Artificial intelligence in head and neck cancer: a bibliometric and visualization analysis (1995–2025)

Abstract

Objectives

Head and neck cancers (HNCs) pose significant challenges for clinical diagnosis and treatment due to their complex anatomical structures, atypical early symptoms, and the considerable morbidities associated with treatment. The rapid development of artificial intelligence (AI) technologies in medicine has introduced a new paradigm for the precise diagnosis and management of HNCs.

Materials and methods

This study used bibliometric methods to systematically analyze the research landscape of AI applications in HNCs from 1995 to 2025. The aim was to identify research trends, collaboration networks, and emerging directions, thereby providing a reference for future investigations.

Results

A total of 230 AI-related publications on HNCs were retrieved from the Web of Science database. Tools such as CiteSpace and VOSviewer were used to analyze temporal publication trends, national and institutional contributions, core author groups, journal distribution, and keyword clustering. Key milestone studies and the evolution of research hotspots were identified through co-citation analysis and burst keyword detection.

Conclusion

AI research in HNCs has evolved into a multimodal and multi-task field, with deep learning playing a central role in image analysis. However, challenges persist regarding model interpretability and generalizability.

Clinical relevance

In the future, AI applications in HNCs are expected to further enhance diagnostic and therapeutic strategies. Strengthening interdisciplinary collaboration is essential to translate AI algorithms into comprehensive, end-to-end clinical applications. Such integration will optimize the entire care pathway for head and neck cancer patients.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03992-0.

Introduction

Artificial intelligence (AI), including machine learning (ML) and deep learning (DL), has transformed medical research and clinical practice by enabling data-driven decision-making [1]. AI algorithms are particularly effective at identifying complex patterns in multidimensional datasets such as medical imaging, genomic data, and electronic health records [2–4]. In oncology, AI applications span from automated tumor segmentation in radiology to predictive modeling of treatment outcomes [5, 6]. For instance, convolutional neural networks (CNNs) have achieved human-level accuracy in detecting lymph node metastases from computed tomography (CT) scans [7], while survival prediction models like DeepSurv outperform traditional statistical methods by capturing nonlinear interactions among clinical variables [7]. These advances underscore the potential of AI to address persistent challenges in cancer care, especially in settings requiring the rapid integration and analysis of heterogeneous data.

Head and neck cancers (HNCs) are the sixth most common cancers worldwide. These malignancies of the oral cavity, larynx, and pharynx account for over 900,000 new annual cases [8, 9]. The management of HNCs is particularly challenging for three main reasons: the anatomical complexity of the region, frequent diagnostic delays due to nonspecific early symptoms, and the high morbidity of aggressive treatments [10]. Current diagnostic methods like endoscopy and biopsy are invasive and operator-dependent. In contrast, radiotherapy planning requires extremely precise tumor targeting near critical structures like optic nerves and salivary glands [11]. Moreover, treatment-related complications like sensorineural hearing loss and dysphagia can severely impact patients’ quality of life [12]. These challenges underscore the urgent need for precision tools that facilitate early detection, improve treatment planning, and reduce adverse effects.

Bibliometrics, the quantitative analysis of scholarly publications, offers valuable insights into research trends, collaboration patterns, and emerging frontiers through metrics like citation counts, co-occurrence networks, and keyword clustering [13, 14]. Tools such as CiteSpace and VOSviewer facilitate the visualization of knowledge structures, identification of influential studies, and detection of paradigm shifts within a given field [15, 16]. In oncology, bibliometric analyses have revealed the trajectory of immunotherapy research and the emergence of radiomics as a distinct subdiscipline [17]. By examining publication volume, geographic distribution, and interdisciplinary linkages, bibliometrics provides a strategic roadmap for policymakers and researchers to guide resource allocation and prioritize innovation.

