Bibliometric Analysis of Artificial Intelligence Use in Small Animal Thoracic Radiography Authors: Vildan Aslan Canatan, Ender Uzabaci Published: 2026 ABSTRACT ABSTRACTBackgroundArtificial intelligence (AI) is increasingly used in veterinary diagnostic imaging, particularly in thoracic radiology, to support image interpretation.ObjectiveThe aim of this study was to evaluate research trends in AI applications for small animal thoracic radiography through bibliometric analysis, identifying key contributors, leading journals, thematic focuses and collaboration patterns.MethodsStudies published between 2018 and 2025 were retrieved from the Web of Science (WoS) and PubMed databases using predefined Boolean search terms. After duplicate removal and screening based on inclusion and exclusion criteria, 27 articles were included. Data were analysed using RStudio (Biblioshiny) to evaluate publication trends, citation impact, scientific productivity, collaboration networks and thematic structures.ResultsPublications showed an increasing trend, with an annual growth rate of 10.41% and a peak in 2023 (n = 8). Studies published in 2020 showed the highest citation impact, averaging 46.33 citations. T. Banzato emerged as the most prolific author, while S. Burti authored the most cited study. The United States (n = 32) and Italy (n = 23) were the leading contributing countries. Thematic analyses demonstrated a primary focus on convolutional neural network (CNN)‐ and deep learning‐based automated detection of cardiomegaly, pleural effusion and pulmonary diseases in dogs, reflecting the clinically driven nature of early studies. A significant positive correlation (p < 0.001) was found between WoS citation counts and those from other databases, whereas other bibliometric indicators showed no significant associations.ConclusionsAI shows considerable potential to support diagnostic processes in veterinary thoracic radiology; however, dataset heterogeneity and species diversity remain major limitations, highlighting the need for methodological standardization and stronger international collaboration.Keywords: artificial intelligence, bibliometric analysis, cat, dog, thorax radiography Introduction Thoracic radiography is one of the fundamental diagnostic imaging methods in veterinary medicine, particularly in small animal practice, for evaluating cardiopulmonary and thoracic cavity pathologies (Prather et al. 2005; Tahghighi, Appleby, et al. 2023). This method, which provides rapid, accessible and cost‐effective imaging of thoracic structures in cats and dogs, plays a significant role in clinical decision‐making processes (Burti et al. 2020; Tahghighi, Norena, et al. 2023). However, the diagnostic effectiveness of the method depends on the accurate evaluation of the obtained radiographic images. The interpretation of radiographic images largely relies on observer experience, which can lead to limitations in diagnostic stability, inter‐observer variability and variability in diagnostic accuracy (Banzato, Wodzinski, Tauceri, et al. 2021; Tahghighi et al. 2024). With the digital transformation of imaging technologies, artificial intelligence (AI)‐based approaches, especially machine learning (ML) and deep learning (DL) algorithms, are increasingly being used in clinical imaging practice (Appleby and Parminder 2024). These technologies have the potential to increase diagnostic accuracy, standardize interpretation processes and reduce error margins in applications such as image segmentation, pattern recognition, classification and decision support systems (Pereira et al. 2023). Widely adopted in human medicine, these systems are also being adapted for veterinary diagnostic imaging in recent years; they are applied across modalities such as radiography, ultrasonography, computed tomography (CT) and magnetic resonance imaging (MRI) (Burti et al. 2024). Although AI applications in veterinary imaging have expanded across various modalities, recent studies have shown a noticeable increase in those specifically focused on thoracic radiography in small animals (Rubini et al. 2025). These studies often utilize DL approaches, particularly convolutional neural networks (CNNs), to enable a range of clinical tasks such as automated cardiac size measurement (Jeong and Sung 2022; Solomon et al. 2023), diagnosis of mitral valve disease (Valente et al. 2023, Valente et al. 2024), detection of pulmonary oedema (Kim et al. 2022), identification of lung lesions (Arsomngern et al. 2019), detection of pleural effusion (Müller et al. 2022), identification of pulmonary masses (Pomerantz et al. 2023), classification of pulmonary patterns (Yoon et al. 2019), interpretation of thoracic radiographs (Banzato et al. 2023; Celniak et al. 2023) and detection of thoracic pathologies (Dumortier et al. 2022). Moreover, some studies have compared AI systems with veterinary radiologists in interpreting thoracic radiographs (Adrien‐Maxence et al. 2022; Banzato, Wodzinski, Burti, et al. 2021). However, the literature lacks comprehensive studies that systematically reveal research trends, methodological diversity and the structure of international collaboration in this field. Bibliometrics is a method for quantitatively evaluating production trends, collaboration networks and thematic orientations in a particular scientific field by examining published academic works using statistical data (Serin and Körez 2025). The interaction between specific research elements (institutions, authors, journals, keywords, etc.) is examined to evaluate bibliometric indicators such as h‐index, g‐index, m‐index, citation count and publication count (Wang et al. 2020). It is used to discover patterns in the literature, summarize research focuses and identify development trends in the analysed disciplines. In this context, bibliometric analysis stands out as a comprehensive and robust method for objectively examining the field's development dynamics in AI‐based veterinary diagnostic imaging (Elasan and Yilmaz 2025). Although there are some bibliometric analyses of AI applications in veterinary medicine, these generally address general veterinary medicine, education or broad AI application areas (Elasan and Yilmaz 2025; Santana et al. 2025; Serin and Körez 2025). The lack of studies evaluating AI‐based research trends in small animal thoracic radiography using bibliometric methods indicates a gap in this field, both thematically and methodologically. This situation underscores the need for a systematic analysis of the dynamics of scientific production in the field of veterinary thoracic imaging. The aim of this study was to apply bibliometric analytical techniques to examine trends in scientific knowledge production on AI applications in small animal thoracic radiography. The scope of this study is to provide an overview of current research progress and to identify core areas of interest, highlight research topics, influential authors, leading academic journals, global research distribution and patterns of collaboration, thereby