Generative and Foundation-Based Artificial Intelligence in Medical Imaging: A Bibliometric Analysis of Global and United Kingdom Research, 2017-2025 Authors: Jatin Naidu, Sonia Naidu, Vasanth Baskaradoss Published: 2026 ABSTRACT Artificial intelligence (AI), including generative and foundation-based methods, has rapidly expanded within medical imaging research, but the structure, citation impact, collaboration patterns, and thematic orientation of the United Kingdom national research ecosystem remain incompletely characterised. This bibliometric analysis examined Scopus-indexed publications from 2017 to 2025 using a predefined search strategy targeting generative and foundation-based AI methods in medical imaging. Records were analysed globally and filtered for United Kingdom affiliation. Descriptive indicators, including total publications, total citations, citations-per-paper, and year-on-year growth, were calculated. Co-authorship and keyword co-occurrence networks were generated using VOSviewer version 1.6.19. A total of 13,452 publications were identified globally, receiving 194,650 citations and a global citations-per-paper value of 14.47. Of these, 889 publications were United Kingdom-affiliated, representing 6.6% of global output. The United Kingdom ranked fourth by publication volume yet demonstrated higher unadjusted citations per paper than several higher-volume countries, with a value of 21.00. United Kingdom output increased approximately 18-fold between 2017 and 2025, with evidence of a citation-lag effect in recent years. The leading UK institutions had the highest full-count publication totals, although institution-level counts were not mutually exclusive because individual publications could be attributed to multiple organisations. Countries with more journal-dominant dissemination profiles also had higher unadjusted citations-per-paper values, although this descriptive comparison could not determine whether document-type composition explained the differences. Keyword analysis identified three principal thematic clusters: generative and deep learning methodologies, MRI- and diffusion-focused applications, and broader diagnostic imaging workflows. More recent highly cited publications included diffusion- and foundation-model architectures, although citation-age differences limit temporal interpretation. United Kingdom-affiliated research represents a rapidly expanding and highly cited component of the global generative and foundation-based AI in medical imaging literature. These findings provide a transparent bibliometric reference point for monitoring research activity, collaboration patterns, and potential translational priorities while recognising that citation-based indicators do not directly measure clinical implementation, methodological quality, or real-world impact. Introduction and background Artificial intelligence (AI) refers broadly to computational systems capable of performing tasks that have traditionally depended on human expertise, including pattern recognition, prediction, and decision support [1]. Advances in machine learning and deep learning have substantially expanded the capacity of AI systems to learn complex representations from large datasets, driving rapid research activity across many areas of healthcare. Within medical imaging, AI research has explored applications in image acquisition, reconstruction, segmentation, synthesis, triage, and interpretation, with the potential to support diagnostic efficiency and consistency [2]. These developments have positioned AI as a central area of investigation in radiology and other imaging-intensive specialities. In recent years, generative AI has emerged as a particularly influential class of methods within medical imaging. Generative models, including generative adversarial networks (GANs), diffusion models, and variational autoencoders, are designed primarily to generate or transform data, such as synthesising images, reducing noise, or translating between imaging modalities. Foundation models, by contrast, are large-scale models pretrained on broad datasets and subsequently adapted to multiple downstream tasks; they may be generative, discriminative, or multimodal. Generative and foundation-based architectures have therefore enabled image synthesis and transformation while also supporting more generalisable applications across imaging tasks [3]. These methods have been applied to accelerated magnetic resonance imaging reconstruction, image enhancement, data augmentation, anomaly detection, multimodal image synthesis, image interpretation, and report generation [4,5]. Collectively, they extend beyond conventional task-specific discriminative modelling and support broader manipulation, representation, and interpretation of imaging data. The rapid growth of generative and foundation-based AI research has produced a large, heterogeneous body of literature spanning clinical journals, engineering venues, and conference proceedings [6]. As this field expands, it becomes increasingly important to understand how research activity is distributed, which countries and institutions contribute most substantially, how collaborative networks are structured, and which thematic areas dominate scholarly output. Such insights are essential for informing research strategy, funding allocation, and translational planning [7]. Bibliometric analysis provides a systematic and reproducible framework for quantitatively assessing research output, citation impact, collaboration patterns, and thematic evolution within large scientific corpora. By applying mathematical and statistical techniques to bibliographic metadata, bibliometric methods enable the identification of influential publications, emerging research topics, and structural relationships among authors, institutions, and countries [8]. Previous bibliometric studies have examined AI broadly across medicine and radiology. Chogtu et al. evaluated the bibliometric development of AI in radiology during 2018-2022 [9]. Lin et al. analysed the wider medical AI literature from 2019 to 2023, encompassing areas such as digital health, precision medicine, public health, and medical imaging [10], while Baalla and Jellouli mapped the broader medical imaging AI literature, with particular emphasis on hospital performance, governance, and regulation [11]. However, these analyses treated AI as a broad technological field and did not specifically delineate the rapidly emerging literature on generative AI and foundation-based methods, including GANs, diffusion models, large language models, vision-language models, and general-purpose foundation models. They also did not systematically compare UK research performance with the global literature as we do in the present study. The United Kingdom represents a particularly relevant context for such an analysis. The UK hosts a strong academic medical imaging community, maintains a nationally integrated healthcare system, and has articulated strategic priorities for responsible AI development and deployment in healthcare [12,13]. These features position the UK as both a major contributor to methodological innovation and a relevant context for translational research