AI applications in HNC are advancing rapidly. However, a detailed bibliometric analysis that captures the specific trajectory of this field is still lacking. While previous bibliometric studies in oncology have highlighted broader trends(e.g., such as the rise of immunotherapy and radiomics) and investigations of AI in medicine have predominantly focused on radiology or general clinical applications. This study presents the first dedicated, longitudinal bibliometric analysis of AI in HNC research. We analyzed 230 AI-related HNC publications (Web of Science, 1995–2025) to reveal the field’s intellectual structure and evolution. This analysis used co-citation, keyword burst detection, and thematic mapping to account for the unique clinical challenges of HNC. Our fine-grained analysis identifies foundational works and collaborative networks, reveals the evolution of HNC-specific research themes, and pinpoints clinically significant but underexplored areas like explainable AI and multi-omics integration. Through synthesizing these domain-specific insights, our work aims to inform targeted research efforts, foster interdisciplinary collaboration between AI experts and HNC clinicians, and accelerate the clinical translation of AI innovations in head and neck oncology.

Methods

The Web of Science database was selected for data retrieval. “Head and neck cancer” and “artificial intelligence” were used as search keywords, with the detailed search strategy provided in the Supplementary Materials. Supplementary Fig. 1 shows the specific screening process for included literature. Inclusion criteria were restricted to English-language publications classified as reviews and articles. Two researchers, Chuyi Cai and Kai Li, independently screened the titles and abstracts to exclude studies not related to AI and HNC. The date of retrieval is March 8, 2025. The final results were exported in plain text format. The analysis and visualization were conducted using Citespace, VOSviewer, SCImago, Prism 10.0, and the Bibliometrix package in R Studio. Duplicate references were removed using Citespace (Fig. 1).

Fig. 1.

Flow chart of the study design and selection process

Results

Literature overview and development trend analysis

Bibliometric analysis using the Bibliometrix package in R Studio indicates that research on AI in the field of head and neck tumors has reached an initial stage of development. A total of 230 relevant publications, comprising 199 original articles (86.5%) and 31 reviews (13.5%), were published across 114 academic journals between 1995 and 2025. The collaboration network includes 1545 researchers from 47 countries and regions, reflecting substantial international engagement. The temporal distribution of publications reveals distinct developmental stages (Fig. 2A). Following the first publication in 1995, the field underwent a 16-year period of academic dormancy (1996–2010). From 2011 to 2020, annual output remained in the single digits (mean: 3.7 articles per year). Notably, the field entered a phase of rapid growth after 2021, with annual publications exceeding 30, peaking at 64 articles in 2024, representing more than a 20-fold increase compared to the early years. In terms of academic influence (Fig. 2B), key inflection points occurred in 2013, 2015, 2018, and 2019. Publications from these years received significantly higher cumulative citation counts, suggesting the emergence of milestone studies or technological breakthroughs that have had a lasting impact on subsequent research.

Fig. 2.

A Annual publications of AI research in head and neck cancers. The horizontal axis is the year and the vertical axis is the number of publications. B Citation trends over time for AI studies in head and neck cancers. The horizontal axis is the year, and the vertical axis is the number of citations per year

Country/region contribution and institutional collaboration network

Collaboration network analysis using VOSviewer, combined with national research influence data from SCImago, reveals that AI research in the field of head and neck tumors has formed 6 major international collaboration clusters, with the United States, China, Finland, the Netherlands, Italy, and India serving as central nodes (Fig. 3A, B).Temporal evolution analysis highlights a significant shift in regional research focus over time (Fig. 3C). In the early stage (1995–2010), the field was primarily driven by Dutch scholars. As the research matured, a dual-center pattern emerged, dominated by North American and European institutions, particularly those in Canada and Germany.

Fig. 3.

Visualization of country-level research contributions in AI for head and neck neoplasms (A-D) The circles represent different countries, and the size of the circles represents the number of publications

Table 1 shows the top 10 countries by publication volume, led by the United States (n = 60), followed by China (n = 49), India (n = 30), Italy (n = 23), the Netherlands (n = 20), the United Kingdom (n = 14), Germany (n = 14), Canada (n = 13), Japan (n = 12), and Finland (n = 9). The United States not only leads in absolute publication count (accounting for 26% of the total) but also demonstrates strong academic influence, with the highest citation frequency per article (28.45 citations/article), underscoring its dominant role in the field. Institutional productivity analysis (Fig. 4A) identifies 13 institutions worldwide that meet the high-output threshold (≥ 5 publications). Affiliation refers to the attribution of the article at the time of publication. The geographic distribution of these institutions is relatively concentrated (Fig. 4B): 46.15% are located in Europe (6/13), 23.08% in North America (3/13), and 15.38% in China (2/13). This institutional pattern is consistent with national-level contributions shown in Fig. 3D, reaffirming that current AI research in HNCs is primarily driven by three major scientific hubs: the European cluster led by Germany and the Netherlands, the North American alliance centered on the United States and Canada, and a growing Chinese research team.