offering insights into regional challenges and knowledge transfer dynamics. Material and Methods Data Collection and Search Strategy To identify the current literature on AI in the field of veterinary medicine, the WoS and PubMed databases were utilized. These databases were selected due to their comprehensive indexing of high‐impact, peer‐reviewed journals in the fields of veterinary medicine, biomedical sciences and diagnostic imaging, as well as their inclusion of studies adhering to established publication ethics (Falagas et al. 2008; Tarazi 2024). The literature search was conducted on 22 October 2025, without initial restrictions regarding publication type or language. However, only studies published in English and meeting the predefined inclusion criteria were included in the final bibliometric analysis. The following Boolean operators were used in the search strategy: ((“small animal” OR “dog” OR “dogs” OR “canine” OR “cat” OR “cats” OR “feline”) AND (“thorax” OR “thoracic” OR “chest”) AND (“radiography” OR “x‐ray” OR “radiology” OR “veterinary diagnostic imaging” OR “diagnostic imaging” OR “thorax radiography” OR “radiographic parameters”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “CNN” OR “convolutional neural network” OR “neural network model” OR “LLM” OR “large language models” OR “HRNet” OR “high resolution network”)). This keyword combination was carefully designed to ensure comprehensive and accurate coverage of studies involving AI applications in thoracic radiography of cats and dogs. Inclusion and Exclusion Criteria The retrieved studies were evaluated based on the following inclusion and exclusion criteria: Inclusion criteria: Studies involving AI, ML or DL algorithms applied to thoracic radiography in small animals (cats and dogs). Studies involving AI, ML or DL algorithms applied to thoracic radiography in small animals (cats and dogs). Imaging studies utilizing AI to evaluate thoracic structures, including pathological conditions as well as technical and qualitative aspects of thoracic radiography in small animals. Imaging studies utilizing AI to evaluate thoracic structures, including pathological conditions as well as technical and qualitative aspects of thoracic radiography in small animals. AI research focused on intrathoracic organ pathologies (e.g., heart, lungs, diaphragm) in small animals. AI research focused on intrathoracic organ pathologies (e.g., heart, lungs, diaphragm) in small animals. Exclusion criteria: Studies focused on anatomical regions outside the thorax. Studies focused on anatomical regions outside the thorax. Irrelevant studies, including those involving animals other than cats or dogs, studies using imaging modalities other than radiography (e.g., ultrasound, CT, MRI) or those not incorporating any AI‐related methodology. Irrelevant studies, including those involving animals other than cats or dogs, studies using imaging modalities other than radiography (e.g., ultrasound, CT, MRI) or those not incorporating any AI‐related methodology. To ensure the reliability of the article selection process, a sequential consensus approach was adopted. In the first stage, the first researcher, an expert in the field, identified articles based on the predefined inclusion and exclusion criteria. In the second stage, the selected articles were independently re‐evaluated by the second researcher for methodological suitability. Any disagreements were resolved through discussion until consensus was reached, and the final list of articles included in the analysis was established accordingly. Duplicate records identified across databases were removed prior to the screening process. Irrelevant studies were excluded following title, abstract and full‐text evaluations. Any discrepancies between reviewers were resolved through discussion and consensus. As a result of this systematic screening process, a final dataset of 27 articles that fulfilled all inclusion criteria was compiled, forming the foundation for the subsequent bibliometric analysis (Figure 1). A PRISMA‐style flow diagram was used to illustrate the study selection and screening process. Flow diagram illustrating the data selection process for the study. Data Extraction and Analysis Bibliometric data, including the publication title, first author's name, corresponding author's country, publication year, journal name, citation count, keywords and references, were extracted from the included studies. Bibliometric analyses were performed using the Biblioshiny web interface of the bibliometrix package in R Studio (version 4.3.1) (Aria and Cuccurullo 2017). Network analyses were generated using the full counting method and association‐strength normalization. The inclusion thresholds applied for network construction were as follows: authors with at least two published documents for the co‐authorship network, countries with at least two publications for the country collaboration network, cited references with at least 20 citations for the co‐citation network and keywords occurring at least five times for the keyword co‐occurrence network. These thresholds were selected to improve the readability of the visualizations while retaining the most influential nodes. Country productivity was calculated using the full counting method based on the institutional affiliations of authors. Publications involving authors from multiple countries were counted once for each represented country. Consequently, country counts may exceed the total number of publications included in the analysis. As part of the bibliometric analysis, author, country and institutional collaboration networks were assessed using co‐authorship, co‐citation and keyword co‐occurrence techniques. These analyses aimed to visualize the structural relationships and thematic clusters within the research domain. The results were presented as visual network maps, where each element is represented by a node (circle). The size of each node reflects its relative weight or impact level, while the connecting lines (edges) indicate the strength of co‐occurrence relationships. In this way, the interconnectedness of authors, countries and keywords was evaluated within a network structure. Author citation counts were obtained using the full counting approach. In multi‐authored publications, all citations received by a publication were assigned in full to each co‐author, consistent with the default counting method of the bibliometric analysis software. In addition, key bibliometric indicators, including the number of publications, citation density, author and country distributions, were analysed to identify productivity trends, collaborative patterns and the dynamics of the field's scientific impact. Citation density was calculated as the average number of citations per year since publication. The normality of the distribution for article characteristics was tested using the Shapiro–Wilk test. Relationships between bibliometric indicators were examined using Spearman's correlation analysis in IBM SPSS Statistics 24.0 (IBM Corp., Chicago, IL, USA). Non‐parametric analyses were preferred due to the limited sample size and non‐normal distribution characteristics of the bibliometric variables. A p value < 0.05 was considered statistically significant. Ethics Approval This study is a bibliometric analysis based on publication data obtained from publicly accessible scientific databases (Web of Science and PubMed). During the research process, there was no contact with animals, and animals were not used for research purposes. Therefore, ethical committee approval was not required. Results A total of 76 records were identified through the literature search; 36 from Web of Science and 40 from PubMed. After removing duplicates, 47 unique studies remained; 27 met the inclusion criteria and were included in the bibliometric analysis (the list of relevant articles is included in the Supporting Information). These studies were published between 1 January 2018 and 22 October 2025 across 11 different journals or edited books. Collectively, the 27 studies cited 632 references and used 59 distinct keywords. A total of 155 authors contributed to these publications, with no single‐authored works identified. The rate of international co‐authorship was calculated at 37.04%. Among the included publications, 24 were original research articles, two were review papers, and one was a proceedings paper. Publication Volume and Citation Impact According to bibliometric indicators, the annual growth rate of publications was 10.41%, the average document age was 3.85 years, and the average number of citations per document was 15.56. The data presented in Figure 2 reveal a clear relationship between the number of publications, the average citations per article and the average annual citation rates over the years. In particular, the year 2020 stands out as the most impactful year, with a remarkably high average of 46.33 citations per article. Although the total number of publications during the 2018–2020 period was relatively low, the high citation rates observed during this period are noteworthy. Starting from 2021, a steady increase in the number of publications is observed; however, this is accompanied by a noticeable decline in the average number of citations per article. The year 2023, with eight publications, was the most productive in terms of output. Nevertheless, due to the shorter citation window, the average number of citations per article in 2023 remained lower compared to previous years. In 2024 and 2025, both the number of publications and the average citation rates declined, reflecting the limited time available for citation accumulation. Publication and citation trends of artificial intelligence research in chest radiography over time. The figure illustrates annual scientific production and citation impact across years. Bars represent the number of publications (left axis), while lines indicate mean TC per year and mean TC per article values (right axis). Analysis of Journals Leading academic journals in the fields of AI and thoracic radiology stand out for both publication volume and scientific impact. Among the top 10 journals identified (Table 1), the highest number of publications was observed in Veterinary Radiology & Ultrasound (n = 8). This journal is classified in the Q2 quartile, with both its 2024 and 5‐year impact factors (IFs) reported as 1.50. This is followed by Scientific Reports, which includes five publications. The journal is ranked in the Q1 category, with an IF of 3.90 and a 5‐year IF of 4.30. Frontiers in Veterinary Science ranks third with three publications and is also classified as a Q1 category, with an IF of 2.90. The Veterinary Journal (IF: 3.10), Veterinary Sciences (IF: 2.30) and Research in Veterinary Science (IF: 1.80) each contributed two publications. In contrast, the American Journal of Veterinary Research, Journal of Small Animal Practice, Journal of Veterinary Science and Journal of Medical Imaging each published one article. The IFs of these journals range between 1.40 and 1.90. Overall, the majority of the journals are classified in the Q1 or Q2 quartiles, and their IFs range from 1.40 to 4.30. TABLE 1Distribution of article counts, quartile classifications and impact indicators for the top 10 journals identified in the study.SourcesArticlesRank by JIFRank by JCIIF (2024) IF (5 years) Veterinary Radiology & Ultrasound 8Q2Q21.501.50 Scientific Reports 5Q1Q13.904.30 Frontiers in Veterinary Science 3Q1Q12.903.30 Research in Veterinary Science 2Q2Q11.802.29 Veterinary Journal 2Q1Q13.102.70 Veterinary Sciences 2Q1Q12.302.40 American Journal of Veterinary Research 1Q2Q21.401.20 Journal of Medical Imaging 1Q3Q21.702.50 Journal of Small Animal Practice 1Q2Q21.901.67 Journal of Veterinary Science 1Q2Q21.501.61Abbreviations: IF, Impact Factor; JCI, Journal Citation Indicator; JIF, Journal Impact Factor. Distribution of article counts, quartile classifications and impact indicators for the top 10 journals identified in the study. IF (5 years) Abbreviations: IF, Impact Factor; JCI, Journal Citation Indicator; JIF, Journal Impact Factor. Analysis of Authors Quantitative analyses of author performance reveal significant differences in publication count, fractional contribution and citation impact (Table 2). In terms of publication output, T. Banzato (n = 8), A. Zotti (n = 7) and M. Wodzinski (n = 6) emerge as the most prolific authors. Regarding fractional contribution, T. Banzato holds the highest value (1.11), followed by A. Zotti (0.99) and M. Wodzinski (0.71). Despite a lower publication count (n = 4), S. Burti stands out with a relatively high fractional contribution of 0.65. With respect to citation indicators, T. Banzato and A. Zotti are the most cited authors, each receiving 28 citations. S. Burti and V. L. Osti follow, with 17 citations each. Authors such as M. Wodzinski, T. Hwang, H. Lee and Y. Yoon demonstrate moderate citation impact, with 11 citations each. C. Chiara and Dona, on the other hand, had relatively lower citation counts, each receiving 9 citations. TABLE 2Bibliometric analysis of publications and citations of the top 10 authors.Author Number of articles Articles fractionalizedAuthorNumber of citationsT. Banzato81.11T. Banzato28A. Zotti70.99A. Zotti28M. Wodzinski60.71S. Burti17S. Burti40.65V. L. Osti17A. Komeili30.50T. Hwang11N. Norena30.50H. Lee11P. Tahghighi30.50M. Wodzinski11R. B. Appleby20.33Y. Yoon11M. Atzori20.22C. Chiara9D. Chiavegato20.22C. Dona9 Note: Citation counts were calculated using the full counting method. For multi‐authored publications, the total number of citations received by each publication was fully attributed to each co‐author. Bibliometric analysis of publications and citations of the top 10 authors. Number of articles Note: Citation counts were calculated using the full counting method. For multi‐authored publications, the total number of citations received by each publication was fully attributed to each co‐author. An examination of the author collaboration network reveals a structure composed of