Bibliometric map shows AI research surge in bone metastasis

Artificial intelligence research focused on bone metastasis expanded rapidly after 2020, reaching 540 publications by June 26, 2026, according to a bibliometric and visualization analysis of English-language articles and reviews in the Web of Science Core Collection, PubMed and Scopus. The study mapped the field’s publication, citation, collaboration and thematic structure from January 2006 through June 2026. It found a transition from computer-assisted diagnosis and bone-scintigraphy analysis toward deep learning, radiomics, prognostic modeling, multimodal prediction and a limited but emerging use of language-based AI. Imaging remained the dominant foundation, while the literature did not provide direct evidence that AI has become established in routine clinical care.

The investigators combined terms for bone or skeletal metastasis with artificial intelligence and related methods, including machine learning, deep learning, radiomics, computer vision, computer-aided diagnosis, neural networks, support vector machines, random forests, image segmentation and large language models. The databases initially returned 1,756 records: 495 from Web of Science, 424 from PubMed and 837 from Scopus. After hierarchical removal of 832 duplicates, 924 unique records underwent title-and-abstract screening by two reviewers, with a third resolving uncertain cases. Another 384 records were excluded, leaving 540 publications. R, VOSviewer and CiteSpace were used to examine annual trends, journals, countries, institutions, authors, citation bursts, keyword co-occurrence, collaboration networks and thematic evolution.

Annual output was limited before 2020 and then rose sharply, producing an overall annual growth rate of 24.01%. The cumulative total rose from 61 publications in 2019 to 540 by the retrieval date, with 101 publications in 2025, the most recent complete calendar year. Seventy-four papers had been included by June 26, 2026, so this partial-year figure should not be compared directly with complete years or interpreted as a decline. The publications involved 3,439 authors, 1,118 author keywords and 16,020 references. The mean publication age was 3.46 years, the average was 10.06 citations, and publications had 8.13 coauthors on average. International coauthorship accounted for 19.63% of the dataset.

China was the largest national contributor, with 233 publications, followed by Japan and the United States with 51 each. Italy produced 19 papers, South Korea 18, Canada 15, Germany 14 and the Netherlands 13. The collaboration network centered on China, the United States, Japan, Italy and South Korea, with additional links among European research groups. Contributions from Africa, South America and parts of Central Asia were limited. China had 204 single-country and 29 multiple-country publications; Japan had 44 and seven, while the United States had a more balanced split of 31 and 20. These results describe the structure of scholarly cooperation, not the quality, generalizability or clinical effectiveness of models developed in any country.

Institutional output was concentrated in several collaborative hubs rather than one fully integrated global network. Northwest Minzu University contributed 64 publications, China Medical University 47, Harvard University 34 and Harvard University Medical Affiliates 31. China Medical University in Taiwan, National Taiwan University and the University of Toronto each produced 28. Network maps identified connections among Chinese institutions and international partners, a separate cluster centered on Harvard Medical School and Massachusetts General Hospital, and smaller groups involving Lund University, the University of Gothenburg, Sahlgrenska University Hospital and institutions in Taiwan. Among authors, Cao YC had 22 publications, Lin Q 21 and Zeng XW 20, while Edenbrandt L had 353 total citations and 96 local citations from 10 publications.

Citation-burst analysis traced a shift in the references attracting attention. The strongest burst, with a strength of 7.69 from 2013 to 2017, was associated with computer-aided diagnosis of bone scintigrams. A review of deep learning in medical image analysis registered a burst strength of 6.42 during 2020–2021, and two deep-learning studies of bone-metastasis classification remained active from 2021 to 2023. The most locally cited work was a 2021 global cancer statistics paper with 46 local citations. The foundational U-Net paper and a study of computer-assisted interpretation of whole-body bone scans each received 41 local citations. Such citation patterns map attention to research themes but do not rank individual methods by clinical effectiveness, methodological rigor or readiness for implementation.

Keyword networks separated the literature into four broad themes: bone metastasis with prognostic and treatment assessment; deep learning, bone scintigraphy and computer-aided diagnosis; prostate cancer with MRI and PET/CT; and broader methods involving convolutional neural networks, segmentation and natural language processing. Bone metastasis was the most frequent author keyword, appearing 216 times, followed by machine learning with 106, deep learning with 103, radiomics with 69, prostate cancer with 68, artificial intelligence with 57 and bone scintigraphy with 51. The chronology mapped an early emphasis on computer-aided diagnosis and quantitative skeletal assessment, followed by deep-learning image analysis, predictive modeling and cancer-specific applications. Prostate cancer formed a sustained branch, whereas research concerning breast-cancer bone metastasis remained comparatively smaller.

Language-based AI appeared only as a limited extension beyond conventional imaging applications. Recent studies used fine-tuned or general-purpose language models to structure radiology reports, extract electronic-health-record variables, assemble cohorts and provide accessible information about bone metastasis and cancer-induced bone pain. In a Chinese multicenter workflow, clinical-logic-guided prompting reduced manual processing time by 74.5%–82.6% while supporting an atlas of 18,331 patients. Another study reported an accuracy of 0.85 for extracting bone-metastasis status from electronic health records with Llama 3.1 8B. These studies demonstrate feasibility for particular information-processing tasks, not autonomous clinical decision-making, and their performance cannot be directly compared because datasets, validation strategies and outcome definitions differed.

The analysis identified data heterogeneity, limited external validation, reproducibility concerns and insufficient prospective evidence as barriers to clinical translation. Most systems address specific tasks such as detection, segmentation, classification, risk prediction or outcome estimation, and many are evaluated with retrospective datasets. Transferring such systems to other hospitals may expose them to differences in scanners, acquisition protocols, reconstruction settings, radiotracers, patient populations and clinical workflows. Retrospective cohorts can also underrepresent some cancers, stages or healthcare settings, while variations in reference standards and annotation procedures can introduce measurement bias. The authors emphasize representative multicenter datasets, transparent reporting of cohort selection and annotation, and evaluation across clinically relevant subgroups.

Prospective evaluation would need to assess more than accuracy, area under the curve or segmentation scores. Relevant measures include calibration, clinical utility, workflow effects and meaningful patient outcomes, alongside the stability of performance across institutions and time. The dataset itself also has limitations: relevant studies outside the three databases or excluded by the English-language restriction may have been missed; network maps depend on term normalization, thresholds and clustering choices; citation indicators favor older publications; and the 2026 record covers only January 1 through June 26. Most importantly, the bibliometric study maps research activity and themes rather than directly comparing model quality or clinical performance. Its central finding is that AI publication activity in bone metastasis is growing quickly, imaging still dominates, and language-based research remains early and limited.

Subject of Research: Bibliometric analysis of artificial intelligence applications in bone metastasis research

Article Title: Mapping the evolution of artificial intelligence applications in bone metastasis: a bibliometric and visualization analysis

Article References: Lin, Y., Ye, S., Wu, J., Hu, Y., Wu, H., He, H., Yang, J., Wang, Y., Ni, B., Zhu, L., Ding, C., Liu, C., Musaev, E. R., Wang, J., Zhang, C., & Zheng, M. (2026). Mapping the evolution of artificial intelligence applications in bone metastasis: a bibliometric and visualization analysis. Frontiers in Oncology, 16, Article 1971840. https://doi.org/10.3389/fonc.2026.1971840

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1971840

Keywords: bone metastasis, artificial intelligence, bibliometric analysis, medical imaging, deep learning, radiomics, clinical language models, oncology, prognostic modeling, research collaboration, Mapping, evolution