Table 1.

Top 10 publishing countries in head and neck cancer AI research

Rank	Country	Publication	Citation	Citation/Publication
1	USA	60	1707	28.45
2	China	49	805	16.43
3	India	30	647	21.57
4	Italy	23	304	13.22
5	Netherlands	20	449	22.45
6	UK	14	315	22.5
7	Germany	14	90	6.43
8	Canada	13	129	9.92
9	Japan	12	176	14.67
10	Finland	9	104	11.56

Fig. 4.

A Institutions with the most publications and their collaborative relationships. B Top 10 publishing institutions categorized by country. The circles represent different countries, and the colors represent different clusters. In B, the horizontal axis is the number of published articles, and the vertical axis represents different countries

Core author group and academic network

Using the productivity evaluation metrics in bibliometrics, this study constructed a core author identification model within the field via the Bibliometrix package in R Studio. Figure 5A highlights three core scholars with six or more publications: Professor Langendijk JA from the University of Groningen, the Netherlands, leads with eight papers, followed by Professor Fuller CD from Canada and Professor Stogbauer F from the Technical University of Munich, Germany, each with six papers. Analysis of sustained academic contributions (Fig. 5B) reveals that Professor Langendijk JA has maintained continuous scholarly activity over 15 years (2011–2025). His research trajectory spans from the nascent phase to the explosive growth stage of the field, positioning him as a key witness and driver of its evolving research paradigm. Data presented in Table 2 further demonstrate a marked regional concentration within the core author group: 40% (4/10) of the top ten prolific authors are affiliated with the Technical University of Munich. The institution’s disciplinary strengths in intelligent brain neuroimaging analysis provide a solid foundation supporting the continued productivity of its scholars.

Fig. 5.

Top 10 authors by publication count and their publication timeline; A Total publications per author. B Annual publication output over time. The vertical axis represents different authors, the horizontal axis of A represents the number of publications, and the vertical axis of B represents the year

Table 2.

Top 10 most productive authors in head and neck cancer AI research

Rank	Author	Country	Organizations
1	Langendijk, JA	Netherlands	University of Groningen
2	Fuller CD	Canada	University of Toronto
3	Stogbauer F	Germany	Technical University of Munich
4	Peretti G	Italy	University of Genoa
5	Benedikt Schmidl	Germany	Technical University of Munich
6	Wahid, Kareem A	USA	University of Texas
7	Wang Y	USA	Mississippi State University
8	Wirth M	Germany	Technical University of Munich
9	Wollenberg B	Germany	Technical University of Munich
10	Adeoye J	China	University of Hong Kong

Analysis of published journals

Journal analysis highlights the primary platforms for knowledge dissemination in the field, offering valuable guidance for new authors seeking suitable submission venues. Figure 6 lists the top ten journals by publication volume, with Frontiers in Oncology leading the field with 16 articles. Table 3 shows the Journal Citation Reports (JCR) classifications and impact factors of these journals. The impact factors predominantly range between 2 and 5. Regarding JCR rankings, one journal is classified in Q2, two in Q3, and the remaining seven are ranked in Q1, reflecting a strong presence of high-impact journals in this research area.

Fig. 6.

Top 10 journals by publication volume in AI research on head and neck cancers. The horizontal axis and the size of the circle represent the number of publications, and the vertical axis represents different authors

Table 3.

