Abstract
Background
Artificial intelligence (AI) has become a major methodological driver in computational pathology, but the field’s temporal growth, collaboration structure, intellectual foundations, and emerging translational priorities remain incompletely mapped.
Objective
To characterize the development of AI-related computational pathology from 2009 to 2025 and identify literature-level signals of research maturation, knowledge structure, and emerging frontiers.
Methods
The Science Citation Index Expanded within the Web of Science Core Collection was searched for AI-related computational pathology publications from 1 January 2009 to 31 December 2025. The initial Topic Search identified 2,480 records; after document-type and language filtering, 2,216 English-language articles and reviews were included. CiteSpace, VOSviewer, and bibliometrix were used to analyse publication trends, collaboration networks, citation and co-citation structures, keyword co-occurrence, and burst patterns. Model-fit, network-structure, and cluster-quality metrics were used to support temporal and network interpretation.
Results
Annual publications increased from 1 in 2009 to 540 in 2025, with cumulative citations reaching 20,910. A logistic model provided the primary descriptive fit for publication growth (R2 = 0.990; adjusted R2 = 0.988), capturing rapid post-2019 expansion with attenuation of the growth rate. The United States led in publication output and co-authorship connectivity, China showed rapid volume growth, and the United Kingdom had the highest country-level betweenness centrality. Co-citation and journal analyses indicated an interdisciplinary knowledge base linking computational imaging, pathology, oncology, and biomedical research. Keywords and bursts shifted from classification and segmentation toward analysis of whole-slide images, transformers, biomarker inference, prognostic modelling, multimodal integration, and workflow-oriented research questions.
Conclusions
AI-related computational pathology has developed into a rapidly expanding and interdisciplinary literature. These findings are bibliometric signals of scholarly development, not evidence of clinical effectiveness or real-world implementation. Future work should prioritize externally validated, interpretable, calibrated, workflow-compatible, and ethically governed AI systems across diverse pathology settings.
Keywords
1. Introduction
Pathology plays a central role in disease diagnosis, prognostic stratification, and therapeutic decision-making, serving as a critical bridge between basic biomedical research and clinical practice. 1 Traditionally, pathological assessment has relied heavily on expert visual interpretation of histomorphological features, a process that is labor-intensive, time-consuming, and subject to substantial inter- and intra-observer variability.2,3 The widespread adoption of whole-slide imaging has transformed glass slides into high-resolution, multi-scale digital data, enabling large-scale computational analysis.4,5 Similar digital transitions have also been described in cytopathology, where whole-slide imaging and AI-assisted digital cytology systems are being explored to address image-acquisition, screening, and workflow challenges. 6 Within this context, digital pathology primarily refers to the digitization, management, and visualization of pathological slides, whereas computational pathology represents a distinct analytical paradigm that leverages artificial intelligence (AI) and advanced computational methods to extract quantitative, reproducible, and clinically meaningful information from digitized histopathological data.7–9
Over the past decade, AI—particularly machine learning and deep learning—has emerged as a major methodological driver in computational pathology.10,11 AI-based approaches have demonstrated strong performance across a wide range of tasks, including tissue and cell segmentation, tumor detection and classification, grading, biomarker quantification, and outcome prediction, in some cases achieving accuracy comparable to that of experienced pathologists.12–14 Recent AI-assisted tools in subspecialty pathology, such as pre-implantation kidney biopsy assessment, further illustrate the extension of AI methods beyond tumour-centred image analysis into other diagnostic contexts. 15 More recently, advances in representation learning and model scalability have extended computational pathology beyond image-centric analysis toward the integration of multi-modal data, such as genomics, transcriptomics, radiomics, and clinical variables, supporting a shift toward mechanism-informed and precision pathology.16–18 However, the rapid growth of this field has also generated a highly interdisciplinary and heterogeneous body of literature spanning pathology, oncology, computer science, engineering, and data science.19–21 As a result, fundamental questions regarding the intellectual structure, thematic evolution, and cross-disciplinary knowledge flows of AI-driven computational pathology are increasingly difficult to address through individual studies or traditional narrative reviews.7,19,22
Bibliometric and scientometric approaches provide a systematic and quantitative framework to address these challenges by enabling large-scale analysis of publication patterns, citation relationships, and collaboration networks to reveal intellectual foundations, research trajectories, and emerging paradigms that are not readily discernible at the level of individual studies.23–25 Although bibliometric methods have been widely applied to map the development of rapidly evolving scientific domains, existing studies related to AI in medicine or digital pathology have often been constrained by limited time windows, narrow analytical scopes, or a primary focus on publication productivity rather than on knowledge structure, disciplinary interactions, and translational dynamics.26,27 At a time when computational pathology is transitioning from proof-of-concept studies toward broader clinical and translational consideration—raising persistent challenges related to generalizability, reproducibility, and disease-specific applicability—a comprehensive and longitudinal mapping of its knowledge landscape remains lacking. 28 To address this gap, the present study conducts a bibliometric and knowledge visualization analysis of AI-related research in computational pathology published between 2009 and 2025. By integrating multiple scientometric techniques, we aim to (i) characterize temporal growth patterns and global collaboration networks; (ii) identify major research themes, intellectual bases, and their dynamic evolution; (iii) map interdisciplinary knowledge flows and paradigm shifts; and (iv) highlight emerging research frontiers with potential clinical and translational relevance. Through this integrative analysis, we seek to provide a structured overview of the evolution of AI in computational pathology and to inform future methodological development and translational research.
