Mapping the global landscape of artificial intelligence in pancreatic cancer research: A bibliometric and visualization analysis

Pegah Rashidian; Aseel Smerat; Mohammad Ahmar Khan; Pankaj Bansal; Chandan Sharma; Djaloliddin Mansurov; Seyedsina Moghimnejadhosseini; Amirmahdi Mojtahedzadeh; Mahsa Talebzadeh; Ehsan Amini-Salehi

doi:10.1097/MD.0000000000050681

. 2026 Sep 25;105(39):e50681. doi: 10.1097/MD.0000000000050681

Mapping the global landscape of artificial intelligence in pancreatic cancer research

A bibliometric and visualization analysis

Pegah Rashidian ^a, Aseel Smerat ^b, Mohammad Ahmar Khan ^c, Pankaj Bansal ^d, Chandan Sharma ^e, Djaloliddin Mansurov ^f, Seyedsina Moghimnejadhosseini ^g,^h, Amirmahdi Mojtahedzadeh ^g, Mahsa Talebzadeh ^g, Ehsan Amini-Salehi ^i,^*

PMCID: PMC13619209 PMID: 42798094

Abstract

Background:

Artificial intelligence (AI) applications in pancreatic cancer are expanding rapidly, but the field’s global structure and emerging priorities remain incompletely characterized.

Methods:

The Web of Science Core Collection was searched through July 5, 2026. English-language original articles and reviews addressing AI in pancreatic cancer were eligible. After manual screening, 583 publications from 1998 to 2026 were analyzed using Biblioshiny, VOSviewer, and CiteSpace to assess publication trends, contributors, collaboration networks, co-citation patterns, keyword evolution, thematic clusters, and citation bursts.

Results:

Publication output accelerated after 2019 and reached 164 publications in 2025; the lower total in 2026 reflected partial-year coverage. The literature involved 59 countries, 1230 institutions, and 4038 authors. China produced the largest number of publications (n = 222), whereas the United States ranked second (n = 150) and had the highest country-level centrality (0.39). Shanghai Jiao Tong University was the most productive institution (n = 20), while Harvard University had the highest institutional centrality (0.22). Frontiers in Oncology was the most productive journal (n = 32). Influential studies focused on deep learning-based computed tomography, electronic health record-based risk prediction, exosome-based machine learning, and endoscopic ultrasonography. Research themes evolved from neural-network classification and texture analysis toward radiomics, deep learning, early detection, risk prediction, liquid biopsy, tumor biology, treatment-response prediction, and precision oncology.

Conclusion:

AI research in pancreatic cancer is growing rapidly and becoming increasingly multidisciplinary, although productivity and collaboration remain uneven. Clinical translation will require prospective multicenter validation, representative datasets, transparent reporting, calibration, fairness assessment, workflow evaluation, and evidence of improved patient outcomes.

Keywords: artificial intelligence, bibliometric analysis, deep learning, machine learning, pancreatic cancer, precision oncology, radiomics

1. Introduction

Pancreatic cancer remains one of the most lethal solid malignancies, largely because of its aggressive biology, nonspecific early manifestations, and frequent diagnosis at an advanced stage. Contemporary evidence indicates that overall 5-year survival remains approximately 13%, while the opportunity for curative-intent treatment is concentrated among patients diagnosed with localized or otherwise resectable disease.^[1] Population-based screening is not recommended for asymptomatic average-risk adults, and current surveillance strategies are primarily directed toward individuals with substantial inherited or familial risk.^[1,2] These limitations have intensified efforts to identify approaches capable of recognizing high-risk individuals, subtle early disease, and clinically relevant tumor characteristics before they are detected through conventional diagnostic pathways.

Artificial intelligence (AI) has consequently emerged as an important research direction in pancreatic cancer. Machine learning, deep learning, radiomics, and related computational methods can integrate complex information from computed tomography (CT), endoscopic ultrasonography (EUS), pathology, electronic health records (EHRs), and molecular or circulating biomarkers.^[3,4] Recent studies illustrate the breadth of this potential: deep learning systems have detected pancreatic cancer on noncontrast CT at scale,^[5] radiomics-based models have identified disease-associated features on prediagnostic CT examinations,^[6] and longitudinal deep learning models have estimated future pancreatic cancer risk from disease trajectories recorded in national health data.^[7] More recently, the international PANORAMA study provided a large-scale comparison of AI systems and radiologists for pancreatic ductal adenocarcinoma (PDAC) detection on routine CT, reflecting the field’s movement toward clinically oriented external evaluation.^[8] Nevertheless, rapid technical expansion has produced a heterogeneous literature spanning multiple disciplines, data types, clinical tasks, institutions, and geographic regions, making the overall structure and evolution of the field increasingly difficult to assess from individual studies or conventional narrative reviews alone.

Bibliometric and science-mapping approaches can address this problem by quantitatively characterizing publication growth, influential contributors, collaboration networks, knowledge foundations, and emerging research themes. A previous pancreatic cancer-specific bibliometric analysis by Yin et al evaluated AI-related publications from 1997 to 2021 and identified diagnosis, treatment, risk assessment, biology, and EUS among the major areas of interest.^[9] However, that analysis preceded several influential developments in large-scale CT detection, prediagnostic imaging, longitudinal risk prediction, and more recent clinically oriented validation studies. Given the substantial expansion of AI research since 2021, an updated and methodologically broader assessment is needed to clarify how the field has changed and which directions are gaining prominence.

Accordingly, the present study maps the global landscape of AI research in pancreatic cancer using 583 manually screened publications indexed in the Web of Science Core Collection from 1998 through July 5, 2026. By integrating Biblioshiny, VOSviewer, and CiteSpace, we examine scientific production, leading countries, institutions, authors, and journals, as well as international collaboration, network centrality, co-citation structure, citation bursts, keyword evolution, thematic clusters, and temporal research trends. In addition to updating the period covered by the earlier bibliometric literature, this multidimensional approach is intended to distinguish research productivity from collaborative influence, identify the intellectual foundations of the field, and reveal emerging priorities relevant to early detection, risk prediction, biomarkers, treatment-response assessment, and precision oncology.

2. Methods

2.1. Data collection

Bibliographic records were retrieved from the Web of Science Core Collection on July 5, 2026. The Web of Science Core Collection was selected for its broad coverage of peer-reviewed literature, standardized citation indexing, and suitability for bibliometric analyses.

The search strategy targeted publications concerning the application of AI to pancreatic cancer and comprised 2 conceptual domains: AI and pancreatic cancer. AI-related terms included machine learning, deep learning, neural networks, computer vision, natural language processing, predictive modeling, and related computational techniques. Pancreatic cancer terms included pancreatic cancer, pancreatic carcinoma, pancreatic adenocarcinoma, and PDAC. The 2 domains were combined using appropriate Boolean operators. The complete search strategy, including keywords, field identifiers, truncation symbols, and Boolean combinations, is provided in Table S1, Supplemental Digital Content 1.

2.2. Eligibility criteria and study selection

Eligibility criteria were predefined before screening. Studies were included if they were indexed in the Web of Science Core Collection, published in English, and classified as original research articles or review articles. Eligible publications were required to address the development, evaluation, implementation, or clinical application of AI in pancreatic cancer, including detection, diagnosis, classification, prognosis, treatment planning, outcome prediction, imaging, pathology, or other relevant applications.

Conference proceedings, meeting abstracts, editorials, letters, book chapters, preprints, news items, early-access publications, retracted articles, and non-English publications were excluded. Restriction to full-length articles and reviews was applied to ensure a consistent and methodologically reliable bibliometric dataset.

Study selection was conducted independently by 2 reviewers using a predefined sequential process. Records were first filtered by publication type and language, followed by manual title and abstract screening for relevance to AI applications in pancreatic cancer. Disagreements were resolved through discussion or, when necessary, consultation with a third reviewer. No automated screening system was used.

The initial search identified 2249 records. During preliminary refinement, 316 records were excluded, including 304 publications that were neither original articles nor reviews and 12 non-English publications. The remaining 1933 records underwent title and abstract screening, during which 1350 records were excluded for insufficient relevance. Ultimately, 583 publications met all eligibility criteria and were included in the final bibliometric analysis. The study-selection process is presented in Figure 1.

Figure 1. — Flow diagram of the study selection process. The figure summarizes the stepwise process used to identify, screen, assess, and select publications for inclusion in the final bibliometric analysis. Each box represents a specific stage of the selection procedure, whereas the arrows indicate the progression of records through the screening workflow and the principal reasons for exclusion.

2.3. Data analysis

Bibliometric analyses were performed using 3 complementary platforms: VOSviewer, CiteSpace, and Biblioshiny. Complete bibliographic records were exported from the Web of Science Core Collection in comma-separated values and plain-text formats to ensure compatibility across software platforms.

VOSviewer was primarily used to visualize bibliometric networks, including co-authorship and keyword relationships, and to identify collaboration patterns and conceptual linkages within the field. Association-strength normalization was applied to standardize network connections. Clustering used a resolution parameter of 1.00 and a minimum cluster size of 1. Network layout generation used a single random start and a maximum of 1000 iterations, while cluster optimization used 10 random starts and 10 iterations. Other parameters remained at their default settings.^[10,11]

CiteSpace was used to investigate the intellectual structure, citation dynamics, and temporal development of research on AI in pancreatic cancer. Betweenness centrality was calculated to identify influential bridging nodes, while co-citation clustering was used to characterize major research themes. Cluster labels were generated using the log-likelihood ratio method, and cluster quality was assessed using modularity and silhouette statistics.

Temporal analyses covered 1998 to 2026 using 1-year time slices. Terms were extracted from titles, abstracts, author keywords, and Keywords Plus, and network similarity was calculated using the cosine coefficient. Node selection was performed using the g-index with k = 25, without network pruning. CiteSpace was additionally used for co-citation analysis, temporal visualization, dual-map overlays, citation-burst detection, and identification of emerging research fronts.^[12,13]

Biblioshiny, the web-based interface of the Bibliometrix R package, was used for descriptive and network-based bibliometric analyses. It supported assessment of annual scientific production, citation patterns, productive sources, journal performance, country contributions, international collaboration, co-citation relationships, and thematic development. Network, clustering, and dimensionality-reduction approaches, including Louvain clustering, Walktrap community detection, and multidimensional scaling, were used to characterize scientific and thematic relationships within the dataset.