Top 10 journals by import factor and JCR ranking in artificial intelligence research on head and neck cancers

Rank	Journal	Import Factor (2023)	JCR (2024)
1	Frontiers in Oncology	3.5	Q2(ONCOLOGY)
2	Cancers	4.5	Q2(ONCOLOGY)
3	Diagnostics	3.0	Q1(MEDICINE, GENERAL & INTERNAL)
4	Head and Neck-Journal for the Science and Special	2.3	Q1(OTORHINOLARYNGOLOGY)
5	Oral Oncology	4.0	Q1(DENTISTRY, ORAL SURGERY & MEDICINE)
6	Laryngoscope	2.2	Q3(MEDICINE, RESEARCH & EXPERIMENTAL)
7	Scientific Reports	3.8	Q1(MULTIDISCIPLINARY SCIENCES)
8	Journal of Oral Pathology and Medicine	2.7	Q2(DENTISTRY, ORAL SURGERY & MEDICINE)
9	BMC oral health	2.6	Q1(DENTISTRY, ORAL SURGERY & MEDICINE)
10	Clinical and Translational Radiation Oncology	2.7	Q3(ONCOLOGY)

JCR, Journal Citation Reports

Literature analysis

Literature analysis enables rapid identification of the most influential papers in the field, helping new researchers to quickly grasp the foundational literature. Table 4 lists the top 10 most cited references in this research area. The most cited article is Radiomic Machine-Learning Classifiers for Prognostic Biomarkers of Head and Neck Cancer, published in Frontiers in Oncology in 2015, with a total of 272 citations.

Table 4.

Top 10 most citated publications

Rank	Title	Journal	Years	Citations	Citations/per year
1	Radiomic Machine-Learning Classifiers for Prognostic Biomarkers of Head and Neck Cancer [18]	Front Oncol	2015	272	27.2
2	3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture [19]	Phys Med Biol	2019	240	40
3	Deep learning-based survival prediction of oral cancer patients [20]	Sci Rep	2019	226	37.7
4	Computer-assisted medical image classification for early diagnosis of oral cancer employing deep learning algorithm [21]	J Cancer Res Clin Oncol	2019	170	28.3
5	Pretreatment Identification of Head and Neck Cancer Nodal Metastasis and Extranodal Extension Using Deep Learning Neural Networks [22]	Sci Rep	2018	124	17.7
6	Contrast-enhanced computed tomography image assessment of cervical lymph node metastasis in patients with oral cancer by using a deep learning system of artificial intelligence [23]	Oral Surg Oral Med Oral Pathol Oral Radiol	2019	121	20.2
7	Automated Detection and Classification of Oral Lesions Using Deep Learning for Early Detection of Oral Cancer [24]	IEEE Access	2020	121	24.2
8	Oral cancer prognosis based on clinicopathologic and genomic markers using a hybrid of feature selection and machine learning methods [25]	BMC Bioinformatics	2013	99	8.25
9	Machine Learning-Guided Adjuvant Treatment of Head and Neck Cancer [26]	JAMA Netw Open	2020	97	19.4
10	Computer-aided diagnosis of laryngeal cancer via deep learning based on laryngoscopic images [10]	EBioMedicine	2019	86	14.3

Keyword analysis

Keyword analysis provides insight into the key topics and trends receiving attention in this research field. As shown in the keyword density map (Fig. 7A), the most frequently occurring keywords include HNC, ML, AI, and DL. The keyword overlay map (Fig. 7B) illustrates the temporal appearance of keywords, with purple indicating earlier terms and yellow highlighting more recent ones. Keywords such as ChatGPT, toxicity, therapy, quality, and performance have emerged in recent years. To further explore keyword groupings, we performed clustering analysis using CiteSpace software. As shown in Fig. 8A, keywords cluster into distinct categories: AI research methods (#2 deep learning, #6 machine learning, #4 artificial intelligence), disease types (#3 laryngeal cancer), and application areas (#0 diagnosis, #1 Risk, #5 adaptive radiotherapy, #7 meta analysis, #9 accuracy, #10 classifier fusion).The keyword burst analysis in Fig. 8B indicates that recent AI research in head and neck tumors has predominantly focused on employing ML and DL techniques to enhance diagnostic accuracy and minimize treatment-related risks.

Fig. 7.

Keyword analysis maps of AI research in head and neck cancers. A Keyword density map showing the frequency of key terms. B Keyword overlay map illustrating the temporal emergence of keywords, with purple indicating earlier appearance and yellow indicating more recent terms. The circles represent different keywords, and the size of the circles indicates the frequency of the keywords

Fig. 8.