multiple clusters. Within the main collaboration cluster, reciprocal co‐authorship connections are evident among T. Banzato, A. Zotti and M. Wodzinski. Each of these authors maintains direct links with several other researchers within the cluster. Authors such as S. Burti, A. Komeili and P. Tahghighi are also positioned within this main cluster and demonstrate interconnectedness through direct collaborations. Additionally, the network contains smaller peripheral clusters and isolated nodes, representing authors with limited collaborative links or single‐study contributions (Figure 3). An author collaboration network showing co‐authorship relationships among the most active researchers in the field. Node size reflects the number of publications per author, while line thickness indicates the strength of co‐authorship links. Colours represent distinct collaboration clusters. Analysis of Key Publications Table 3 presents the characteristics of the top 10 most cited articles on AI applications in thoracic radiology for small animals. The most cited study was conducted by Burti et al. (2020) and published in The Veterinary Journal. This research investigated the potential of AI to detect cardiomegaly in dogs using thoracic radiographs. With 50 citations, it achieved the highest citation count, reflecting a stable and long‐term academic impact, supported by an average of 7.14 annual citations and a normalized citation value of 1.08. Banzato, Wodzinski, Burti, et al. (2021), in a study published in Scientific Reports, aimed to classify thoracic radiographic findings in dogs automatically. Comparing the ResNet‐50 and DenseNet‐121 models, the authors found that ResNet‐50 was more effective at identifying conditions such as cardiomegaly, pleural effusion and alveolar patterns. The study received 47 citations, with 7.83 annual citations and a normalized citation value of 1.38, indicating strong scientific visibility and suggesting its prominent position within the emerging literature on clinical AI applications. In another highly cited study, Li et al. (2020), published in Veterinary Radiology & Ultrasound, developed two CNN models for detecting left atrial enlargement and compared their performance with that of veterinary radiologists. One model was optimized for accuracy and the other for sensitivity. The accuracy‐focused model demonstrated an 85.19% agreement with radiologists. The study earned 47 citations, 6.71 annual citations and a normalized value of 1.01, highlighting its consistent influence. Boissady et al. (2020) evaluated the PicoxAI system using three CNN models to detect 15 different thoracic lesions in cats and dogs. Only the model pre‐trained with ImageNet achieved optimal performance. Its outputs were compared with veterinary classifications, and error rates were assessed. Published in Veterinary Radiology & Ultrasound, the study received 42 citations, with 6.00 annual citations and a normalized value of 0.91, reflecting a steady but moderate impact. Yoon et al. (2018) compared a CNN model with traditional ML approaches in evaluating cardiomegaly, pneumothorax, pleural effusion, pulmonary patterns and mediastinal shifts. Their findings indicated superior performance of the CNN model. The study, published in The Veterinary Journal, received 25 citations, with an annual citation rate of 2.78 and a normalized IF of 1.00, reflecting a balanced and enduring impact. Kim et al. (2022) assessed the Vetology AI system against veterinary radiologists in diagnosing cardiogenic pulmonary oedema in dogs. The model showed >90% accuracy, sensitivity, and specificity, but a relatively lower positive predictive value (56%). The study, published in Veterinary Radiology & Ultrasound, garnered 23 citations, with 4.60 annual citations and a normalized citation value of 1.34, indicating high early impact. In a separate study, Banzato, Wodzinski, Tauceri, et al. (2021) classified canine thoracic radiographs into 8 categories using ResNet‐50 and DenseNet‐121, and again found that ResNet‐50 yielded higher accuracy. Published in Frontiers in Veterinary Science, the study accumulated 21 citations, 3.50 annual citations and a normalized value of 0.62, showing a relatively limited influence compared to their earlier work. Pereira et al. (2023) authored a review to introduce veterinarians to AI‐related concepts, including DL, CNNs and transfer learning. Covering imaging applications across various organ systems, the review was published in Veterinary Sciences and garnered early traction, with 19 citations, 4.75 annual citations and a normalized impact of 2.34, notable for its educational and conceptual contributions. Burti et al. (2024) published a comprehensive review of the current status and future directions of AI in veterinary diagnostic imaging, encompassing multiple modalities, including radiology, ultrasound, CT and MRI. Featured in Research in Veterinary Science, the review received 19 citations, 6.33 annual citations and the highest normalized citation impact (2.92) among all studies, marking it as a particularly significant and timely contribution. Finally, Müller et al. (2022) evaluated the performance of an AI model in diagnosing pleural effusion, reporting an accuracy of 88.7%, a sensitivity of 90.2% and a specificity of 81.8%. The study, published in Veterinary Radiology & Ultrasound, received 18 citations, 3.60 annual citations and a normalized value of 1.05, indicating consistent academic visibility. TABLE 3Analysis of the top 10 most cited articles based on bibliometric findings.PaperDOITotal citations (TC)TC per yearNormalized TCBurti et al. (2020)10.1016/j.tvjl.2020.105505507.141.08Banzato, Wodzinski, Burti, et al. (2021)10.1038/s41598‐021‐83515‐3477.831.38Li et al. (2020)10.1111/vru.12901476.711.01Boissady et al. (2020)10.1111/vru.12912426.000.91Yoon et al. (2018)10.1016/j.tvjl.2018.05.009252.781.00Kim et al. (2022)10.1111/vru.13062234.601.34Banzato, Wodzinski, Tauceri, et al. (2021)10.3389/fvets.2021.731936213.500.62Pereira et al. (2023)10.3390/vetsci10050320194.752.34Burti et al. (2024)10.1016/j.rvsc.2024.105317196.332.92Müller et al. (2022)10.1111/vru.13089183.601.05 Analysis of the top 10 most cited articles based on bibliometric findings. Keyword and Thematic Analysis When the ten most frequently occurring keywords were examined, it was found that the terms ‘artificial intelligence’, ‘convolutional neural network’ and ‘dogs’ each appeared eight times. These were followed by ‘classification’ and ‘radiology’, with seven occurrences each. The terms ‘deep learning’ and ‘machine learning’ appeared six times each, while ‘dog’ and ‘error’, related to species type and performance evaluation, occurred five times. The keyword ‘thorax’ was noted four times. This distribution aligns with the visual patterns observed in the word cloud representation. The word cloud analysis (Figure 4) clearly highlights the most dominant concepts in the research field. The terms ‘convolutional neural network’, ‘artificial intelligence’, ‘classification’, ‘machine