Together, the 3 platforms provided complementary perspectives: VOSviewer emphasized network visualization, CiteSpace enabled structural and temporal analysis, and Biblioshiny supported descriptive performance and thematic analyses. Their combined use enabled assessment of research productivity, influence, collaboration patterns, intellectual structure, and emerging trends in AI research related to pancreatic cancer.^[14,15]

3. Results

3.1. Publication trends

The examination of publication patterns demonstrated a marked expansion in scholarly output on AI in pancreatic cancer over the study period. As shown in Figure 2, publication activity was very limited during the early years. Between 1998 and 2016, annual production remained minimal, ranging from 0 to 2 articles per year, with only occasional publications appearing during this period. A clearer increase began in 2017, when the annual number of publications rose to 3 articles. This number remained unchanged in 2018, followed by a more noticeable increase to 14 articles in 2019.

Figure 2. — Annual publication output on AI in pancreatic cancer research. The figure shows the annual number of publications related to AI and pancreatic cancer from 1998 to 2026. Publication activity remained limited during the early years, followed by a clear increase after 2019 and a marked acceleration in recent years. The highest annual output was observed in 2025, whereas the apparent decline in 2026 should be interpreted cautiously because the year was still ongoing at the time of data retrieval. AI = artificial intelligence.

Growth became more pronounced from 2020 onward. The number of publications increased from 22 articles in 2020 to 43 in 2021, 55 in 2022, and 84 in 2023. Although a slight decrease to 81 publications was observed in 2024, the overall upward trajectory remained evident. The highest annual output was recorded in 2025, with 164 publications, indicating a substantial acceleration of academic interest in this field. In 2026, the number of publications declined to 104. This decrease should be interpreted cautiously, as it may reflect incomplete indexing or the fact that the year was still ongoing at the time of data retrieval. Despite this fluctuation, the fitted polynomial trendline showed a clear upward trajectory, with an R² value of 0.8126, indicating a strong fit to the observed publication data.

The cumulative publication pattern, illustrated in Figure 3, further highlights the progressive development of the field. During the early phase, cumulative output increased slowly, reaching only 10 publications by 2016. Thereafter, growth became more evident, with the cumulative number of publications rising to 13 in 2017, 16 in 2018, 30 in 2019, and 52 in 2020. This expansion accelerated considerably in the following years, reaching 95 publications in 2021, 150 in 2022, 234 in 2023, and 315 in 2024.

A particularly sharp increase was observed during the most recent period, with cumulative production rising to 479 publications in 2025 and ultimately reaching 583 publications in 2026. Overall, the cumulative curve demonstrates a sustained and accelerating increase in research activity related to AI in pancreatic cancer. The polynomial trendline had an R² value of 0.799, indicating a strong overall fit and supporting a consistent upward growth pattern across the study period.

3.2. Countries and institutions

Research on AI in pancreatic cancer involved contributions from 59 countries and 1230 institutions, indicating a broad international research landscape. The country-level collaboration network is illustrated in Figure 4. In terms of publication volume, China emerged as the leading contributor, producing 222 documents. The United States ranked second with 150 documents, followed by India with 38, Japan with 33, and Germany with 30. Several other countries also made notable contributions to the field. England produced 26 documents, followed by Türkiye with 22, Italy with 21, South Korea with 19, and the Netherlands with 16.

Figure 4. — Worldwide country-level collaboration network in AI and pancreatic cancer research. The map was generated using Biblioshiny and shows the geographic distribution of countries contributing to research on AI in pancreatic cancer, as well as their international collaboration links. Country shading represents publication output, with darker red indicating a higher number of publications and lighter red indicating lower output. Blue curved lines represent international co-authorship links between countries, with thicker or more prominent lines indicating stronger collaborative relationships. AI = artificial intelligence.

Additional contributors included France with 14 documents, Spain with 13, Taiwan with 12, and Romania with 11. Saudi Arabia, Pakistan, and Iran each contributed 10 documents, whereas Australia and Canada each produced 9 (Table 1).

Table 1.

Top 10 countries and institutions by number of publications.

Rank of Country	Country	Rank of Institution	Institution
1	China	1	Shanghai Jiao Tong University
2	United States	2	Zhejiang University
3	India	3	Mayo Clinic
4	Japan	4	Johns Hopkins University
5	Germany	5	University of Texas MD Anderson Cancer Center
6	England	6	Sichuan University
7	Türkiye	7	Cedars-Sinai Medical Center
8	Italy	8	Chinese Academy of Medical Sciences and Peking Union Medical College
9	South Korea	9	Guangxi Medical University
10	Netherlands	10	Sun Yat-sen University

Open in a new tab

Processed using VOSviewer.

To evaluate the structural importance of countries within the international collaboration network, centrality was examined. This indicator reflects the extent to which a country functions as a bridge connecting different parts of the research network. Countries with higher centrality values were therefore interpreted as occupying more influential and connective positions in international scientific collaboration.

According to Figure 5 and Table 2, the United States recorded the highest centrality value, at 0.39, indicating its prominent role in linking international research communities in the field of AI in pancreatic cancer. England held the second-highest centrality value, at 0.28, followed by India at 0.25 and Italy at 0.21. Although China had the highest publication output, with 222 documents, its centrality value was 0.14, suggesting that its contribution was more strongly characterized by research productivity than by network bridging.

Table 2.

Top 10 countries and institutions by centrality.

Country	Centrality	Institution	Centrality
USA	0.39	Harvard University	0.22
England	0.28	University of California System	0.13
India	0.25	Mayo Clinic	0.13
Italy	0.21	Johns Hopkins University	0.13
China	0.14	Assistance Publique–Hôpitaux de Paris (APHP)	0.12
Pakistan	0.13	Zhejiang University	0.09
Germany	0.10	Shanghai Jiao Tong University	0.07
France	0.05	Cornell University	0.07
Brazil	0.05	Imperial College London	0.06
Türkiye	0.04	Chinese Academy of Medical Sciences–Peking Union Medical College	0.04

Open in a new tab

Processed using CiteSpace. Centrality values are reported as rounded values.

Additional countries also demonstrated meaningful centrality within the collaboration network. Pakistan had a notable centrality value of 0.13, followed by Germany at 0.10. France and Brazil each recorded a centrality value of 0.05, whereas Türkiye and the Netherlands each had a value of 0.04. Spain and Jordan also contributed to network connectivity, each with a centrality value of 0.03. Saudi Arabia, Norway, and the United Arab Emirates each had lower but still observable centrality values of 0.02.

The institutional analysis showed that Shanghai Jiao Tong University was the most productive organization in research on AI in pancreatic cancer, contributing 20 publications. Zhejiang University ranked second with 19 publications, followed by the Mayo Clinic with 17. Johns Hopkins University also demonstrated substantial output, producing 13 publications, whereas the University of Texas MD Anderson Cancer Center and Sichuan University each contributed 12.

A further group of institutions made notable contributions to this body of literature. Cedars-Sinai Medical Center, the Chinese Academy of Medical Sciences and Peking Union Medical College, and Guangxi Medical University each produced 10 publications. In addition, Sun Yat-sen University and Harvard Medical School each contributed 9 publications, whereas Fudan University produced 8. Several institutions, including Memorial Sloan Kettering Cancer Center, China Medical University, Southern Medical University, National Taiwan University, Sungkyunkwan University, Capital Medical University, Naval Medical University, Zhengzhou University, and Soochow University, each contributed 7 publications.

Institutional centrality was also assessed to determine the extent to which organizations served as connectors within the collaboration network. This measure reflects the bridging role of institutions and helps identify organizations that connect different research groups within the field.

According to Figure 6 and Table 2, Harvard University recorded the highest centrality value, at 0.22, indicating that it occupied the most prominent bridging position in the institutional collaboration network. The University of California System, the Mayo Clinic, and Johns Hopkins University each had a centrality value of 0.13, suggesting that these institutions also played important roles in linking different collaborative clusters. Assistance Publique–Hôpitaux de Paris also had a notable centrality value of 0.12.

Several other institutions demonstrated meaningful centrality within the network. Zhejiang University had a centrality value of 0.09, followed by Shanghai Jiao Tong University and Cornell University, each with a value of 0.07. Imperial College London recorded a centrality value of 0.06, whereas the Chinese Academy of Medical Sciences and Peking Union Medical College and Sichuan University each had a value of 0.04.

A broader group of organizations had lower but still observable centrality values of 0.03, including Harvard Medical School, the Helmholtz Association, Harvard University Medical Affiliates, the German Cancer Research Center, Capital Medical University, National Taiwan University, Wenzhou Medical University, the University of Liverpool, and Columbia University. Additional institutions, including the State University System of Florida, Cedars-Sinai Medical Center, Zhengzhou University, China Medical University, Southwest Medical University, the University of Copenhagen, Brigham and Women’s Hospital, and Beihang University, had centrality values of 0.02.

3.3. Journals and co-cited sources

The journal analysis identified the leading publication outlets contributing to research on AI in pancreatic cancer. As shown in Figure 7, Frontiers in Oncology was the most prolific journal, publishing 32 articles in this area. Cancers ranked second with 26 publications, followed by Scientific Reports with 21. These 3 journals represented the most active sources for disseminating research at the intersection of AI and pancreatic cancer.

Several additional journals also made notable contributions. BMC Cancer and Pancreatology each published 10 articles, whereas Abdominal Radiology, Diagnostics, and Discover Oncology each contributed 9 publications. Pancreas ranked next with 8 publications, followed by the World Journal of Gastroenterology with 7. Overall, the most productive journals were primarily focused on oncology, cancer, gastroenterology, radiology, and diagnostics, reflecting the multidisciplinary nature of AI applications in pancreatic cancer research.

The cited-source analysis further highlighted the journals that exerted the greatest influence within the dataset. As illustrated in Figure 8, Cancers was the most frequently locally cited source, with 574 local citations, followed closely by Scientific Reports with 556. Frontiers in Oncology ranked third, with 508 local citations, indicating that the most productive journals were also among the most influential sources within the analyzed literature.

Other highly cited sources included Gastroenterology, with 470 local citations; PLOS ONE, with 356; Pancreas, with 344; and the World Journal of Gastroenterology, with 339. Radiology-related journals also exerted substantial influence, including European Radiology, with 336 local citations, and Radiology, with 331. CA: A Cancer Journal for Cliniciansalso appeared among the most frequently cited sources, with 326 local citations, reflecting the importance of high-impact oncology literature in shaping this research field.