Keyword clustering and burst analysis in AI research on head and neck cancers. A Keyword categories identified through clustering analysis. B Keyword burst (peak) diagram showing emerging and trending keywords over time. Different colors represent different clusters. The horizontal axis represents time

Discussion

General information

AI offers a transformative approach for HNC research, addressing the disease’s anatomical complexity and heterogeneity. Current research, led by the US and China within a collaborative international network, was analyzed via 230 core publications. Deep learning dominates, primarily applied to three clinical areas: optimizing image-based diagnosis, supporting treatment decisions, and prognostic prediction. These models demonstrate high performance in detecting lymph node metastasis, planning radiotherapy, and predicting survival. Integrating multi-modal data further enhances diagnostic and prognostic accuracy.

Table 5.

Summary of the top 10 most cited publications

Artificial Intelligence Methods	Tumor type	Research direction	Research outcomes	Citation
Machine Learning (mRMR/Mutual Information Feature Selection)	Head and Neck Cancer	Imaging Biomarkers	Prognostic prediction AUC 0.61–0.69, stability RSD 7.36–11.28	[18]
DeepSurv/Survival Forest Model	Head and Neck Cancer	Treatment Optimization	The HR of the recommended chemotherapy group was 0.79–0.90 (P < 0.01), which can avoid unnecessary chemotherapy.	[26]
Deep Learning Image Classification	Oral squamous cell carcinoma	Diagnostic imaging (CT)	Accuracy 78.2%, AUC 0.80 (comparable to radiologists)	[23]
DeepSurv	Oral squamous cell carcinoma	Prognostic Model	The test set c-index is 0.781, which is better than the traditional model.	[20]
U-Net	Head and Neck Cancer	Radiation therapy dose planning	The maximum OAR dose prediction error was 6.3% and the mean error was 5.1%.	[19]
Partitioned Convolutional Neural Network	Oral squamous cell carcinoma	Pathology Image Classification	The accuracy rate of distinguishing benign from malignant is 91.4% (sensitivity 0.94, specificity 0.91)	[21]
Deep Convolutional Neural Network	Laryngeal cancer	Imaging diagnosis (endoscopy)	LCA/PRELCA detection accuracy was 86.7%, AUC 0.922	[10]
3D Convolutional Neural Networks	Head and Neck Cancer	Diagnostic imaging (CT)	Lymph node metastasis/extranodal extension detection AUC 0.91 (95%CI 0.85–0.97)	[22]
ReliefF-GA-ANFIS hybrid model	Oral squamous cell carcinoma	Biomarkers	Prognostic accuracy: 93.81% (AUC = 0.90)	[25]
ResNet-101/Faster R- Convolutional Neural Network	Oral squamous cell carcinoma	Pathology + imaging combined analysis	Lesion detection F1 = 87.07%, lesion identification requiring referral F1 = 78.30%	[27]

AUC, Area Under the Curve; RSD, Relative Standard Deviation; LCA, Laryngeal Cancer; PRELCA, Precancerous Laryngeal Lesions; GA, Genetic Algorithm; ANFIS, Adaptive Neuro-Fuzzy Inference System

Frontiers and hotspot

Application of AI in imaging diagnosis of HNCs

AI has demonstrated considerable promise in the imaging diagnosis of head and neck tumors. Peretti et al. introduced the SegMENT-Plus model, designed to assist clinicians in delineating the superficial extent of laryngeal cancer. Their model delivered reliable performance comparable to that of two otolaryngology residents, with the advantage of faster computational speed [28]. Yang et al. developed an automatic recognition algorithm utilizing optical coherence tomography (OCT) images, successfully differentiating normal mucosa, precancerous lesions, and oral squamous cell carcinoma, achieving a classification accuracy of 96.76%, significantly outperforming traditional ML approaches [29]. Similarly, Song’s study demonstrated the effectiveness of the ResNet50 model applied to Raman spectroscopy data, attaining an accuracy of 92.81%, and led to the development of a prototype intelligent detection system [30]. Sundari and colleagues developed the XceptionAttnV1 model for automatic oral cancer detection, achieving an average accuracy of 96.59%. This innovative model combines CLAHE preprocessing, a channel attention mechanism, and custom convolutional layers, with its robustness confirmed by five-fold cross-validation [31]. Collectively, these findings highlight significant advances in AI-assisted diagnosis using optical imaging modalities.