learning’ and ‘deep learning’ are especially prominent. Notably, ‘convolutional neural network’ stands out as the most visually emphasized term, indicating its central role in the literature. In addition, the frequent use of domain‐specific and anatomical terms such as ‘radiology’ and ‘thorax’ suggests that the studies are firmly grounded in radiological imaging. The recurrence of ‘dogs’ and ‘dog’ as keywords reflects the emphasis on species‐specific analysis, particularly in canine subjects. Furthermore, the inclusion of ‘error’ highlights the focus on performance evaluation and diagnostic accuracy within the reviewed studies. Overall, the word cloud effectively illustrates that the concepts in the reviewed literature are predominantly concentrated on AI, image analysis and species‐specific terminology. Word cloud representing the keywords used in the analysed studies. Country‐Level Analysis An evaluation of publication and citation data by country indicates that the literature on AI applications in small animal thoracic radiography is predominantly concentrated in specific geographic regions. North America leads in publication output, with the United States producing the most publications (n = 32). In Europe, multiple countries contribute to the literature, indicating a broader geographic spread. Notable contributors include Italy (n = 23), Portugal (n = 11), Switzerland (n = 10), Poland (n = 8), France (n = 7) and Spain (n = 6). Regarding total citations, Italy ranks first with 160, followed by the United States (n = 152) and South Korea (n = 45). These findings suggest that countries with higher publication output tend to exhibit a correspondingly high citation impact. When assessing average citations per article, the United States (21.7 citations/article) and Italy (20.0) rank highest in average citation impact. Despite having fewer publications, Portugal (19.0) and France (17.0) also exhibit relatively high average citation rates. In contrast, countries such as Canada (4.3) and Thailand (3.5) show comparatively lower average citation values. When viewed alongside the map visualization (Figure 5), it is evident that both publication output and citation impact are predominantly concentrated in North America and Europe. Asia is represented more modestly by a few countries, while South America and Oceania contribute to a lesser extent. This numerical and spatial distribution highlights a notable geographical disparity in scientific productivity and impact within the examined literature. Geographic distribution of scientific production based on authors' country affiliations. Colour intensity represents publication frequency, with darker blue indicating higher scientific production and lighter blue indicating lower publication output. Countries shown in grey had no publications included in the present dataset. Co‐Citation Network Analysis The resulting co‐citation network reveals several distinct clusters in the literature (Figure 6). The first cluster (represented in red), which forms the dominant structure of the network, includes the works of authors such as T. Banzato, A. Zotti and M. Wodzinski. This cluster is characterized by dense internal linkages and strong co‐citation relationships, suggesting that these publications constitute one of the central co‐citation structures within the field. Within the same cluster, authors such as S. Burti, M. Atzori and H. Mueller have contributed publications that reinforce this central body of literature, enhancing the internal coherence of the group. Other clusters in the network consist of smaller sets of publications organized around more specialized themes. These peripheral clusters exhibit weaker connections to the central literature but reflect the thematic diversity and varied research foci within the broader field. Overall, the co‐citation network illustrates the existence of a knowledge structure concentrated around key foundational works, alongside the emergence of thematically distinct yet internally cohesive sub‐literatures that contribute to the field's intellectual development. Co‐citation network analysis of the analysed studies. Relationship between Citation Indicators and Usage Metrics According to the correlation analysis of selected bibliometric features of the articles, a strong relationship was identified among key citation metrics (Table 4). A very high, statistically significant correlation was observed between the number of Web of Science citations and the total number of citations across all databases (r = 0.995; p < 0.001; 95% CI: 0.985, 0.998), indicating near‐perfect consistency across platforms. In contrast, the correlations between the number of cited references and both the Web of Science citation count (r = 0.411) and the total number of citations (r = 0.348) were low and not statistically significant, indicating that the volume of references in an article does not strongly predict its citation impact. Additionally, the number of articles used in the past 180 days showed no statistically significant correlations with any other bibliometric indicators. This implies that short‐term usage activity may not directly reflect long‐term citation impact within the analysed literature. TABLE 4Statistical analysis results of the relationship between citation indicators and usage metrics.Cited reference countTimes cited, WOSTimes cited, all databases180‐day usage countCited reference count r 1.000 0.165 (95% CI: −0.189, 0.508) 0.188 (95% CI: −0.156, 0.516) −0.063 (95% CI: −0.537, 0.392) p —0.4110.3480.756Times cited, WOS r 0.165 (95% CI: −0.189, 0.508) 1.000 0.995 (95% CI: 0.985, 0.998) 0.048 (95% CI: −0.332, 0.428) p 0.411—< 0.0010.811Times cited, all databases r 0.188 (95% CI: −0.156, 0.516) 0.995 (95% CI: 0.985, 0.998) 1.000 0.026 (95% CI: −0.354, 0.407) p 0.348< 0.001—0.898180‐day usage count r −0.063 (95% CI: −0.537, 0.392) 0.048 (95% CI: −0.332, 0.428) 0.026 (95% CI: −0.354, 0.407) 1.000 p 0.7560.8110.898— Statistical analysis results of the relationship between citation indicators and usage metrics. r 0.165 (95% CI: −0.189, 0.508) 0.188 (95% CI: −0.156, 0.516) −0.063 (95% CI: −0.537, 0.392) r 0.165 (95% CI: −0.189, 0.508) 1.000 0.995 (95% CI: 0.985, 0.998) 0.048 (95% CI: −0.332, 0.428) r 0.188 (95% CI: −0.156, 0.516) 0.995 (95% CI: 0.985, 0.998) 0.026 (95% CI: −0.354, 0.407) r −0.063 (95% CI: −0.537, 0.392) 0.048 (95% CI: −0.332, 0.428) 0.026 (95% CI: −0.354, 0.407) Discussion In recent years, rapid advancements in AI have significantly impacted diagnostic and decision‐support processes in veterinary medicine and the health sciences. Clinical decision support systems, treatment planning and medical image analysis are among the most common application areas of AI‐based methods (Vickram et al. 2025). Early ML studies in radiographic imaging in veterinary medicine demonstrated the applicability of AI across different anatomical regions and provided a methodological basis for subsequent research (McEvoy and Amigo 2013). Following this accumulation of knowledge, one of the most