2. Method
2.1. Search strategy and data collection
The Science Citation Index Expanded (SCIE) of the Web of Science Core Collection was selected as the data source due to its comprehensive journal coverage, rigorous citation indexing, and well-established reputation in citation-based bibliometric analysis.29,30 Because the present study focused on citation-based network mapping rather than exhaustive literature retrieval, a single-source WoSCC/SCIE strategy was used to preserve comparability of citation counts, cited-reference records, author affiliations, and country-level metadata. Although Scopus, PubMed, and Embase provide valuable complementary coverage, cross-database integration may introduce non-equivalent citation metrics, inconsistent cited-reference formats, and duplicate records, which can distort co-citation, burst-detection, and collaboration-network analyses. This database ensures a robust, reproducible, and longitudinal mapping of research across interdisciplinary domains.
To avoid bias from subsequent database updates, all searches and downloads were completed on 12 January 2026. A systematic search was conducted in SCIE covering publications from January 1, 2009, to December 31, 2025, to capture the integration of whole-slide imaging and machine learning techniques in computational pathology. Searches were performed in the Topic Search (TS) field, including titles, abstracts, author keywords, and Keywords Plus. Document types were restricted to “article” and “review,” and the language was restricted to English. Records in other languages were excluded. The complete search strategy is provided in Supplementary Table 1. The initial WoSCC Topic Search identified 2,480 records. After document-type filtering and language restriction, 2,216 English-language articles and reviews were included in the final bibliometric analysis, comprising 1,814 original articles and 402 reviews (Figure 1). The dataset encompasses 12,203 authors from 3,576 institutions across 95 countries and regions. All records were exported from SCIE in plain text format, including bibliographic metadata, abstracts, author keywords, and cited references, and were subsequently used for bibliometric and visualization analyses. Potential limitations include the exclusion of non-English publications and the reliance on a single database, which may limit the scope of the data. PRISMA-style flow diagram of record selection. The initial WoSCC topic search identified 2,480 records. After document-type filtering and language restriction, 2,216 English-language articles and reviews were included in the final bibliometric analysis.
2.2. Bibliometric analysis framework and visualization strategy
Bibliometric and knowledge visualization analyses were performed using CiteSpace
31
(version 6.4.R2), VOSviewer
32
(version 1.6.20), the bibliometrix package in RStudio
33
(Version 2026.01.0+392), and additional software tools, each selected for its ability to address specific dimensions of the study. CiteSpace was used for co-citation analysis, burst detection, and cluster visualization, enabling the identification of foundational research, paradigm shifts, and emerging topics. VOSviewer facilitated the construction of co-authorship and keyword co-occurrence networks, while the bibliometrix package provided descriptive statistics on publication trends, country and journal productivity, and citation indicators. Microsoft Excel was used for data management and cross-validation of bibliometric indicators, and Origin was employed for high-quality graphical rendering of temporal trends. This multi-tool, multi-level approach enabled a comprehensive, reproducible mapping of the research landscape, focusing on intellectual structures, collaborative patterns, and the evolution of AI in computational pathology (Figure 2). All bibliometric indicators were interpreted as proxies of scholarly communication patterns within the retrieved literature, rather than as direct evidence of causality, policy influence, regulatory effects, or clinical implementation. Overview of the bibliometric analysis process. This flowchart outlines the process of bibliometric analysis in this study, beginning with data retrieval from the web of science core collection, followed by data cleaning and normalization. It then progresses through the stages of bibliometric and knowledge mapping analysis, including descriptive analysis, collaboration network analysis, citation/co-citation analysis, and conceptual co-occurrence analysis. The final stage involves the synthesis of research patterns and the identification of intellectual structures, research frontiers, and evolutionary trajectories in the AI-based pathology field.
2.3. Statistical validation and network-structure assessment
To reduce reliance on visual interpretation, temporal and network analyses were supplemented with quantitative validation metrics. For annual publication trends, the quadratic polynomial and logistic models were evaluated using R2, adjusted R2, RMSE, MAE, MAPE, AIC, BIC, and residual diagnostics. CiteSpace-generated networks were assessed using node and edge counts, network density, largest connected component, modularity Q, weighted mean silhouette, harmonic mean Q/S, pruning strategy, and node-selection criteria. VOSviewer-generated maps were documented using item counts, clusters, links, total link strength, threshold settings, and node-weighting metrics. Complete validation metrics are provided in Supplementary Table S2A–C.
3. Results
3.1. Global trend in publication outputs and citations
Between 2009 and 2025, publications on artificial intelligence in pathology increased markedly, accompanied by a sustained rise in cumulative citations (Figure 3(a)). Annual outputs remained low and stable from 2009 to 2014 (1–6 publications per year), with limited citation accrual. A gradual expansion was observed between 2015 and 2018, during which publications increased from 8 to 34 and cumulative citations rose from 129 to 870. From 2019 onward, publication activity accelerated substantially, with annual outputs increasing from 100 in 2019 to 540 in 2025. In parallel, cumulative citations grew from 2,352 to 20,910. The annual H-index increased steadily and reached a maximum of 54 in 2021, followed by a decline in subsequent years, coinciding with the rapid influx of recent publications that had not yet accumulated long-term citations. Global trends in publication outputs, citations, and fitted growth patterns. (a) Global trends in publication outputs and citations. Annual publication counts and cumulative citations from 2009 to 2025 are shown. The annual H-index is included to reflect the temporal distribution of citation impact across years. (b) Fitted trend and projected trajectory of annual publication outputs. Observed annual publication counts from 2009 to 2025 are shown together with a regression-based fitted curve. The fitted and short-term extrapolated trajectory illustrates the continuation of the post-2019 growth pattern based on the fitted model. The projection represents a model-based extrapolation of observed trends rather than a deterministic prediction.
