A bibliometric analysis of artificial intelligence in ovarian cancer research from 2006 to 2025

Abstract

Background

Ovarian cancer is a gynecological malignancy associated with high mortality and poses significant clinical challenges in early diagnosis and precision treatment. Although the rapid advancement of artificial intelligence (AI) has introduced novel approaches to this field, a comprehensive bibliometric overview remains lacking. This study aims to fill this gap by providing a systematic bibliometric analysis of this rapidly evolving domain.

Methods

In this study, the Web of Science Core Collection (WoSCC) was used to retrieve literature on AI applications in ovarian cancer research published from 2006 to the search date (November 19, 2025). Using CiteSpace and VOSviewer, we conducted visual and quantitative analyses of publication trends, countries/regions, institutions, authors, journals, highly cited papers, and keywords.

Results

A total of 786 publications were included in the analysis. The annual publication output showed pronounced exponential growth, with a marked acceleration after 2019. China, the United States, and the United Kingdom were the leading contributing countries. Research hotspots centered on AI-assisted diagnosis, prognostic prediction models, radiomics, and biomarker discovery. The evolution of keywords indicated that frontier research has shifted from basic classification toward more advanced areas, including high-grade serous ovarian carcinoma, multimodal learning, and explainable AI.

Conclusion

Research on AI in ovarian cancer has progressed rapidly, with international collaboration concentrated among leading contributors such as China, the USA, and the UK. Future efforts should prioritize the development of explainable and robust clinical AI systems, deeper integration of multimodal data, closer collaboration between clinicians and AI researchers, and high-quality data sharing to facilitate the translation of research findings into precise clinical practice.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05483-2.

Introduction

Ovarian cancer ranks among the most lethal gynecological malignancies. Due to its insidious early symptoms, over 70% of patients are diagnosed at an advanced stage (III/IV), which leads to complex treatment and poor prognosis [1]. The 5-year survival rate has long remained around 30–50% [2, 3]. Although surgery combined with platinum-based chemotherapy remains the standard treatment, significant clinical challenges persist, primarily owing to tumor heterogeneity, high recurrence rates, and the emergence of drug resistance [2, 4]. Therefore, exploring novel approaches to achieve early and precise diagnosis, prognostic stratification, and personalized treatment is of urgent clinical importance for improving survival outcomes in patients with ovarian cancer.

In recent years, the rapid advancement of AI technology has provided transformative tools for addressing these challenges. Techniques such as machine learning, deep learning, and radiomics enable the automatic extraction of complex patterns and deep correlations from high-throughput, multidimensional medical data, including medical imaging, genomics, histopathology slides, and clinical text [5, 6]. In ovarian cancer research, AI applications have been extensively explored across multiple stages of disease management. For diagnostic support, AI models based on CT, MRI, or ultrasound imaging can enhance the detection of minute lesions and improve the accuracy of benign or malignant differentiation [7]. In prognosis assessment, multimodal predictive models integrating clinical, molecular, and imaging features hold promise for more refined risk stratification [8]. For treatment guidance, AI can predict chemotherapy sensitivity, assist in surgical planning, and identify potential therapeutic targets [9, 10]. Collectively, these studies demonstrate the immense potential of AI to drive the transformation of ovarian cancer management toward precision and intelligence.

With the rapid accumulation of research, the application of AI in ovarian cancer has evolved into a vast and complex knowledge domain. However, existing bibliometric studies have specifically examined machine learning applications in ovarian cancer [11]. Despite this growth, a systematic panoramic analysis encompassing the full spectrum of AI—including machine learning, deep learning, natural language processing, computer vision, and other subfields—in ovarian cancer research remains lacking. Prior bibliometric analyses have examined narrower aspects of this field. Zeng et al. analyzed 777 publications on machine learning and deep learning in ovarian cancer (2004–2024), focusing on ML-driven biomarker discovery and imaging-based diagnosis [12]. Wang et al. examined 149 radiomics-specific articles (2010–2024), mapping imaging biomarker development but excluding other AI methodologies [13]. Leng et al. analyzed 5,958 medical imaging publications in ovarian cancer (2000–2022) without AI-specific stratification, making it impossible to distinguish AI-based from conventional imaging studies [14]. These studies were also completed before the emergence of large language models (late 2022 onwards). Therefore, a systematic analysis encompassing the full AI spectrum—including NLP, computer vision, radiomics as an independent methodology, and emerging LLMs—remains lacking. Identifying core authors and institutions, international collaboration patterns, the distribution of academic influence, and the evolution of research hotspots is crucial for understanding the field’s dynamics, pinpointing research gaps, guiding resource allocation, and fostering interdisciplinary collaboration.

