulously preserving the natural distribution ratio between well/moderately differentiated and poorly differentiated tumors (approximately 8:2) observed in real-world clinical practice. In contrast to previous studies that combined multiple anatomical subsites of HNSCC, our targeted study design focusing on this specific subtype enabled more accurate characterization of LHSCC-specific deep learning features. Although challenged by class imbalance, this real-world data-driven research strategy effectively mitigated potential model biases introduced by artificial balancing methods such as oversampling, thereby significantly enhancing the clinical translational value of our findings. Validation results demonstrated that our model exhibited superior robustness in both internal and external validation cohorts, with performance significantly outperforming conventional radiomics approaches. Notably, the traditional radiomics model showed substantially degraded predictive efficacy in the external test set, revealing poor generalizability. This study has several limitations. First, the manual tumor segmentation method is time-consuming, and future clinical applications may require the development of semi-automated or fully automated segmentation techniques. Second, the model does not incorporate clinical information. Although existing literature suggests clinical parameters provide limited improvement to model performance [ 30 , 39 ], their impact on LHSCC remains unclear. Subsequent studies will incorporate clinical indicators (e.g., alcohol consumption history, smoking history) and blood biomarkers to further evaluate their potential for model optimization. Third, there is an imbalance between the number of patients with LHSCC within our cohort. However, to better approximate real-world clinical scenarios, we preserved the authentic distribution ratio of histological differentiation grades in LHSCC. Conclusion In conclusion, our large-scale, multicenter ViT model demonstrated strong performance in predicting histological differentiation grades (well/moderately differentiated vs. poorly differentiated) in patients with LHSCC. In addition to its predictive accuracy, the model also showed good generalizability across data from different centers, highlighting that the proposed non-invasive approach has significant potential to support prognostic assessment and inform clinical decision-making in LHSCC management. Declarations Author contributions R.G. and X.Q. contributed equally to this study. Conception and design of the study: R.G., X.Q., J.X. Data collection: R.G., X.W., Z.S., R.X. Analysis and interpretation of data: R.G., S.T., Z.L. Writing - review and editing: R.G., X.Q., and J.X. All authors contributed to the article and approved the submitted version. Funding This work was supported by the Beijing Municipal Administration of Hospitals’ Ascent Plan (DFL20190203) and National Natural Science Foundation of China (82471951). Data availability The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request. Conflict of interest The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Ethical approval The study conformed to the ethical principles for medical research involving human participants as outlined in the World Medical Association's Declaration of Helsinki. This study was approved by the Ethics Committee of Hospital 1 (name anonymized per journal requirements). Informed consent The informed consent requirement was waived by the Ethics Committee due to the retrospective nature of the study. References Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I et al (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA Cancer J Clin 74:229–263. https://doi.org/10.3322/caac.21834 G Wiernik 1, P R Millard, J L Haybittle (1991) The predictive value of histological classification into degrees of differentiation of squamous cell carcinoma of the larynx and hypopharynx compared with the survival of patients. Histopathology 19:411–417. https://doi.org/10.1111/j.1365-2559.1991.tb00230.x Coca‐Pelaz A, Rodrigo JP, Suárez C (2012) Clinicopathologic analysis and predictive factors for distant metastases in patients with head and neck squamous cell carcinomas, Head Neck 34:771–775. https://doi.org/10.1002/hed.21804 Qian X, Sinikovic B, Schreiber F, Ochsenreither S, Klinghammer K, Wollenberg B et al (2018) pN status predicts outcomes in surgically treated pT1–pT2 patients of various disease stages with squamous cell carcinoma of the head and neck: a 17-year retrospective single center cohort study. Eur Arch Otorhinolaryngol 275:2787–2795. https://doi.org/10.1007/s00405-018-5108-z Ding S, Guo W, Yin G, Li N, Liu H, Huang J et al (2022) Clinical study of poorly differentiated head and neck squamous cell carcinoma: a pros-pective cohort study in China, Ann Transl Med 10: 703. https://doi.org/10.21037/atm-22-2630 Zhu Y, Shi X, Zhu X, Diao W, Chen X (2022) Association between pathological differentiation and survival outcomes of patients with laryngeal squamous cell carcinoma, Eur Arch Otorhinolaryngol 279:4595–4604. https://doi.org/10.1007/s00405-022-07420-9 Tang D, Tao L, Zhou L, Zhang M, Wu H, Li X et al (2018) Retrospective analysis of 659 laryngeal squamous cell carcinoma patients treated with open laryngeal function-preserving operations, Acta Otolaryngol 138:1043–1050. https://doi.org/10.1080/00016489.2018.1500711 Perez-Lopez R, Ghaffari Laleh N, Mahmood F, Kather JN (2024) A guide to artificial intelligence for cancer researchers, Nat Rev Cancer 24:427–441. https://doi.org/10.1038/s41568-024-00694-7 Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Madabhushi (2022) Predicting cancer outcomes with radiomics and artificial intelligence in radiology, Nat Rev Clin Oncol 19:132–146. https://doi.org/10.1038/s41571-021-00560-7 Zhang G, Xu L, Zhao L, Mao L, Li X, Jin Z et al (2020). CT-based radiomics to predict the pathological grade of bladder cancer, Eur Radiol 30:6749–6756. https://doi.org/10.1007/s00330-020-06893-8 Ye J-Y, Fang P, Peng Z-P, Huang X-T, Xie J-Z, Yin X-Y (2023) A radiomics-based interpretable model to predict the pathological grade of pancreatic neuroendocrine tumors, Eur Radiol 34:1994–2005. https://doi.org/10.1007/s00330-023-10186-1 Liu H-F, Wang M, Wang Q, Lu Y, Lu Y-J, Sheng Y et al (2024) Multiparametric MRI-based intratumoral and peritumoral radiomics for predicting the pathological differentiation of hepatocellular carcinoma, Insights Imaging 15: 97. https://doi.org/10.1186/s13244-024-01623-w Zheng Y, Che J, Yuan M, Wu Z, Pang J, Zhou R et al (2023) A CT-Based Deep Learning Radiomics Nomogram to Predict Histological Grades of Head and Neck Squamous Cell Carcinoma, Acad Radiol 30:1591–1599. https://doi.org/10.1016/j.acra.2022.11.007 Fujima N, Homma A, Harada T, Shimizu Y, Tha KK, Kano S et al (2019) The utility of MRI histogram and texture analysis for the prediction of histological diagnosis in head and neck malignancies, Cancer Imaging 19:5. https://doi.org/10.1186/s40644-019-0193-9 Chan JYK, Zhen G, Agrawal N (2019) The role of tumor DNA as a diagnostic tool for head and neck squamous cell carcinoma, Semin. Cancer Biol 55: 1–7. https://doi.org/10.1016/j.semcancer.2018.07.008 Mody MD, Rocco JW, Yom SS, Haddad RI, Saba NF (2021) Head and neck cancer, Lancet 398:2289–99. https://doi.org/10.1016/S0140-6736(21)01550-6 Vaswani, A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez, AN et al (2017). Attention is all you need. Adv Neural Inf Process Syst 30:5998-6008.https://doi.org/10.48550/arXiv.1706.03762 Shmatko A, Ghaffari Laleh N, Gerstung M, Kather JN (2022) Artificial intelligence in histopathology: enhancing cancer research and clinical oncology, Nat Cancer 3:1026–1038. https://doi.org/10.1038/s43018-022-00436-4 Wagner SJ, Reisenbüchler D, West NP, Niehues JM, Zhu J, Foersch S et al (2023) Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study, Cancer Cell 41:1650-1661.e4. https://doi.org/10.1016/j.ccell.2023.08.002 Khan A, Rauf Z, Sohail A, Khan AR, Asif H, Asif A et al (2023) A survey of the vision transformers and their CNN-transformer based variants, Artif Intell Rev 56:2917–2970. https://doi.org/10.1007/s10462-023-10595-0 Wang H, Guo S, Ye J, Deng Z, Cheng J, Li T et al (2023) SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images. arXiv