The temporal distribution of journal publications is presented in Figure 9. The results show that growth among the leading sources became particularly evident after 2019, with a sharp increase in cumulative publication output from 2021 onward. Frontiers in Oncology showed the steepest increase and reached the highest cumulative output by 2026. Cancers and Scientific Reports also demonstrated marked growth in recent years, whereas BMC Cancer and Pancreatology showed more moderate but consistent increases. These patterns suggest that journal-level activity in AI and pancreatic cancer research has accelerated substantially in recent years, paralleling the broader expansion of the field.

A dual-map overlay was used to illustrate citation relationships between citing and cited journal domains. As shown in Figure 10, the left side of the map represents the journals in which the analyzed articles were published, whereas the right side represents the journals most frequently cited by those articles. The most prominent citation pathways originated primarily from journals in the “Medicine, Medical, Clinical” and “Molecular, Biology, Immunology” domains and extended toward cited journals in the “Health, Nursing, Medicine” and “Molecular, Biology, Genetics” domains. Additional citation links involved domains related to systems, computing, and computer science, as well as chemistry, materials, physics, and environmental sciences.

3.4. Authors and co-cited authors

The authorship analysis showed that 4038 authors contributed to the literature on AI in pancreatic cancer. Among these contributors, Yun Bian, Jianping Lu, and Chengwei Shao were the most prolific authors, each contributing 10 publications. Jieyu Yu, Jing Li, and Xu Fang followed, each producing 9 publications, whereas Fang Liu and Ajit H. Goenka each contributed 8.

A further group of authors also demonstrated notable productivity. Mengmeng Zhu, Shanyu Qin, Sovanlal Mukherjee, and Wansu Chen each published 7 articles. Other active contributors included Qi Li, Hao Zhang, Shuangyang Mo, Yingwei Wang, Suresh T. Chari, and Adrian Saftoiu, each with 6 publications. These findings indicate that the most productive authors were primarily researchers working in AI-assisted imaging, pancreatic cancer diagnosis, prognosis, and clinical decision support. Collaboration patterns among the leading authors are visualized in Figure 11 and summarized in Table 3.

Figure 11. — Co-authorship network of prominent authors in AI and pancreatic cancer research. Each node corresponds to an author contributing to the field. Node size represents publication productivity, whereas the thickness of the connecting lines reflects the strength of co-authorship ties. Distinct colors identify collaboration clusters, highlighting groups of researchers who frequently publish together. Authors positioned closer to one another indicate stronger or more intensive collaborative relationships. The network was generated using VOSviewer. AI = artificial intelligence.

Table 3.

Top authors by publication productivity, citation impact, and co-citation frequency.

Number	Author with high number of publications	Number of publications	Author with high number of citations	Number of citations	The most co-cited authors	Number of co-citations
1	Yun Bian	10	Adrian Saftoiu	287	R. L. Siegel	159
2	Jianping Lu	10	Suresh T. Chari	273	L. C. Chu	83
3	Chengwei Shao	10	Ajit H. Goenka	255	L. Rahib	79
4	Jieyu Yu	9	Shounak Majumder	249	A. Saftoiu	78
5	Jing Li	9	Sovanlal Mukherjee	248	Y. Bian	68
6	Xu Fang	9	Yun Bian	200	T. Kuwahara	65
7	Fang Liu	8	Jianping Lu	200	T. Conroy	58
8	Ajit H. Goenka	8	Chengwei Shao	200	V. Granata	57
9	Mengmeng Zhu	7	Xu Fang	193	S. T. Chari	57
10	Shanyu Qin	7	Jieyu Yu	189	P. T. Chen	56

Open in a new tab

Processed using VOSviewer. The table reports 3 conceptually distinct bibliometric measures at the author level. Publication count indicates the total number of works attributed to each researcher in the dataset under review. Citation count represents the cumulative number of citations received by those works within the same dataset. Co-citation frequency indicates how often a particular author is cited alongside another author in the reference lists of the analyzed publications. Because these indicators reflect different dimensions of scholarly activity, namely research productivity, citation impact, and intellectual connectedness, their values should be interpreted independently rather than compared directly.

The author citation analysis revealed that Adrian Saftoiu had the highest citation count, with 287 citations, followed by Suresh T. Chari with 273. Ajit H. Goenka ranked third, with 255 citations, followed by Shounak Majumder with 249 and Sovanlal Mukherjee with 248. Several highly productive authors also accumulated substantial numbers of citations, including Yun Bian, Jianping Lu, and Chengwei Shao, each with 200. Xu Fang and Jieyu Yu followed with 193 and 189 citations, respectively.

The co-citation analysis identified 15,408 co-cited authors, reflecting the intellectual foundations of the field. R. L. Siegel was the most frequently co-cited author, with 159 co-citations, indicating the substantial influence of the cancer epidemiology and statistics literature on this research area. L. C. Chu ranked second with 83 co-citations, followed by L. Rahib with 79 and A. Saftoiu with 78. Other highly co-cited authors included Y. Bian with 68 co-citations, T. Kuwahara with 65, T. Conroy with 58, and V. Granata and S. T. Chari, each with 57.

Additional influential co-cited authors included P. T. Chen with 56 co-citations, A. P. Klein with 55, P. Lambin with 54, S. Mukherjee with 53, and W. S. Chen with 50.

3.5. Most highly cited articles

The citation analysis of key references in AI and pancreatic cancer research identified a group of highly influential publications. The most frequently cited article was “Large-scale pancreatic cancer detection via non-contrast CT and deep learning” by Cao et al,^[5] published in Nature Medicine, which received 258 citations. The second most frequently cited article was “A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories” by Placido et al,^[7] also published in Nature Medicine, with 241 citations.

The third most frequently cited study was “Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes” by Ko et al,^[21] which received 204 citations. This was followed by “Neural network analysis of dynamic sequences of EUS elastography used for the differential diagnosis of chronic pancreatitis and pancreatic cancer” by Saftoiu et al,^[20] with 200 citations. The fifth most frequently cited article was “Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic CT scans at a substantial lead time before clinical diagnosis” by Mukherjee et al,^[6] which received 151 citations.

Several other highly cited publications also made important contributions to the field. Hussein et al^[24] received 149 citations for their study, “Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches.” Kuwahara et al^[25] followed closely with 148 citations for “Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas.” Nasief et al^[26] received 135 citations for their study on machine learning-based delta radiomics for the early prediction of treatment response in pancreatic cancer.

The ninth most frequently cited article was “Pancreatic cancer detection on CT scans with deep learning: a nationwide population-based study” by Chen et al,^[27] with 114 citations. The tenth most frequently cited article was “AI and early detection of pancreatic cancer: 2020 summative review” by Kenner et al,^[28] which received 108 citations.

The 10 most highly cited articles are summarized in Table 4.

Table 4.

Top 10 most cited references in AI and pancreatic cancer research.

Rank	Title of the most cited paper	Year	Journal	Citations
1	Large-scale pancreatic cancer detection via non-contrast CT and deep learning	2023	Nature Medicine	258
2	A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories	2023	Nature Medicine	241
3	Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes	2017	ACS Nano	204
4	Neural network analysis of dynamic sequences of EUS elastography used for the differential diagnosis of chronic pancreatitis and pancreatic cancer	2008	Gastrointestinal Endoscopy	200
5	Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before clinical diagnosis	2022	Gastroenterology	151
6	Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches	2019	IEEE Transactions on Medical Imaging	149
7	Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas	2019	Clinical and Translational Gastroenterology	148
8	A machine learning based delta-radiomics process for early prediction of treatment response of pancreatic cancer	2019	NPJ Precision Oncology	135
9	Pancreatic cancer detection on CT scans with deep learning: a nationwide population-based study	2023	Radiology	114
10	Artificial intelligence and early detection of pancreatic cancer: 2020 summative review	2021	Pancreas	108

Open in a new tab

This table summarizes the 10 most cited articles identified in the VOSviewer output. Citation counts indicate global citation impact.

3.6. References with the strongest citation bursts

In bibliometric analysis, citation burst detection is used to identify references that experience a sudden and substantial increase in citation frequency during a defined period. Such references are considered to have attracted concentrated scholarly attention and often represent influential studies, methodological advances, or emerging research directions. In the present analysis, the 25 references with the strongest citation bursts in AI and pancreatic cancer research were identified, as shown in Figure 12. The detected burst periods extended from 2019 to 2026, reflecting the rapid development of this research field during the most recent phase of the study period.

Figure 12. — References with the strongest citation bursts in AI and pancreatic cancer research. This figure presents the references that experienced the most pronounced increases in citation frequency during specific time intervals. The blue lines indicate the overall citation period for each reference, whereas the red segments show the periods of active citation bursts. Higher burst strengths indicate more intense scholarly attention, whereas longer bursts indicate more sustained attention. These publications may have played important roles in shaping emerging research directions in AI-related pancreatic cancer research. The burst patterns highlight a transition from foundational radiomics and imaging-based machine learning studies to more recent work focused on early detection, risk prediction, clinical implementation, and AI-supported pancreatic cancer management. AI = artificial intelligence.

The strongest citation burst was observed for van Griethuysen et al,^[23] published in Cancer Research, with a burst strength of 6.55 from 2019 to 2022. This reference reflects the importance of radiomics methodology in the development of quantitative image analysis and machine learning-based cancer research. Kuwahara et al^[25] also showed a prominent burst, with a strength of 5.91 from 2021 to 2022, highlighting the growing role of deep learning in the assessment of pancreatic lesions.

Several other references with strong citation bursts were closely related to radiomics, medical imaging, cancer epidemiology, and machine learning. Gillies et al^[22] had a burst strength of 5.35 from 2019 to 2021.

The more recent references with citation bursts indicate a shift toward contemporary AI applications, risk prediction, cancer biology, and clinically oriented decision-support research. Huang et al^[4] showed the strongest ongoing burst, with a strength of 5.09 from 2024 to 2026. Placido et al^[7] also demonstrated an ongoing burst from 2024 to 2026, with a strength of 4.91, reflecting recent attention to deep learning-based prediction of pancreatic cancer risk from disease trajectories.

As shown in Figure 12, the blue lines represent the overall citation period for each reference, whereas the red segments indicate the specific periods during which citation bursts occurred. Longer red segments indicate more sustained periods of unusually high citation activity, whereas higher burst-strength values reflect a greater intensity of scholarly attention.