In the domain of radiological image analysis, AI has achieved significant breakthroughs. A novel DL–based enhanced Mask R-CNN model has been developed to identify laryngeal cancer and related manifestations in real time by leveraging multiple imaging datasets, including CT scan. This model demonstrates a marked improvement in detecting small malignant tumors within the laryngeal cavity during real-time patient screening [32]. Huynh et al. reported that CNNs outperform traditional radiomics approaches by directly processing fluorodeoxyglucose positron emission tomography/computed tomography (FDG PET/CT) images to predict overall survival (OS) and disease-free survival (DFS) in HNC patients [33]. Furthermore, Park and colleagues validated a combined model integrating magnetic resonance imaging (MRI) radiomics with ML algorithms, which achieved an AUC of 0.786 in predicting postoperative recurrence of oropharyngeal cancer [34]. The DeepLab V3 + model was notably effective in segmenting hypopharyngeal tumors on MRI, achieving a Dice score of 0.77 and excelling particularly at identifying small tumors [35]. Troise et al. showed that ML-based radiomic features can effectively predict occult neck metastases in oral cavity cancer prior to surgery. Their model achieved a sensitivity of 78%, specificity of 83%, and an AUC of 0.86 [36]. Collectively, these findings underscore the superior capability of DL models in radiological image feature extraction and classification for HNCs.

Intelligent analysis of pathological images represents a crucial application area for AI. Chen et al. developed the innovative CANet model, which achieved a remarkable classification accuracy of 97.00% for oral cancer histopathological images [37]. Similarly, Klein’s team successfully predicted HPV status directly from standard H&E-stained slides, achieving an AUC of 0.83–0.88. This non-invasive approach outperformed traditional detection methods [38]. A novel pathological image–based model predicting CXCL8 expression revealed that elevated CXCL8 levels were significantly associated with poor prognosis in head and neck squamous cell carcinoma [39]. Beyond imaging data, clinical feature sets have also been used to train AI models for prognostic prediction in nasopharyngeal carcinoma, demonstrating the expanding scope of AI applications in this field [40].

Application of AI in treatment planning for HNCs

AI is increasingly integral to treatment planning in head and neck oncology. In the automatic delineation of radiotherapy targets, Kihara et al. demonstrated that a U-Net model utilizing combined CT and gross tumor volume (GTV) inputs achieved a significantly higher DSC (0.80 ± 0.03) compared to models relying solely on CT images [41]. Similarly, the UNet + MV model developed by Strijbis et al. has proven effective for optimizing lymph node target delineation [42] Kawahara’s team introduced a generative adversarial network (GAN)-based MRI segmentation model exhibiting excellent performance, particularly suited for MR-guided radiotherapy applications [43]. A recent validation study reported that 3D CNN–based automatic delineation enhanced time efficiency by 33%, with DSC values ranging from 0.84 to 0.90, ensuring stable and reliable contouring quality [44]. More importantly, beyond enhancing efficiency and consistency, the high precision of AI-driven contouring holds direct clinical promise for reducing radiotherapy-related toxicities, such as xerostomia and dysphagia, by minimizing irradiation to adjacent critical structures like salivary glands and swallowing muscles. This directly contributes to the paramount goal of preserving patients’ quality of life (QoL).​​ The focus on QoL and functional preservation is equally critical in the surgical management of head and neck cancer, particularly with the rise of transoral robotic surgery (TORS). While TORS for oropharyngeal cancer aims to minimize morbidity, functional outcomes—especially swallowing and speech—are highly variable, and the need for adjuvant therapy can significantly compromise QoL benefits.​​ [45] This clinical challenge represents a major opportunity for AI. ​​The same deep learning architectures that excel in image segmentation for radiotherapy could be leveraged to predict post-TORS functional outcomes. For example, by analyzing pre-operative imaging to quantify the extent of muscular invasion or to model the functional impact of planned resections, AI models could help identify patients at high risk for prolonged feeding tube dependence or severe dysphagia.​​ This predictive capability would enable proactive interventions, personalized patient counseling, and refined surgical decision-making, thereby directly linking AI’s analytical power to addressing the core clinical implications of toxicity and QoL raised by our bibliometric findings.