and evaluation. A global bibliometric analysis by Chogtu et al. identified England among the leading contributors to AI research in radiology but assessed its publication output only as part of an international country-level comparison and did not examine UK institutions, collaboration networks, citation performance, or thematic structure in detail [9]. More recently, a bibliometric study evaluated generative AI research in healthcare imaging in India [14]. However, to our knowledge, no study has provided a detailed characterisation of the UK research ecosystem or directly examined its position within the global literature on generative and foundation-based AI in medical imaging. Therefore, this study undertook a comprehensive global bibliometric analysis alongside a focused assessment of UK-affiliated research, including publication output, citation impact, institutional productivity, collaboration patterns, and thematic development. Accordingly, the present study aims to conduct a bibliometric analysis of United Kingdom-affiliated research in generative and foundation-based AI in medical imaging between 2017 and 2025. The year 2017 was selected to capture the early expansion of modern generative approaches in medical imaging, particularly GAN-based methods, while allowing the subsequent emergence of diffusion models, large language models, and foundation-model architectures to be examined within the same longitudinal period. The endpoint of 2025 represented the most recent complete calendar year available before the Scopus dataset was exported on 6 February 2026. By examining publication trends, citation impact, country and institutional contributions, collaboration networks, and keyword co-occurrence patterns, this study seeks to map the structure and evolution of this research domain and to contextualise the UK contribution within the global literature. These findings provide an empirical baseline to support future research planning and policy discussion while recognising that bibliometric indicators do not directly establish clinical translation or implementation. This article was previously posted to the medRxiv preprint server on June 29, 2026. Review Materials and methods Search Strategy The Scopus database was used to identify relevant publications because of its broad coverage of biomedical, engineering, computer science, and conference proceedings literature, which are important dissemination routes for AI research in medical imaging. Searches were conducted in the title, abstract, and keyword fields for publications indexed between 2017 and 2025. No restrictions were applied by language or document type. The search combined terms related to generative and foundation-based AI with those related to medical imaging. It was designed to capture the principal methodological categories represented during the study period, including GANs, diffusion models, foundation models, large language models, ChatGPT, and Gemini AI. The complete Boolean search strategy for our global search was: TITLE-ABS-KEY("generative AI" OR "generative artificial intelligence" OR "generative adversarial network*" OR GAN OR "diffusion model*" OR "foundation model*" OR "large language model*" OR ChatGPT OR "Gemini AI") AND TITLE-ABS-KEY("medical imag*" OR "diagnostic imaging" OR radiolog* OR CT OR MRI OR ultrasound OR "computed tomography" OR "magnetic resonance") The strategy used broad umbrella concepts and the dominant model families represented across the study period rather than an exhaustive list of architecture-specific terms. The terms “generative AI”, “foundation model*”, and “large language model*” were intended to capture the wider field, while GANs and diffusion models were included separately because of their established prominence in medical imaging. The UK dataset was generated by applying LIMIT-TO (AFFILCOUNTRY, "United Kingdom") as stated below: TITLE-ABS-KEY("generative AI" OR "generative artificial intelligence" OR "generative adversarial network*" OR GAN OR "diffusion model*" OR "foundation model*" OR "large language model*" OR ChatGPT OR "Gemini AI") AND TITLE-ABS-KEY("medical imag*" OR "diagnostic imaging" OR radiolog* OR CT OR MRI OR ultrasound OR "computed tomography" OR "magnetic resonance") AND (LIMIT-TO (AFFILCOUNTRY, "United Kingdom")) The search was executed, and the datasets were exported on 6 February 2026. As Scopus is continuously updated, subsequent execution of the same query may produce slightly different record counts. All analyses were based on the datasets exported on 6 February 2026. Eligibility and Screening Publications were eligible if they were indexed in Scopus between 2017 and 2025 and retrieved by the predefined search strategy. This required at least one AI term and at least one medical imaging term to appear in the Scopus title, abstract, or keyword fields. All Scopus-indexed publication languages and document types were eligible. For inclusion in the UK dataset, at least one author affiliation had to be indexed by Scopus as being located in the United Kingdom. First author or corresponding author status was not required. Non-English publications were retained where they met the search and affiliation criteria and contained sufficient indexed metadata for the relevant analysis. Eligibility was determined by the predefined Scopus query and, for the UK analysis, the affiliation country filter. No manual title or abstract screening or additional relevance exclusions were undertaken after retrieval. To assess the precision of the search strategy, a stratified random sample of 400 records was selected from the complete search results. Records were stratified by publication year into pre-2023, 2023, 2024, and 2025 groups, with proportional allocation reflecting the distribution of records across these periods. Each sampled record was manually reviewed and assigned to one of four mutually exclusive categories according to its principal methodological focus: generative AI in medical imaging, foundation-based AI in medical imaging, potentially relevant, or unrelated. Records were classified as potentially relevant when the title, abstract, or indexed keywords indicated a plausible application of generative or foundation-based AI to medical imaging, but the available bibliographic information was insufficient to confirm that this represented the principal methodological focus of the publication. Search precision was estimated using two definitions. The primary strict definition treated only records classified as generative AI or foundation-based AI in medical imaging as relevant. A secondary inclusive definition additionally treated potentially relevant records as relevant. Wilson 95% confidence intervals were calculated for each classification category and for both precision estimates. The validation exercise was used to assess the performance of the search strategy and did not result in record-level exclusions from the bibliometric datasets. Data Cleaning and Normalisation Bibliographic metadata were exported from Scopus, including publication year, document type, author names and identifiers, affiliations, citation counts, funding information, and indexed keywords. Exported records were