3.7. Keyword trends

Keyword analysis is a central component of bibliometric studies because it helps identify the major research themes, conceptual priorities, and evolving directions within a field. In the literature on AI in pancreatic cancer, “pancreatic cancer” emerged as the most frequently used keyword, appearing 214 times. This was followed closely by “machine learning,” which occurred 208 times, indicating the prominent role of machine learning-based approaches in this research area. “Artificial intelligence” ranked third, with 145 occurrences, followed by “cancer” with 126 occurrences and “diagnosis” with 108.

Several other keywords also appeared frequently and reflected important thematic domains. “Pancreatic ductal adenocarcinoma” appeared 93 times, whereas “deep learning” occurred 88 times and “radiomics” was recorded 79 times. Additional high-frequency terms included “survival” with 64 occurrences, “prognosis” with 57, “ductal adenocarcinoma” with 52, “CT” with 43, “adenocarcinoma” with 41, “prediction” with 39, “biomarkers” with 38, and “classification” with 34. These findings suggest that the literature is strongly centered on AI-based diagnosis, imaging, classification, prognosis, survival prediction, and biomarker-related research in pancreatic cancer (Fig. 13).

Figure 13. — Treemap of the most frequent keywords in AI and pancreatic cancer research. The figure displays the keywords that appeared most frequently in the analyzed literature. The relative size of each rectangle corresponds to keyword frequency, with larger areas representing terms that occurred more frequently in the dataset. The treemap was generated using Biblioshiny. AI = artificial intelligence.

Figure 14 presents the trend topics identified across the study period. The figure shows the temporal distribution of frequently occurring terms, with the size of each circle indicating term frequency and the horizontal line representing the period during which each term appeared prominently in the literature. Earlier trend terms included “artificial neural network,” which was prominent around 2018, suggesting early interest in neural network-based approaches. Around 2020 to 2021, terms such as “discovery,” “system,” “endoscopic ultrasound elastography,” and “CT scans” became prominent, reflecting early applications of AI in imaging-based assessment and diagnostic support.

From 2022 onward, the thematic focus expanded to include “classification,” “endoscopic ultrasound,” “differential diagnosis,” “model,” “adenocarcinoma,” and “CT.” These terms indicate increasing attention to diagnostic differentiation, imaging interpretation, and model development. By 2024 to 2025, larger and more frequent terms such as “pancreatic cancer,” “artificial intelligence,” “cancer,” “machine learning,” “pancreatic ductal adenocarcinoma,” and “ductal adenocarcinoma” became dominant, indicating the maturation and rapid expansion of the field. The most recent trend topics, extending through 2026, included “liquid biopsy,” “prognostic model,” and “differentiation,” suggesting growing interest in AI-supported biomarker analysis, prognostic prediction, and refined tumor characterization.

Figure 15 illustrates the cumulative frequency of the most frequently occurring keywords over time. The temporal pattern shows that keyword activity remained relatively low during the early years of the study period, with only limited cumulative growth before approximately 2018 to 2019. From 2020 onward, keyword occurrence began to increase more clearly, followed by a sharp acceleration after 2022. By the end of the study period, “pancreatic cancer” and “machine learning” showed the steepest cumulative increases, reaching the highest frequencies among all analyzed terms. “Artificial intelligence,” “cancer,” and “diagnosis” also demonstrated strong upward trajectories, particularly in recent years.

Other terms, including “pancreatic ductal adenocarcinoma,” “deep learning,” “radiomics,” “survival,” and “prognosis,” showed more moderate but consistent growth. Overall, the cumulative keyword pattern indicates that the field has shifted from limited early keyword activity to rapid thematic expansion in recent years. The strongest recurring themes were related to machine learning, AI, pancreatic cancer diagnosis, radiomics, deep learning, prognosis, and survival prediction.

Figure 16 presents the temporal overlay map of keyword co-occurrence generated using VOSviewer. In this visualization, node size reflects keyword frequency, link thickness indicates the strength of co-occurrence relationships, and node color represents the average publication year associated with each keyword. Earlier keywords are shown in darker blue or purple tones, whereas more recent keywords are shown in green to yellow tones.

The map shows that “pancreatic cancer,” “machine learning,” “artificial intelligence,” “diagnosis,” “cancer,” “deep learning,” and “radiomics” were the most prominent and highly connected keywords. These terms were strongly linked to keywords such as “survival,” “prognosis,” “biomarkers,” “classification,” “CT,” “endoscopic ultrasound,” “prediction model,” “risk prediction,” and “pancreatic ductal adenocarcinoma.” This pattern indicates that AI research in pancreatic cancer is organized around several interconnected themes, including imaging-based diagnosis, radiomics, machine learning model development, prognostic prediction, biomarker discovery, and clinical decision support.

The temporal overlay also suggests a gradual shift in research focus. Earlier themes were more closely associated with endoscopic ultrasound, classification, texture analysis, surgery, and diagnostic imaging. More recent keywords, shown in yellow-green to yellow tones, included “early detection,” “risk prediction,” “biomarker,” “tumor microenvironment,” “immunotherapy,” “subtypes,” “metastasis,” “PDAC,” and “gemcitabine.” This pattern suggests that the field is expanding beyond diagnostic imaging to broader applications involving early detection, molecular and biomarker-based prediction, treatment response, tumor biology, and personalized pancreatic cancer management.

3.8. Cluster analysis

Cluster analysis was performed using CiteSpace to identify the main thematic groupings within the literature on AI in pancreatic cancer (Fig. 17). The cluster labels shown in the visualization were generated automatically by CiteSpace using the log-likelihood ratio method and should therefore be interpreted as software-derived thematic indicators rather than manually assigned topic names. To improve interpretability, the labels were reviewed in relation to the surrounding keywords, connected nodes, and broader thematic context of the clusters.

Figure 17. — Cluster analysis of keywords in AI and pancreatic cancer research. Processed using CiteSpace. AI = artificial intelligence.

The analysis identified 13 principal clusters: cluster #0, “chronic pancreatitis”; cluster #1, “artificial intelligence-driven approaches”; cluster #2, “pancreatic ductal adenocarcinoma”; cluster #3, “artificial neural networks approach”; cluster #4, “routine data”; cluster #5, “deep learning”; cluster #6, “using exosome”; cluster #7, “conversational artificial intelligence”; cluster #8, “predicting neuroendocrine tumor”; cluster #9, “perineural invasion”; cluster #10, “pancreatic adenocarcinoma”; cluster #11, “patient outcome”; and cluster #12, “diagnostic performance.” These clusters indicate that the field is organized around several major thematic directions, including AI-assisted diagnosis, deep learning and artificial neural network methods, PDAC, pancreatic adenocarcinoma, differentiation from chronic pancreatitis, routine clinical data, exosome-based diagnostic approaches, neuroendocrine tumor prediction, perineural invasion, patient outcomes, and diagnostic performance assessment.

The largest and most central thematic areas were related to chronic pancreatitis, AI-driven approaches, PDAC, and artificial neural network approaches. This finding suggests that a major focus of the literature has been the use of AI methods to distinguish pancreatic cancer from benign pancreatic disease, support the diagnosis of PDAC, and improve clinical interpretation through neural network-based models. The presence of clusters such as deep learning, diagnostic performance, and routine data further reflects the importance of imaging-based classification, model evaluation, and routinely collected clinical or imaging data in AI research.

Other clusters point to more specialized or emerging research directions. The “using exosome” cluster suggests growing interest in AI-supported biomarker and liquid biopsy approaches, whereas the “predicting neuroendocrine tumor” cluster indicates the extension of AI methods beyond PDAC to other pancreatic tumor types. The “perineural invasion” and “patient outcome” clusters reflect increasing attention to prognostic modeling, tumor behavior, and clinically meaningful outcome prediction. The cluster labeled “conversational artificial intelligence” appears to represent a more recent development and may reflect the growing relevance of generative AI, large language models, and AI-supported communication or decision-support tools in pancreatic cancer research.

The developmental trajectories of these clusters are shown in the CiteSpace timeline visualization (Fig. 18). Earlier research activity was primarily associated with themes such as chronic pancreatitis, pancreatic adenocarcinoma, perineural invasion, and neuroendocrine tumor prediction, suggesting that early AI-related research was strongly connected to diagnostic differentiation, tumor classification, and pathological or imaging-based prediction. During the middle phase of the field’s development, clusters related to deep learning, artificial neural networks, routine data, and PDAC became more prominent, indicating a shift toward more advanced computational models and broader clinical applications.

Figure 18. — Time trend analysis of keyword clusters in AI and pancreatic cancer research. Processed using CiteSpace. AI = artificial intelligence.

In the most recent period, increased activity was observed in clusters such as AI-driven approaches, conversational AI, using exosome, patient outcome, and diagnostic performance. These patterns suggest that the field has gradually expanded from earlier diagnostic imaging and tumor classification studies to more diverse themes involving early detection, biomarker integration, outcome prediction, model performance evaluation, and emerging AI tools. Because the cluster labels were generated automatically by CiteSpace, these results should be interpreted as a broad thematic overview rather than a definitive classification of the field.

4. Discussion

This bibliometric and science mapping analysis characterizes a rapidly expanding but still translationally immature research domain at the intersection of AI and pancreatic cancer. Across 583 eligible publications produced by 4038 authors from 1230 institutions in 59 countries, scientific output remained sparse until the late 2010s and accelerated sharply after 2019, reaching 164 publications in 2025. The geographic and institutional analyses revealed an important distinction between research volume and network position: China produced the greatest number of publications, whereas the United States occupied the most prominent bridging position in the international collaboration network. A similar divergence was observed institutionally, with Shanghai Jiao Tong University leading in productivity and Harvard University displaying the highest betweenness centrality. The journal, co-citation, and dual-map analyses further showed that the field is not confined to medical imaging or computer science but is situated at the convergence of oncology, gastroenterology, radiology, diagnostics, molecular biology, genetics, epidemiology, and computational research. Finally, the citation-burst and keyword analyses suggest an intellectual progression from neural-network-assisted lesion differentiation and radiomics toward early detection, longitudinal risk prediction, biomarker integration, treatment-response assessment, and clinically oriented precision oncology.