Beyond target delineation, AI models offer promising capabilities in predicting treatment response and toxicity. Cauvin et al. identified the maximum plasma concentration (C_max, 2.4–4.1 µg/mL) as the optimal biomarker for predicting cisplatin toxicity and efficacy, achieving a model accuracy of 0.71 [46]. Chu’s team developed an innovative 3D DL normal tissue complication probability model that effectively predicted advanced xerostomia, with an AUC of 0.78–0.79 [47]. Furthermore, Agheli et al. reported a radiation-induced oral mucositis prediction model integrating multimodal features, which achieved a high predictive accuracy with an AUC of 0.917 [48].

AI plays a key role in personalized treatment decisions for head and neck oncology. For instance, Howard et al., analyzing 33,527 patients, demonstrated that about half of intermediate-risk HNSCC patients gain significant survival benefits from chemoradiotherapy [26]. Zhang et al. found that a support vector machine (SVM) best predicted xerostomia risk after proton/heavy ion radiotherapy. The model showed a balanced accuracy of 0.66 and was associated with a low rate (5.8%) of severe (grade 2) xerostomia [49]. Furthermore, a recently developed multiparametric ML model enables simultaneous prediction of HPV status and survival outcomes, offering a promising tool for non-invasive clinical decision support [50]. Collectively, these findings highlight AI’s potential to stratify patients based on likely treatment benefit, optimize therapeutic strategies, and ultimately enhance clinical outcomes and quality of life.

Application of AI in the prognosis prediction of HNCs

In survival prediction studies, Kim et al. demonstrated that the DeepSurv model achieved the highest c-index for predicting oral squamous cell carcinoma survival, with a c-index of 0.810 on the training set, outperforming both random survival forests and traditional Cox proportional hazards models [20]. Similarly, Peng and colleagues found that the random forest model yielded the best prognostic performance [51]. Additionally, Mansouri et al. developed an innovative fusion feature model that showed superior predictive accuracy for OS in HNC, with a c-index of 0.73 [52].

Significant progress has also been made in predicting recurrence risk. Bourdillon et al. evaluated 12 ML models and found that a combined decision tree and SVM approach achieved an accuracy of 80.8% in predicting oral squamous cell carcinoma recurrence within one year, aiding personalized follow-up planning [53]. Bos and colleagues demonstrated that a ML model integrating clinical and MRI features effectively predicted local control of oropharyngeal cancer after chemoradiotherapy, with a combined model AUC of 0.745 [54]. Additionally, Tseng developed a model combining clinicopathological and genetic variation data, which performed well in survival risk stratification for advanced oral cancer, achieving a cancer-specific survival c-index of 0.689 [55].

The exploration of new prediction algorithms continues to drive breakthroughs. Somyanonthanakul’s fuzzy DL model demonstrated exceptional performance in classifying oral cancer survival time, achieving an accuracy of 97% and an AUC of 1.00, significantly outperforming traditional DL models [56]. Parola’s team developed an interpretable oral cancer screening system capable of handling imperfect images, attaining an informed DL accuracy of 85% and providing explanations better aligned with clinical needs [57]. Joshi’s enhanced CNN, combined with autoencoder technology, effectively mitigated overfitting in oral cancer image classification, improving accuracy by 5 to 5.5% on average [58]. These innovative approaches demonstrate that advanced AI models can enhance interpretability and robustness while maintaining high accuracy, better meeting the practical demands of clinical applications.

Application of AI in the study of specific subtypes of HNCs

In laryngeal cancer research, Mohamed et al. developed a DL system combining EfficientNet-B0, multi-head bidirectional gated recurrent units (GRU), and the mongoose optimization algorithm, achieving strong performance in automatic detection and classification of laryngeal cancer [59]. Achieving 96.02% accuracy, a model developed by Irowais’s team excelled at detecting laryngeal cancer. The approach used InceptionV3 for feature extraction, a deep belief network for classification, and Aquila optimization [60]. Alabi’s interpretable ML model robustly predicted overall survival in laryngeal cancer, achieving an AUC of 0.76–0.77. The model also identified key prognostic factors: age, tumor/node stage, and marital status [61].