To characterize the growth dynamics of research output, annual publication counts were fitted using both a quadratic polynomial model (blue) and a logistic growth model (red) (Figure 3(b)). The logistic model closely tracked the observed data across the study period, capturing the early low-output phase, the rapid acceleration after 2019, and a subsequent attenuation in the rate of increase. Specifically, annual publications increased from 1 in 2009 to 100 in 2019 and further to 540 by 2025, forming an S-shaped trajectory that is well described by the logistic curve. In contrast to the quadratic fit, which implies continued unbounded growth, the logistic model suggests a gradual deceleration of annual publication increases toward an upper limit under the current trajectory. Model-fit comparisons are summarized in Supplementary Table S2A. Compared with the quadratic polynomial model, the logistic model showed higher R2 (0.990 vs 0.982) and adjusted R2 (0.988 vs 0.978), as well as lower RMSE (17.2 vs 22.8), MAE (11.5 vs 19.5), AIC (153 vs 163), and BIC (156 vs 166). Accordingly, the logistic curve was used as the primary descriptive fit for the observed annual publication trajectory, while the fitted curves were interpreted as descriptive representations rather than deterministic forecasts.
3.2. National contributions and geospatial distribution of research output
Country-level productivity, impact and collaboration metrics.
TLS: total link strength, PY_start: Publication Year Start; SCP: single country publications; MCP: multiple country publication.
Temporal analysis of publication outputs (Figure 4(a)) revealed markedly different national trajectories in the development of AI-driven computational pathology. The United States demonstrated sustained activity from the earliest stage of the field (PY_start = 2009), whereas China showed a pronounced surge after 2014, particularly following 2019. In contrast, several European countries—including the United Kingdom, Germany and the Netherlands—entered the field with more modest publication volumes but engaged earlier in international research. Temporal evolution of national publication outputs and international collaboration networks. (a) Temporal evolution of national publication outputs. Annual publication outputs of the top contributing countries from 2009 to 2025. The heatmap illustrates country-specific publication trajectories over time, highlighting sustained early activity in the United States and a pronounced post-2014 acceleration in China, alongside earlier international engagement among several European countries. Color intensity corresponds to the number of publications per year. (b) International collaboration patterns among countries. Chord diagram depicting international co-authorship relationships among major contributing countries. The width of each chord represents the frequency of collaborative publications between country pairs, illustrating dense bilateral and multilateral ties among the United States, China, and European research systems. (c) Country-level co-authorship network. CiteSpace visualization showing country-level co-authorship ties, with node size proportional to publication output. The right panel highlights the United Kingdom, which had the highest betweenness centrality, indicating an intermediary position within the observed network.
Patterns of collaboration intensity are visualized in the chord diagram (Figure 4(b)), which highlights dense bilateral and multilateral ties among the United States, China and major European partners. Notably, despite a lower publication output than the United States and China, the United Kingdom showed a relatively prominent intermediary position within the observed country-level co-authorship network. At the network level, the CiteSpace country-level co-authorship map comprised 92 country nodes and 803 co-authorship links, with a density of 0.1918 and a largest connected component of 86 nodes (93.0%) (Supplementary Table S2B). Within this largely connected network, the United Kingdom had the highest betweenness centrality (0.18), together with a high proportion of multi-country publications (MCP = 87; MCP% = 58.0) and strong collaborative connectivity (TLS = 484), suggesting a comparatively intermediary position among high-output and internationally connected countries. Germany (125 publications; centrality = 0.12) and France (55 publications; centrality = 0.14) display similar but less pronounced intermediary roles, while smaller systems such as the Netherlands (59 publications; MCP% = 62.7) and Switzerland (35 publications; MCP% = 71.4) demonstrate high levels of international integration relative to output volume.
Collectively, these findings indicate heterogeneous national profiles in publication output and international co-authorship, with some countries contributing a larger share of publications and others showing higher betweenness centrality within the observed country-level co-authorship network.
3.3. Institutional collaboration networks and research contributions
Figure 5(a) and Table 2 visualize the institutional co-authorship network generated by VOSviewer. The network included 100 institutions, eight clusters, 868 links, and a total link strength of 1,718, with full threshold, density, and node-weighting metrics reported in Supplementary Table S2C. A densely interconnected core dominated by Emory University, Case Western Reserve University, Harvard Medical School, and Mayo Clinic occupied a central position in the network, suggesting intensive cross-institutional collaboration and a prominent bridging position across clusters. In parallel, institutions such as the University of Warwick and the Alan Turing Institute in the light-blue cluster, Tsinghua University and the Chinese Academy of Sciences in the red cluster, and Radboud University Nijmegen and Linköping University in the blue cluster form tightly knit regional collaboration groups that are extensively linked to the central hub, highlighting a multi-polar yet highly integrated global collaboration structure. Institutional collaboration and temporal evolution in pathology research. (a) Clustering analysis of institutions based on collaborative patterns, visualized using VOSviewer. Institutions are grouped into clusters based on their research collaboration, with node size reflecting publication volume. (b) Temporal evolution of institutional activity, with node color indicating the average publication year and node size proportional to the total number of publications. Top institutions contributing to AI research in pathology. TLS: total link strength.
Figure 5(b), generated by VOSviewer, maps the average publication year of institutions, with node colour encoding temporal engagement (blue indicating earlier entry and red indicating more recent activity). Established centres such as Harvard Medical School, Case Western Reserve University, and Emory University are characterised by predominantly blue to light-blue nodes, reflecting sustained involvement since the earlier phase of the field, whereas institutions including University of Warwick and Alan Turing Institute show intermediate colours, consistent with consolidation during the expansion period. In contrast, several rapidly emerging contributors—such as Shanghai Jiao Tong University and Sun Yat-sen University—are marked by warmer hues, indicating more recent entry into AI-related pathology research and suggesting a growing cohort of late-joining institutions that may shape subsequent phases of development.
3.4. Author productivity, collaboration networks, and citation impact
Leading authors by publication output and co-citation impact.
TLS: total link strength.