These findings extend, rather than inaugurate, the bibliometric literature on this topic. Yin et al previously evaluated AI research in pancreatic cancer published between 1997 and 2021 and identified rapid growth after 2018, with diagnosis, survival, classification, management, risk assessment, and EUS among the major themes.^[9]

The present analysis adds approximately 5 years of rapidly evolving literature, applies manual relevance screening, and combines Biblioshiny, VOSviewer, and CiteSpace to examine centrality, citation bursts, dual-map pathways, temporal overlays, and cluster timelines. It therefore provides an updated and methodologically broader map of the field, particularly of developments that occurred after the emergence of large-scale CT detection systems, longitudinal EHR models, and biologically informed AI applications. Direct numerical comparisons with the earlier analysis should nevertheless be avoided because the search terminology, screening thresholds, document eligibility criteria, database export dates, and bibliographic preprocessing were not identical.

The publication curves in Figures 2 and 3 indicate that AI in pancreatic cancer changed from an intermittent research topic into a recognizable scientific field during the past decade. The increase from only 14 publications in 2019 to 164 in 2025 is consistent with broader bibliometric studies showing steep expansion of AI research across oncology and other cancer-specific domains.^[16–19]

This parallel growth suggests that the pancreatic cancer literature is participating in a larger transition in oncology from conventional statistical modeling and manually engineered imaging features toward deep learning, multimodal analysis, and data-intensive prediction. However, the growth observed here is particularly clinically relevant because pancreatic cancer presents problems for which pattern-recognition methods are intuitively attractive: subtle abnormalities may precede conventional radiological diagnosis, early symptoms are nonspecific, screening the average-risk population is not currently feasible, and therapeutic outcomes vary substantially even within apparently similar clinical stages.

Several developments may plausibly have supported the post-2019 acceleration. These include greater availability of digitized CT, magnetic resonance imaging, EUS, histopathology, and routinely collected clinical data; advances in computing infrastructure; increased accessibility of open-source machine-learning frameworks; and growing recognition that imaging and longitudinal health records contain information not fully captured by conventional interpretation. These explanations remain contextual rather than causal because bibliometric data cannot establish why individual studies were initiated. Moreover, a rising publication count does not by itself demonstrate increasing methodological quality. Rapid expansion may include exploratory, retrospective, single-center, or incompletely validated models alongside rigorous multicenter investigations.

The 104 publications recorded in 2026 should not be interpreted as evidence that interest is declining. Data were retrieved on July 5, 2026, and both incomplete calendar-year coverage and indexing delays are likely to affect the observed count. Likewise, the polynomial curves and their coefficients of determination in Figures 2 and 3 summarize the historical pattern but do not validate a forecasting model. Extrapolating them would assume that publication behavior, database indexing, terminology, funding priorities, and technological development remain stable, which is unlikely in a rapidly changing field. The defensible interpretation is therefore that recent output has accelerated substantially, not that a particular future publication trajectory has been established.

The country-level findings demonstrate that productivity and network-bridging influence represent different dimensions of scientific activity. China contributed 222 publications, compared with 150 from the United States, but the United States had substantially higher betweenness centrality than China, at 0.39 versus 0.14. England, India, and Italy also displayed relatively high centrality despite lower absolute publication volumes. In bibliometric terms, publication count reflects contribution volume, whereas centrality estimates how often a country connects otherwise less-connected parts of the collaboration network. Thus, China’s position indicates a major capacity for producing AI–pancreatic cancer research, while the United States appears to have played a more prominent role in linking national research communities. Neither measure directly evaluates methodological rigor, originality, clinical relevance, or study quality.

The prominence of China and the United States is consistent with broader AI-oncology analyses, which have repeatedly identified these countries as major contributors.^[16–18] However, the present ranking differs from the earlier pancreatic cancer-specific analysis by Yin et al, in which the United States had the highest publication volume through 2021.^[9]

One possible interpretation is that Chinese publication activity increased more rapidly during the subsequent period. Differences in search strategy, manual screening, country attribution, and the treatment of multicountry publications could also contribute; consequently, the change should not be framed as a definitive transfer of scientific leadership.

The collaboration map in Figure 4 and centrality network in Figure 5 also have implications beyond bibliometric prominence. AI models are especially vulnerable to dataset shift caused by differences in disease prevalence, scanner manufacturers, acquisition protocols, referral patterns, annotation practices, and patient characteristics. International partnerships can therefore provide access to scientifically important heterogeneity rather than merely increasing publication counts. Collaboration between high-output and high-centrality systems could facilitate geographically external validation, comparison across health systems, and assessment in populations that differ from the model-development cohort. The participation of 59 countries demonstrates broad international interest, but the distribution remains uneven. Because the study did not normalize output by population size, pancreatic cancer burden, research expenditure, or number of active institutions, the analysis cannot determine whether particular regions are underperforming relative to their resources or disease burden.

The institutional results reinforce the distinction between concentration of production and network connectivity. Shanghai Jiao Tong University, Zhejiang University, and the Mayo Clinic were the most productive institutions, with 20, 19, and 17 publications, respectively. In contrast, Harvard University had the highest institutional centrality at 0.22, while the University of California System, Mayo Clinic, and Johns Hopkins University each had a value of 0.13. This pattern suggests that the field combines high-output institutional groups, particularly within China, with major American academic medical centers that frequently connect collaborative communities. Because university systems, medical schools, affiliated hospitals, and cancer centers may be indexed inconsistently, small differences between institutional rankings should not be overinterpreted.

The journal distribution further illustrates the field’s interdisciplinary character. Frontiers in Oncology published the largest number of articles, followed by Cancers and Scientific Reports, whereas Cancers, Scientific Reports, and Frontiers in Oncology were the most locally cited sources. These measures describe different relationships: publication count captures the volume disseminated by a journal, while local citations measure how often each source was cited within the 583-publication corpus. Neither is equivalent to journal impact factor, peer-review quality, or the methodological quality of individual studies. The prominence of Gastroenterology, Radiology, European Radiology, Pancreas, and CA: A Cancer Journal for Clinicians among locally cited sources indicates that the intellectual base extends beyond the journals publishing the greatest number of AI-specific articles.

The dual-map overlay in Figure 10 adds conceptual information that cannot be obtained from journal rankings alone. Its principal pathways connect clinical medicine and molecular biology on the citing side with health, medicine, molecular biology, and genetics on the cited side and include additional contributions from computing and other scientific domains. This indicates bidirectional dependence: computational approaches are being applied to clinically defined problems, while model development is informed by disease biology, epidemiology, imaging, and clinical outcome literature. Future progress is therefore likely to depend on sustained collaboration among data scientists, radiologists, gastroenterologists, pathologists, surgeons, oncologists, molecular scientists, epidemiologists, and implementation researchers.

Author-level indicators should also be interpreted independently. Yun Bian, Jianping Lu, and Chengwei Shao were the most productive authors, but Adrian Săftoiu, Suresh T. Chari, and Ajit H. Goenka had the highest citation totals. R. L. Siegel position as the most frequently co-cited author reflects the importance of cancer epidemiology and disease-burden literature to the intellectual framing of the field, rather than direct authorship of the greatest number of pancreatic AI models. Figure 11 shows several relatively distinct co-authorship clusters, including a dense group around Bian, Lu, and Shao and smaller groups around other investigators. This is compatible with the existence of productive local research teams and partial fragmentation between groups, although the displayed network alone cannot establish that collaboration is inadequate. It does identify an opportunity for broader cross-cluster partnerships, especially where complementary expertise or independent datasets could support external validation.

Table 4 and Figure 12 together describe a sequence of methodological and translational milestones rather than a simple hierarchy of citation counts. Early work by Săftoiu et al demonstrated that neural-network analysis of dynamic EUS elastography could help distinguish pancreatic malignancy from chronic pancreatitis.^[20] Its continued citation prominence reflects the longstanding clinical importance of differentiating malignant masses from benign inflammatory disease and shows that AI in pancreatic cancer did not originate exclusively with modern CT-based deep learning. Ko et al subsequently combined machine learning with nanofluidic exosome analysis, extending computational classification beyond imaging and providing an early link between AI and liquid-biopsy research.^[21]

The strong citation bursts for Gillies et al^[22] and van Griethuysen et al^[23] indicate that the field’s growth depended heavily on general radiomics infrastructure and concepts. Gillies et al articulated the view that medical images can be analyzed as quantitative data rather than solely as visual pictures, while the PyRadiomics system developed by van Griethuysen et al provided an open-source platform for standardized feature extraction. Their prominence in Figure 12 therefore reflects methodological dependence on transferable tools developed outside pancreatic oncology itself. At the same time, radiomic features can be sensitive to image acquisition, reconstruction, segmentation, and preprocessing; methodological influence should not be interpreted as proof that every downstream radiomics model is reproducible or clinically valid.

The influential 2019 studies mark a second transition. Hussein et al examined supervised and unsupervised deep-learning approaches for pancreatic and pulmonary tumor characterization^[24]; Kuwahara et al applied deep learning to malignancy assessment in IPMNs^[25]; and Nasief et al used delta-radiomics to predict treatment response.^[26] Collectively, these studies broadened the field from binary differentiation of solid masses toward representation learning, cystic-lesion risk assessment, and longitudinal evaluation of treatment-related imaging changes.

The most recent highly cited studies represent movement toward earlier and more scalable detection. Mukherjee et al investigated prediagnostic CT examinations, asking whether disease-associated patterns could be identified before clinical diagnosis.^[6] Cao et al developed a large-scale noncontrast CT system,^[5] while Chen et al evaluated deep-learning detection in a nationwide population-based setting.^[27]

Placido et al moved beyond image interpretation by modeling longitudinal disease trajectories in national health-registry data.^[7] These works collectively shift the central question from “Can AI classify an established pancreatic lesion?” toward “Can AI identify individuals or subtle abnormalities before conventional diagnosis?” Kenner et al summative review also became highly cited because it consolidated early-detection applications across imaging, biomarkers, and risk prediction.^[28]

The ongoing citation bursts from 2024 through 2026 for Placido et al and for publications addressing pancreatic biology, epidemiology, and clinical applications suggest that scholarly attention is broadening beyond technical model development. Nevertheless, citation bursts identify concentrated attention, not successful clinical adoption. Recent publications also have less time to accrue citations, while older methodological papers may be cited across many cancer types. Table 4 and Figure 12 should therefore be interpreted as complementary representations of visibility, foundational dependence, and changing scientific priorities.