frequently studied areas in the veterinary literature using CNN‐based models has been the automated evaluation of thoracic radiographs (Hennessey et al. 2022). A large portion of current studies focus on dogs, while research on cats is limited (Burti et al. 2024). In the field of veterinary thoracic radiology, Yoon et al. (2018) presented one of the early studies in this field by comparing a bag‐of‐features‐based ML algorithm with CNN in the evaluation of heart silhouette, lung pattern, presence of mediastinal shift and pleural space using a binary classification approach of normal and abnormal. Following this study, a significant increase has been observed in the number of studies evaluating cardiac and pulmonary structures on thoracic radiographs using AI (Burti et al. 2024; Hennessey et al. 2022; Rho et al. 2023). This trend indicates that research interest in AI applications in this field is gaining momentum, as reflected in the growing number of publications and their increasing visibility in the literature (Rubini et al. 2025; Vickram et al. 2025). The number of bibliometric studies examining AI applications in veterinary medicine is quite limited today, and the existing literature lacks comprehensive, field‐specific evaluations (Elasan and Yilmaz 2025). A large portion of existing bibliometric studies addresses AI applications in general medicine (Serin and Körez 2025) and does not provide a systematic bibliometric analysis specific to veterinary radiology, particularly in thoracic radiography. Given the increasing importance of AI technologies in clinical and research areas of veterinary thoracic radiology, this situation highlights an emerging knowledge gap in the literature. In this context, examining research trends, scientific collaboration networks and publication dynamics related to AI applications in veterinary thoracic radiography is important for contextualizing the field's early development and informing future research. Bibliometric analysis is a method that goes beyond systematic reviews, aiming to reveal the structural characteristics, developmental trajectory and research trends of a specific research area through quantitative data, allowing for the visualization and mapping of the accumulated knowledge in the literature (Donthu et al. 2021; Hennessey et al. 2022). In this study, bibliometric analyses were performed using the bibliometrix package developed in the RStudio environment and its graphical interface, Biblioshiny. This open‐source, flexible and comprehensive tool was chosen as a methodologically suitable option because it supports various bibliometric analysis methods, including mapping, and can integrate with other statistical R extensions (Yürük Kayapınar 2023). Within the scope of this study, scientific publications on AI applications in thoracic radiology for small animals were analysed using bibliometric methods to examine the field's development, identify prominent themes and identify existing research gaps. Accordingly, the present analysis provides an initial bibliometric overview of the field rather than a definitive assessment of a mature research domain. The publication and citation patterns observed in this study suggest that research on AI applications in small animal thoracic radiography is still evolving while demonstrating steady growth. The annual increase of 10.41% in publications and the marked rise in publication output after 2021, peaking in 2023, are consistent with the growing scientific interest in this field. However, this increase in publication output was not accompanied by a comparable increase in the average number of citations per article. This pattern suggests that while publication output reflects the development of a research field, it is not sufficient on its own to evaluate scientific impact and visibility (Huang et al. 2025). The comparatively high average citation rate observed in 2020 may be interpreted in the context of both citation dynamics and the characteristics of the studies published during that period. In particular, three studies published in 2020, including those focusing on the radiographic evaluation of heart diseases (Burti et al. 2020; Li et al. 2020) and the role of AI in veterinary decision‐making (Boissady et al. 2020), addressed clinically relevant topics that may have enhanced their visibility within the literature. At the same time, these publications have had a longer period to accumulate citations than more recent studies. Consequently, the lower citation counts observed for publications from 2024 and 2025 are likely related, at least in part, to their shorter citation window rather than to differences in scientific value. The higher citation counts observed for 2020 publications should be interpreted with caution, as older articles have had more time to accumulate citations than recent ones. Previous studies have highlighted that the length of the citation time window is an important consideration in the evaluation of scientific impact (Chen et al. 2022). Therefore, the prominence of 2020 in our dataset may partly reflect this temporal advantage rather than greater scientific impact alone. Citation counts should therefore be interpreted alongside publication year rather than as direct indicators of scientific influence. Identifying suitable journals and journal classifications provides researchers with reliable and guiding reference information, while also contributing to a better understanding of the existing publication ecosystem. Such analyses help researchers determine the most suitable target journals for publishing their work. The academic visibility and impact level of journals are generally evaluated through indicators such as total citation count, IF and Journal Citation Reports (JCR) categories (Garfield 2006). Evaluations based on these criteria reveal that publications in AI and thoracic radiology are concentrated in academically reputable journals with moderate to high‐impact, indicating increasing scientific interest in this interdisciplinary field. The fact that most of the journals examined are in the Q1 or Q2 category and have IFs ranging from 1.20 to 4.30 shows that these studies are published in journals with medium‐to‐high impact and possess both intra‐field and interdisciplinary visibility. The journal Veterinary Radiology & Ultrasound, which has the highest number of publications, is noteworthy because, despite being in the Q2 category, it is a specialized journal specific to the field. Furthermore, publications in general science journals such as Scientific Reports, which are in the Q1 category, demonstrate that AI applications are not limited to veterinary literature but have a broader academic impact. This distribution indicates that AI resonates across disciplines, both in applied clinical contexts and in methodological contexts (Burti et al. 2024). Additionally, the higher visibility and citation potential of studies published in such journals are significant factors influencing research trends and publication strategies in