The VOSviewer author collaboration network comprised 200 authors, 11 clusters, 727 links, and a total link strength of 1,950, whereas the co-cited author network comprised 303 cited authors, three clusters, 38,780 links, and a total link strength of 322,667; full threshold, density, and node-weighting metrics are reported in Supplementary Table S2C. The co-authorship network exhibited a modular, colour-segmented structure, in which authors formed multiple tightly knit clusters with limited cross-cluster integration (Figure 6(a)). Each colour-coded group represents a relatively stable collaborative community, often centred on a small number of highly productive investigators, indicating that research activity in AI-related pathology is organised around localized teams rather than a single, globally integrated author network. Prominent contributors such as Madabhushi, A, Rajpoot, N, and Pantanowitz, L occupy central positions within their respective clusters, functioning as intra-cluster coordinators rather than universal collaboration hubs. By contrast, the co-cited author network reveals a markedly more consolidated intellectual landscape (Figure 6(b)). Three dominant colour-defined clusters structure the citation space, each anchored by a distinct methodological lineage. The blue cluster, centred on He, K., reflects foundational deep learning architectures that underpin a broad range of computational pathology applications. The green cluster, led by Kather JN, captures pathology-oriented machine learning frameworks that adapt generic AI methods to histopathological and disease-specific contexts. The red cluster, anchored by Bejnordi, BE, represents validation-driven and benchmark-focused contributions that emphasize performance assessment and clinical robustness. Author productivity, collaboration networks, and citation impact. (a) Author collaboration network visualized using VOSviewer, with nodes representing authors and colored clusters indicating distinct collaboration groups. The size of each node reflects the number of publications, while the links between nodes represent collaborative ties. (b) Co-citation network of authors, with different node colors representing distinct clusters of co-cited authors. Larger nodes indicate higher citation frequency or stronger intellectual influence, with edges reflecting the strength of co-citation relationships.
Notably, dense interconnections among these clusters indicate that contemporary studies rarely rely on a single methodological tradition in isolation. Instead, AI-related pathology research integrates general-purpose deep learning models, pathology-specific adaptations, and rigorously validated analytical pipelines. Taken together, the contrast between a fragmented co-authorship structure and a concentrated co-citation core highlights a field in which collaborative practice is distributed across multiple teams, whereas intellectual influence is shaped by a small number of method-defining contributions that transcend collaborative boundaries.
3.5. Key journals and citation patterns
Major journals contributing to artificial intelligence research in pathology.
The VOSviewer journal citation network included 89 sources, four clusters, 1,389 links, and a total link strength of 5,492, whereas the journal co-citation network included 108 cited sources, five clusters, 5,765 links, and a total link strength of 1,050,105; full network-structure, threshold, and node-weighting metrics are provided in Supplementary Table S2C. Journal citation and co-citation analyses suggest a differentiated publication and knowledge-base structure rather than a self-contained disciplinary core (Figure 7(a)–(c)). Citation mapping (Figure 7(a)) resolves the publication landscape into three tightly coupled but functionally distinct journal communities. A pathology-centred cluster (blue), anchored by Modern Pathology, Journal of Pathology and Histopathology, constitutes the primary application and validation layer in which AI methods are assessed against diagnostic and prognostic tasks. A methodological cluster (green), dominated by Medical Image Analysis, IEEE Transactions on Medical Imaging and Computers in Biology and Medicine, represented a major methodological publication cluster, concentrating algorithmic development, modelling strategies and benchmarking practices. Bridging these domains, an oncology-oriented cluster (red), led by Cancers and npj Precision Oncology, highlights cancer pathology as the dominant clinical context in which computational advances are operationalised. The dense interconnections among clusters indicate coordinated co-evolution driven by clinical demand rather than isolated disciplinary development. Journal-level citation structure and knowledge flows. (a) Journal citation mapping showing the distribution and clustering of journals based on publication volume. Node size reflects the number of publications, and colours denote journal clusters identified by citation similarity. (b) Co-citation network of journals, in which nodes represent cited journals and links indicate co-citation relationships. Larger nodes correspond to higher co-citation frequency, highlighting journals that constitute the intellectual core of the field. (c) Dual-map overlay of journals illustrating directional citation flows between citing and cited domains. Coloured trajectories represent major knowledge transfer pathways, with line thickness indicating citation intensity between disciplinary clusters.
The intellectual foundations and directional logic of this structure are further clarified by the co-citation network and dual-map overlay analyses (Figure 7(b) and (c)). Co-citation patterns position high-impact multidisciplinary journals—Nature, Nature Medicine, Nature Communications and Cell—at the conceptual core despite their comparatively low publication frequencies, underscoring their role in supplying paradigm-setting frameworks and evaluation norms. Interpreted from the cited side, the dual-map overlay reveals that these upstream knowledge sources are concentrated within Systems/Computing/Computer Science and Molecular Biology/Genetics domains, which are systematically imported by journals in Medicine/Medical/Clinical and Molecular Biology/Immunology. This asymmetric knowledge flow is consistent with a clinically driven convergence model, in which pathological research integrates externally generated computational and biological advances to address domain-specific diagnostic challenges.
3.6. Reference burst analysis and co-citation networks
Highly cited articles shaping the intellectual foundation of the field.