The keyword and timeline analyses support a 3-stage interpretation of thematic development. The earliest phase centered on artificial neural networks, EUS, elastography, diagnostic differentiation, and relatively focused lesion-classification tasks. The second phase was characterized by CT, texture analysis, radiomics, deep learning, and prediction models. The most recent phase includes early detection, longitudinal risk prediction, liquid biopsy, exosomes, prognostic modeling, tumor microenvironment, immunotherapy, molecular subtypes, metastasis, treatment response, and patient outcomes. This progression is consistent with movement from modality-specific computer-aided diagnosis toward broader precision-oncology questions.

The dominant keywords nevertheless show continuity as well as change. “Pancreatic cancer,” “machine learning,” “artificial intelligence,” “diagnosis,” “pancreatic ductal adenocarcinoma,” “deep learning,” and “radiomics” remained central and highly interconnected in Figures 13 to 16. Newer terms have therefore not replaced imaging and diagnosis; rather, they have expanded the field around an established diagnostic core. Prognosis, survival, biomarkers, classification, and prediction link the earlier imaging-focused literature to newer efforts involving biological characterization and clinical outcomes.

The 13 CiteSpace clusters provide additional detail but require restrained interpretation. Clusters relating to chronic pancreatitis, artificial neural networks, deep learning, diagnostic performance, and pancreatic adenocarcinoma are coherent with the earlier emphasis on differentiation and classification. “Routine data” may represent increasing use of EHRs, clinical histories, laboratory results, or routinely acquired imaging, although study-level inspection is needed to establish its exact composition. “Using exosome,” “perineural invasion,” and “patient outcome” plausibly indicate expansion toward biomarkers, biological behavior, and clinically meaningful prediction. The pancreatic neuroendocrine tumor cluster also shows that the retrieved literature is not limited exclusively to PDAC.

The label “conversational artificial intelligence” warrants particular caution. It was generated automatically through the CiteSpace log-likelihood ratio procedure and appears as a relatively recent, small cluster in Figures 17 and 18. Without examining its constituent publications, it should not be treated as evidence that large language models, generative AI, or conversational clinical systems have become an established pancreatic cancer research front. At most, it identifies a candidate emerging theme requiring study-level confirmation.

Overall, the thematic expansion may indicate increasing clinical ambition, but it does not by itself establish clinical maturity. A literature can become more multimodal and outcome-oriented while remaining predominantly retrospective. Similarly, frequent use of terms such as “personalized management” or “early detection” does not demonstrate that models improve decisions or patient outcomes. The bibliometric evidence supports a change in research emphasis, whereas clinical effectiveness requires prospective evaluation.

The mapped research priorities closely correspond to major unresolved clinical problems in pancreatic cancer. Routine screening is not recommended for asymptomatic average-risk adults,^[2] while surveillance is primarily directed toward selected individuals with substantial inherited or familial risk.^[6,29] AI-based risk models could enrich a population for targeted investigation, but this role differs fundamentally from deploying an algorithm as a population screening test. Risk thresholds would need to reflect disease prevalence, downstream imaging capacity, false-positive harms, and the consequences of invasive investigation.

Imaging AI may be particularly relevant to opportunistic analysis of CT examinations performed for other indications, detection of subtle prediagnostic abnormalities, characterization of pancreatic cystic lesions, and differentiation of PDAC from chronic pancreatitis. EUS-based systems may support interpretation of diagnostically uncertain masses, whereas pathology and molecular models could contribute to subtype classification, perineural invasion assessment, or treatment selection. Prognostic and delta-radiomics studies suggest further applications in recurrence estimation and response monitoring. The bibliometric shift toward exosomes and other biomarkers raises the possibility that imaging, longitudinal clinical data, and blood-based signals may eventually be combined rather than evaluated in isolation.

However, the central translational gap is between retrospective discrimination and prospective clinical utility. A high area under the receiver operating characteristic curve does not establish adequate positive predictive value in a low-prevalence population, reliable calibration, acceptable workload, or improved survival. Models can appear highly accurate because of spectrum bias, inappropriate control selection, data leakage, overlapping patients, acquisition-specific artifacts, or unrecognized differences between development and test cohorts. Even multicenter retrospective validation may not reproduce the intended workflow because clinicians’ responses to model outputs can alter performance and downstream care.

PANORAMA represents an important step toward clinically oriented evaluation because it compared AI and radiologists in an international, confirmatory observational design.^[8] Nevertheless, diagnostic performance in an observational study is not equivalent to evidence that AI changes management appropriately or improves patient outcomes. Subsequent evaluations should assess incremental benefit over expert interpretation, performance in consecutive and representative patients, calibration at clinically meaningful thresholds, time to diagnosis, unnecessary investigations, resectability at detection, treatment decisions, quality of life, cost, and survival.

Methodological and reporting standards can provide a structured pathway for this work. Prediction-model studies should follow Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis–AI^[30] and be appraised using Prediction model Risk of Bias ASsessment Tool–AI.^[31] Imaging studies should address the Checklist for AI in Medical Imaging 2024,^[32] while diagnostic-accuracy investigations should follow Standards for Reporting Diagnostic Accuracy Studies–AI.^[33] Early live clinical evaluations should consider Developmental and Exploratory Clinical Investigations of Decision Support Systems Driven by AI^[34] and clinical trial protocols and reports should follow Standard Protocol Items: Recommendations for Interventional Trials–AI^[35] and Consolidated Standards of Reporting Trials–AI.^[36] Fair, Universal, Traceable, Usable, Robust, and Explainable AI additionally emphasizes trustworthy and deployable systems across the full life cycle, including fairness, universality, traceability, usability, robustness, and explainability.^[37]

Beyond transparent reporting, clinical translation requires independent external validation, harmonized acquisition and preprocessing, explicit prevention of data leakage, reproducible code and model documentation, subgroup-specific performance, and calibration assessment. Decision-curve or related utility analyses should examine whether predictions produce net clinical benefit at plausible intervention thresholds. Explainability should be evaluated according to its clinical purpose rather than included as a purely visual add-on. Fairness analyses should examine performance across sex, age, ethnicity, socioeconomic circumstances, comorbidity, disease stage, scanner type, hospital setting, and geographic region where data permit.

Implementation also raises questions that conventional model-performance studies rarely answer. These include how radiologists or gastroenterologists should respond to discrepant human and algorithmic assessments; whether automation bias or alert fatigue occurs; who is responsible for monitoring performance drift; how patient consent, privacy, and secondary data use are governed; and whether the intervention is affordable relative to its clinical benefit. The relevant endpoint is therefore not simply whether an algorithm discriminates cases from controls, but whether a human–AI care pathway performs more safely, equitably, efficiently, and effectively than current practice.

4.1. Future research priorities

Near-term research should prioritize reproducible external validation and prospective multicenter evaluation of existing high-performing models rather than continued proliferation of closely related single-center algorithms. Consecutive cohorts, prespecified analysis plans, locked models, transparent exclusions, calibration measures, and clinically defined thresholds are required. International studies should intentionally include variation in scanner equipment, imaging protocols, referral pathways, patient demographics, and disease prevalence. Silent prospective deployment, in which model outputs are recorded without influencing care, may provide a practical intermediate step before interventional evaluation.

A second near-term priority is harmonization. Common data dictionaries, standardized imaging and pathology pipelines, transparent segmentation procedures, documented outcome definitions, and accessible benchmark datasets would improve comparison and replication. Code sharing, model cards, detailed reporting of hyperparameter selection, and clear identification of overlapping patient cohorts would reduce avoidable duplication. Where unrestricted data sharing is impossible, federated or other privacy-preserving approaches may support multicenter learning, although these methods still require governance, harmonization, and evaluation for site-specific bias.

Longer-term work should test genuinely multimodal systems combining imaging with EHR trajectories, pathology, genomics, transcriptomics, proteomics, metabolomics, circulating biomarkers, or exosome profiles. Potentially valuable use cases include surveillance of genetically or familially high-risk individuals, enrichment of patients with new-onset diabetes or concerning clinical trajectories, characterization of cystic lesions, prediction of neoadjuvant response, recurrence assessment, and treatment selection. These systems should be compared with simpler models to determine whether additional modalities provide meaningful incremental benefit rather than technical complexity alone.

Ultimately, prospective impact studies and, where justified, randomized trials should determine whether AI changes clinical decisions and improves outcomes. Economic evaluation should be incorporated early, particularly for low-prevalence detection pathways in which false-positive investigations may be costly or harmful. Future studies should also report algorithmic failures, subgroup inequities, workflow effects, and performance drift, not only average discrimination. Foundation models and conversational systems may eventually contribute to data integration or clinical communication, but the present cluster evidence is insufficient to identify them as mature priorities in pancreatic cancer.

Strengths, interpretive cautions, and transition to limitations. A major strength of the present study is the integration of 3 complementary bibliometric platforms with manual screening of the retrieved literature. This enabled simultaneous assessment of performance indicators, collaboration structures, centrality, co-citation, citation bursts, journal pathways, keyword evolution, and cluster timelines through July 5, 2026. The analysis therefore captures both the historical foundations and the recent expansion of the field.

Several cautions directly affect interpretation. Citation counts indicate visibility rather than quality; centrality depends on network construction; automatically generated cluster labels are approximate; and partial-year data preclude interpreting 2026 as a downturn. Exclusion of non-English publications, conference proceedings, and preprints may be especially consequential in AI, where technical work is frequently disseminated outside conventional clinical journals. Most importantly, bibliometric methods cannot determine risk of bias, calibration, fairness, reproducibility, transportability, or patient benefit. These issues are addressed more fully in the dedicated Section 4.2.

4.2. Limitations

Several limitations should be considered when interpreting the findings of this bibliometric study. First, the analysis was based exclusively on the Web of Science Core Collection and included only English-language original research articles and reviews. Consequently, relevant publications indexed in other databases, non-English studies, conference proceedings, preprints, and other gray literature may have been missed. This is particularly important in AI research, where novel technical developments are often first reported in conference proceedings or on preprint platforms.

Second, the completeness of the dataset depended on the terminology used in the search strategy. Studies using related terms such as computational imaging, computer-aided diagnosis, digital pathology, pattern recognition, or foundation models may not have been retrieved if explicit AI-related terminology was absent. Conversely, broad search terms may have captured studies in which AI represented only a secondary component.