preprint arXiv:2310.15161. https://doi.org/10.48550/arXiv.2310.15161 Daneshi N, Fararouei M, Mohammadianpanah M, Zare-Bandamiri M, Parvin S, Dianatinasab M (2018) Effects of Different Treatment Strategies and Tumor Stage on Survival of Patients with Advanced Laryngeal Carcinoma: A 15-Year Cohort Study, J Cancer Epidemiol 9678097 https://doi.org/10.1155/2018/9678097 Zhu X, Heng Y, Zhou L, Zhang M, Li W, Tao L (2020) Survival prediction and treatment strategies for patients with advanced laryngeal carcinoma: a population-based study, Int J Clin Oncol 25: 1483–1491. https://doi.org/10.1007/s10147-020-01688-9 Wang S-X, Ning W-J, Zhang X-W, Tang P-Z, Li Z-J, Liu W-S (2019) Predictors of Occult Lymph Node Metastasis and Prognosis in Patients with cN0 T1–T2 Supraglottic Laryngeal Carcinoma: A Retrospective Study, ORL J Otorhinolaryngol Relat Spec 81:317–326. https://doi.org/10.1159/000503007 Wang Z, Zeng Q, Li Y, Lu T, Liu C, Hu G (2020) Extranodal Extension as an Independent Prognostic factor in Laryngeal Squamous Cell Carcinoma Patients, J Cancer 11:7196–7201. https://doi.org/10.7150/jca.47700 Meng Z, Zhang L, Huang C, Piao Y, Chen X, Xian J (2022) Quantitative parameters derived from 18F-fluorodeoxyglucose positron emission tomography/magnetic resonance imaging can accurately estimate the histologic grade of hypopharyngeal squamous cell carcinoma preoperatively, Neuroradiology 64:2153–2162. https://doi.org/10.1007/s00234-022-03052-2 Geng D, Chen X, Zhao X-G, Xu X-Q, Su G-Y, Zhou Y et al (2023) Laryngeal and hypopharyngeal squamous cell carcinoma: association between quantitative parameters derived from dual-energy CT and histopathological prognostic factors, Acta Radiol. ;64: 2268–2276. https://doi.org/10.1177/02841851221095237 Arthur A, Orton MR, Emsley R, Vit S, Kelly-Morland C, Strauss D et al (2023) A CT-based radiomics classification model for the prediction of histological type and tumour grade in retroperitoneal sarcoma (RADSARC-R): a retrospective multicohort analysis, Lancet Oncol 24:1277–1286. https://doi.org/10.1016/S1470-204523)00462- Li Z, Liu Z, Guo Y, Wang S, Qu X, Li Y et al (2022) Dual-energy CT-based radiomics nomogram in predicting histological differentiation of head and neck squamous carcinoma: a multicenter study, Neuroradiology 64:361–369. https://doi.org/10.1007/s00234-021-02860-2 Wu W, Ye J, Wang Q, Luo J, Xu S (2019) CT-Based Radiomics Signature for the Preoperative Discrimination Between Head and Neck Squamous Cell Carcinoma Grades, Front Oncol 9:821. https://doi.org/10.3389/fonc.2019.00821 Zhao Y, Chang M, Wang R, Xi IL, Chang K, Huang RY et al (2020) Deep Learning Based on MRI for Differentiation of Low‐ and High‐Grade in Low‐Stage Renal Cell Carcinoma, J Magn Reson Imaging 52:1542–1549. https://doi.org/10.1002/jmri.27153 Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z et al (2021) Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 10012-10022. https://doi.org/10.1109/ICCV48922.2021.00986 Shamshad F, Khan S, Zamir SW, Khan MH, Hayat M, Khan FS et al (2023) Transformers in medical imaging: A survey, Med Image Anal 88: 102802. https://doi.org/10.1016/j.media.2023.102802 Khan RF, Lee B-D, Lee MS (2023) Transformers in medical image segmentation: a narrative review, Quant Imaging Med Surg 13:8747–8767. https://doi.org/10.21037/qims-23-542 Yang M, He X, Xu L, Liu M, Deng J, Cheng X et al (2022) CT-based transformer model for non-invasively predicting the Fuhrman nuclear grade of clear cell renal cell carcinoma, Front Oncol 12:961779. https://doi.org/10.3389/fonc.2022.961779 Cicalese PA, Mobiny A, Shahmoradi Z, Yi X, Mohan C, Van Nguyen H (2021) Kidney Level Lupus Nephritis Classification Using Uncertainty Guided Bayesian Convolutional Neural Networks, IEEE J Biomed Health Inform 25:315–324. https://doi.org/10.1109/JBHI.2020.3039162 Li J, Chen J, Tang Y, Wang C, Landman BA, Zhou SK (2023) Transforming medical imaging with Transformers? A comparative review of key properties, current progresses, and future perspectives, Med Image Anal 85:102762. https://doi.org/10.1016/j.media.2023.102762 Liu Y, Zhang Y, Wang Y, Hou F, Yuan J, Tian J et al (2024) A Survey of Visual Transformers, IEEE Trans Neural Netw Learn Sys 35:7478–7498. https://doi.org/10.1109/TNNLS.2022.3227717 Ye G, Wu G, Li K, Zhang C, Zhuang Y, Liu H et al (2024) Development and Validation of a Deep Learning Radiomics Model to Predict High-Risk Pathologic Pulmonary Nodules Using Preoperative Computed Tomography, Acad Radiol 31:1686–1697. https://doi.org/10.1016/j.acra.2023.08.040 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 27 Dec, 2025 Read the published version in Neuroradiology → Version 1 posted 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-7249038","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":494884148,"identity":"c0e51e9e-e434-4609-b263-f0dcca8c8df2","order_by":0,"name":"Ran Guo","email":"","orcid":"","institution":"Beijing Tongren Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ran","middleName":"","lastName":"Guo","suffix":""},{"id":494884149,"identity":"c0d87ef4-cfe8-497c-84c6-3b31417fa908","order_by":1,"name":"Xiaoxia Qu","email":"","orcid":"","institution":"Beijing Tongren Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoxia","middleName":"","lastName":"Qu","suffix":""},{"id":494884150,"identity":"f50670d8-0e27-46b3-b3c6-f37803bcfc4b","order_by":2,"name":"Song Tian","email":"","orcid":"","institution":"Philips Healthcare","correspondingAuthor":false,"prefix":"","firstName":"Song","middleName":"","lastName":"Tian","suffix":""},{"id":494884151,"identity":"c281b9eb-d007-420c-9226-52de9b85306c","order_by":3,"name":"Zheng Li","email":"","orcid":"","institution":"Beijing Tongren Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zheng","middleName":"","lastName":"Li","suffix":""},{"id":494884152,"identity":"bd1caafe-b6a6-46d1-b07f-8bf826fc05d5","order_by":4,"name":"Xinyan Wang","email":"","orcid":"","institution":"Beijing Tongren Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xinyan","middleName":"","lastName":"Wang","suffix":""},{"id":494884153,"identity":"9eb3ded3-632a-404b-a289-060fd2f804ac","order_by":5,"name":"Zhenchao Sun","email":"","orcid":"","institution":"Linyi People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zhenchao","middleName":"","lastName":"Sun","suffix":""},{"id":494884154,"identity":"78d392f1-8a0e-49ef-ac2a-2e6488c3c929","order_by":6,"name":"Ruiqiang Xin","email":"","orcid":"","institution":"Beijing Luhe Hospital, Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ruiqiang","middleName":"","lastName":"Xin","suffix":""},{"id":494884155,"identity":"38a24978-f7d0-4fa6-8ccb-3e36b7392d03","order_by":7,"name":"Junfang Xian","email":"data:image/png;base64,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","orcid":"","institution":"Beijing Tongren Hospital, Capital Medical University","correspondingAuthor":true,"prefix":"","firstName":"Junfang","middleName":"","lastName":"Xian","suffix":""}],"badges":[],"createdAt":"2025-07-30 06:08:42","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7249038/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7249038/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00234-025-03876-8","type":"published","date":"2025-12-27T15:57:20+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":88408745,"identity":"079d37f9-e60b-4b7e-8514-c1cd6286a781","added_by":"auto","created_at":"2025-08-06 08:14:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1061968,"visible":true,"origin":"","legend":"\u003cp\u003eRecruitment pathway for eligible patients with laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) in this study.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7249038/v1/dd69d91659cc71927832c479.png"},{"id":88408743,"identity":"729dfceb-04e2-46cf-90a9-9a583c6aec11","added_by":"auto","created_at":"2025-08-06 08:14:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1962446,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow of the Vision Transformer (ViT) model construction for histological grade prediction of LHSCC\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7249038/v1/fdb1a95323513b131b1a49ae.png"},{"id":88410341,"identity":"cdada0aa-e9de-447b-8634-2900f848cdeb","added_by":"auto","created_at":"2025-08-06 08:22:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":380817,"visible":true,"origin":"","legend":"\u003cp\u003eEvaluation of the ViT and radiomics models for LHSCC histological grade prediction: ROC curves, decision curve analysis, and calibration curves in the internal (a-c) and external validation cohorts (d-f). ViT: Vision Transformer; LHSCC: Laryngeal and hypopharyngeal squamous cell carcinoma\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7249038/v1/eb094fbadeb2996ae8a1a7c3.png"},{"id":88408755,"identity":"7406a87d-24ca-410b-be99-314cdb81cab7","added_by":"auto","created_at":"2025-08-06 08:14:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":597808,"visible":true,"origin":"","legend":"\u003cp\u003eAxial contrast-enhanced CT images (left side of each panel) and the corresponding Grad-CAM visualizations (right side of each panel) generated from the final layer of the Vision Transformer (ViT) encoder, are shown for six patients. Activation regions are color-coded: red indicating higher attention weights. (a) A 61-year-old male patient with poorly differentiated squamous cell carcinoma (SCC), showing markedly increased attention weights concentrated on the lesion. (b) A 48-year-old male patient with poorly differentiated SCC. Mild focal activation is observed in the anterior portion of the lesion, with a ViT-predicted probability of 0.658. (c) A 65-year-old male patient with poorly differentiated SCC. No significant activation regions are detected, and the ViT model predicts a probability of 0.246. (d) A 65-year-old male patient with well/moderately differentiated SCC, showing increased attention focused on the lesion area. (e) A 72-year-old male patient with well/moderately differentiated SCC. Multiple activation regions are noted in the anterior portion of the lesion, and the ViT model predicts a probability of 0.780 for well/moderately differentiated histology. (f) A 61-year-old male patient with well/moderately differentiated SCC. No significant activation is observed, with a final prediction probability of 0.291\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7249038/v1/a3123de9f1afd41b7f07a0b5.png"},{"id":99172264,"identity":"65cf3db1-2c78-4c16-ac3b-238ec35029c0","added_by":"auto","created_at":"2025-12-29 16:06:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4700071,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7249038/v1/dca65b6e-1430-440f-bb2c-044fd53ba06b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Leveraging Vision Transformer for Histological Grade Prediction in Laryngeal and Hypopharyngeal Squamous Cell Carcinoma: A Large-Scale Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHead and neck squamous cell carcinoma (HNSCC) ranks as the sixth most prevalent cancer worldwide. Globally, approximately 890,000 new cases of head and neck cancer are diagnosed annually, of which laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) account for 189,000 and 86,000 new cases respectively, with corresponding mortality rates of 103,000 and 41,000 deaths per year [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Squamous cell carcinoma can be further classified into well-differentiated, moderately differentiated, and poorly differentiated types based on the degree of cell differentiation. Among these, well/moderately differentiated squamous cell carcinomas are the most common. Notably, although poorly differentiated tumors account for a smaller proportion, they exhibit more aggressive biological behavior: demonstrating significantly greater propensity for local invasion, distant metastasis [\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], and higher sensitivity to chemoradiotherapy, yet generally poorer treatment outcomes and unfavorable prognosis [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Earlier study showed a substantial difference in 5-year overall survival rates between well/moderately-differentiated and poorly differentiated patients, ranging from 12\u0026ndash;30% [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Currently, the assessment of histological differentiation grades primarily relies on biopsy and surgical resection specimens. Nevertheless, for LHSCC, invasive biopsies have limited predictive value and often fail to fully capture the tumor's comprehensive biological characteristics and heterogeneity. Therefore, developing a novel preoperative, non-invasive method to accurately predict tumor differentiation status is of significant clinical importance-this would optimize clinical decision-making and facilitate personalized treatment strategies.\u003c/p\u003e\u003cp\u003eCurrent studies demonstrate the promising applications of radiomics in oncology [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], with preoperative features capable of predicting pathological differentiation grades in various tumors including hepatocellular carcinoma, pancreatic neuroendocrine tumors, bladder cancer, and HNSCC [\u003cspan additionalcitationids=\"CR11 CR12\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Additionally, histogram and texture analysis of Fs-T2WI enables noninvasive prediction of differentiation grades in HNSCC [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. As a cornerstone of computer vision, convolutional neural networks (CNNs) can automatically extract hierarchical image features through multilayer convolution operations and have been effectively applied to assess HNSCC differentiation [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Zheng et al. developed an integrated model combining radiomics and deep learning features, achieving an AUC of 0.822 in the test cohort [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. However, current research has several limitations: 1) Significant heterogeneity in study populations, often mixing distinct anatomical subtypes of HNSCC (e.g., oropharyngeal, laryngeal, and hypopharyngeal SCC) with divergent treatment approaches and prognoses [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. 2) Insufficient sample sizes hindering the model's ability to learn complex feature-differentiation relationships; 3) Lack of well-established pretrained weights for 3D medical imaging of LHSC.\u003c/p\u003e\u003cp\u003eIn recent years, the Vision Transformer (ViT), a Transformer-based deep learning model, has shown strong knowledge transfer capabilities [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Originally successful in natural language processing, ViT has since been applied to medical image analysis with promising results. Studies have shown that ViT outperforms conventional CNNs in various medical imaging tasks [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], mainly due to its self-attention mechanism, which captures global context and long-range dependencies. In contrast, CNNs are limited by their local receptive fields, restricting their ability to model global features [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The Segment Anything Model for 3D medical images (SAM-Med3D) is a foundation model for automatic segmentation across different imaging modalities such as CT and MRI [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Notably, its encoder shares a similar architecture with ViT.\u003c/p\u003e\u003cp\u003eThis study explored an early and innovative application of ViT models fine-tuned with pre-trained weights from SAM-Med3D for predicting histological differentiation grades in LHSCC. Leveraging large-scale, multicenter contrast-enhanced CT (CECT) datasets, we developed a deep learning framework and conducted comparative analyses with conventional radiomics approach.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cp\u003e\u003cb\u003ePatients\u0026rsquo; selection\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe retrospectively collected patients with LHSCC confirmed by postoperative pathology between March 2009 and May 2024 from three medical institutions. The inclusion criteria were as follows: (1) LHSCC confirmed by pathology; (2) Histopathological examination performed within 2 weeks post-scanning; (3) No prior tumor-related treatment before surgery. Exclusion criteria included: (1) Recurrent patient; (2) Patient was diagnosed with a second primary tumor; (3) Lession maximum diameter less than 5 mm; (4) Inadequate CT imaging quality, which could not be used for the imaging analysis, or the loss of images. The patient recruitment pathway is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Ultimately, 1,648 eligible LHSCC cases were included in this study from the three hospitals. Next, they were randomly divided into a training cohort (n\u0026thinsp;=\u0026thinsp;1,239) and an internal validation cohort (n\u0026thinsp;=\u0026thinsp;310) according to the 8:2 ratio from Hospital 1, with the external validation cohort consisted of 99 patients from Hospital 2 and 3.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCECT image acquisition, segmentation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eCECT images were obtained using five different scanners: SOMATOM Definition Flash and SOMATOM Force CT (Siemens Healthineers, Germany), the Brilliance 64 and iCT 256 CT (Philips Medical Systems, Nederland), and Revolution CT (GE Healthcare, USA). CECT scans were performed in the supine position. Patients were instructed to remain motionless and avoid swallowing during image acquisition, with both slice thickness and interval set at 1 mm.\u003c/p\u003e\u003cp\u003eThe region of interest delineation was performed using ITK-SNAP (version 3.8.0, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.itksnap.org/\u003c/span\u003e\u003cspan address=\"http://www.itksnap.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) software by a radiologist with 8 years of clinical experience, employing manual slice-by-slice annotation to comprehensively encompass both cystic and necrotic components of the lesions. All delineations were subsequently reviewed and validated by a senior radiologist with 20 years of professional experience. Notably, both radiologists were blinded to the patients' definitive pathological diagnoses throughout the entire annotation process.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Preprocessing\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the workflow of our study. All scans were preprocessed using a window level of 40 HU and a window width of 350 HU. Before being fed into the model, each scan underwent Z-score normalization to ensure consistency in intensity distribution.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eRadiomics and deep learning model development\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe first trained a radiomics-based model as a baseline, following the Image Biomarker Standardization Initiative (IBSI) guidelines. The extracted radiomics features included first-order statistics, Gray Level Co-occurrence Matrix, Gray Level Dependence Matrix, Gray Level Run Length Matrix, Gray Level Size Zone Matrix, Neighbouring Gray Tone Difference Matrix, and shape features. Additionally, wavelet transformation was applied, resulting in a total of 1,454 radiomics features per sample. For feature selection, we used Pearson correlation (threshold\u0026thinsp;=\u0026thinsp;0.8) and Least absolute shrinkage and selection operator (LASSO) regression. Ultimately, nine radiomics features were retained, consisting of five texture features, three shape features, and one first-order feature.\u003c/p\u003e\u003cp\u003eFor deep learning, we used a large-scale 3D ViT model with 90\u0026nbsp;million parameters to extract high-order texture features of the tumor. The CT images were first processed through a 3D embedding layer to convert them into a sequence of flattened patch embeddings (a transformer-compatible format). These transformed features were then passed through ViT, yielding a final set of 384 extracted features. To leverage pre-trained knowledge, we incorporated SAM-Med3D as the initial pre-trained weights for 3D ViT encoder. Although SAM-Med3D is originally designed for instance segmentation, its encoder shares the architecture and has generalizable representations of 3D anatomical structures and textures from a massive and diverse medical image dataset. Therefore, we used its encoder weights as pre-trained encoder in our model.\u003c/p\u003e\u003cp\u003eFinally, we used EXtreme Gradient Boosting (XGBoost) as the classifier to train models based on both radiomics and ViT-extracted features. We defined a hyperparameter search space with a maximum tree depth of 2 to 3 and estimators ranging from 50 to 200. A five-fold cross-validation strategy was applied to the training set to optimize and evaluate the final model.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eStatistical analysis was performed using SPSS software (version 25.0, IBM) and Python software (version 3.5.6; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.python.org\u003c/span\u003e\u003cspan address=\"http://www.python.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Continuous variables were assessed for normality using the Shapiro-Wilk test. Normally distributed data are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (x̄\u0026plusmn;s) and compared using independent samples t-tests. Categorical variables are reported as frequencies and compared using chi-square tests. Model predictive performance was evaluated using receiver operating characteristic curve analysis with area under the curve (AUC), and compared by DeLong's test. Calibration was assessed via calibration curves and Hosmer-Lemeshow test. Clinical utility was evaluated using decision curve analysis (DCA). A \u003cem\u003ep\u003c/em\u003e-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eCharacteristic of patients\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis study ultimately enrolled 1,648 patients. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e lists the details of the clinical data of all patients. The training cohort included 1,239 patients from Hospital 1 (1,190 males and 49 females; mean age 61.64\u0026thinsp;\u0026plusmn;\u0026thinsp;8.82 years), consisting of 946 well/moderately differentiated cases (313 well differentiated and 633 moderately differentiated) and 293 poorly differentiated cases. The internal validation cohort included 310 patients from Hospital 1 (294 males and 16 females; mean age 62.63\u0026thinsp;\u0026plusmn;\u0026thinsp;9.24 years; 264 well/moderately differentiated including 80 well and 184 moderately differentiated, 46 poorly differentiated). The external validation cohorts comprised patients from two centers: 80 from Hospital 2 (78 males; mean age 65.63\u0026thinsp;\u0026plusmn;\u0026thinsp;8.90 years; 69 well/moderately differentiated including 26 well and 43 moderately differentiated, 11 poorly differentiated), 19 from Hospital 3 (all males; mean age 64.26\u0026thinsp;\u0026plusmn;\u0026thinsp;10.36 years; 15 well/moderately differentiated including 6 well and 9 moderately differentiated, 4 poorly differentiated), Compared to the training and internal validation cohorts, the external validation cohort has a slightly older age distribution (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).The T stages showed significant differences among the three cohort. There were no significant differences in gender, tumor location, or lymph node metastasis across all the study sets (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;0.05).\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\u003eClinical characteristics of patients with LHSCC in the training, internal validation, and external validation cohorts\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCharacteristics\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTraining cohort\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;1,239)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInternal validation\u003c/p\u003e\u003cp\u003ecohort (n\u0026thinsp;=\u0026thinsp;310)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eExternal validation\u003c/p\u003e\u003cp\u003ecohort (n\u0026thinsp;=\u0026thinsp;99)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cem\u003ep\u003c/em\u003e value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge (years)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e61.64\u0026thinsp;\u0026plusmn;\u0026thinsp;8.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e62.63\u0026thinsp;\u0026plusmn;\u0026thinsp;9.24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e65.36\u0026thinsp;\u0026plusmn;\u0026thinsp;9.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSex /n(%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.358\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1,190(96.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e294 (94.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e97(98.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e49 (4.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16 (5.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2(2.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003elocation/n(%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.112\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLarynx\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1,102 (88.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e283(91.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e83(83.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHypopharynx\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e137 (11.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e27 (8.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e16 (16.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eT stage/n(%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001*\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eT1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e412 (33.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e137(44.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e19 (19.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eT2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e400 (32.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e94 (30.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e41 (41.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eT3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e307 (24.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e54 (17.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e28 (28.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eT4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e120(9.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e25 (8.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e11 (11.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLNM/n(%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.962\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAbsent\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e489 (61.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e102 (60.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e55 (59.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePresent\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e310 (38.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e66 (39.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e37 (40.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUnknown\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e440\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e142\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"5\"\u003e*Represents \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05; LNM: lymph node metastasis; Data of age are shown as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation; other data are number of patients with percentage in parentheses.Unknown: No lymphadenectomy was done\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eComparison of the ViT and radiomics models\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe ViT model demonstrated robust performance in discriminating well/moderately differentiated from poorly differentiated grades across validation cohorts, achieving an AUC of 0.887 [95% (confidence interval,CI): 0.848\u0026ndash;0.927]in the internal validation cohort and 0.796 (95%CI: 0.693\u0026ndash;0.899) in the external validation cohort. In contrast, the radiomics model showed significantly lower performance, with AUCs of 0.775 (95%CI: 0.714\u0026ndash;0.837) and 0.544 (95%CI: 0.388\u0026ndash;0.699) in the internal and external validation cohorts, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea and d). The results of the ROC analysis of each model were shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The DeLong test confirmed the superior predictive efficacy of the ViT model over the radiomics model in both the internal (\u003cem\u003ep\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;1x10\u003csup\u003e\u0026minus;\u0026thinsp;7\u003c/sup\u003e) and external validation cohorts (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.002). DCA (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb and e) revealed that the ViT model had a higher clinical net benefit in predicting the histological differentiation grades of LHSCC. The calibration curves (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec and f) further evidenced its excellent calibration performance indicating strong concordance between predicted and observed risks. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the Gradient-weighted Class Activation Mapping (Grad-CAM), which is used to visualize salient regions in CT images that contribute most to the model\u0026rsquo;s predictions.\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\u003eDiagnostic performance of the ViT and radiomics models in the validation cohorts\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\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\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003emodel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAUC (95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003cp\u003e(95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003cp\u003e(95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003cp\u003e(95%CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003cp\u003e(95%CI)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e\u003cp\u003eInternal validation cohort\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eViT model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.887\u003c/p\u003e\u003cp\u003e(0.848\u0026ndash;0.927)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003e0.784 (0.733\u0026ndash;0.832)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e0.825 (0.714\u0026ndash;0.926)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u003cp\u003e0.805\u003c/p\u003e\u003cp\u003e(0.743\u0026ndash;0.859)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.864\u003c/p\u003e\u003cp\u003e(0.829\u0026ndash;0.896)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRadiomics model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.775\u003c/p\u003e\u003cp\u003e(0.714\u0026ndash;0.837)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003e0.731 (0.676\u0026ndash;0.783)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e0.673 (0.532\u0026ndash;0.809)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u003cp\u003e0.702 (0.619\u0026ndash;0.722)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.817\u003c/p\u003e\u003cp\u003e(0.780\u0026ndash;0.854)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e\u003cp\u003eExternal validation cohort\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eViT model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.796\u003c/p\u003e\u003cp\u003e(0.693\u0026ndash;0.899)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003e0.774\u003c/p\u003e\u003cp\u003e(0.682\u0026ndash;0.860)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e0.602 (0.333\u0026ndash;0.857)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u003cp\u003e0.688 (0.549\u0026ndash;0.824)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.838 (0.772\u0026ndash;0.897)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRadiomics model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.544\u003c/p\u003e\u003cp\u003e(0.388\u0026ndash;0.699)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003e1.000 (1.000\u0026ndash;1.000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e0.000\u003c/p\u003e\u003cp\u003e(0.000\u0026ndash;0.000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e\u003cp\u003e0.500\u003c/p\u003e\u003cp\u003e(0.500\u0026ndash;0.500)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.919 (0.875\u0026ndash;0.958)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eViT: Vision Transformer; CI: confidence interval\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe ViT model developed in our study demonstrated excellent performance in predicting histological grades of LHSCC, with an AUC of 0.887 and 0.796 in the internal and external validation cohort, respectively. To the best of our knowledge, this is the largest LHSCC cohort analyzed by radiomics and deep learning.\u003c/p\u003e\u003cp\u003ePrevious studies have established the significant correlation between histological differentiation and disease prognosis in LHSCC. Daneshi et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] demonstrated a 59% increased mortality risk in advanced-stage poorly differentiated LSCC compared to well-differentiated tumors. Zhu et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] identified tumor differentiation as an independent prognostic factor for 5-year survival in LSCC patients. Wang et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] reported significantly higher rates of lymph node metastasis in early-stage poorly differentiated LSCC versus moderately/well-differentiated cases. Poorly differentiated histology correlates with extranodal extension, an established predictor of poor survival in LHSCC [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. These findings highlight the critical importance of accurately assessing tumor differentiation before treatment. Although advanced imaging techniques, such as the \u003csup\u003e18\u003c/sup\u003eF-FDG PET/MRI model proposed by Meng et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] achieved excellent preoperative prediction of LHSCC differentiation (AUC\u0026thinsp;=\u0026thinsp;0.936), and dual-energy CT has also demonstrated promising diagnostic performance [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], these high-end imaging modalities remain largely inaccessible in primary care settings. Our study developed a more universally applicable differentiation prediction model through advanced analytical methods based on routinely performed neck CECT (the standard imaging modality for LHSCC staging). This approach aims to provide a widely accessible and precise diagnostic tool for institutions with varying levels of healthcare resources.\u003c/p\u003e\u003cp\u003eMultiple studies have demonstrated that radiomics can be used to assess tumor histological differentiation grades. Arthur et al reported that a CT-based radiomics model can predict the histological grades of retroperitoneal sarcomas with excellent performance (AUC of 0.88 on an external test set) [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Liu et al. successfully constructed an MRI-based radiomics model that integrates intratumoral and peritumoral features for distinguishing the pathological differentiation of hepatocellular carcinoma, achieving an AUC of 0.86 on the validation cohorts [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. It is noteworthy that this approach also demonstrates potential utility in predicting the differentiation degree of HNSCC. Li et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] (n\u0026thinsp;=\u0026thinsp;178) and Wu et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] (n\u0026thinsp;=\u0026thinsp;206) reported that CECT-based radiomics features demonstrated excellent performance in predicting the histological differentiation grades of HNSCC. In the current study, the performance of our radiomics model was not as good as that reported in previous studies, with the AUC of our radiomics model being only 0.776 and 0.544 for the validation cohort. This difference may be related to different sample sizes and disease subtypes.\u003c/p\u003e\u003cp\u003eDeep learning has also demonstrated significant value in assessing tumor histological differentiation grades. Zhao et al. developed a deep learning model based on conventional MRI that enables noninvasive preoperative differentiation of histological differentiation grades in early-stage renal cell carcinoma [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Notably, ViT, originally developed for natural language processing, has emerged as a foundational architecture for models like GPT. In computer vision, ViT have demonstrated effectiveness for non-medical image classification [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], with growing applications in medical imaging that show promising performance [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], Currently, ViT represent the state-of-the-art in both image and language processing tasks [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Yang et al. used Transformer-based model to classify the grade of clear cell renal cell carcinoma with high performance[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e], and Yu et al. proposed a novel 3D block aggregation Transformer for efficient segmentation of kidney substructures with small datasets[\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. In ViT, images are divided into fixed-size patches that are sequentially fed into the Transformer as input tokens. This architecture leverages self-attention mechanisms to capture global relationships among image patches, enabling comprehensive image modeling-a particularly valuable feature for processing high-dimensional, high-resolution medical images requiring anatomical context [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. However, it should be noted that Transformers typically demand larger training datasets compared to conventional CNNs, presenting challenges in medical imaging where annotated data is often scarce. To address this, we used SAM-Med3D as a pre-trained model, which was extensively trained on millions of high-quality medical image\u0026ndash;mask pairs, demonstrating strong generalizability and accurate segmentation across diverse anatomical structures and multimodal images. By applying transfer learning with SAM-Med3D encoder weights to initialize our 3D ViT architecture, we enhanced CT image feature representation. Unlike models pre-trained on natural images (e.g., ImageNet), SAM-Med3D\u0026rsquo;s weights are tailored for medical imaging, mitigating domain shift and improving performance. Additionally, the use of an XGBoost classifier helped reduce overfitting in predicting histological differentiation of LHSCC. Our full 3D feature extraction preserved the spatial and biological characteristics of tumors more comprehensively than traditional 2D approaches, enabling richer tumor characterization.\u003c/p\u003e\u003cp\u003eIn this study, we developed a ViT-based deep learning model specifically for LHSCC, utilizing the largest multicenter cohort to date in this field (n\u0026thinsp;=\u0026thinsp;1,648) while meticulously preserving the natural distribution ratio between well/moderately differentiated and poorly differentiated tumors (approximately 8:2) observed in real-world clinical practice. In contrast to previous studies that combined multiple anatomical subsites of HNSCC, our targeted study design focusing on this specific subtype enabled more accurate characterization of LHSCC-specific deep learning features. Although challenged by class imbalance, this real-world data-driven research strategy effectively mitigated potential model biases introduced by artificial balancing methods such as oversampling, thereby significantly enhancing the clinical translational value of our findings. Validation results demonstrated that our model exhibited superior robustness in both internal and external validation cohorts, with performance significantly outperforming conventional radiomics approaches. Notably, the traditional radiomics model showed substantially degraded predictive efficacy in the external test set, revealing poor generalizability.\u003c/p\u003e\u003cp\u003eThis study has several limitations. First, the manual tumor segmentation method is time-consuming, and future clinical applications may require the development of semi-automated or fully automated segmentation techniques. Second, the model does not incorporate clinical information. Although existing literature suggests clinical parameters provide limited improvement to model performance [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], their impact on LHSCC remains unclear. Subsequent studies will incorporate clinical indicators (e.g., alcohol consumption history, smoking history) and blood biomarkers to further evaluate their potential for model optimization. Third, there is an imbalance between the number of patients with LHSCC within our cohort. However, to better approximate real-world clinical scenarios, we preserved the authentic distribution ratio of histological differentiation grades in LHSCC.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, our large-scale, multicenter ViT model demonstrated strong performance in predicting histological differentiation grades (well/moderately differentiated vs. poorly differentiated) in patients with LHSCC. In addition to its predictive accuracy, the model also showed good generalizability across data from different centers, highlighting that the proposed non-invasive approach has significant potential to support prognostic assessment and inform clinical decision-making in LHSCC management.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003eR.G. and X.Q. contributed equally to this study. Conception and design of the study: R.G., X.Q., J.X. Data collection: R.G., X.W., Z.S., R.X. Analysis and interpretation of data: R.G., S.T., Z.L. Writing - review and editing: R.G., X.Q., and J.X. All authors contributed to the article and approved the submitted version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e This work was supported by the Beijing Municipal Administration of Hospitals\u0026rsquo; Ascent Plan (DFL20190203) and National Natural Science Foundation of China (82471951).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interest\u0026nbsp;\u003c/strong\u003eThe author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval\u0026nbsp;\u003c/strong\u003eThe study conformed to the ethical principles for medical research involving human participants as outlined in the World Medical Association\u0026apos;s Declaration of Helsinki. This study was approved by the Ethics Committee of Hospital 1 (name anonymized per journal requirements).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed consent\u003c/strong\u003e The informed consent requirement was waived by the Ethics Committee due to the retrospective nature of the study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I et al (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA Cancer J Clin 74:229\u0026ndash;263. https://doi.org/10.3322/caac.21834\u003c/li\u003e\n\u003cli\u003eG Wiernik 1, P R Millard, J L Haybittle (1991) The predictive value of histological classification into degrees of differentiation of squamous cell carcinoma of the larynx and hypopharynx compared with the survival of patients. Histopathology 19:411\u0026ndash;417. https://doi.org/10.1111/j.1365-2559.1991.tb00230.x\u003c/li\u003e\n\u003cli\u003eCoca‐Pelaz A, Rodrigo JP, Su\u0026aacute;rez C (2012) Clinicopathologic analysis and predictive factors for distant metastases in patients with head and neck squamous cell carcinomas, Head Neck 34:771\u0026ndash;775. https://doi.org/10.1002/hed.21804\u003c/li\u003e\n\u003cli\u003eQian X, Sinikovic B, Schreiber F, Ochsenreither S, Klinghammer K, Wollenberg B et al (2018) pN status predicts outcomes in surgically treated pT1\u0026ndash;pT2 patients of various disease stages with squamous cell carcinoma of the head and neck: a 17-year retrospective single center cohort study. Eur Arch Otorhinolaryngol 275:2787\u0026ndash;2795. https://doi.org/10.1007/s00405-018-5108-z\u003c/li\u003e\n\u003cli\u003eDing S, Guo W, Yin G, Li N, Liu H, Huang J et al (2022) Clinical study of poorly differentiated head and neck squamous cell carcinoma: a pros-pective cohort study in China, Ann Transl Med 10: 703. https://doi.org/10.21037/atm-22-2630\u003c/li\u003e\n\u003cli\u003eZhu Y, Shi X, Zhu X, Diao W, Chen X (2022) Association between pathological differentiation and survival outcomes of patients with laryngeal squamous cell carcinoma, Eur Arch Otorhinolaryngol 279:4595\u0026ndash;4604. https://doi.org/10.1007/s00405-022-07420-9\u003c/li\u003e\n\u003cli\u003eTang D, Tao L, Zhou L, Zhang M, Wu H, Li X et al (2018) Retrospective analysis of 659 laryngeal squamous cell carcinoma patients treated with open laryngeal function-preserving operations, Acta Otolaryngol 138:1043\u0026ndash;1050. https://doi.org/10.1080/00016489.2018.1500711\u003c/li\u003e\n\u003cli\u003ePerez-Lopez R, Ghaffari Laleh N, Mahmood F, Kather JN (2024) A guide to artificial intelligence for cancer researchers, Nat Rev Cancer 24:427\u0026ndash;441. https://doi.org/10.1038/s41568-024-00694-7\u003c/li\u003e\n\u003cli\u003eBera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Madabhushi (2022) Predicting cancer outcomes with radiomics and artificial intelligence in radiology, Nat Rev Clin Oncol 19:132\u0026ndash;146. https://doi.org/10.1038/s41571-021-00560-7\u003c/li\u003e\n\u003cli\u003eZhang G, Xu L, Zhao L, Mao L, Li X, Jin Z et al (2020). CT-based radiomics to predict the pathological grade of bladder cancer, Eur Radiol 30:6749\u0026ndash;6756. https://doi.org/10.1007/s00330-020-06893-8\u003c/li\u003e\n\u003cli\u003eYe J-Y, Fang P, Peng Z-P, Huang X-T, Xie J-Z, Yin X-Y (2023) A radiomics-based interpretable model to predict the pathological grade of pancreatic neuroendocrine tumors, Eur Radiol 34:1994\u0026ndash;2005. https://doi.org/10.1007/s00330-023-10186-1\u003c/li\u003e\n\u003cli\u003eLiu H-F, Wang M, Wang Q, Lu Y, Lu Y-J, Sheng Y et al (2024) Multiparametric MRI-based intratumoral and peritumoral radiomics for predicting the pathological differentiation of hepatocellular carcinoma, Insights Imaging 15: 97. https://doi.org/10.1186/s13244-024-01623-w\u003c/li\u003e\n\u003cli\u003eZheng Y, Che J, Yuan M, Wu Z, Pang J, Zhou R et al (2023) A CT-Based Deep Learning Radiomics Nomogram to Predict Histological Grades of Head and Neck Squamous Cell Carcinoma, Acad Radiol 30:1591\u0026ndash;1599. https://doi.org/10.1016/j.acra.2022.11.007\u003c/li\u003e\n\u003cli\u003eFujima N, Homma A, Harada T, Shimizu Y, Tha KK, Kano S et al (2019) The utility of MRI histogram and texture analysis for the prediction of histological diagnosis in head and neck malignancies, Cancer Imaging 19:5. https://doi.org/10.1186/s40644-019-0193-9\u003c/li\u003e\n\u003cli\u003eChan JYK, Zhen G, Agrawal N (2019) The role of tumor DNA as a diagnostic tool for head and neck squamous cell carcinoma, Semin. Cancer Biol 55: 1\u0026ndash;7. https://doi.org/10.1016/j.semcancer.2018.07.008\u003c/li\u003e\n\u003cli\u003eMody MD, Rocco JW, Yom SS, Haddad RI, Saba NF (2021) Head and neck cancer, Lancet 398:2289\u0026ndash;99. https://doi.org/10.1016/S0140-6736(21)01550-6\u003c/li\u003e\n\u003cli\u003eVaswani, A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez, AN et al (2017). Attention is all you need. Adv Neural Inf Process Syst 30:5998-6008.https://doi.org/10.48550/arXiv.1706.03762\u003c/li\u003e\n\u003cli\u003eShmatko A, Ghaffari Laleh N, Gerstung M, Kather JN (2022) Artificial intelligence in histopathology: enhancing cancer research and clinical oncology, Nat Cancer 3:1026\u0026ndash;1038. https://doi.org/10.1038/s43018-022-00436-4\u003c/li\u003e\n\u003cli\u003eWagner SJ, Reisenb\u0026uuml;chler D, West NP, Niehues JM, Zhu J, Foersch S et al (2023) Transformer-based biomarker prediction from colorectal cancer histology: A large-scale multicentric study, Cancer Cell 41:1650-1661.e4. https://doi.org/10.1016/j.ccell.2023.08.002\u003c/li\u003e\n\u003cli\u003eKhan A, Rauf Z, Sohail A, Khan AR, Asif H, Asif A et al (2023) A survey of the vision transformers and their CNN-transformer based variants, Artif Intell Rev 56:2917\u0026ndash;2970. https://doi.org/10.1007/s10462-023-10595-0\u003c/li\u003e\n\u003cli\u003eWang H, Guo S, Ye J, Deng Z, Cheng J, Li T et al (2023) SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images. arXiv preprint arXiv:2310.15161. https://doi.org/10.48550/arXiv.2310.15161\u003c/li\u003e\n\u003cli\u003eDaneshi N, Fararouei M, Mohammadianpanah M, Zare-Bandamiri M, Parvin S, Dianatinasab M (2018) Effects of Different Treatment Strategies and Tumor Stage on Survival of Patients with Advanced Laryngeal