CiteSpace-based co-citation analysis generated a sparse but well-partitioned reference network comprising 1,243 nodes and 2,941 edges, with a density of 0.0038 (Figure 8(a)). The network was divided into 15 clusters, with cluster silhouette values ranging from 0.817 to 1.000, modularity Q = 0.7597, weighted mean silhouette = 0.8698, and harmonic mean Q/S = 0.8109. Detailed CiteSpace validation metrics are provided in Supplementary Table S2B. Beyond thematic separation, node colours represented publication time, and inter-cluster links denoted citation dependencies among clusters; together, these visual features were consistent with a temporally layered co-citation structure in AI-related pathology research. At the base of the network, early clusters—most notably computer-aided diagnosis (#13) and annotation-efficient AI (#9)—are characterised by cooler colours and occupy upstream positions with extensive outgoing dependencies. These clusters predominantly addressed algorithmic feasibility, feature representation, and data-efficiency constraints, and were connected to later diagnostic and prognostic research fronts within the co-citation network. In the co-citation map, early clusters such as computer-aided diagnosis (#13) and annotation-efficient AI (#9) were positioned upstream of later application-oriented clusters and were connected to subsequent research fronts related to diagnostic and prognostic applications; this pattern reflects citation dependence within the retrieved literature rather than evidence of necessary technical preconditions. A second tier of clusters, including uncertainty-aware AI (#8), augmented pathology (#0), image adequacy (#4) and generalizable AI (#3), marks a consolidation phase in which methodological reliability, robustness to domain shift and human–AI interaction become central organising principles. These clusters occupied an intermediate citation layer between earlier method-enabling work and more recent diagnostic or prognostic research fronts, and were primarily associated with model reliability, robustness, and human–AI interaction within the retrieved literature. In contrast, the most recent clusters—AI-based diagnosis (#5), automated morphometry (#2) and AI prognostication (#1)—are dominated by warmer colours and occupy downstream positions with predominantly incoming dependencies. Clinically oriented inference tasks, including quantitative phenotyping and outcome prediction, were mainly represented by more recent clusters that were linked to earlier work on annotation efficiency, uncertainty modelling, and generalisation within the co-citation network. The inter-cluster dependency links showed connections between recent application-oriented themes and earlier methodological clusters; these connections are best interpreted as citation dependence within the retrieved literature rather than evidence of conceptual inheritance. Reference co-citation clusters, dependency structure, and bridging references. (a) Reference co-citation network generated using CiteSpace, with clusters identified by log-likelihood ratio (LLR) labeling. Left, the clustered reference network shows major thematic groups, with node size proportional to citation frequency and node colour indicating the average publication year. Directed arrows between clusters represent cluster dependence, indicating asymmetric citation flows from earlier, method-enabling clusters toward downstream, application-oriented clusters, showing a temporally layered structure of citation-based knowledge accumulation. Centre, the global reference network highlights the vertical stratification from foundational methodological studies to clinically oriented research fronts. Right, a magnified view of high-betweenness-centrality references (purple-ringed nodes) illustrates key bridging publications that connect otherwise weakly linked clusters; these references function as intellectual pivots linking methodological advances with subsequent diagnostic, prognostic, and workflow-oriented research directions. (b) Reference burst detection highlighting the key milestones in the development of the field. Burst intensity is represented by line thickness, with early bursts indicating foundational methodological studies and later bursts indicating validation-oriented and application-oriented work.
The citation-burst map reveals a clear stage-wise evolution of artificial intelligence research in pathology (Figure 8(b)). Early bursts (2016–2018) are dominated by foundational methodological studies, including Janowczyk et al. (2016; burst strength = 36.92) and Ronneberger et al. (2015; 32.86), reflecting rapid consolidation of deep-learning architectures and computational workflows. A second phase (2018–2022) is characterised by clinically transformative references with sustained bursts, most notably Bejnordi et al. (2017, JAMA; 44.30) and Esteva et al. (2017, Nature; 19.09), marking the first large-scale demonstrations of specialist-level diagnostic performance. In contrast, the most recent bursts (2024–2025) are driven almost exclusively by transformer-based and weakly supervised models (e.g. Lu et al. 2021; Dosovitskiy et al. 2021; Wang et al. 2022), indicating a shift toward scalability, generalisation, and translation-oriented evaluation.
Collectively, the co-citation clustering, temporal colouring, and dependency structure are consistent with a cumulative pattern of citation-based knowledge accumulation within the retrieved literature. In this descriptive bibliometric sense, later diagnostic and prognostic research fronts were positioned downstream of, and linked to, earlier methodological clusters. This pattern reflects citation dependence within the retrieved literature, rather than evidence that clinically oriented applications are causally or necessarily conditioned on specific methodological milestones.
3.7. Keyword evolution and emerging research trends
High-frequency keywords reflecting major research themes.
TLS: total link strength.
The VOSviewer keyword co-occurrence network included 238 keywords, three clusters, 3,240 links, and a total link strength of 6,145, based on a minimum occurrence threshold of five and node weighting by occurrences (Supplementary Table S2C). Keyword clustering identified three major thematic groups in computational pathology (Figure 9(a)). The red cluster was centred on foundational image-analysis terms, including “computational pathology” and “whole slide image”. The green cluster connected molecular and tumour-microenvironment-related terms, including “biomarker” and “tumor-infiltrating lymphocytes”, indicating a literature-level association between computational pathology and precision-medicine-oriented research. The blue cluster was dominated by segmentation- and architecture-related terms, including “nuclei segmentation” and “transformer”, reflecting continued attention to automated image-analysis methods. The temporal gradient, alongside inter-cluster dependencies, reflects a cumulative knowledge structure, where foundational algorithmic advancements serve as the basis for increasingly clinical and transformative AI applications in pathology. Keyword evolution and temporal dynamics. (a) Keyword co-occurrence network visualizing relationships between major research topics. Node size corresponds to term frequency, while edge thickness reflects the strength of co-occurrence, revealing the key interconnected research areas over time. (b) Temporal evolution of keywords in computational pathology research. A colour gradient indicates the increasing prominence of each keyword over time, with earlier terms in blue and more recent terms in red, showing the shifting focus in the field.
The timeline of keyword evolution in computational pathology reveals distinct research trajectories driven by advancements in AI technologies (Figure 9(b)). Early keywords such as computational pathology and whole slide image mark foundational stages in image analysis, reflecting the initial integration of AI into histopathology. As the field progressed, terms like transformer models, image segmentation, and feature extraction emerged, signaling a shift towards more sophisticated AI techniques in diagnostic accuracy. This evolution is further emphasized by the growing prominence of biomarker, tumor-infiltrating lymphocytes, and immunohistochemistry, indicating a convergence of AI with molecular pathology and precision medicine. The rise of terms like prognosis and training reflects a growing focus on predictive models and personalized care, showcasing AI’s increasing role in clinical decision support. The colour gradient from blue to red indicates a temporal shift from foundational image-analysis terms toward more recent themes related to segmentation, transformers, biomarker inference, prognosis, and prediction-oriented modelling.