the field. In evaluating author contributions, considering publication count, fractional contribution and citation indicators together offers a more balanced analysis, especially in multi‐authored and collaborative research areas. The fractional approach stands out as a complementary measure that allows for interpreting relationships among individual contributions, publication volume and citation impact within a single framework (Demaine 2022). The presented findings show that T. Banzato, A. Zotti and M. Wodzinski have high values ​​across publication count, fractional contribution and citation indicators. The author collaboration network also supports this, revealing that these authors are connected within the same collaboration cluster. In contrast, S. Burti, despite having a relatively limited number of publications, demonstrates a distinct profile in the table in terms of fractional contribution and citation indicators and is represented by connections in the collaboration network. These findings demonstrate that authors can exhibit different profiles in terms of publication count, fractional contribution and citation indicators, and that the structure of the author collaboration network supports the contextual interpretation of these differences. Accordingly, evaluating author performance through a combined assessment of multiple indicators provides a more robust and nuanced perspective than reliance on a single metric. Research at the intersection of thoracic radiology and AI is characterized by technical depth and methodological diversity, spanning a wide range of DL architectures, evaluation protocols and imaging tasks (Mekonnen et al. 2025; Suksangvoravong et al. 2024). An examination of the 10 most influential studies published in this field revealed a comparative assessment of DL models such as Banzato et al.’s (Banzato, Wodzinski, Burti, et al. 2021) ResNet‐50 and DenseNet, focusing on the automated diagnosis of specific thoracic and lung pathologies like cardiomegaly (Burti et al. 2020), pleural effusion and alveolar pattern, among those with high citation counts. These studies, by authors such as Banzato et al. and Burti et al., have demonstrated notable academic impact not only through their productivity but also through their normalized total citation values. Normalized total citation indicators offer a more balanced measure of the relative effects of publications within a temporal context compared to absolute citation counts. This approach is critical in revealing the early academic visibility of newly published studies (Bornmann and Wohlrabe 2019; Ioannidis et al. 2016). The high normalized citation values of the review and synthesis studies published by Pereira et al. (2023) and Burti et al. (2024) demonstrate that publications that guide clinical practice can quickly gain reference status in the field literature, particularly during the early phases of field development. This shows that the need for information in the field is also directed towards publications that explain fundamental concepts and clarify application methods. Similarly, it is noteworthy that the normalized citation values ​​in the publications analysed in this study are closely related to clinical focus and methodological scope. The direct comparison of model performance with veterinarians in most studies reflects a recurring methodological pattern in the literature, indicating that researchers consistently frame AI evaluation within a clinical reference context. Keyword analysis is a widely employed method in bibliometric research to uncover the thematic structure and knowledge‐production orientations within a given scientific domain (Feng et al. 2023). This study finds that the literature on AI applications in small animal thoracic radiography exhibits a distinctly method‐driven thematic orientation. The frequent use of technical terms such as ‘convolutional neural networks’ and ‘deep learning’ underscores a strong emphasis on algorithmic approaches to enhance image analysis performance. Conversely, the prominence of keywords such as ‘classification’ and ‘error’ indicates that many studies assess these models within a clinical diagnostic framework, focusing on diagnostic accuracy, performance evaluation and error profiling. The repeated use of the terms ‘radiology’ and ‘thorax’ confirms the literature's concentration on radiological imaging applications, while the dominance of ‘dog’ and ‘dogs’ as keywords reflects a species‐specific orientation, with a notable focus on canine subjects. Overall, this thematic structure suggests that the field is undergoing methodologically driven technological advancement, particularly in the development and application of ML models. However, it also indicates ongoing challenges in clinical integration and limited generalizability, stemming from species variation and dataset heterogeneity. The literature on AI applications in small animal thoracic radiography shows a geographical concentration in specific areas, which parallels the scientific impact. The United States’ significant lead in publication volume (n = 32) points to its prominent position in technological infrastructure and academic productivity in this field. On the other hand, Italy's highest total citation count (n = 160) indicates that research conducted in that country is influential in shaping the field and has substantial citation visibility within the field. The high citation values achieved by some Asian countries, such as South Korea, despite a limited number of publications, suggest that certain research groups can have a significant impact on the literature. Similarly, the presence of countries such as Australia, Brazil and Ireland in the literature, despite lower publication volumes, indicates that AI applications in small animal thoracic radiography have become a global ​​research area of interest. However, this regional concentration in publication and citation distribution indicates that scientific production in the field is not evenly distributed geographically. This situation highlights the importance of increasing multi‐regional collaborations for a more inclusive and equitable development at the global level. The co‐citation network shows that the literature on the use of AI in small animal thoracic radiography exhibits a citation pattern clustered around standard references and structured across different publication groups (Figure 6). Such analyses are among the commonly used methods in bibliometric research to reveal the intellectual structure of the literature (Huang et al. 2025). The formation of multiple clusters within the network structure indicates that the field is structured around specific thematic and methodological focuses. The fact that the works of researchers such as T. Banzato, S. Burti, A. Zotti and M. Wodzinski are associated with different parts of the network suggests that these publications function as common reference points among different citation groups. This indicates that the literature on the use of AI in small