checked for the completeness of variables required for each planned analysis. Because Scopus was the sole bibliographic database, cross-database duplicate removal was not required. Potential duplicate records within the exported datasets were checked using Scopus record identifiers, DOIs, titles, and publication years. Records with incomplete metadata were excluded only from analyses requiring the missing variable and were retained for all other applicable analyses. Funding-body names and counts were retained as indexed by Scopus and were not hierarchically consolidated. Consequently, UK Research and Innovation and its constituent councils, including the Engineering and Physical Sciences Research Council and Medical Research Council, could appear as separate entities, and their counts were not combined. Institutional affiliations were generally retained as indexed by Scopus. However, directly identifiable hierarchical duplicates within the reported productivity ranking, such as the University of Oxford and University of Oxford Medical Sciences Division, were consolidated under the parent institution to avoid presenting the same institutional structure as separate ranked entities. No comprehensive institution-wide normalisation was otherwise undertaken. Organisations with incomplete or unresolvable affiliation metadata were not assigned to a separate institutional category. Generic indexing terms, including “article”, “human”, and “humans”, were removed before keyword-network analysis. No formal VOSviewer thesaurus file was applied. Synonymous terms, such as “MRI” and “magnetic resonance imaging,” were therefore retained as separate terms when indexed separately by Scopus. Author identities were determined using Scopus author identifiers and indexed author names. No additional manual verification or disambiguation using ORCID records, institutional profiles, or personal webpages was undertaken. Bibliometric Indicators Descriptive bibliometric analyses were performed using Microsoft Excel (Microsoft Corporation, Redmond, WA, USA). The principal indicators were total publications, total citations, citations per paper, and annual publication growth. Citations per paper were calculated as total citations divided by total publications. Year-on-year growth in UK publications was calculated as: \[ \left[\frac{\mathrm{UK}\ TP_{n}-\mathrm{UK}\ TP_{n-1}}{\mathrm{UK}\ TP_{n-1}}\right]\times 100 \] where UK TPₙ represents the number of UK-affiliated publications in the current year and UK TPₙ₋₁ represents the number in the preceding year. Historical annual citation counts for highly cited publications were obtained using the Scopus analyze search results function by sequentially updating the date filters for each relevant year. Citation counts used in the highly cited publication analysis represented citations accrued through 31 December 2025. The approximate annual citation rate was calculated by dividing total citations by the number of calendar years from the publication year through 2025 inclusive (2025−publication year+1). This year-based calculation did not account for the exact publication month or date and was therefore interpreted as an approximate rather than exact annualised rate. For publications with a complete five-year citation window, mean annual citations during the first five calendar years were calculated by summing citations accrued in the publication year and the subsequent four calendar years and dividing by five. Full counting was used for country- and institution-level publication indicators. Each publication was credited to every contributing country and every Scopus-indexed institutional affiliation represented in its author affiliations. Multinational publications were therefore attributed to each unique participating country. As a result, country and institutional publication counts were non-mutually exclusive and could exceed the number of unique publications when summed. Country percentages could similarly sum to more than 100%. VOSviewer Network Analysis Keyword co-occurrence and author co-authorship analyses were performed using VOSviewer version 1.6.19, released on 31 October 2023 and developed by Nees Jan van Eck and Ludo Waltman at Leiden University’s Centre for Science and Technology Studies. Networks were constructed using association strength normalisation and full counting. Full counting was selected because the analysis aimed to describe participation and network visibility, with each publication credited fully to every contributing author or keyword. For both networks, a fixed seed of 1 was used for the advanced layout and clustering parameters; complete layout and clustering settings are provided in Supplementary Material 1. For author co-authorship mapping, documents with more than 25 authors were excluded. Authors were required to have contributed to at least three documents. Among authors meeting these criteria, the 50 authors with the highest total link strength were retained for network construction. For keyword co-occurrence analysis, a minimum occurrence threshold of 10 was applied. Where more than 40 keywords met this threshold, the 40 keywords with the highest number of occurrences were retained for network construction. VOSviewer clusters were generated algorithmically according to the strength of co-authorship or keyword co-occurrence relationships. Cluster themes were subsequently assigned qualitatively by reviewing the titles, abstracts, and keywords of publications associated with the most prominent authors or terms within each cluster. The resulting labels were interpretive summaries produced by the study authors and were not generated automatically by VOSviewer. The Scopus dataset and VOSviewer files supporting the analyses are available in the accompanying Zenodo repository: https://doi.org/10.5281/zenodo.18659245. Results Dataset Identification and Search Strategy Validation The predefined Scopus search identified 13,452 records published between 2017 and 2025, of which 889 had at least one United Kingdom affiliation (Figure 1). No duplicate records were identified. PRISMA style flow diagram of bibliometric record identification, search strategy validation, dataset construction, and analysis-specific metadata availability Blue indicates record identification; green indicates dataset construction and search strategy validation; orange indicates analysis-specific metadata availability; and purple indicates the final UK bibliometric dataset. Dashed borders denote metadata completeness assessments within individual analyses. The figure was created using Canva (Canva Pty Ltd, Sydney, New South Wales, Australia). n, number of records; PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-analyses; UK, United Kingdom. Among the 400 records included in the stratified validation sample, 360 were classified as definitively relevant, giving a strict precision of 90.0% (95% CI: 86.7%-92.6%). A further 36 records were classified as potentially relevant and four as unrelated. When potentially relevant records were included, the estimated relevance proportion was 99.0% (95% CI: 97.5%-99.6%). The validation exercise was not used to exclude records from either dataset. Overall