The keyword burst analysis (Figure 10) reveals key temporal shifts in computational pathology, reflecting both technological advancements and their clinical applications. Classification (strength: 8.8065) saw its highest burst between 2015-2019, coinciding with the widespread adoption of AI for automated cancer diagnosis. From 2017 to 2020, terms such as convolutional neural networks (strength: 8.3187) and nuclei (strength: 7.1598) marked the transition to more sophisticated image-specific methods, particularly for tumor segmentation. More recently, computational modeling (strength: 4.7658), which surged in 2024-2025, highlights the growing shift toward predictive AI models for personalized medicine. Additionally, lymph node metastasis has shown sustained bursts since 2023, reflecting its increasing significance in cancer staging and prognosis. These burst patterns suggest a literature-level shift from foundational classification and segmentation topics toward prediction-oriented modelling, metastatic assessment, and prognosis-related research questions. Temporal distribution of keyword bursts. The top 25 keywords with the strongest citation bursts are shown across the study period from 2009 to 2025. Burst periods (red) mark intervals of accelerated citation activity, indicating shifts in research emphasis.
4. Discussion
4.1. General information
The longitudinal publication and citation patterns indicate that AI-driven computational pathology has entered a phase of rapid consolidation since 2019, moving from sporadic exploratory studies toward a larger and more methodologically organized literature. Early output between 2009 and 2014 remained limited, a pattern that is consistent with the early-stage availability of digital pathology infrastructure, annotated whole-slide image datasets, and mature visual-recognition architectures.38–40 The acceleration observed after 2015 coincided with the broader adoption of whole-slide imaging and the maturation of deep learning methods for large-scale analysis of digitized histopathological data.3,41 Importantly, the logistic model offered a more conservative description of the observed publication trajectory than polynomial extrapolation, capturing a possible attenuation in the rate of annual publication increase rather than unbounded growth. This pattern is more appropriately interpreted as a bibliometric signal of validation-constrained maturation, rather than as evidence of unlimited field expansion. As computational pathology moves beyond algorithmic proof-of-concept studies, further progress is likely to depend increasingly on external validation, data standardisation, annotation quality, workflow integration, and regulatory-grade evidence.42,43 These factors were not directly measured in the present bibliometric dataset and should therefore be regarded as contextual interpretations rather than direct bibliometric findings. Model-fit and network-structure metrics strengthened the quantitative basis for interpreting temporal, collaboration, and co-citation patterns, reducing reliance on visual inspection alone. However, these metrics characterize the internal organization of the retrieved literature and should not be interpreted as evidence of causal mechanisms, clinical effectiveness, workflow impact, or real-world implementation.
At the global level, the findings suggest that publication volume, citation visibility, and co-authorship embeddedness represent related but distinct dimensions of scholarly influence in AI-related pathology (Table 1 and Figure 4). The United States showed high publication output and citation visibility, consistent with a prominent contribution to method-development and validation-oriented literature. 44 China showed rapid growth in publication volume but comparatively weaker international co-authorship connectivity, whereas the United Kingdom occupied a comparatively intermediary position within the observed country-level co-authorship network despite lower publication output. This contrast illustrates a scale–connectivity distinction: publication volume captures the scale of research activity, whereas network embeddedness reflects the extent to which a research system is linked to cross-national collaboration and knowledge-exchange structures.45,46 This distinction is particularly relevant for computational pathology AI, because models intended for clinical translation require validation across heterogeneous scanners, staining protocols, tissue-processing procedures, disease populations, annotation practices, and institutional workflows.47,48 In this context, internationally embedded research systems may be better positioned to support multi-institutional validation ecosystems in which diverse datasets, technical expertise, annotation standards, and clinical perspectives intersect. Thus, the translational bottleneck for AI-related pathology may lie not only in producing more models, but also in building collaborative infrastructures for shared benchmarks, interoperable data governance, standardized annotation, and external validation.
The author-level, journal-level, and reference-level patterns further suggest that AI-related pathology is socially distributed but intellectually concentrated. Publication activity was dispersed across many contributors, whereas co-citation influence converged around a smaller set of methodological and validation-oriented works. The prominence of citation clusters anchored by pathology-specific machine learning frameworks and validation-oriented benchmark studies indicates that the field’s intellectual base is organized around reusable analytical methods, benchmark tasks, and evaluation norms.49,50 This concentration has a constructive function: shared methodological references support comparability, cumulative knowledge building, and communication across pathology, computer vision, oncology, and data science. At the same time, intellectual convergence may also carry risks. Repeated reliance on a limited set of datasets, model architectures, and retrospective performance metrics may encourage benchmark dependence and methodological narrowing if generalizability is not tested across tissue types, staining variability, scanner platforms, demographic groups, and clinical workflows.51–53 Journal and co-citation analyses further indicate that the field’s publication outlets and intellectual reference base are partially separated. Methodologically oriented outlets, particularly Medical Image Analysis and IEEE Transactions on Medical Imaging, accounted for a substantial share of publications, whereas high-impact biomedical journals, including Nature Medicine, Nature Communications, and Nature, occupied prominent positions in the co-citation structure. Together, these findings suggest that AI-related pathology has developed at the interface of computational imaging, pathology, and clinically oriented biomedical research, while remaining a citation-based pattern rather than evidence of clinical implementation. Foundational contributions, notably the task-defining survey by Litjens et al. and the multicentre validation study by Bejnordi et al., continue to structure citation behaviour, indicating that intellectual influence in this field accrues primarily to works that establish shared analytical standards and validation-oriented evaluation frameworks.2,54 Taken together, the author, journal, and reference-level patterns suggest that computational pathology has developed as a method-intensive literature in which recurring methodological and validation-oriented works provide shared reference points across distributed research communities. These citation patterns help characterize the field’s intellectual organization, but they do not establish whether methodological convergence has translated into clinical deployment or institutional implementation.
4.2. Hotspots and frontiers
The keyword co-occurrence and burst analyses suggest that the frontier of AI-related computational pathology is not a simple linear progression from algorithm development to clinical use, but a layered research agenda organized around image-scale computation, cancer-centred applications, and prediction-oriented modelling. High-frequency terms such as “computational pathology,” the indexed keyword “whole slide image,” and “histopathology,” together with method-oriented terms including “convolutional neural network,” “image segmentation,” and “nuclei segmentation,” indicate that the field remains anchored in the conversion of digitized tissue morphology into computable representations. This image-analysis foundation is important because segmentation, classification, and feature extraction provide the methodological basis on which later biomarker-, prognosis-, and modelling-oriented studies are built.55–57 At the same time, the prominence of “breast cancer,” “colon cancer,” and “cancer” indicates a strong oncology concentration within the retrieved literature. Such concentration may have facilitated the development of task-specific models for tumour detection, grading, metastatic assessment, and tumour-microenvironment analysis; however, it also defines an important boundary condition. Hotspot patterns derived primarily from cancer-rich datasets may not generalize directly to non-neoplastic, inflammatory, infectious, renal, dermatopathological, or cytopathological settings.58–60
The co-occurrence structure further suggests that AI-related pathology research is moving from isolated image-analysis tasks toward a more biologically informed and prediction-oriented research agenda. The clustering of terms such as “biomarker,” “tumor-infiltrating lymphocytes,” “immunohistochemistry,” “prognosis,” and “computational modelling” indicates increasing literature-level interest in linking image-derived morphology with molecular, immune, and outcome-related information. This shift is scientifically important because it reframes computational pathology from a predominantly image-classification field toward a research space in which quantitative morphology may be connected with biological interpretation and prognostic inference.42,43 However, these bibliometric signals should not be equated with clinical readiness. Biomarker inference and prognostic modelling require assay harmonization, calibration, external validation, prospective evaluation, and workflow-compatible reporting before they can support patient management or diagnostic decision-making.61,62 Thus, the keyword frontier exposes a gap between the expanding ambition of the research agenda and the level of evidence required for reliable implementation in routine pathology environments.
The temporal burst results identify two interacting frontiers. Earlier bursts around “classification,” “convolutional neural networks,” and “nuclei” are consistent with the consolidation of deep learning-based image recognition and segmentation. More recent bursts around “computational modelling” and “lymph node metastasis” suggest increasing literature-level attention to higher-order prediction, metastatic assessment, and clinically oriented endpoints. In parallel, the emergence of “transformer” and “feature extraction” points to a methodological shift toward scalable representation learning and data-efficient modelling.63–65 This development may broaden the types of information extractable from whole-slide images, but it also intensifies familiar digital-health challenges, including model opacity, dataset shift, reproducibility, generalizability across scanners and laboratories, and the need for transparent evaluation standards.47,66,67 Therefore, the main implication of the hotspot analysis is not that AI has already transformed pathology practice, but that the literature is converging on the problems that must be solved for translation: robust representation learning, biologically meaningful prediction, external validation, and workflow-compatible evidence generation.
4.3. Bibliometric signals of precision-medicine-oriented computational pathology
The present findings do not demonstrate that AI has already reshaped pathology practice; rather, they indicate that AI-related computational pathology is increasingly framed around diagnostic workflow support, prognostic modelling, biomarker inference, multimodal interpretation, and precision-medicine-oriented research questions. Across citation centrality, co-citation clusters, burst detection, and keyword evolution, the field appears to be moving from isolated image-classification tasks toward research questions that connect quantitative morphology with risk stratification, biological interpretation, and outcome-oriented modelling. These patterns should therefore be interpreted as bibliometric signals of research orientation, not as evidence that AI systems have already generated validated treatment recommendations, routine personalized care, or real-world clinical implementation.
Centrality analysis helps identify references that connect otherwise separated research streams. Veta et al. appeared as a high-centrality reference linking clusters related to image analysis, image adequacy, and AI-based diagnostic research 68 Within Cluster #8 (uncertainty-aware AI), this position suggests literature-level interest in connecting image-quality assessment, uncertainty modelling, and diagnostic workflow support. Similarly, Sirinukunwattana et al. was linked to segmentation-oriented histopathology research, illustrating how nuclear segmentation methods became connected to later diagnostic and workflow-oriented research themes. 69 These bridging references show how foundational image-analysis methods were incorporated into later discussions of diagnostic reliability, image quality, and workflow-oriented evaluation. In this context, centrality should not be interpreted as clinical impact; rather, it identifies references that connect methodological and application-oriented citation streams.
Cluster-dependence patterns further suggest that diagnostic, prognostic, and biomarker-oriented themes are increasingly connected within the citation network. Cluster #1 (AI prognostication) and Cluster #11 (biomarker inference) were linked within the co-citation structure, consistent with a literature-level association between prognostic modelling and biomarker-oriented research [ ]. These citation links are best understood as signals that diagnostic AI research is broadening toward prognostic modelling and biomarker inference, rather than as evidence of treatment recommendation, personalized care, or routine deployment. Citation burst analysis also identified temporally concentrated attention to both methodological foundations and emerging application-oriented themes. For example, Janowczyk et al. showed a strong citation burst, reflecting concentrated literature attention to early deep learning approaches in computational pathology. 70 LeCun et al. remained an important foundational reference, reflecting the methodological influence of deep learning architectures on subsequent computational pathology research. 41 These bursts suggest that foundational deep learning methods continue to shape the field, while recent attention has shifted toward more scalable, data-efficient, and application-oriented modelling strategies.
