Bibliometric and Latent Dirichlet Allocation (LDA) analysis of artificial intelligence for thyroid nodules

Abstract

Artificial intelligence (AI) has rapidly integrated into thyroid nodule ultrasound diagnosis and malignant risk stratification, fostering a fast-growing interdisciplinary domain encompassing endocrinology, medical imaging, and computational intelligence. Nevertheless, existing bibliometric studies in this field rely on single-database datasets and short time-series windows, lacking systematic elaboration of the dynamic linkage between clinical guideline iterations and AI algorithm evolution. This study conducted a comprehensive bibliometric analysis to fill this research gap, clarify the global intellectual landscape, and reveal the evolutionary rules and emerging frontiers of AI-assisted thyroid nodule diagnosis. Peer-reviewed articles and reviews on AI in thyroid nodule diagnosis were retrieved from Web of Science and PubMed (2006-2025). After strict data cleaning, 821 valid publications were analyzed using CiteSpace, VOSviewer, and online platforms, examining publication trends, contributions, collaboration networks, citations, and keyword evolution. Latent Dirichlet Allocation (LDA) was applied for topic modeling. The field showed rapid and sustained growth, with a compound annual growth rate of 24.1%, and a notable acceleration around 2017. In total, 2,870 authors from 75 countries contributed. China led global output, while the United States and South Korea showed higher citation quality. Shanghai Jiao Tong University was the most productive institution. Keyword analysis revealed close coupling with clinical guidelines (ATA, TI-RADS, Bethesda). Core hotspots included ultrasound image analysis, deep learning optimization, and malignant risk stratification. Emerging frontiers involve explainable AI, multimodal data fusion, and standardized clinical validation. This study establishes a long-time-series, multi-database knowledge map of AI applications in thyroid nodule diagnosis, systematically revealing the global research pattern, spatiotemporal evolution characteristics, and guideline-driven developmental logic of this field. Different from previous single-dimensional bibliometric summaries, this work clarifies how standardized clinical diagnostic criteria chronologically correspond with the iterative progress of radiomics and deep learning research. The findings provide reliable theoretical references for subsequent algorithm optimization, interdisciplinary innovation, and large-scale multicenter prospective validation, facilitating the standardized clinical translation and high-quality development of AI-assisted thyroid diagnostic models and precision endocrinology.