Publication Trends Global publication output increased substantially across the study period, particularly after 2022 (Figure 2). The complete global dataset comprised 13,452 publications and received 194,650 citations, corresponding to an unadjusted citations-per-paper value of 14.47. Annual global publication trends in generative and foundation-based artificial intelligence applied to medical imaging (2017-2025) The line plot shows the number of Scopus-indexed publications per year. The chart was generated using Microsoft Excel for Microsoft 365, 2026 (Microsoft Corporation, Redmond, WA, USA), from bibliographic data exported from Scopus. Country-level publication and citation indicators are presented in Figure 3 and Table 1. China had the largest publication output, followed by the United States, India, and the United Kingdom. The United States received the highest total number of citations, whereas Canada had the highest unadjusted citations per paper among the nine countries presented. The United Kingdom ranked fourth by publication volume and had an unadjusted citations-per-paper value of 21.00. Country counts were derived using full counting and were therefore non-mutually exclusive. Country-wise distribution of global research output in generative and foundation-based artificial intelligence applied to medical imaging Bars represent the number of publications associated with each country under full counting. Multinational publications were attributed to every country represented in the author affiliations; country counts are therefore non-mutually exclusive. The chart was generated using Microsoft Excel for Microsoft 365, 2026 (Microsoft Corporation, Redmond, WA, USA), from bibliographic data exported from Scopus. Table 1Contribution of the top nine countries to global research output in medical imaging, with comparison of publication volume and citation impact metrics relative to the United KingdomMetrics include total publications (TP), total citations (TC), citations per paper (CPP), and percentage representation within the global dataset (TP%). TP% was calculated using the global total number of records (n=13,452) as the denominator. Under full counting, each multinational publication was attributed to every country represented in its author affiliations. TP and TP% values are therefore non-mutually exclusive and may sum to more than the number of unique publications or 100%. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. RankName of countryTPTCCPPTP (%)1China400852,13013.0029.82United States324973,31022.5624.23India148011,1557.5411.04United Kingdom88918,67221.006.65Germany75811,20814.785.66South Korea55910,61218.984.27Canada51314,40428.073.88Japan472599912.713.59Australia41910,57625.253.1 Contribution of the top nine countries to global research output in medical imaging, with comparison of publication volume and citation impact metrics relative to the United Kingdom Metrics include total publications (TP), total citations (TC), citations per paper (CPP), and percentage representation within the global dataset (TP%). TP% was calculated using the global total number of records (n=13,452) as the denominator. Under full counting, each multinational publication was attributed to every country represented in its author affiliations. TP and TP% values are therefore non-mutually exclusive and may sum to more than the number of unique publications or 100%. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. United Kingdom Publication and Citation Trends Annual UK-affiliated output increased from 17 publications in 2017 to 305 in 2025, representing an approximately 18-fold increase (Table 2). From 2019 onwards, the UK generally contributed approximately 6%-7% of annual global output. More than one-third of all UK-affiliated publications in the dataset were published in 2025. Table 2Annual trends in global and United Kingdom-affiliated research output in generative and foundation-based artificial intelligence in medical imaging (2017-2025)Global TP represents the annual number of Scopus-indexed records in the global dataset. UK TP represents the annual number of records with at least one United Kingdom-affiliated author. TP of global (%) was calculated using annual global TP as the denominator for each year. TP of UK (%) was calculated using the total UK-affiliated dataset as the denominator (n=889). UK TC represents total citations received by UK-affiliated publications in each year. CPP was calculated as TC divided by UK TP. YoY growth was calculated as (current-year UK TP - previous-year UK TP/previous-year UK TP) x 100. The 2017 YoY value was not applicable because it was the first year in the study period. Values are descriptive bibliometric indicators; no inferential statistical testing was performed.CPP, citations per paper; TC, total citations; TP, total publications; YoY, year-on-year.PeriodGlobal TPUK TPTP of global (%)TP of UK (%)UK TCUK CPPYoY growth (%)20171311713.01.9109864.59-2018256207.82.3194797.3517.72019525356.73.9150342.9475.02020753486.45.4247251.5037.120211011696.87.8297043.0443.820221138756.68.4171622.888.7202316901207.113.5346328.8660.0202430382006.622.5263813.1966.7202549103056.234.38652.8452.5Total13,452889--18,672-- Annual trends in global and United Kingdom-affiliated research output in generative and foundation-based artificial intelligence in medical imaging (2017-2025) Global TP represents the annual number of Scopus-indexed records in the global dataset. UK TP represents the annual number of records with at least one United Kingdom-affiliated author. TP of global (%) was calculated using annual global TP as the denominator for each year. TP of UK (%) was calculated using the total UK-affiliated dataset as the denominator (n=889). UK TC represents total citations received by UK-affiliated publications in each year. CPP was calculated as TC divided by UK TP. YoY growth was calculated as (current-year UK TP - previous-year UK TP/previous-year UK TP) x 100. The 2017 YoY value was not applicable because it was the first year in the study period. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. CPP, citations per paper; TC, total citations; TP, total publications; YoY, year-on-year. Unadjusted citations per paper were higher for earlier publication years and lower for recent years, consistent with the shorter citation window available to newer publications. Annual growth in UK publication output was variable but accelerated again from 2023 onwards. Document Types Journal articles constituted the majority of UK-affiliated publications, followed by conference papers and reviews (Figure 4). Comparative document-type distributions are presented in Supplementary Material 1. The United States had the highest proportion of journal articles among the four leading countries, while India had the highest proportion of conference papers. The UK profile included a majority of journal articles, alongside a substantial conference-paper component. Distribution of document types among United Kingdom-affiliated publications in generative and foundation-based artificial intelligence applied to medical imaging Proportions are shown for journal articles, conference papers, reviews, and other document types indexed in Scopus, with percentages calculated using the total UK-affiliated publication denominator (n=889). The chart was generated using Microsoft Excel for Microsoft 365, 2026 (Microsoft Corporation, Redmond, WA, USA), from bibliographic data exported from Scopus. Institutional Productivity The eight most productive UK organisations are presented in Table 3. Imperial College London had the largest publication output, followed by King’s College London, University College London, and the University of Oxford. The National Heart and Lung Institute had the highest unadjusted citations per paper among the organisations presented. Institution-level counts were based on full counting and were therefore non-mutually exclusive. Table 3The eight most productive United Kingdom organisations in generative and foundation-based artificial intelligence in medical imaging research, ranked by total publicationsTP% represents the proportion of UK-affiliated publications contributed by each institution, calculated using the total UK-affiliated dataset as the denominator (n=889). Values are descriptive bibliometric indicators; no inferential statistical testing was performed. Eight organisations are presented because incomplete or unresolvable affiliation metadata prevented reliable identification of the top 10 organisations.CPP, citations per paper; TC, total citations; TP, total publications.RankName of organisationTPTP (%)TCCPP1Imperial College London13114.7331025.272King’s College London11012.4346231.473University College London10111.4417141.304University of Oxford9710.9366637.795University of Cambridge758.4168922.526The University of Edinburgh384.3125433.007National Heart and Lung Institute374.2171346.308University of Leeds333.772121.85 The eight most productive United Kingdom organisations in generative and foundation-based artificial intelligence in medical imaging research, ranked by total publications TP% represents the proportion of UK-affiliated publications contributed by each institution, calculated using the total UK-affiliated dataset as the denominator (n=889). Values are descriptive bibliometric indicators; no inferential statistical testing was performed. Eight organisations are presented because incomplete or unresolvable affiliation metadata prevented reliable identification of the top 10 organisations. CPP, citations per paper; TC, total citations; TP, total publications. Author Productivity and Collaboration A total of 3942 authors contributed to the UK-affiliated dataset. Author-level productivity, citation impact, and collaboration metrics are presented in Tables 4, 5. Yang G had both the highest publication count and the highest total link strength among the listed authors, while Alexander DC had the highest unadjusted citations per paper. Table 4The 10 most productive authors in the United Kingdom-affiliated generative and foundation-based artificial intelligence in medical imaging research and their collaboration strengthTP% represents the proportion of UK-affiliated publications associated with each author, calculated using the total UK-affiliated dataset as the denominator (n=889). TLS was calculated in VOSviewer and represents the overall strength of co-authorship links for each author within the analysed network. Because publications may include multiple authors, author publication counts are not mutually exclusive, and percentages may sum to more than 100%. Values are descriptive bibliometric indicators; no inferential statistical testing was performed.CPP, citations per paper; TC, total citations; TLS, total link strength; TP, total publications.RankName of authorName of organisationTPTP (%)TCCPPTLS1Yang GImperial College London424.7196546.791542Schnabel JAKing’s College London161.8815.06573Frangi AFUniversity of Manchester141.6966.86524Rueckert DImperial College London141.6906.43245Ourselin SKing’s College London131.584765.15416Tsaftaris SAUniversity of Edinburgh131.549638.15287Alexander DCUniversity College London111.21114101.27158Botchu RThe Royal Orthopaedic Hospital NHS Foundation Trust111.220818.91259Huang JNational Heart and Lung Institute111.228025.457210Iyengar KPSouthport and Ormskirk Hospital NHS Trust111.220818.9125 The 10 most productive authors in the United Kingdom-affiliated generative and foundation-based artificial intelligence in medical imaging research and their collaboration strength TP% represents the proportion of UK-affiliated publications associated with each author, calculated using the total UK-affiliated dataset as the denominator (n=889). TLS was calculated in VOSviewer and represents the overall strength of co-authorship links for each author within the analysed network. Because publications may include multiple authors, author publication counts are not mutually exclusive, and percentages may sum to more than 100%. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. CPP, citations per paper; TC, total citations; TLS, total link strength; TP, total publications. Table 5Top 10 most connected authors in the co-authorship network based on collaboration strength in generative and foundation-based artificial intelligence in medical imaging researchAuthors are ranked by TLS, derived from VOSviewer co-authorship network analysis. TP represents the absolute number of publications associated with each author in the Scopus-exported UK-affiliated dataset. TLS represents the overall strength of co-authorship links for each author within the analysed network. Because publications may include multiple authors, author-level publication counts are not mutually exclusive. Values are descriptive bibliometric indicators; no inferential statistical testing was performed.TLS, total link strength; TP, total publications.RankAuthorTPTLS1Yang G421542Wang C7793Huang J11724Zhang Y8625Schnabel JA16576Pinaya WHL9547Frangi AF14528Zhang J13519Hu Y84810Cardoso MJ845 Top 10 most connected authors in the co-authorship network based on collaboration strength in generative and foundation-based artificial intelligence in medical imaging research Authors are ranked by TLS, derived from VOSviewer co-authorship network analysis. TP represents the absolute number of publications associated with each author in the Scopus-exported UK-affiliated dataset. TLS represents the overall strength of co-authorship links for each author within the analysed network. Because publications may include multiple authors, author-level publication counts are not mutually exclusive. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. TLS, total link strength; TP, total publications. For the co-authorship analysis, 327 authors met the minimum threshold of three publications (Supplementary Material 2), and the 50 authors with the highest total link strength were visualised (Figure 5). The network comprised five clusters, including one smaller component that was relatively separate from the principal interconnected network. Co-authorship network of authors publishing on generative and foundation-based artificial intelligence in medical imaging, with emphasis on generative and foundation-based methods Node size represents publication volume, link thickness indicates co-authorship strength, and colours denote collaboration clusters identified through VOSviewer network analysis. Authors positioned closer together have stronger co-authorship relationships within the analysed dataset. The network shows a large interconnected component on the left, containing several overlapping collaboration clusters, and a smaller separate component on the right, indicating a more distinct co-authorship group with limited linkage to the main network. Only authors meeting the predefined inclusion threshold were visualised. The network was generated using full counting in VOSviewer version 1.6.19, released in 2023. External Funding Of the 889 UK-affiliated publications, 552 (62.1%) reported at least one external funding source. A total of 160 funder-name entries were identified as indexed in Scopus. Because funder names were not normalised and umbrella organisations were retained alongside constituent councils, these entries should not be interpreted as 160 unique funding organisations. The Engineering and Physical Sciences Research Council was the most frequently acknowledged funder, followed by the National Natural Science Foundation of China, UK Research and Innovation, the Horizon 2020 Framework Programme, the Wellcome Trust, and the Medical Research Council (Figure 6). Publications could acknowledge more than one funder. Leading funding organisations acknowledged in United Kingdom-affiliated publications on generative and foundation-based artificial intelligence in medical imaging Bars represent the number of Scopus-indexed documents reporting financial support from each funding body. The chart was generated using Microsoft Excel for Microsoft 365, 2026 (Microsoft Corporation, Redmond, WA, USA), from bibliographic data exported from Scopus. Leading Publishing Sources The UK-affiliated publications were distributed across 160 indexed sources. The 20 most productive sources accounted for 337 publications, or 37.9% of the UK dataset. Lecture Notes in Computer Science had the largest publication output, while NeuroImage and IEEE Transactions on Medical Imaging had among the highest unadjusted citations-per-paper values. Keyword Analysis A total of 7685 keywords were identified. After application of the minimum occurrence threshold, 332 keywords met the threshold (Supplementary Material 3). After exclusion of selected generic indexing terms (Supplementary Material 4) and ranking by frequency, 40 keywords were included in the network analysis (Supplementary Material 5). The most frequent terms were “deep learning”, “medical imaging”, “diagnostic imaging”, and “GANs.” The co-occurrence network comprised three principal thematic clusters (Figure 7). The blue cluster was centred on generative and deep-learning methods, including GANs, medical imaging, reconstruction, enhancement, and segmentation. The red cluster primarily represented MRI, diffusion-related imaging, neuroimaging, image quality, and diagnostic imaging. The green cluster included broader AI, machine learning, radiology, image analysis, ChatGPT, large language models, and diagnostic-evaluation terms. Keyword co-occurrence network of United Kingdom-affiliated publications in generative and foundation-based artificial intelligence applied to medical imaging Node size reflects keyword frequency, link thickness indicates co-occurrence strength, and node colour denotes thematic clusters. Keywords positioned closer together co-occurred more frequently within the same publications. The blue cluster primarily represents generative and deep learning methods applied to medical imaging, including generative adversarial networks, deep learning, image reconstruction, image enhancement, and image segmentation. The red cluster is centred on magnetic resonance imaging and diffusion-related applications, including diffusion, diffusion-weighted imaging, neuroimaging, image quality, algorithms, and diagnostic imaging. The green cluster represents broader artificial intelligence and diagnostic evaluation themes, including artificial intelligence, machine learning, convolutional neural networks, image analysis, diagnostic accuracy, sensitivity and specificity, major clinical studies, and ChatGPT. Highly Cited Publications The 20 most highly cited UK-affiliated publications are presented in Table 6. Together, they received 5389 citations. The most highly cited publication was the 2018 DAGAN study by Yang et al, followed by studies of MedGAN, retinal foundation models, NiftyNet, and GANs for medical image analysis. Table 6The 20 most highly cited United Kingdom-affiliated publications in generative and foundation-based artificial intelligence applied to medical imaging, with emphasis on generative and foundation-based methodsPublications are ranked by total citation count in the Scopus-exported dataset. Citation counts represent citations accrued through 31 December 2025, as recorded in Scopus. The approximate annual citation rate was calculated by dividing total citations by the number of calendar years from the publication year through 2025 inclusive (2025−publication year+1). This year-based calculation did not account for the exact publication month or date and should therefore be interpreted as an approximate rather than exact annualised rate. Mean annual citations during the first five calendar years were calculated by summing citations accrued in the publication year and the subsequent four calendar years and dividing by five. N/A indicates that a complete five-calendar-year citation window was unavailable by the end of 2025. Reference numbers in the title column correspond to the full reference list. Values are descriptive bibliometric indicators; no inferential statistical testing was performed.RankCitationsAverage citations per yearAverage citations per year in first five yearsAuthor citationYearTitle1923115.38138.6Yang et al [15]2018DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction 250884.6797Armanious et al [16]2020MedGAN: medical image translation using GANs3492164N/AZhou et al [17]2023A foundation model for generalizable disease detection from retinal images 447859.7558Gibson et al [18]2018NiftyNet: a deep-learning platform for medical imaging 536460.6771.4Kazeminia et al [19]2020GANs for medical image analysis 6297148.5N/AHuang et al [20]2024Segment anything model for medical images?727167.75N/AWyatt et al [21]2022AnoDDPM: Anomaly Detection with Denoising Diffusion Probabilistic Models using Simplex Noise821854.5N/APinaya et al [22]2022Brain imaging generation with latent diffusion models 920228.8636.2Han et al [23]2019Combining noise-to-image and image-to-image GANs: brain MR image augmentation for tumour detection 1019364.33N/AJiang et al [24]2023Deep Learning for Medical Image-Based Cancer Diagnosis1118326.1430.4Ben-Cohen et al [25]2019Cross-modality synthesis from CT to PET using FCN and GAN networks for improved automated lesion detection121703434Han et al [26]2021MADGAN: unsupervised medical anomaly detection GAN using multiple adjacent brain MRI slice reconstruction 1316518.3322Chartsias et al [27]2017Adversarial image synthesis for unpaired multi-modal cardiac data 1416332.632.6Ma et al [28]2021Structure and illumination constrained GAN for medical image enhancement 151442428.8Dar et al [29]2020Prior-guided image reconstruction for accelerated multi-contrast MRI via generative adversarial networks1613869N/AWu et al [30]2024MedSegDiff-V2: diffusion-based medical image segmentation with transformer1712517.8622.4Rau et al [31]2019Implicit domain adaptation with conditional generative adversarial networks for depth prediction in endoscopy 1812124.224.2Karakanis et al [32]2020Lightweight deep learning models for detecting COVID-19 from chest X-ray images 1911859N/ADayarathna et al [33]2024Deep learning-based synthesis of MRI, CT and PET: Review and analysis2011616.5719Han et al [34]2019Synthesizing diverse lung nodules wherever massively: 3D multi-conditional GAN-based CT image augmentation for object detection The 20 most highly cited United Kingdom-affiliated publications in generative and foundation-based artificial intelligence applied to medical imaging, with emphasis on generative and foundation-based methods Publications are ranked by total citation count in the Scopus-exported dataset. Citation counts represent citations accrued through 31 December 2025, as recorded in Scopus. The approximate annual citation rate was calculated by dividing total citations by the number of calendar years from the publication year through 2025 inclusive (2025−publication year+1). This year-based calculation did not account for the exact publication month or date and should therefore be interpreted as an approximate rather than exact annualised rate. Mean annual citations during the first five calendar years were calculated by summing citations accrued in the publication year and the subsequent four calendar years and dividing by five. N/A indicates that a complete five-calendar-year citation window was unavailable by the end of 2025. Reference numbers in the title column correspond to the full reference list. Values are descriptive bibliometric indicators; no inferential statistical testing was performed. Earlier highly cited publications were predominantly concerned with GAN-based reconstruction, synthesis, augmentation, and image analysis, whereas the more recent highly cited set included diffusion and foundation-model architectures. Because citation rankings are strongly influenced by publication age, this descriptive pattern should not be interpreted as definitive evidence of a thematic transition. Discussion Principal Findings This bibliometric analysis characterised global and United Kingdom-affiliated research on generative and foundation-based AI in medical imaging between 2017 and 2025. The findings position the UK as a major contributor to the field, with comparatively high unadjusted citation impact, strong participation from research-intensive institutions, extensive co-authorship, and a thematic profile spanning generative methods, diffusion-related imaging, and broader diagnostic applications. The United Kingdom ranked fourth by full-count publication volume and had an unadjusted citations-per-paper value of 21.00. This suggests that UK-affiliated work achieved substantial visibility relative to several higher-output countries. However, such comparisons should be interpreted cautiously because raw citation indicators are affected by publication age, document type, disciplinary composition, collaboration patterns, and highly cited outliers. The institutional and author-level findings indicate that UK activity is anchored in a group of research-intensive organisations and interconnected collaborative networks. This is consistent with the multidisciplinary nature of medical imaging AI, which commonly requires expertise in clinical imaging, computer science, engineering, statistics, and data governance. Differences in institutional citation impact further show that productivity and influence are related but distinct dimensions of research performance. The mixture of journal articles and conference papers reflects the dual clinical and technical character of the field. Journal publication remains important for clinical and translational dissemination, while conference proceedings continue to provide a major route for rapidly evolving methodological research. Cross-country differences in publication format may therefore reflect different disciplinary and dissemination structures rather than differences in research quality alone. The funding profile similarly highlights the interdisciplinary basis of UK-affiliated research. Support from engineering and computational funders alongside biomedical, charitable, European, and international organisations suggests that progress in this area depends on combined investment in methodological development, clinical evaluation, and translational infrastructure. The presence of both domestic and international funders also reflects the collaborative and cross-border nature of medical imaging AI research. Comparison With Previous Studies The sustained increase in publication volume mirrors broader bibliometric analyses showing rapid expansion of AI research in medicine and radiology since 2019 [10]. Generative and deep learning methods were prominent within this growth, reflected by frequent keywords such as “deep learning”, “medical imaging”, “diagnostic imaging”, and “GANs”, as well as by GAN-, diffusion-, and foundation model-related papers among highly cited UK-affiliated publications. This is consistent with a previous top-cited analysis in which deep learning was the most frequent keyword and neuroimaging and oncology were prominent clinical subjects [35]. A recent India-focused bibliometric analysis of generative AI in healthcare imaging similarly identified rapid national research growth, supporting the relevance of country-level bibliometric mapping [14]. The keyword and highly cited publication analyses also suggest thematic evolution over time. Earlier highly cited publications were predominantly concerned with GAN-based reconstruction, synthesis, augmentation, anomaly detection, and cross-modality translation, particularly in MRI [15,16,19,23,25-29]. More recent highly cited publications included foundation models, general-purpose segmentation approaches, diffusion-based models, and multimodal synthesis reviews [17,20-22,30,33]. However, because citation rankings are strongly influenced by publication age, this pattern should not be interpreted as definitive evidence of a temporal transition in the wider literature. Clinical and Translational Implications The prominence of MRI-related applications may reflect the suitability of MRI for reconstruction, denoising, synthesis, acceleration, and image-quality enhancement. These applications are clinically relevant because acquisition time, image quality, and workflow efficiency remain important practical challenges. However, bibliometric visibility reflects research attention rather than evidence of methodological quality, routine clinical adoption, or patient benefit. These outcomes require direct evaluation through clinical and implementation studies. Deep learning reconstruction has shown potential to reduce MRI acquisition times [36], while workforce shortages and rising imaging demand provide a relevant service context [37]. However, prospective evaluation and deployment