Third, citation-based indicators should not be interpreted as direct measures of scientific quality or clinical relevance. Citation counts are influenced by publication age, journal visibility, database coverage, self-citation, and disciplinary citation practices. Likewise, publication output, citation frequency, and network centrality reflect different dimensions of scientific activity and should not be considered interchangeable. High productivity does not necessarily indicate greater methodological quality, while high centrality primarily reflects a bridging position within collaboration networks.

Fourth, country, institution, and author-level analyses may be affected by inconsistencies in affiliation data and author-name disambiguation. Variations in institutional names, hospital or university affiliations, transliteration, organizational restructuring, shared surnames, initials, and changes in author affiliations may lead to inappropriate splitting or merging of contributors and thereby influence productivity and collaboration rankings.

Fifth, thematic and network analyses were influenced by the bibliographic fields, preprocessing decisions, and software parameters used. Related terms such as “pancreatic cancer,” “pancreatic ductal adenocarcinoma,” and “PDAC” may have been treated separately, potentially fragmenting closely related themes. In addition, CiteSpace cluster labels were generated automatically and should therefore be interpreted as approximate thematic descriptions rather than definitive classifications. Network centrality estimates are similarly sensitive to node-selection criteria, thresholds, time intervals, and other analytical settings.

Sixth, data for 2026 represent only part of the calendar year and may also be affected by indexing delays. The lower publication count in 2026 should therefore not be interpreted as evidence of declining research activity. Similarly, the polynomial trend lines were used only to describe historical publication patterns and should not be regarded as validated forecasting models.

Finally, bibliometric analysis cannot determine whether individual AI models are methodologically robust, reproducible, unbiased, externally validated, or clinically effective. The present study did not assess risk of bias, data leakage, model calibration, fairness, reporting quality, external validation, or prospective clinical performance. In addition, multiple publications derived from overlapping patient cohorts or successive versions of similar AI models may have been counted separately, potentially inflating the apparent productivity or influence of particular research groups. Accordingly, the findings should be viewed as a map of the structure and evolution of the literature rather than as an assessment of the clinical validity or methodological quality of individual studies.

5. Conclusion

AI research in pancreatic cancer has expanded rapidly, particularly since 2019, and has progressed from isolated diagnostic applications toward broader uses in early detection, prognosis, treatment-response prediction, and precision oncology. The field is highly multidisciplinary and internationally active, although scientific productivity and collaborative influence remain uneven across countries and institutions.

Despite promising technical performance, most AI models are not yet ready for routine clinical use. Future studies should emphasize prospective multicenter validation, standardized and representative datasets, transparency, calibration, fairness, cost-effectiveness, and measurable improvements in patient care. The clinical value of AI will ultimately depend on whether it can safely support earlier diagnosis, personalized treatment, and better patient outcomes.

Author contributions

Conceptualization: Pegah Rashidian, Ehsan Amini-Salehi.

Data curation: Pegah Rashidian.

Formal analysis: Seyedsina Moghimnejadhosseini, Ehsan Amini-Salehi.

Investigation: Seyedsina Moghimnejadhosseini, Amirmahdi Mojtahedzadeh.

Methodology: Pegah Rashidian, Djaloliddin Mansurov, Mohammad Ahmar Khan.

Software: Seyedsina Moghimnejadhosseini, Mohammad Ahmar Khan, Ehsan Amini-Salehi.

Supervision: Ehsan Amini-Salehi.

Validation: Mahsa Talebzadeh, Ehsan Amini-Salehi.

Visualization: Seyedsina Moghimnejadhosseini, Amirmahdi Mojtahedzadeh, Mahsa Talebzadeh.

Writing – original draft: Pegah Rashidian, Aseel Smerat, Mohammad Ahmar Khan, Pankaj Bansal, Chandan Sharma, Djaloliddin Mansurov, Seyedsina Moghimnejadhosseini, Ehsan Amini-Salehi.

Writing – review & editing: Pankaj Bansal, Chandan Sharma, Djaloliddin Mansurov, Amirmahdi Mojtahedzadeh, Mahsa Talebzadeh, Ehsan Amini-Salehi.

medi-105-e50681-s001.docx^{(14.9KB, docx)}

Abbreviations:

AI: artificial intelligence
CT: computed tomography
EHR: electronic health record
EUS: endoscopic ultrasonography
PDAC: pancreatic ductal adenocarcinoma

The authors have no funding and conflicts of interest to declare.

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Supplemental Digital Content is available in the online version of this article (http://dx.doi.org/10.1097/MD.0000000000050681).

How to cite this article: Rashidian P, Smerat A, Ahmar Khan M, Bansal P, Sharma C, Mansurov D, Moghimnejadhosseini S, Mojtahedzadeh A, Talebzadeh M, Amini-Salehi E. Mapping the global landscape of artificial intelligence in pancreatic cancer research: A bibliometric and visualization analysis. Medicine 2026;105:39(e50681).

ChatGPT (OpenAI) was utilized exclusively to assist with linguistic editing of the manuscript, including improvement of grammar, sentence structure, punctuation, and overall readability. The artificial intelligence tool did not contribute to the conception or design of the study, data acquisition, bibliometric analyses, interpretation of results, scientific decision-making, or reference selection and citation generation. All scientific content, interpretations, conclusions, references, and citations were independently developed, verified, and approved by the authors. The authors reviewed the manuscript in its entirety and assume full responsibility for the authenticity, accuracy, and integrity of the submitted work.

Contributor Information

Pegah Rashidian, Email: pegah.rashidian@rocketmail.com.

Aseel Smerat, Email: smerat.2020@gmail.com.

Mohammad Ahmar Khan, Email: mkhan@du.edu.om.

Pankaj Bansal, Email: pankaj.bansal1@sharda.ac.in.

Chandan Sharma, Email: chandansharma51@outlook.com.

Djaloliddin Mansurov, Email: djaloliddinmansurov@proton.me.

Seyedsina Moghimnejadhosseini, Email: seyedsina.moghimnejadhosseini@stud.semmelweis.hu.

Amirmahdi Mojtahedzadeh, Email: mojtahedzadeh.amirmahdi@gmail.com.

Mahsa Talebzadeh, Email: mahsa.talebzadeh@stud.semmelweis.hu.

References

[1].Stoop TF, Javed AA, Oba A, et al. Pancreatic cancer. Lancet. 2025;405:1182–202. [DOI] [PubMed] [Google Scholar]
[2].US Preventive Services Task Force; Owens DK, Davidson KW, Krist AH, et al. Screening for Pancreatic Cancer: US Preventive Services Task Force Reaffirmation Recommendation Statement. JAMA. 2019;322:438–44. [DOI] [PubMed] [Google Scholar]
[3].Hayashi H, Uemura N, Matsumura K, et al. Recent advances in artificial intelligence for pancreatic ductal adenocarcinoma. World J Gastroenterol. 2021;27:7480–96. [DOI] [PMC free article] [PubMed] [Google Scholar]
[4].Huang B, Huang H, Zhang S, et al. Artificial intelligence in pancreatic cancer. Theranostics. 2022;12:6931–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
[5].Cao K, Xia Y, Yao J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. 2023;29:3033–43. [DOI] [PMC free article] [PubMed] [Google Scholar]
[6].Mukherjee S, Patra A, Khasawneh H, et al. Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before clinical diagnosis. Gastroenterology. 2022;163:1435–46.e3. [DOI] [PMC free article] [PubMed] [Google Scholar]
[7].Placido D, Yuan B, Hjaltelin JX, et al. A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nat Med. 2023;29:1113–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
[8].Alves N, Schuurmans M, Rutkowski D, et al. Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study. Lancet Oncol. 2026;27:116–24. [DOI] [PubMed] [Google Scholar]
[9].Yin H, Zhang F, Yang X, et al. Research trends of artificial intelligence in pancreatic cancer: a bibliometric analysis. Front Oncol. 2022;12:973999. [DOI] [PMC free article] [PubMed] [Google Scholar]
[10].van Eck NJ, Waltman L. Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics. 2010;84:523–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
[11].Waltman L, van Eck NJ, Noyons ECM. A unified approach to mapping and clustering of bibliometric networks. J Inform. 2010;4:629–35. [Google Scholar]
[12].Chen C. CiteSpace II: detecting and visualizing emerging trends and transient patterns in scientific literature. J Am Soc Inf Sci Technol. 2006;57:359–77. [Google Scholar]
[13].Chen C, Ibekwe-SanJuan F, Hou J. The structure and dynamics of cocitation clusters: a multiple-perspective cocitation analysis. J Am Soc Inf Sci Technol. 2010;61:1386–409. [Google Scholar]
[14].Aria M, Cuccurullo C. bibliometrix: an R-tool for comprehensive science mapping analysis. J Inform. 2017;11:959–75. [Google Scholar]
[15].Mondal H. A technical note on bibliometric analysis by biblioshiny and VOSviewer. Indian J Radiol Imaging. 2026;36:167–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
[16].Wu T, Duan Y, Zhang T, Tian W, Liu H, Deng Y. Research trends in the application of artificial intelligence in oncology: a bibliometric and network visualization Study. Front Biosci (Landmark Ed). 2022;27:254. [DOI] [PubMed] [Google Scholar]
[17].Wang SH, Chen G, Zhong X, et al. Global development of artificial intelligence in cancer field: a bibliometric analysis range from 1983 to 2022. Front Oncol. 2023;13:1215729. [DOI] [PMC free article] [PubMed] [Google Scholar]
[18].Lyu PF, Wang Y, Meng QX, et al. Mapping intellectual structures and research hotspots in the application of artificial intelligence in cancer: a bibliometric analysis. Front Oncol. 2022;12:955668. [DOI] [PMC free article] [PubMed] [Google Scholar]
[19].Sun L, Zhang R, Gu Y, Huang L, Jin C. Application of Artificial Intelligence in the diagnosis and treatment of colorectal cancer: a bibliometric analysis, 2004–2023. Front Oncol. 2024;14:1424044. [DOI] [PMC free article] [PubMed] [Google Scholar]
[20].Saftoiu A, Vilmann P, Gorunescu F, et al. Neural network analysis of dynamic sequences of EUS elastography used for the differential diagnosis of chronic pancreatitis and pancreatic cancer. Gastrointest Endosc. 2008;68:1086–94. [DOI] [PubMed] [Google Scholar]
[21].Ko J, Bhagwat N, Yee SS, et al. Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes. ACS Nano. 2017;11:11182–93. [DOI] [PubMed] [Google Scholar]
[22].Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278:563–77. [DOI] [PMC free article] [PubMed] [Google Scholar]
[23].van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77:e104–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
[24].Hussein S, Kandel P, Bolan CW, Wallace MB, Bagci U. Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches. IEEE Trans Med Imaging. 2019;38:1777–87. [DOI] [PubMed] [Google Scholar]
[25].Kuwahara T, Hara K, Mizuno N, et al. Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas. Clin Transl Gastroenterol. 2019;10:1–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
[26].Nasief H, Zheng C, Schott D, et al. A machine learning based delta-radiomics process for early prediction of treatment response of pancreatic cancer. npj Precis Oncol. 2019;3:25. [DOI] [PMC free article] [PubMed] [Google Scholar]
[27].Chen PT, Wu T, Wang P, et al. Pancreatic cancer detection on CT scans with deep learning: a nationwide population-based study. Radiology. 2023;306:172–82. [DOI] [PubMed] [Google Scholar]
[28].Kenner B, Chari ST, Kelsen D, et al. Artificial intelligence and early detection of pancreatic cancer: 2020 summative review. Pancreas. 2021;50:251–79. [DOI] [PMC free article] [PubMed] [Google Scholar]
[29].Cheng HH, Giri VN, Goggins M, et al. NCCN guidelines(R) insights: genetic/familial high-risk assessment: breast, ovarian, pancreatic, and prostate, version 2.2026. J Natl Compr Canc Netw. 2026;24:2–10. [DOI] [PubMed] [Google Scholar]
[30].Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. [DOI] [PMC free article] [PubMed] [Google Scholar]
[31].Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505. [DOI] [PMC free article] [PubMed] [Google Scholar]
[32].Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for artificial intelligence in medical imaging (CLAIM): 2024 update. Radiol Artif Intell. 2024;6:e240300. [DOI] [PMC free article] [PubMed] [Google Scholar]
[33].Sounderajah V, Guni A, Liu X, et al. The STARD-AI reporting guideline for diagnostic accuracy studies using artificial intelligence. Nat Med. 2025;31:3283–9. [DOI] [PubMed] [Google Scholar]
[34].Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28:924–33. [DOI] [PubMed] [Google Scholar]
[35].Cruz Rivera S, Liu X, Chan AW, et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med. 2020;26:1351–63. [DOI] [PMC free article] [PubMed] [Google Scholar]
[36].Liu X, Cruz Rivera S, Moher D, et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med. 2020;26:1364–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
[37].Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ. 2025;388:e081554. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