Multimodal integration emerged as a recurring precision-medicine-oriented theme in the retrieved literature. Clusters such as computational phenotyping (#7) and biomarker inference (#11) point to increasing interest in linking pathology images with genomic, molecular, immune, and clinical variables. References associated with generalizable AI (#3), including work by Madabhushi and Kumar, further emphasize the need for models that can be evaluated across diverse cohorts, scanner platforms, staining protocols, tissue-processing procedures, and institutional settings.36,71,72 AI-based prognostication, represented by studies such as Yu et al. and Chen et al., constitutes another prominent research frontier.73,74 These studies illustrate efforts to integrate histopathological imaging with genomic, molecular, or clinical variables for outcome-oriented modelling. Nevertheless, prognostic and biomarker-oriented AI models require calibration, external validation, assay harmonization, clinically meaningful endpoints, and prospective evaluation before they can support clinical decision-making.
Workflow optimization (#12) and annotation-efficient AI (#9) indicate a parallel research focus on reducing practical barriers to model development and evaluation. Segmentation and survey-based methodological works by Sirinukunwattana et al. and Litjens et al. are frequently cited in relation to approaches that reduce annotation burden and standardize image-analysis tasks.2,69 However, whether such approaches can support routine clinical use depends on transparent reporting, workflow integration, data governance, and prospective validation in real diagnostic environments.
Taken together, the citation networks, burst patterns, cluster relationships, and keyword evolution indicate that AI-related computational pathology is increasingly oriented toward precision-medicine-relevant research questions. This orientation is most evident in the movement from image-classification-centred studies toward prognostic modelling, biomarker inference, multimodal integration, generalizable AI, annotation-efficient methods, and workflow-supportive analysis. However, these patterns remain signals of scholarly development rather than direct evidence of clinical implementation. The central challenge for the field is therefore not only to design more complex algorithms, but to generate externally validated, interpretable, calibrated, and workflow-compatible evidence across diverse pathology settings. In this sense, the bibliometric trajectory clarifies the evidentiary requirements for responsible translation, rather than demonstrating that AI has already transformed diagnostic practice, therapeutic decision-making, or personalized treatment.
5. Limitations
This study used scientometric methods to examine AI-driven computational pathology research from 2009 to 2025. Several limitations should be acknowledged. First, the dataset was retrieved exclusively from the Web of Science Core Collection/Science Citation Index Expanded (WoSCC/SCIE). This single-source strategy was adopted because the present study required a consistent cited-reference corpus, standardized citation records, and comparable affiliation metadata for co-citation analysis, burst detection, and collaboration-network mapping. Although Scopus, PubMed, and Embase provide valuable complementary coverage, cross-database integration may introduce duplicate records, non-equivalent citation metrics, inconsistent cited-reference formats, and heterogeneous metadata fields, which may affect citation-network analyses. Therefore, the findings should be interpreted as a WoSCC/SCIE-based citation-network map rather than an exhaustive inventory of all AI-related pathology publications. Second, citation-based indicators may be affected by self-citation, citation circles, publication bias, and time-lag effects. In addition, although model-fit, network-structure, and cluster-quality metrics were reported to improve analytical transparency and reproducibility, these indicators characterize the internal structure of the retrieved bibliometric corpus. They should not be interpreted as evidence of causal relationships, clinical effectiveness, workflow impact, or real-world implementation. Third, the restriction to English-language records and the coverage characteristics of WoSCC/SCIE may have led to underrepresentation of non-English or regionally indexed research.
6. Conclusion
This bibliometric analysis mapped the temporal growth, collaboration patterns, intellectual structure, and emerging research themes of AI-related computational pathology from 2009 to 2025. The findings indicate a rapidly expanding and increasingly interdisciplinary literature, with major research attention centred on analysis of whole-slide images, deep learning, segmentation, biomarker inference, prognostic modelling, and multimodal integration. Model-fit and network-structure metrics supported the internal consistency of the observed bibliometric patterns. However, these findings should be interpreted as literature-level signals rather than evidence of clinical effectiveness or real-world implementation. Future research should prioritize externally validated, interpretable, workflow-compatible, and ethically governed AI systems across diverse pathology settings.
Supplemental material
Supplemental material - Mapping the evolving landscape of artificial intelligence in pathology: A bibliometric analysis of research trends and emerging frontiers (2009-2025)
Supplemental material for Mapping the evolving landscape of artificial intelligence in pathology: A bibliometric analysis of research trends and emerging frontiers (2009-2025) by Ke Chai, Kun Wang, Canbin Chen, Shan Zeng, Jinming Zhang, Xiaguang Duan in DIGITAL HEALTH
Footnotes
Author contributions
(I) Ke Chai: Conceptualization, Methodology, Investigation, Writing – Original Draft Preparation, Writing – Review & Editing. (II) Kun Wang: Conceptualization, Formal Analysis, Investigation, Writing – Review & Editing (equal). (III) Canbin Chen: Investigation, Data Curation, Writing – Original Draft Preparation. (IV) Shan Zeng: Investigation, Data Curation, Writing – Review & Editing. (V) Jinming Zhang: Supervision, Validation, Writing – Review & Editing, corresponding author. (VII) Xiaguang Duan: Supervision, Funding Acquisition, Writing – Review & Editing, corresponding author.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by Science and Technology Program of the Joint Fund of Scientific Research for the Public Hospitals of Inner Mongolia Academy of Medical Sciences (Grant No. 2024GLLH0632), China Society for Metals, Metallurgical Safety and Health Branch, Health Research Project (Grant No. jkws202433), Aerospace Medical and Health Technology Group Co., Ltd. Research Project (Grant Nos. 2024YK10 and 2025YK17), Natural Science Foundation of Inner Mongolia Autonomous Region (Grant Nos. 2024MS08058 and 2025QN08075), and Inner Mongolia Medical University Joint Project (Grant No. YKD2024LH011).
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding authors.
Guarantor
Xiaguang Duan.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