To address this gap, this study employs bibliometric methods to quantitatively and visually analyze the literature. It aims to delineate the research landscape by examining publication trends, collaboration networks, core knowledge sources, and evolving hotspots. This work seeks to provide researchers, clinicians, and policymakers with a clear blueprint of the field’s development while offering empirical evidence and strategic references for future research directions.

Materials and methods

Data source

This study used the Web of Science Core Collection (WoSCC) as the primary data source. WoSCC was selected due to its authoritative coverage of high-impact journals across the biomedical sciences, engineering, and interdisciplinary fields, which aligns with the cross-domain nature of AI in ovarian cancer research. Additionally, its structured citation data format ensures compatibility with bibliometric visualization tools such as CiteSpace and VOS viewer.

Search strategy

A comprehensive search strategy was designed to capture the breadth of AI applications in ovarian cancer. The search query combined two conceptual blocks using the “Topic” (TS) field in WoSCC. The first block encompassed AI and related technologies (e.g., “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “computer vision”). The second block included terms for ovarian cancer (e.g., “ovarian cancer”, “ovarian neoplasm*”). The specific Boolean search string was as follows: TS = (“artificial intelligence” OR “intelligence, artificial” OR “artificial-intelligence” OR “computer reasoning” OR “reasoning, computer” OR “AI (artificial intelligence)” OR “AI” OR “machine intelligence” OR “intelligence, machine” OR “machine learning” OR “computational intelligence” OR “intelligence, computational” OR “computer vision system*” OR “system*, computer vision” OR “vision system, computer” OR “knowledge acquisition (computer)” OR “acquisition, knowledge (computer)” OR “knowledge representation* (computer)” OR “representation, knowledge (computer)” OR “deep learning” OR “large language model*” OR “multimodal learning” OR “computer-aided diagnosis” OR “radiomics” OR “chatgpt*” OR “gpt*” OR “transfer learning” OR “unsupervised learning”) AND TS = (“ovarian cancer” OR “ovarian neoplasm” OR “neoplasm, ovarian” OR “ovary neoplasms” OR “neoplasm, ovary” OR “ovary cancer” OR “cancer, ovary” OR “cancer of ovary” OR “cancer of the ovary” OR “cancer, ovarian”).

The publication year was set from 2006 to 2025 (November 19, 2025), to specify the cutoff date for the search. Only English-language articles and reviews were included. The initial search yielded 1,046 records, which were exported in plain text format for further analysis. The information retrieval and screening process is shown in Fig. 1.

Fig. 1.

Flowchart of the bibliometric analysis process

Data selection and standardization

Following the initial retrieval, the dataset underwent a multi-step selection and standardization process to ensure high-quality data for analysis. First, all retrieved records were imported into EndNote X9 software for management and initial screening based on titles and abstracts. Full-text assessments were conducted when necessary to confirm relevance. Second, records were excluded based on the following criteria: (1) studies that did not involve both AI and ovarian cancer; (2) studies in which ovarian cancer was not the primary subject; (3) studies whose core methodology was not based on AI (e.g., relying solely on traditional statistical analysis). After applying these criteria, 786 records were selected for inclusion, comprising 709 research articles and 77 reviews.

To minimize bias and ensure reproducibility, authors and keywords with varying expressions were manually unified into standardized forms following a predefined protocol. For author names, we adhered to the following hierarchical rules: (1) WoSCC canonical format priority: when multiple variants existed (e.g., “Wang J”, “Wang J.“, “Ji Wang”), we used the format most frequently indexed in WoSCC author records; (2) Full name verification: for ambiguous initials, we cross-referenced with institutional affiliations and ORCID identifiers where available; (3) Consistency rule: identical names from the same institution were merged, while homonyms from different institutions were retained as separate entries with institutional suffixes (e.g., “Li Wei Harvard” vs. “Li Wei_Fudan”).

For keywords, standardization followed these rules: (1) MeSH term alignment: uncontrolled author keywords were mapped to their closest Medical Subject Headings (MeSH) equivalents using the NLM MeSH Browser (e.g., “ovarian tumor” → “Ovarian Neoplasms”, “machine learning” → “Machine Learning”); (2) Hierarchical consolidation: specific subtypes were grouped under broader MeSH categories where appropriate (e.g., “deep neural network”, “CNN”, “ResNet” → “Deep Learning”), with original terms retained as sub-entries; (3) Temporal consistency: abbreviations were expanded to full forms for first occurrence (e.g., “AI” → “Artificial Intelligence”), with acronyms retained in parentheses for subsequent use; (4) Exclusion of non-informative terms: generic terms (e.g., “patient”, “outcome”, “analysis”) were removed from keyword lists unless they formed part of specific methodological phrases.

Two researchers independently conducted the entire process and cross-verified their results. Any discrepancies were resolved through discussion with a third researcher until consensus was reached.

Bibliometric and visualization analysis

Bibliometric and visual analyses were performed using CiteSpace (version 6.4.R1) and VOSviewer (version 1.6.20). VOSviewer was primarily employed to construct and visualize static networks of collaboration (e.g., countries, institutions, authors) and keyword co-occurrence, leveraging its strength in handling large datasets and creating clear, interpretable maps [15]. CiteSpace was used for its dynamic analysis capabilities, specifically to identify emerging trends through keyword burst detection and to map the evolution of knowledge domains over time via journal dual-map overlays [16–19]. Journal impact factors were obtained from the 2025 Journal Citation Reports (JCR).

Results

Annual publication trends

Annual publications on AI in ovarian cancer research exhibited a clear exponential growth trend from 2006 to 2025 (Fig. 2). Publication volumes remained low (1–5 papers per year) between 2006 and 2015, then began to rise steadily from 2016 onward. This growth accelerated markedly after 2019, with annual outputs surging from 30 in 2020 to 228 in 2025 (data current as of November 19, 2025, representing a partial year). The trend was statistically validated by an exponential curve fit (y = e^^0.238x, R² = 0.9651), confirming an excellent model fit.

Fig. 2.

Yearly publication outputs in AI in ovarian cancer research

Countries or regions analysis

The top 10 countries in terms of paper output on AI application in ovarian cancer research are detailed in Supplementary Table S1. China leads with 321 papers (40.8%), followed by the United States with 204 papers (26.1%) and the United Kingdom with 80 papers (10.2%). In terms of total citation frequency, the United States ranked first with 4,490 citations, followed by China (3,615 citations) and the United Kingdom (1,665 citations). Regarding the intensity of international collaboration, as measured by total link strength, the United States ranked first (187), followed by the United Kingdom (171) and Italy (144). Despite leading in output, China ranked lower on this measure.

Figure 3A displays the international scientific collaboration network generated using VOSviewer, with node size weighted according to the number of papers published by each country or region. The thickness and number of connecting lines indicate the intensity of collaborative relationships.

Fig. 3.

Visual Analysis Chart of Countries/Regions, Authors, and Institutions. (A visual analysis of countries/regions; B visual analysis of authors; C visual analysis of institutions)

Authors analysis

Publication records of the top 10 most productive scholars are detailed in Supplementary Table 2. As shown in Figure 3B, where node size indicates publication volume, the most prolific scholar is Sala Evis (14 papers) from the United States, followed by Alexandros Laios (13 papers). David Nugent and Amudha Thangavelu (both with 12 papers) are tied for third place. Among these 10 high-output authors, eight are from the UK, reflecting the prominent contributions of British scholars in this field.

Citation frequency serves as a core metric for measuring academic influence. Supplementary Table S3 presents the top 10 scholars ranked by citation count. American scholar Sala Evis leads the list with 350 citations, demonstrating exceptional academic influence. Singaporean scholar U. Rajendra Acharya (235 citations) and Italian scholar Giacomo Avesani (229 citations) rank second and third, respectively. Notably, Sala Evis appears on both the high-output and highly cited scholars lists, underscoring her significant contributions to the field.

Institutions analysis

Figure 3C illustrates the collaborative network among institutions. Node size corresponds to the number of publications, and line thickness reflects collaboration intensity. Based on publication counts (Supplementary Table S4), Memorial Sloan Kettering Cancer Center (USA) and Fudan University (China) jointly lead with 25 publications each, followed closely by Shanghai Jiao Tong University (China) with 20 publications. In terms of academic influence measured by citation count (Supplementary Table S5), the University of Texas MD Anderson Cancer Center leads with 983 citations, followed by the University of Pennsylvania (951 citations) and the Mayo Clinic (906 citations). Among the top 10 institutions by citation impact, U.S. institutions occupy the top six positions, indicating their strong academic influence in this field.

Notably, Memorial Sloan Kettering Cancer Center, Shanghai Jiao Tong University, and Fudan University rank among the top 10 in both publication volume and citation impact, demonstrating their comprehensive strength in this field.

Journal analysis

The top 10 journals by publication volume are listed in Supplementary Table S6. Gynecologic Oncology leads with 73 papers, followed by Frontiers in Oncology (41 papers) and Cancers (38 papers), demonstrating their strong academic cohesion in the field. Supplementary Table S7 lists the top 10 journals by citation impact. Cancers ranked first with 38 papers receiving 537 citations. European Radiology placed second (10 papers, 525 citations), and Nature Communications ranked third (7 papers, 409 citations).

Figure 4 depicts the journal citation network. Nodes in the diagram represent academic journals, with node size reflecting the number of published papers. Connections between nodes indicate citation frequency among journals, where higher citation counts are represented by thicker lines. All journals belong to the Q1 or Q2 category.

Fig. 4.

Visual analysis of journals

To further analyze knowledge flow, a dual-map overlay visualization was employed (Fig. 5). The left side represents citing journals, and the right side indicates cited journals, with labels specifying the subject areas. Colored curves denote primary citation pathways. Four main pathways are discernible: from “Molecular, Biology, Immunology” to “Molecular, Biology, genetics” (z = 4.644208, f = 1872); from “Molecular, Biology, Immunology,” to “Health, Nursing, Medicine” (z = 1.9220096, f = 845); from “Medicine, Medical, Clinical” to “Molecular, Biology, Genetics” (z = 5.49241, f = 2192); from “Medicine, Medical, Clinical” to “Health, Nursing, Medicine” (z = 4.031912, f = 1641). These pathways reflect the broad scope and interdisciplinary nature of research in this field.

Fig. 5.

Dual-map overlay analysis of journals

Highly cited studies

The 10 most frequently cited papers, each with over 110 citations, are detailed in Supplementary Table S8. The most cited article is “Genomic and Molecular Landscape of DNA Damage Repair Deficiency across The Cancer Genome Atlas,” published in Cell Reports in 2018 (802 citations). Ranking second is “Multimodal Data Integration Using Machine Learning Improves Risk Stratification of High-Grade Serous Ovarian Cancer,” published in Nature Cancer in 2022 (202 citations). Third is the 2019 Clinical Cancer Research study “Application of AI for Preoperative Diagnostic and Prognostic Prediction in Epithelial Ovarian Cancer Based on Blood Biomarkers” (155 citations).

Analysis reveals that cited research is concentrated on AI applications in the diagnosis and prognostic assessment of ovarian cancer, particularly high-grade serous ovarian cancer. Among the top ten, four papers focus on AI or machine learning in diagnostic and risk prediction models [20–23]; two elucidate the molecular biology and pathogenesis of ovarian cancer [24, 25]; two are dedicated to discovering and validating novel prognostic biomarkers [26]; one focuses on correlating quantitative imaging features with clinical outcomes [27]; and one reviews the status and challenges of AI across gynecologic oncology [28].

Keywords analysis

Keywords highlight research hotspots and summarize core content. Supplementary Table S9 lists the 30 most frequently occurring keywords. A keyword co-occurrence network diagram was generated using VOSviewer (Fig. 6A). Node size reflects keyword frequency, and different colors denote distinct thematic clusters. Thirteen clusters were identified. Core themes were determined by analyzing high-frequency and central keywords within each cluster. For example, Cluster 1 (red) includes “ovarian cancer,” “express,” “prognosis,” “chemotherapy,” and “carcinoma.” Cluster 2 (green) includes “machine learning,” “radiomics,” “surgery,” “prediction,” and “classification.” Cluster 3 (blue) includes “cancer antigen 125 (CA-125),” “human epididymis protein 4 (HE-4),” “mass spectrometry,” and “biomarkers.” Cluster 4 (yellow) includes “neoadjuvant chemotherapy,” “epithelial ovarian cancer,” “F-18-FDG Positron Emission Tomography/Computed Tomography (PET/CT),” and “diffusion-weighted mri.” Cluster 5 (purple) contains “patterns,” “pathogenesis,” “oophorectomy,” and “epidemiology.” Cluster 6 (light blue) comprised “gene expression,” “pathways,” “survival analysis,” “family-history,” and “Breast Cancer Gene (BRCA) (1,2).” Cluster 7 (pale yellow) included “quality-of-life,” “functional-assessment,” “early detection,” and “clinical-trial.” Clusters 8 through 13 are detailed in Fig. 6B.

Fig. 6.

Keyword visual analysis. (A Visual analysis of co-keywords; B Visual analysis of keywords Cluster; C Visual analysis of keywords Bursts.)

Keyword burst analysis identifies sudden increases in keyword frequency over time, revealing emerging frontiers and evolutionary trends (Fig. 6C). The top five keywords by burst intensity are: “classification” (intensity = 8.29, 2008–2021), “gene expression” (intensity = 7.03, 1998–2019), “simple summary” (intensity = 6.95, 2021–2022), “breast cancer” (intensity = 6.13, 2009–2019), and “high-grade serous ovarian cancer” (intensity = 6.02, 2018–2023).

Discussion

This study systematically reviewed research progress on AI in ovarian cancer from 2006 to 2025 using bibliometric methods, revealing development trends, collaborative networks, research capacity distribution, and knowledge structure. Unlike earlier analyses focused on single technologies (e.g., machine learning) or specific clinical procedures (e.g., robotic surgery) [11], this study provides a broader perspective. The findings indicate that the field has entered a phase of rapid development, forming a multidisciplinary research ecosystem centered on China, the United States, and the United Kingdom, and demonstrating significant value in addressing key clinical challenges such as diagnosis, prognosis, and biomarker discovery.

Trends and driving factors

This bibliometric study differs fundamentally from recent quantitative syntheses of AI in ovarian cancer [29–33]. While those works provide pooled estimates of diagnostic accuracy, survival prediction performance, and biomarker efficacy, our analysis maps the intellectual landscape of the field itself—examining publication trends, collaboration networks, keyword evolution, and research frontiers. Thus, the present study complements rather than competes with these prior syntheses, offering macro-level structural insights that performance-focused meta-analyses cannot capture, while deferring to them for clinical performance benchmarks.

Our analysis identified a clear and robust exponential growth in publications concerning AI applications in ovarian cancer research, with a marked acceleration after 2019. This trajectory appears to align with the diffusion of AI in other oncological specialties such as melanoma [17] and prostate cancer [18]. However, interpreting this trend requires an appreciation of the unique clinical urgency of ovarian cancer. Although its global annual incidence (~ 299,000 cases) [34] is substantially lower than that of breast or lung cancer, the pace of AI-related publications rivals these better-resourced fields. This disproportionate focus could be partly attributed to a particular clinical urgency: a 70% rate of late-stage diagnosis [1] and a stagnant 30–50% five-year survival rate [2, 3] have created extraordinary clinical tolerance for investment in AI-assisted solutions, despite the disease’s relatively low prevalence. The clinical stakes are widely recognized as high—delayed diagnosis directly leads to unresectable disease and platinum resistance—potentially making ovarian cancer a paradigmatic case where AI is viewed not merely as an efficiency tool, but as a potential escape from a decades-long therapeutic plateau.

This acceleration may be attributed to several factors. First, the COVID-19 pandemic catalyzed demand for digital and remote diagnostic tools [35, 36], which indirectly normalized AI deployment in gynecologic workflows. Second, pivotal algorithmic advances—the transfer of transformer architectures to medical image analysis and improved multimodal data fusion frameworks—have enabled more accurate modeling of highly heterogeneous diseases [37–39]. Third, the maturation of large public databases such as TCGA [40, 41] provided critical training substrates. This may suggest a transition from exploring technical feasibility to a phase increasingly shaped by clinical needs and data availability. Yet, unlike in radiology, where public datasets (e.g., ImageNet, ChestX-ray14) enable broad algorithmic pre-training, ovarian cancer AI demands prospective, multi-center validation across diverse histopathological subtypes (high-grade serous, clear cell, mucinous) with linked genomic and survival data. This specificity means that the growth in publications signals a transition from technical feasibility studies toward a phase driven by profound, unmet clinical needs—but it also foreshadows the validation bottlenecks we discuss below. We acknowledge a limitation in trend interpretation: the 2025 data represent a partial year (through November 19). While this does not invalidate the exponential growth pattern—indeed, the 228 papers recorded in under eleven months already exceed the 2024 total—it means that the apparent acceleration into 2025 may be slightly conservative in our model. The true annual total for 2025 could be 15–20% higher, potentially strengthening, rather than weakening, the exponential trend. Future bibliometric updates should verify whether the post-2019 acceleration represents a sustainable paradigm shift or a pandemic-associated transient spike. While publication volume reflects clinical demand and technical capability, it does not guarantee effective knowledge translation. The following analysis of international collaboration reveals a structural barrier: the very data intensity that drives China’s output leadership simultaneously impedes the cross-border validation required for clinically safe ovarian cancer AI.

International cooperation and regional distribution

Our analysis delineates a landscape dominated by three countries— China, the United States, and the United Kingdom—which collectively contributed 77.1% of publications. China leads in quantitative output (321 papers, 40.8%), which may reflect substantial national investment and clinical data resources [42–44]. The United States, while ranking second in output, leads in both total citation count and international collaboration intensity, suggesting its role as a hub for high-impact research and global networking. The United Kingdom maintains a strong presence with significant output and dense collaborative ties.

However, network analysis reveals a notable pattern: China’s prolific output is not matched by proportionally high international co-authorship compared to the U.S. and some European nations. This gap has practical stakes for ovarian cancer AI, because model performance is sensitive to population-specific factors—including histotype distributions and CA-125 trajectories—that vary across regions. Without cross-population validation, models developed primarily on domestic cohorts may not generalize. Strengthening international cooperation through data-sharing policy coordination is therefore not merely an abstract goal but a practical precondition for ensuring the global safety of ovarian cancer AI tools.

Core research strengths and interdisciplinary nature

Analysis of authors and institutions reveals an interesting pattern: there is limited overlap between the most prolific authors and the most highly cited ones. Sala E is the notable exception, ranking highly in both.

This split has practical implications—researchers focused on problem—specific tool development (e.g., radiomic signatures for platinum resistance [20], surgical outcome forecasting [27]) trend to publish more, whereas methodological generalists develop cross-domain architectures attract more citations despite contributing fewer ovarian cancer-specific papers. Both kinds of work are necessary, but the gap between them may partly explain why so few AI tools have moved into clinical use for ovarian cancer: building a tool that works in one retrospective cohort is very different from building one that meets the evidence threshold for regulatory approval. Without prospective trial infrastructure—which neither group is incentivized to build—this gap is unlikely to close on its own.

One potentially critical missing link may be prospective clinical trial infrastructure. Treatment decisions—primary debulking versus neoadjuvant chemotherapy—carry irreversible consequences that require long-term survival endpoint validation. Neither publication-volume-focused engineering groups nor citation-impact-driven generalists are incentivized to conduct resource-intensive international prospective trials. At the institutional level, U.S. cancer centers (e.g., MD Anderson) hold an advantage in citation impact, a position that may related to their clinical resources and international networks [45–47]. Dual-map overlay analysis reveals knowledge flow from basic research fields to applied domains, demonstrating that AI serves as a bridge molecular discoveries and clinical practice [19]. Yet, without structural incentives to close the trial gap, this bridge remains incomplete for ovarian cancer.

Research hotspots and frontier developments

Clustering identified 13 thematic groups, with diagnosis and prognosis, machine learning and radiomics, and biomarkers as the three primary directions. Convergence across clustering, burst detection, and high-citation analysis confirms that there three areas constitute the enduring core of the field [27, 37, 38]. Within diagnostic applications, ultrasound-based AI models have shown particular promise, achieving reported AUCs of 0.90–0.95 in distinguishing benign from malignant adnexal masses [43]. For prognostic prediction, multimodal approaches integrating clinical, genomic, and radiomic features have demonstrated improved risk stratification for high-grade serous ovarian cancer compared to clinical variables alone [20, 48]. Radiomics, as an independent methodology, has become a powerful tool for extracting quantitative imaging features that correlate with tumor biology and treatment response [27].

Looking at how the terminology has shifted over time, early studies centered on basic classification and gene expression, whereas more recent work clusters around “high-grade serous ovarian cancer,” “multimodal learning,” and “radiomics” [26, 27]. —a trajectory that suggests the field is moving from general methodology exploration toward subtype-specific investigation and from single-modality analysis toward multimodal integration.

However, explainable AI (XAI) appears relatively underdeveloped considering its potential regulatory importance for clinical adoption. Unlike dermatology or radiology AI, where false positives trigger non-invasive follow-up, ovarian cancer AI guides irreversible treatment decisions with high morbidity stakes. Gynecologic oncologists cannot validate or override “black box” predictions without auditable reasoning linking radiomic features to histopathological outcomes. The current “accuracy-first” development paradigm—evident in highly cited papers focusing on AUC optimization—misaligns with gynecologic oncology’s risk-averse clinical culture. Our analysis raises the possibility that regulatory science, rather than algorithmic performance, could become a key bottleneck for deployment in the coming years. Early investment in XAI architectures that natively output clinically interpretable features may be important to bridge this gap.

Generative AI represents an emerging frontier. Although our search included generative AI terms (e.g., “ChatGPT*”), these records were not separately quantified due to their extremely low frequency as of 2025. Peer-reviewed generative AI applications in ovarian cancer remain rare, signaling a nascent frontier. Future research may explore LLM-based clinical decision support, automated report generation from imaging or pathology data, and patient-facing educational tools. Early movers in this space will need to establish domain-specific validation protocols and benchmark datasets.

Translational challenges and clinical implementation gaps

Despite the exponential growth in publications, a critical paradox can be observed: algorithmic performance has improved dramatically, yet clinical translation appears to remain limited. Our bibliometric analysis reveals that the field is overwhelmingly dominated by retrospective, single-center cohort studies. While these studies demonstrate technical feasibility, they fail to address the translational requirements of clinical deployment. For ovarian cancer AI, at least three specific implementation gaps may warrant attention:

First, prospective validation infrastructure is lacking. Ovarian cancer treatment follows dynamic clinical pathways—decisions regarding primary debulking surgery versus neoadjuvant chemotherapy, interval debulking, and maintenance therapy selection are sequential and interdependent. Static models trained on retrospective data often fail to generalize to these evolving pathways. Promisingly, emerging initiatives such as the Ovarian Cancer AI Consortium (OCAIC) and AI-specific arms within established trial networks (e.g., GOG, ENGOT) could serve as templates for prospective validation frameworks. Second, data interoperability and standardization remain inadequate. Unlike the relatively uniform DICOM standards in radiology, ovarian cancer AI must integrate heterogeneous data modalities (ultrasound/CT/Magnetic Resonance Imaging (MRI) radiomics, whole-slide pathology, genomic panels, and longitudinal CA-125 values) locked within non-interoperable hospital information systems. Solutions may include adopting common data models (e.g., OMOP CDM) across participating institutions and developing standardized imaging acquisition protocols specifically for ovarian cancer AI research. Third, regulatory pathways for gynecologic oncology AI are underdeveloped. While radiology AI has established 510(k) and De Novo pathways, ovarian cancer AI—which spans diagnostic, prognostic, and therapeutic decision-support functions—lacks clear regulatory taxonomy. This ambiguity discourages industry investment and academic teams from pursuing the costly clinical trials necessary for approval. Regulatory bodies such as the FDA and EMA may wish to consider developing specific guidance categories for gynecologic oncology AI, potentially including “breakthrough device” designation pathways that facilitate prospective data collection while maintaining safety standards. Future efforts may need to prioritize building international prospective trial consortia specifically for ovarian cancer AI, with standardized endpoints that include not only diagnostic accuracy but also surgical outcome metrics, quality-of-life measures, and long-term survival.

Data heterogeneity and algorithmic fairness in global context

China’s domestic AI development largely proceeds on local cohorts [49, 50], while U.S. and U.K. groups draw on shared resources such as TCGA and UK Biobank [39, 51–53]. This fragmented landscape may obscure a challenge that aggregate population counts alone cannot reveal:

algorithmic fairness across diverse populations with different histotype distributions. Ovarian cancer is not a single biological entity but a constellation of histotypes with distinct molecular drivers and epidemiological patterns. Notably, the distribution of these histotypes varies significantly by geography. Studies indicate that East Asian populations, including China, exhibit a higher proportion of clear cell carcinoma and endometrioid carcinoma—subtypes strongly associated with endometriosis—whereas HGSOC constitutes over 70% of cases in Western cohorts. These subtypes differ fundamentally in chemosensitivity, surgical approach, and prognosis.

Our analysis suggests that the current fragmentation of data governance is fostering parallel, non-interoperable AI development ecosystems. Chinese algorithms optimized on domestic cohorts may exhibit performance drift when applied to Western HGSOC-predominant populations, and conversely, Western models may systematically underperform on clear cell carcinomas prevalent in Asia. Such disparities are invisible in aggregate accuracy metrics but have profound implications for clinical safety. This suggests that federated learning and privacy-preserving computation frameworks could be worth prioritizing, as they could potentially enable cross-border model validation without violating data localization requirements. By training models across distributed datasets without centralizing patient-level data, federated learning could break the current data silos and ensure that ovarian cancer AI systems are robust across diverse histotype distributions. Policymakers and funding agencies may wish to prioritize such infrastructure to prevent AI from inadvertently widening global disparities in gynecologic oncology outcomes.

Research limitations and future directions

Several limitations of this work should be kept in mind. First, data were sourced exclusively from the WoSCC. This study chose WoSCC because its structured citation records support the co-citation and burst detection analyses central to this study—functionality not equivalently available in Scopus or PubMed. However, journal coverage across these databases is far from complete, and some relevant studies published in journals indexed only by Scopus, PubMed, or regional databases would have been missed. Additionally, we restricted our search to articles and reviews, which excluded conference proceedings and other document types. While this is standard bibliometric practice, it may omit very recent work presented at conferences but not yet published as journal articles. Citation counts derived from WoSCC also differ from those in Scopus or Google Scholar, and our rankings of highly cited papers and co-citation clusters should be interpreted with this in mind. Future studies incorporating multiple databases and broader document types could further validate our findings. Second, the literature screening was limited to English-language papers, possibly overlooking research published in Chinese, Korean, or other languages. Third, despite manual standardization, some subjective errors could persist. Finally, the search cutoff date of November 2025 does not cover the full year; however, the 228 papers recorded through mid-November already exceed the 2024 total, so the exponential trend is likely robust.

Based on these findings, we propose the following targeted recommendations. First, researchers may consider strengthening research on multimodal data fusion, but with an explicit focus on clinical interpretability—developing XAI architectures that natively output clinically actionable features rather than post-hoc explanations. Second, funding agencies could encourage transnational, interdisciplinary research networks that specifically include gynecologic oncologists, pathologists, and AI researchers in jointly designed prospective trials, not merely retrospective data mining. Third, journals and institutions may support the publication of innovative, clinically significant research in high-impact journals while promoting open-source sharing of de-identified, standardized datasets and algorithms to enhance transparency and reproducibility; this should include histotype-stratified benchmarks to test for algorithmic bias. Fourth, at the policy level, policymakers may consider increasing support for AI research in gynecologic oncology, with emphasis on data standardization across modalities, ethical frameworks for federated learning, and dedicated clinical validation platforms that can accommodate the longitudinal, multi-modal nature of ovarian cancer care. Fifth, regulatory bodies may wish to develop clear guidance for ovarian cancer AI that distinguishes between diagnostic, prognostic, and therapeutic decision-support tools, accelerating safe clinical translation. Sixth, the bibliometric analysis in this study relies primarily on descriptive visualization and basic statistical fitting provided by CiteSpace and VOSviewer. Advanced quantitative network metrics—including betweenness centrality, clustering coefficients, modularity indices, and formal model selection criteria (AIC/BIC)—were not computed due to software limitations. While this descriptive approach is standard in bibliometric reviews and sufficient for landscape mapping, future studies should employ dedicated network analysis tools (e.g., R igraph, Gephi, Pajek) to validate the structural properties of scientific collaboration networks and apply rigorous statistical diagnostics to temporal trend models. Additionally, while our search strategy included generative AI and LLM terms, these records were not separately flagged during initial data extraction, precluding a quantitative sub-analysis of this emerging frontier. Future bibliometric updates should prospectively track generative AI publications separately given their anticipated exponential growth.

Conclusion

Research on AI in ovarian cancer appears to be undergoing rapid development, with indications of close international collaboration and a landscape dominated by China, the United States, and the United Kingdom. Current studies predominantly focus on AI-assisted diagnosis, prognosis prediction, and radiomics, suggesting a trend toward enhancing interpretability, achieving multimodal fusion, and promoting clinical translation. However, based on the results of existing bibliometric analyses, several potential directions can be considered regarding the future direction of this field. First, the field may see a gradual shift from retrospective proof-of-concept studies toward prospective, multi-center clinical trials in the coming years. Second, explainable AI could transition from a technical desideratum to a regulatory prerequisite. Third, federated learning may emerge as a paradigm for international collaboration in this field. Fourth, the field could potentially bifurcate into two validated tracks: screening and early-detection AI optimized for population-specific imaging protocols, and prognostic and therapeutic decision-support AI requiring global histotype-stratified validation. Attempting to develop universal models without subtype-aware architecture could potentially lead to algorithmic failure and clinical harm. To realize this potential, concerted efforts may be needed to strengthen interdisciplinary collaboration, data sharing, and research into model reliability, with the goal of facilitating the substantive application of AI in the precise diagnosis and treatment of ovarian cancer.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Jiujie He : Writing – original draft, Writing – review & editing, Conceptualization, Data curation, Funding acquisition, Software, Visualization. Wanting Zhou : Writing – original draft, Formal analysis, Data curation, Investigation, Software. Yujun He : Writing – original draft, Conceptualization, Methodology, Resources, Software. Yingjie Nie : Writing – review & editing, Investigation, Validation, Visualization. Hua Qiu : Writing – original draft, Formal analysis, Supervision. Wei Mai : Writing – review & editing, Formal analysis, Funding acquisition, Supervision.

Funding

The research was supported by: (1) Joint Project on Regional High-Incidence Diseases Research of Guangxi Natural Science Foundation under Grant (No. 2024GXNSFBA010165); (2) Project of the Universal Support Policy for the First Batch of Young Talent “Qingmiao” Program in Guangxi Zhuang Autonomous Region (No. 2024-2); (3) Youth Program of Scientific Research Foundation of Guangxi Medical University Cancer Hospital (No. 2023-9); (4) Youth Fund of Guangxi Medical University (No. GXMUYSF202445); (5) Excellent Doctoral Funding Project of Guangxi Medical University Cancer Hospital (No. 2023-1); (6) Guangxi Zhuang Autonomous Region Engineering Research Center for the Development and Application of Snake-Based Anticancer Medicines [Guangxi Development and Reform Commission High-Tech Letter (2023) No. 2727]; (7) Guangxi Key Laboratory for the Prevention and Treatment of Regionally Prevalent Cancers Using Traditional Chinese Medicine (Guangxi Traditional Chinese Medicine Science and Education Development [2023] No. 9); (8) Self-funded scientific research project of the Guangxi Zhuang Autonomous Region Administration of Traditional Chinese Medicine (No. GXZYA20240358).

Data availability

The datasets analyzed in this study were retrieved from the Web of Science Core Collection (WoSCC) database. The complete search strategy and inclusion/exclusion criteria are described in the Methods section. The bibliometric datasets generated during the analysis—including annual publication counts, country/institution/author rankings, journal metrics, keyword co-occurrence data, and supplementary tables—are provided in the Supplementary Materials. The raw bibliographic records exported from WoSCC can be obtained from the corresponding author upon reasonable request, subject to the terms of the Clarivate Analytics data usage agreement.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.