animal thoracic radiography exhibits citation‐based coherence across different research orientations. Overall, the co‐citation network reveals that the field's knowledge production process has a multi‐layered literature structure that develops through complementary theoretical and applied studies. In the correlation analysis of bibliometric indicators, the extraordinarily high and statistically significant relationship between Web of Science citation counts and citation counts across all databases (r = 0.995; p < 0.001) indicates that citation measurements from different data sources are largely consistent. This indicates that citation‐based metrics offer similar trends in evaluating scientific impact. In contrast, the lack of a significant relationship between the number of cited sources and citation levels reveals that the number of references alone is not a determining factor in citation performance. Furthermore, the lack of a significant relationship between short‐term usage metrics and citation indicators suggests that these two sets of measures may reflect different levels of impact. These findings emphasize the importance of considering different metrics together rather than relying on a single indicator in bibliometric evaluations. The publication patterns observed in this study suggest that the field is progressing beyond early feasibility studies towards a more structured phase of methodological development. Nevertheless, several gaps remain. Canine datasets continue to dominate the literature, leaving feline thoracic radiography comparatively underrepresented despite its clinical relevance. Among the ten most cited studies, considerable variation was observed in the evaluation approaches employed, with studies reporting different combinations of model accuracy, sensitivity and veterinarian comparison rates, without a consistent reporting framework (Table 3). This heterogeneity limits direct comparison across studies and complicates the synthesis of evidence. In addition, research output remains concentrated in North America and Europe, indicating that the current evidence base is derived from a relatively limited geographical context. Collectively, these findings suggest that broader species representation, greater consistency in performance reporting, and wider geographical participation may facilitate more robust comparisons across studies and contribute to the continued methodological development of the field. This study has some limitations specific to bibliometric analyses. The inclusion of particular databases such as WoS and PubMed, and of only English‐language literature, may have led to the exclusion of relevant studies published in other languages or indexed in regional databases. Furthermore, as the literature search was limited to WoS and PubMed, relevant studies indexed in other databases (e.g., Scopus and IEEE Xplore) may not have been captured. Therefore, the findings should be interpreted within the scope of the selected databases. The limited number of included publications (n = 27) restricted the range of meaningful statistical analyses; the correlation analysis between WoS and total database citations should therefore be interpreted as a cross‐platform validation rather than a substantive finding regarding field structure. In future studies, additional databases such as Scopus and Google Scholar could be used in conjunction to obtain a more comprehensive and representative overview of the literature. Additionally, variations in author name abbreviations and spelling may have resulted in minor inconsistencies in author‐based analyses. Furthermore, the predominance of dog‐focused studies and the clustering of scientific output in specific geographic regions, such as the United States and Italy, may limit the generalizability of findings. Given the heterogeneity of datasets and breed diversity inherent to veterinary medicine, these factors should be considered when interpreting the current results, rather than being viewed as definitive constraints on the clinical applicability of existing AI models. The fact that AI applications in small animal thoracic radiography are still a developing and early‐stage field may have led to the relatively limited number of publications included; this should be taken into account when interpreting bibliometric patterns. Conclusion This study examined 27 articles published between 2018 and 2025 using the Biblioshiny analysis tool to provide a bibliometric overview of AI applications in veterinary thoracic radiography. The findings indicate that AI is an emerging research area in veterinary thoracic radiography, with the USA and Italy standing out in terms of publication volume and citation impact, while research interest is also increasing in other countries. AI applications in small animal thoracic radiography, especially those based on DL and CNNs, show potential for supporting the detection of thoracic pathologies involving the heart, lungs and pleural structures. However, the limited and heterogeneous nature of veterinary radiology datasets, together with species‐ and breed‐specific differences, may limit the generalizability of these models. Therefore, based on the current literature, AI systems should be considered supportive tools in clinical decision‐making rather than replacements for expert clinical judgment. Overall, this bibliometric analysis provides an exploratory overview of publication trends in this field. Based on only 27 publications, the robustness of network analyses and thematic inferences is limited; network visualizations, in particular, should be approached with appropriate caution. This small corpus reflects the field's early developmental stage rather than a limitation of the bibliometric approach itself — working with constrained datasets is a recognized feature of bibliometric research in emerging areas. Due to the limited number of available studies and the exploratory nature of bibliometric methods, the findings should be interpreted cautiously. Future studies with broader datasets, updated literature coverage, increased international collaboration and improved methodological standardization are warranted to further advance the field. Author Contributions Vildan Aslan Canatan: visualization, methodology, writing – original draft, investigation, resources, project administration, data curation, conceptualization, formal analysis, writing – review and editing. Ender Uzabaci: validation, resources, data curation, formal analysis, visualization, writing – review and editing, methodology, investigation. All authors read and approved the final manuscript. Funding The authors have nothing to report. Ethics Statement The authors have nothing to report. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Supporting File 1: vms371169‐sup‐0001‐SuppMat.docx Acknowledgements The authors have nothing to report. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Supporting File 1: vms371169‐sup‐0001‐SuppMat.docx Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.