Carcinoma: A 15-Year Cohort Study, J Cancer Epidemiol 9678097 https://doi.org/10.1155/2018/9678097\u003c/li\u003e\n\u003cli\u003eZhu X, Heng Y, Zhou L, Zhang M, Li W, Tao L (2020) Survival prediction and treatment strategies for patients with advanced laryngeal carcinoma: a population-based study, Int J Clin Oncol 25: 1483\u0026ndash;1491. https://doi.org/10.1007/s10147-020-01688-9\u003c/li\u003e\n\u003cli\u003eWang S-X, Ning W-J, Zhang X-W, Tang P-Z, Li Z-J, Liu W-S (2019) Predictors of Occult Lymph Node Metastasis and Prognosis in Patients with cN0 T1\u0026ndash;T2 Supraglottic Laryngeal Carcinoma: A Retrospective Study, ORL J Otorhinolaryngol Relat Spec 81:317\u0026ndash;326. https://doi.org/10.1159/000503007\u003c/li\u003e\n\u003cli\u003eWang Z, Zeng Q, Li Y, Lu T, Liu C, Hu G (2020) Extranodal Extension as an Independent Prognostic factor in Laryngeal Squamous Cell Carcinoma Patients, J Cancer 11:7196\u0026ndash;7201. https://doi.org/10.7150/jca.47700\u003c/li\u003e\n\u003cli\u003eMeng Z, Zhang L, Huang C, Piao Y, Chen X, Xian J (2022) Quantitative parameters derived from 18F-fluorodeoxyglucose positron emission tomography/magnetic resonance imaging can accurately estimate the histologic grade of hypopharyngeal squamous cell carcinoma preoperatively, Neuroradiology 64:2153\u0026ndash;2162. https://doi.org/10.1007/s00234-022-03052-2\u003c/li\u003e\n\u003cli\u003eGeng D, Chen X, Zhao X-G, Xu X-Q, Su G-Y, Zhou Y et al (2023) Laryngeal and hypopharyngeal squamous cell carcinoma: association between quantitative parameters derived from dual-energy CT and histopathological prognostic factors, Acta Radiol. ;64: 2268\u0026ndash;2276. https://doi.org/10.1177/02841851221095237\u003c/li\u003e\n\u003cli\u003eArthur A, Orton MR, Emsley R, Vit S, Kelly-Morland C, Strauss D et al (2023) A CT-based radiomics classification model for the prediction of histological type and tumour grade in retroperitoneal sarcoma (RADSARC-R): a retrospective multicohort analysis, Lancet Oncol 24:1277\u0026ndash;1286. https://doi.org/10.1016/S1470-204523)00462-\u003c/li\u003e\n\u003cli\u003eLi Z, Liu Z, Guo Y, Wang S, Qu X, Li Y et al (2022) Dual-energy CT-based radiomics nomogram in predicting histological differentiation of head and neck squamous carcinoma: a multicenter study, Neuroradiology 64:361\u0026ndash;369. https://doi.org/10.1007/s00234-021-02860-2\u003c/li\u003e\n\u003cli\u003eWu W, Ye J, Wang Q, Luo J, Xu S (2019) CT-Based Radiomics Signature for the Preoperative Discrimination Between Head and Neck Squamous Cell Carcinoma Grades, Front Oncol 9:821. https://doi.org/10.3389/fonc.2019.00821\u003c/li\u003e\n\u003cli\u003eZhao Y, Chang M, Wang R, Xi IL, Chang K, Huang RY et al (2020) Deep Learning Based on MRI for Differentiation of Low‐ and High‐Grade in Low‐Stage Renal Cell Carcinoma, J Magn Reson Imaging 52:1542\u0026ndash;1549. https://doi.org/10.1002/jmri.27153\u003c/li\u003e\n\u003cli\u003eLiu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z et al (2021) Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 10012-10022. https://doi.org/10.1109/ICCV48922.2021.00986\u003c/li\u003e\n\u003cli\u003eShamshad F, Khan S, Zamir SW, Khan MH, Hayat M, Khan FS et al (2023) Transformers in medical imaging: A survey, Med Image Anal 88: 102802. https://doi.org/10.1016/j.media.2023.102802\u003c/li\u003e\n\u003cli\u003eKhan RF, Lee B-D, Lee MS (2023) Transformers in medical image segmentation: a narrative review, Quant Imaging Med Surg 13:8747\u0026ndash;8767. https://doi.org/10.21037/qims-23-542\u003c/li\u003e\n\u003cli\u003eYang M, He X, Xu L, Liu M, Deng J, Cheng X et al (2022) CT-based transformer model for non-invasively predicting the Fuhrman nuclear grade of clear cell renal cell carcinoma, Front Oncol 12:961779. https://doi.org/10.3389/fonc.2022.961779\u003c/li\u003e\n\u003cli\u003eCicalese PA, Mobiny A, Shahmoradi Z, Yi X, Mohan C, Van Nguyen H (2021) Kidney Level Lupus Nephritis Classification Using Uncertainty Guided Bayesian Convolutional Neural Networks, IEEE J Biomed Health Inform 25:315\u0026ndash;324. https://doi.org/10.1109/JBHI.2020.3039162\u003c/li\u003e\n\u003cli\u003eLi J, Chen J, Tang Y, Wang C, Landman BA, Zhou SK (2023) Transforming medical imaging with Transformers? A comparative review of key properties, current progresses, and future perspectives, Med Image Anal 85:102762. https://doi.org/10.1016/j.media.2023.102762\u003c/li\u003e\n\u003cli\u003eLiu Y, Zhang Y, Wang Y, Hou F, Yuan J, Tian J et al (2024) A Survey of Visual Transformers, IEEE Trans Neural Netw Learn Sys 35:7478\u0026ndash;7498. https://doi.org/10.1109/TNNLS.2022.3227717\u003c/li\u003e\n\u003cli\u003eYe G, Wu G, Li K, Zhang C, Zhuang Y, Liu H et al (2024) Development and Validation of a Deep Learning Radiomics Model to Predict High-Risk Pathologic Pulmonary Nodules Using Preoperative Computed Tomography, Acad Radiol 31:1686\u0026ndash;1697. https://doi.org/10.1016/j.acra.2023.08.040\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, Vision Transformer, Histological grade, Laryngeal and hypopharyngeal squamous cell carcinoma","lastPublishedDoi":"10.21203/rs.3.rs-7249038/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7249038/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground \u003c/strong\u003ePretreatment determination of histological differentiation grade is critical for prognostic evaluation in laryngeal and hypopharyngeal squamous cell carcinoma (LHSCC) .This study aimed to develop a contrast-enhanced CT (CECT)-based Vision Transformer (ViT) model for noninvasive evaluation of histological grades in LHSCC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e A total of 1,648 LHSCC patients who underwent CECT scans were enrolled from three hospitals in this study. Participants were divided into a training cohort (n=1,239) , an internal validation cohort (n=310) from one hospital, and an external validation cohort (n = 99) from the other two hospitals. The diagnostic model integrates a pre-trained ViT for CECT feature extraction and an XGBoost classifier for prediction. The model’s predictive performance was evaluated using the area under the curve (AUC), decision curve analysis (DCA), and calibration curve.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults \u003c/strong\u003eThe ViT model achieved AUCs of 0.887 (95%CI: 0.848-0.927) in internal validation and 0.796 (95%CI: 0.693-0.899) in external validation cohorts, significantly outperforming the conventional radiomics model (AUCs: 0.775, 95%CI: 0.714-0.837 and 0.544, 95%CI: 0.388-0.699; \u003cem\u003ep\u0026lt;\u003c/em\u003e0.001 and 0.002, respectively). Clinically, DCA demonstrated superior clinical utility, while calibration curves showed excellent prediction reliability. Gradient-weighted Class Activation Mapping visualization identified CT image regions most influential for the model's predictions, providing interpretability for clinical decision-making.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion \u003c/strong\u003eThe ViT-based deep learning model developed in this study using CECT demonstrated excellent predictive performance for histological grading of LHSCC,with promising application for patient prognosis assessment.\u003c/p\u003e","manuscriptTitle":"Leveraging Vision Transformer for Histological Grade Prediction in Laryngeal and Hypopharyngeal Squamous Cell Carcinoma: A Large-Scale Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-06 08:14:50","doi":"10.21203/rs.3.rs-7249038/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a7766204-1e5f-40e4-a28f-81f9f4a36251","owner":[],"postedDate":"August 6th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-29T16:00:00+00:00","versionOfRecord":{"articleIdentity":"rs-7249038","link":"https://doi.org/10.1007/s00234-025-03876-8","journal":{"identity":"neuroradiology","isVorOnly":false,"title":"Neuroradiology"},"publishedOn":"2025-12-27 15:57:20","publishedOnDateReadable":"December 27th, 2025"},"versionCreatedAt":"2025-08-06 08:14:50","video":"","vorDoi":"10.1007/s00234-025-03876-8","vorDoiUrl":"https://doi.org/10.1007/s00234-025-03876-8","workflowStages":[]},"version":"v1","identity":"rs-7249038","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7249038","identity":"rs-7249038","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}