In the study of oropharyngeal cancer, Taku’s DL-CNN model based on the U-Net architecture demonstrated excellent performance in segmenting HPV-related positive lymph nodes, achieving a median Dice coefficient of 0.92 and a detection AUC of 0.98 [62]. Lang’s team applied a transfer learning strategy with a video-pretrained 3D CNN model to successfully classify HPV status in small sample datasets, achieving an AUC of 0.81 [63]. In an analysis of NCDB data, Ahn et al. found that random forests were superior to other ML methods for predicting prolonged radiotherapy risk in oropharyngeal cancer. This approach helps identify high-risk patients for optimized intervention [64]. These studies confirm that tailored AI models addressing specific clinical challenges in oropharyngeal cancer can significantly enhance predictive accuracy and support more precise individualized treatment.

In studies of hypopharyngeal cancer and other subtypes, the standardized HNC radiotherapy data management framework developed by Tryggestad provided a foundation for building automatic segmentation models, with 490 cases undergoing data standardization [65]. Lauwers’ DL model for automatic segmentation of cell nuclei and 53BP1 lesions significantly reduced analysis time from hours to minutes while maintaining high accuracy (Dice = 0.901) [66]. Black’s study demonstrated that an ordered logistic regression model based on demographic and symptom data consistently classified benign and malignant HNCs, achieving an AUC of 0.6697 [67]. These findings illustrate the broad potential of AI technologies across various head and neck tumor subtypes, supporting precise management throughout diagnosis and follow-up.

Limitations

Several limitations must be acknowledged. Firstly, the exclusive use of the Web of Science database may have introduced selection bias, as relevant studies indexed in Scopus, PubMed, or Embase were excluded. This could affect the comprehensiveness of the bibliometric landscape, particularly for region-specific literature. Future bibliometric studies would benefit from a multi-database approach. Secondly, the inclusion of only English publications might have led to an under-representation of non-English research, introducing a potential geographical bias. In addition, there are many types of head and neck tumors, and it may be more scientific to narrow the research field according to anatomical regions. Finally, the fast-paced nature of AI research means our keyword-based search may have had retrieval gaps despite our best efforts. These limitations, however, establish a clear framework for more expansive future updates.

Summary and outlook

AI applications in HNC research span diagnosis, treatment, and prognosis prediction, forming a multimodal, multi-task landscape. DL, particularly CNNs, dominates technically, while emerging methods such as attention mechanisms, Transformers, and GANs are rapidly gaining ground. Image analysis, covering radiological, pathological, and optical imaging, remains a central focus. Meanwhile, clinical decision support systems for treatment planning and toxicity prediction are evolving rapidly. Current research trends emphasize improving model interpretability, tackling small sample learning challenges, and integrating multi-center data.

Future efforts should address key challenges: (1) enhancing model generalizability to manage data distribution shifts via transfer learning and domain adaptation; (2) advancing clinical applicability through development of end-to-end systems beyond algorithm validation; (3) fostering multidisciplinary collaboration with standardized data collection and processing; and (4) reinforcing ethical oversight to ensure AI safety, fairness, and accountability. With ongoing technological progress and clinical validation, AI is poised to become a cornerstone of precision medicine in HNC.

Author contributions

Chuyi Cai and Kai li designed and prepared manuscript. Zachary James Drew, Xiaohui Wang, Qi Dai, Guoping Chen participated in the revision of the manuscript.

Funding

This study is supported by Medical Scientific Research Foundation of Zhejiang Province (2020KY841); Project of National Key Clinical Specialty (Department of Medical Imaging, Grant No. 2024017); Ningbo Clinical Research Center for Medical Imaging (No. 2021L003).

Data availability

The original data can be obtained by contacting the corresponding author.

Declarations

Ethics approval and consent to participate

The data used in this study were obtained from public databases and did not require an ethical approval. Chuyi Cai and Kai li designed and prepared manuscript. Zachary James Drew, Xiaohui Wang, Qi Dai, Guoping Chen participated in the revision of the manuscript.

Consent for publication

All authors agreed to publish this article.

Competing interests

The authors declare no competing interests.