guidance remain necessary before routine clinical implementation [38,39]. Implementation may be affected by data governance, interoperability, computational infrastructure, procurement, regulatory evaluation, model monitoring, and accountability [40-42]. Generative imaging models introduce concerns relating to artefact generation, image authenticity, and deepfakes [43,44]. Medically intended AI software may also fall within medical device regulation [45]. Standardised evaluation and transparent reporting are therefore important, with CONSORT-AI and TRIPOD+AI providing relevant guidance for applicable study designs [46,47]. The concentration of UK research within major academic centres and its dense collaboration networks may support multicentre validation and coordinated evaluation. The identified areas of research activity, including MRI reconstruction, image enhancement, segmentation, multimodal imaging, and foundation models, provide potential priorities for further study, although their clinical value must be assessed against patient need, safety, workflow integration, cost-effectiveness, and equity. UK Policy and Funding Implications The concentration of output within research-intensive institutions indicates an established UK academic base but also suggests potential value in broadening participation across NHS and university centres. Future funding could support multicentre datasets, cross-institutional validation, and collaboration among NHS imaging departments, universities, industry, regulators, and patient representatives. This is consistent with the AI Opportunities Action Plan, which emphasises infrastructure, data access, public sector adoption, and translation of AI capability into societal benefit [48]. The interdisciplinary funding profile suggests that UK medical imaging AI already spans engineering, computational, and biomedical research. UK Research and Innovation funding programmes for AI in biomedical and health research provide a relevant model through their emphasis on partnerships, translational impact, and reusable outputs [49]. Medical Research Council priorities concerning AI and related technologies further reflect the strategic importance of AI within biomedical research [50]. Secure imaging datasets, trusted research environments, computational capacity, interoperability, and clear governance pathways are likely to influence translation. NHS information-governance guidance emphasises lawful data use, transparency, accountability, and appropriate governance [40], while Medicines and Healthcare products Regulatory Agency (MHRA) guidance indicates that medically intended software and AI may fall within medical device regulation [45]. Funding and policy support should therefore balance development of novel models with validation, governance, monitoring, and evaluation across NHS settings. Limitations This study has several limitations inherent to bibliometric analysis. The analysis was restricted to publications indexed in Scopus and may not have captured relevant records indexed exclusively in other databases. Scopus was selected because of its broad coverage of engineering, computer science, biomedical, medical imaging, and conference proceedings literature, which represent central dissemination routes for medical imaging AI research. Although adding databases such as Web of Science or PubMed could increase record capture, cross-database merging introduces challenges, including metadata inconsistencies, duplicate records, and discordant citation counts. A single-database approach was therefore adopted to preserve internal consistency and reproducibility. Although 2025 represented a completed calendar year, indexing of publications released late in 2025 may still have been incomplete when the Scopus data were retrieved in February 2026. Consequently, publication and citation counts for 2025 may be underestimated. The findings were also dependent on the chosen search strategy. Although the strategy was designed to capture generative and foundation-based AI methods in medical imaging, terminology in this field is heterogeneous and rapidly evolving. The search was focused rather than exhaustive and relied on broad umbrella concepts and dominant model families. Consequently, some publications indexed exclusively under narrower terms, such as “variational autoencoder”, “VAE”, “vision-language model”, or “multimodal foundation model”, may not have been retrieved. Conversely, broad terms such as “GAN”, “ChatGPT”, and “Gemini AI” may have retrieved records outside the intended scope. Although manual review of a stratified random sample of 400 records indicated a strict estimated precision of 90.0%, not every retrieved record underwent individual relevance screening, and some false-positive or indeterminate records may therefore remain. Citation-based indicators were unadjusted for publication age, document type, research field, collaboration patterns, and highly cited outliers. Citations per paper may therefore be skewed and should be interpreted as a descriptive indicator rather than a normalised measure of research quality or comparative impact. Citations per year provided partial adjustment for publication age among highly cited publications but did not account for field or document-type differences. Citation counts may also be affected by journal visibility, article type, self-citation, field size, and collaboration structure. The study relied on bibliographic metadata and did not assess the technical performance, clinical implementation, regulatory status, ethical characteristics, or patient-level impact of individual AI models. Full rather than fractional counting may overrepresent contributions from multi-authored or multi-institutional publications. Affiliation-based filtering may also underrepresent UK researchers contributing to international collaborations where their UK affiliations were not indexed. Finally, keyword co-occurrence and co-authorship maps are sensitive to threshold selection, author name disambiguation, indexing quality, and VOSviewer settings. The absence of systematic synonym merging may have divided occurrences across equivalent terms, influenced keyword frequencies, and cluster composition. Conclusions This bibliometric analysis demonstrates rapid global growth in generative and foundation-based AI research applied to medical imaging between 2017 and 2025. UK-affiliated work represented a substantial component of the field and had a comparatively high unadjusted mean citation rate. The UK ranked fourth by publication volume and showed relatively high citation visibility, with research activity concentrated among major academic centres and supported by interdisciplinary funding. Keyword and highly cited publication analyses indicate prominent themes around generative and deep learning methods, MRI reconstruction and enhancement, diffusion-based approaches, and foundation-model architectures. These findings provide a transparent bibliometric reference point for monitoring the evolution of AI medical imaging research while recognising that citation-based indicators do not directly measure clinical implementation, methodological quality, or patient-level impact.