medi-105-e50681-s001.docx^{(14.9KB, docx)}

[R1] [1].Stoop TF, Javed AA, Oba A, et al. Pancreatic cancer. Lancet. 2025;405:1182–202. [DOI] [PubMed] [Google Scholar]

[R2] [2].US Preventive Services Task Force; Owens DK, Davidson KW, Krist AH, et al. Screening for Pancreatic Cancer: US Preventive Services Task Force Reaffirmation Recommendation Statement. JAMA. 2019;322:438–44. [DOI] [PubMed] [Google Scholar]

[R3] [3].Hayashi H, Uemura N, Matsumura K, et al. Recent advances in artificial intelligence for pancreatic ductal adenocarcinoma. World J Gastroenterol. 2021;27:7480–96. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] [4].Huang B, Huang H, Zhang S, et al. Artificial intelligence in pancreatic cancer. Theranostics. 2022;12:6931–54. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R5] [5].Cao K, Xia Y, Yao J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning. Nat Med. 2023;29:3033–43. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R6] [6].Mukherjee S, Patra A, Khasawneh H, et al. Radiomics-based machine-learning models can detect pancreatic cancer on prediagnostic computed tomography scans at a substantial lead time before clinical diagnosis. Gastroenterology. 2022;163:1435–46.e3. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R7] [7].Placido D, Yuan B, Hjaltelin JX, et al. A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nat Med. 2023;29:1113–22. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R8] [8].Alves N, Schuurmans M, Rutkowski D, et al. Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study. Lancet Oncol. 2026;27:116–24. [DOI] [PubMed] [Google Scholar]

[R9] [9].Yin H, Zhang F, Yang X, et al. Research trends of artificial intelligence in pancreatic cancer: a bibliometric analysis. Front Oncol. 2022;12:973999. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R10] [10].van Eck NJ, Waltman L. Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics. 2010;84:523–38. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R11] [11].Waltman L, van Eck NJ, Noyons ECM. A unified approach to mapping and clustering of bibliometric networks. J Inform. 2010;4:629–35. [Google Scholar]

[R12] [12].Chen C. CiteSpace II: detecting and visualizing emerging trends and transient patterns in scientific literature. J Am Soc Inf Sci Technol. 2006;57:359–77. [Google Scholar]

[R13] [13].Chen C, Ibekwe-SanJuan F, Hou J. The structure and dynamics of cocitation clusters: a multiple-perspective cocitation analysis. J Am Soc Inf Sci Technol. 2010;61:1386–409. [Google Scholar]

[R14] [14].Aria M, Cuccurullo C. bibliometrix: an R-tool for comprehensive science mapping analysis. J Inform. 2017;11:959–75. [Google Scholar]

[R15] [15].Mondal H. A technical note on bibliometric analysis by biblioshiny and VOSviewer. Indian J Radiol Imaging. 2026;36:167–74. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] [16].Wu T, Duan Y, Zhang T, Tian W, Liu H, Deng Y. Research trends in the application of artificial intelligence in oncology: a bibliometric and network visualization Study. Front Biosci (Landmark Ed). 2022;27:254. [DOI] [PubMed] [Google Scholar]

[R17] [17].Wang SH, Chen G, Zhong X, et al. Global development of artificial intelligence in cancer field: a bibliometric analysis range from 1983 to 2022. Front Oncol. 2023;13:1215729. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] [18].Lyu PF, Wang Y, Meng QX, et al. Mapping intellectual structures and research hotspots in the application of artificial intelligence in cancer: a bibliometric analysis. Front Oncol. 2022;12:955668. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] [19].Sun L, Zhang R, Gu Y, Huang L, Jin C. Application of Artificial Intelligence in the diagnosis and treatment of colorectal cancer: a bibliometric analysis, 2004–2023. Front Oncol. 2024;14:1424044. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] [20].Saftoiu A, Vilmann P, Gorunescu F, et al. Neural network analysis of dynamic sequences of EUS elastography used for the differential diagnosis of chronic pancreatitis and pancreatic cancer. Gastrointest Endosc. 2008;68:1086–94. [DOI] [PubMed] [Google Scholar]

[R21] [21].Ko J, Bhagwat N, Yee SS, et al. Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes. ACS Nano. 2017;11:11182–93. [DOI] [PubMed] [Google Scholar]

[R22] [22].Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278:563–77. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] [23].van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77:e104–7. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] [24].Hussein S, Kandel P, Bolan CW, Wallace MB, Bagci U. Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches. IEEE Trans Med Imaging. 2019;38:1777–87. [DOI] [PubMed] [Google Scholar]

[R25] [25].Kuwahara T, Hara K, Mizuno N, et al. Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas. Clin Transl Gastroenterol. 2019;10:1–8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R26] [26].Nasief H, Zheng C, Schott D, et al. A machine learning based delta-radiomics process for early prediction of treatment response of pancreatic cancer. npj Precis Oncol. 2019;3:25. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] [27].Chen PT, Wu T, Wang P, et al. Pancreatic cancer detection on CT scans with deep learning: a nationwide population-based study. Radiology. 2023;306:172–82. [DOI] [PubMed] [Google Scholar]

[R28] [28].Kenner B, Chari ST, Kelsen D, et al. Artificial intelligence and early detection of pancreatic cancer: 2020 summative review. Pancreas. 2021;50:251–79. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] [29].Cheng HH, Giri VN, Goggins M, et al. NCCN guidelines(R) insights: genetic/familial high-risk assessment: breast, ovarian, pancreatic, and prostate, version 2.2026. J Natl Compr Canc Netw. 2026;24:2–10. [DOI] [PubMed] [Google Scholar]

[R30] [30].Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] [31].Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R32] [32].Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for artificial intelligence in medical imaging (CLAIM): 2024 update. Radiol Artif Intell. 2024;6:e240300. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] [33].Sounderajah V, Guni A, Liu X, et al. The STARD-AI reporting guideline for diagnostic accuracy studies using artificial intelligence. Nat Med. 2025;31:3283–9. [DOI] [PubMed] [Google Scholar]

[R34] [34].Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28:924–33. [DOI] [PubMed] [Google Scholar]

[R35] [35].Cruz Rivera S, Liu X, Chan AW, et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med. 2020;26:1351–63. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] [36].Liu X, Cruz Rivera S, Moher D, et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med. 2020;26:1364–74. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R37] [37].Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ. 2025;388:e081554. [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Mapping the global landscape of artificial intelligence in pancreatic cancer research

Pegah Rashidian, MD

Aseel Smerat, PhD

Mohammad Ahmar Khan, PhD

Pankaj Bansal, MD

Chandan Sharma, MD

Djaloliddin Mansurov, MD

Seyedsina Moghimnejadhosseini, MD

Amirmahdi Mojtahedzadeh, MD

Mahsa Talebzadeh, MD

Ehsan Amini-Salehi, MD

Abstract

Background:

Methods:

Results:

Conclusion:

1. Introduction

2. Methods

2.1. Data collection

2.2. Eligibility criteria and study selection

Figure 1.

2.3. Data analysis

3. Results

3.1. Publication trends

Figure 2.

Figure 3.

3.2. Countries and institutions

Figure 4.

Table 1.

Figure 5.

Table 2.

Figure 6.

3.3. Journals and co-cited sources

Figure 7.

Figure 8.

Figure 9.

Figure 10.

3.4. Authors and co-cited authors

Figure 11.

Table 3.

3.5. Most highly cited articles

Table 4.

3.6. References with the strongest citation bursts

Figure 12.

3.7. Keyword trends

Figure 13.

Figure 14.

Figure 15.

Figure 16.

3.8. Cluster analysis

Figure 17.

Figure 18.

4. Discussion

4.1. Future research priorities

4.2. Limitations

5. Conclusion

Author contributions

Abbreviations:

Contributor Information

References

Associated Data

Supplementary Materials

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases