The high stakes of pulmonary nodule diagnosis. Lung cancer is one of the most common and deadly cancers worldwide, with about 2.2 million new cases and 1.8 million deaths in 2020. Because survival rates drop dramatically from stage I (92% five-year survival) to stage IV (38%), detecting lung nodules early is critical.
The challenge of benign versus malignant. Pulmonary nodules are small, distinct spots in the lung less than 3 cm in size. Nodules between 10 and 20 mm carry a 33-64% chance of being malignant, while those over 20 mm carry a 64-82% chance. Yet many lesions are difficult to classify by visual inspection alone, and human fatigue or variation in clinical experience creates inconsistency.
AI as a promising solution. Artificial intelligence can overcome limitations of conventional radiological review by converting images into quantifiable data features. Research has shown AI can match or even exceed expert radiologists in detecting and classifying nodules, helping reduce misdiagnosis rates and supporting better clinical decisions.
Scope of this bibliometric study. This paper is the first comprehensive bibliometric analysis of AI applied to pulmonary nodule research, examining 1,657 publications from 2005 to 2024. It maps trends, identifies research hotspots, and offers strategic guidance for researchers and clinicians at the intersection of AI and medical imaging.
Data source and search strategy. The Web of Science Core Collection (WoSCC) was selected as the primary database because of its comprehensive, internationally recognized bibliographic records covering both science and social science citation indexes. The search used Medical Subject Headings plus related terms combining AI-related keywords (artificial intelligence, deep learning, machine learning, CNN, radiomics) with pulmonary nodule terms.
Screening and data cleaning. From an initial retrieval of 2,201 publications, data were cleaned by standardizing author and institution names, unifying country naming conventions, and removing duplicate records. Two independent researchers performed literature screening and keyword clustering, with disagreements resolved by a third reviewer.
Bibliometric tools used. Three complementary tools were employed: CiteSpace for citation burst analysis and keyword timeline clusters, VOSviewer for reference cluster analysis and co-citation mapping, and the Online Analysis Platform of Literature Metrology for publication volume and growth trends. These tools visualize research networks as interconnected nodes, where node size reflects publication count.
Final dataset. After applying eligibility criteria - English-language articles and reviews published between January 2005 and September 2024 - the final dataset comprised 1,657 publications, including 1,478 original articles (89.2%) and 179 reviews (10.8%), collectively receiving 34,344 citations with an H-index of 81.
Rapid growth after 2014. Publications grew steadily from 2005, with a notable surge beginning around 2015. The acceleration after 2014 coincides with the widespread adoption of deep learning techniques and increased global attention to lung cancer screening programs.
China and the United States lead in output. China ranked first with 719 publications, followed by the United States with 358 and India with 143. Together, China and the US account for over 50% of global publications in this field. However, the US outperforms China in research quality metrics such as H-index and total citations, suggesting a contrast between China's high volume and the US's high-impact output.
Smaller nations as key connectors. Despite modest publication counts, the Netherlands (centrality 1.08), Scotland (0.84), and Austria (0.81) ranked highest in network centrality - meaning they served as critical bridges connecting research hubs across different countries. China, though leading in volume, had relatively low centrality, indicating research concentrated within national boundaries.
Top institutions. Shanghai Jiao Tong University led with 60 publications, followed by Fudan University with 58 and the Chinese Academy of Sciences with 32. Stanford University and the Ecole Polytechnique Federale de Lausanne showed the highest institutional centrality, meaning they played key roles in linking geographically distant research groups.
Most prolific and cited authors. Robert J. Gillies from Moffitt Cancer Center, often called the 'father of radiomics,' had the highest publication count (17) and also ranked among the top 10 co-cited authors. Samuel G. Armato from the University of Chicago had the most co-citations (450), reflecting his foundational contributions to computerized pulmonary nodule detection in CT imaging.
Key journals shaping the field. Radiology stands as the most influential journal by co-citation frequency and impact factor (12.1). IEEE Transactions on Medical Imaging and Medical Image Analysis reflect the engineering side of the discipline, while journals like Lung Cancer and Chest represent clinical focus. Notably, 9 of the top 10 co-cited journals are Q1 ranked in the Journal Citation Reports.
Keyword clusters reveal research themes. Keyword analysis identified 17 clusters. Prominent themes included deep convolutional neural network (DCNN) models, false-positive reduction, lung nodule detection, quantitative-semantic models, and clinical utility studies. Ground-glass nodules and malignancy prediction emerged as increasingly important topics in more recent literature.
Most cited single reference. The top-cited paper was 'End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography' by Tse et al., published in Nature Medicine in 2019, which received 750 total citations averaging 150 per year. This underscores the field's interest in practical, fully automated screening pipelines.
From rule-based to deep learning. Early pulmonary nodule detection relied on handcrafted image features and rule-based algorithms, which had low generalizability and high false-positive rates. The introduction of convolutional neural networks (CNNs), particularly 2D and 3D CNNs and U-Net architectures, dramatically improved detection performance, especially for small or irregularly shaped lesions.
Two main architectural strategies. Modern AI detection systems follow two paradigms. Two-stage detectors like Faster R-CNN generate candidate regions first, then classify them - offering high precision but at greater computational cost. One-stage detectors like YOLO perform detection and classification in a single pass, enabling real-time performance suited for clinical workflow triage.
3D and attention-based models. Because nodules exist in three-dimensional space, volumetric 3D CNN models capture spatial context across CT slices more effectively than 2D approaches. More recently, transformer-based architectures borrowed from natural language processing have been adapted to medical imaging, using attention mechanisms to focus on subtle or ambiguous nodules within complex anatomical backgrounds.
Remaining challenges. Most detection models are trained on Western-centric public datasets that exclude complex real-world cases. False-positive rates remain a clinical concern - studies using the LUNA16 benchmark reported up to 0.2 false positives per scan. Integrating clinical metadata such as smoking history and patient age with imaging models has shown promise for reducing unnecessary follow-up procedures.
Two classification approaches. Pulmonary nodule classification into benign or malignant can be approached through radiomics-based machine learning - which extracts predefined features like texture and shape - or through end-to-end deep learning where architectures like ResNet or DenseNet learn patterns directly from CT images. Deep learning generally outperforms traditional methods but requires large annotated datasets and produces difficult-to-explain outputs.
Challenges in real-world deployment. Key obstacles include class imbalance (malignant nodules are rare compared to benign ones), noisy labels based on radiological suspicion rather than pathological confirmation, domain shift between training institutions, and a lack of interpretability. These factors limit the trustworthiness of AI classification tools in high-stakes clinical settings.
Malignancy risk prediction from imaging. AI extends beyond binary classification to estimating the probability of malignancy. Deep learning models trained on low-dose CT data - such as Google's 3D CNN described by Tse et al. - can predict cancer risk with performance comparable to expert radiologists. Radiomics-based models extract quantitative features including texture and edge sharpness to complement traditional risk models like the Mayo Clinic and Brock criteria.
Growth modeling over time. Serial CT scans allow AI to track nodule evolution. Delta-radiomics captures changes in imaging features between time points, while recurrent neural networks and transformer-based temporal models learn progression patterns across multiple scans. These longitudinal models can forecast future malignancy risk, potentially enabling personalized surveillance schedules tailored to individual patients.
The need for non-surgical options. Surgery is the standard treatment for high-risk pulmonary nodules but carries significant risks including reduced lung function and cardiovascular complications, and many patients cannot or will not undergo surgery. Effective pharmacological alternatives require objective methods for evaluating treatment response - a gap that currently limits drug development in this area.
Radiomics for treatment response. Radiomics - the extraction of quantitative imaging features using AI - has demonstrated the ability to assess treatment efficacy more objectively than conventional methods. A study in Radiology showed radiomics enabled accurate survival risk stratification for stage IA non-small cell lung cancer patients. Another study demonstrated radiomics outperformed experienced radiologists in assessing pathological complete response to neoadjuvant therapy in rectal cancer.
Prognosis and survival prediction. Graph neural networks combined with visual transformers have been used to classify overall survival and recurrence risk in early-stage non-small cell lung cancer, outperforming traditional TNM staging on external datasets. These models integrate both imaging and non-imaging data, representing a step toward AI-guided personalized treatment planning.
Potential and limitations. Evidence that early high-risk pure ground-glass nodules can achieve remission through pharmacological treatment makes AI-driven efficacy evaluation highly relevant. However, no universally recognized evaluation method exists yet, and high-level clinical trial evidence is still needed. Developing AI-based efficacy metrics could open a path to non-surgical management for select patients.
China's data-driven advantage. China's large population provides a substantial data resource for AI training, and national policy support - including the 'Three-Year Plan for AI' - has driven rapid growth in publications since 2018. Seven of the top 10 publishing institutions globally are Chinese. However, challenges in data standardization, hospital interconnectivity, and ethical review systems limit large-scale multi-center validation.
The United States' innovation ecosystem. The US benefits from mature data governance frameworks, NIH-funded public datasets like TCIA and NLST, and an 'innovation-driven' model prioritizing foundational algorithm development over rapid deployment. This results in fewer but more highly cited publications and stronger interdisciplinary collaboration between radiology, computer science, and oncology.
Limited international collaboration. Despite the global reach of the field, collaboration across borders remains constrained by economic disparities, inconsistent data policies, and varying research priorities. High-income countries with advanced computing infrastructure dominate, while low- and middle-income countries face barriers to accessing technology, funding, and training data.
The interdisciplinary imperative. The future of AI in pulmonary nodule research depends on training researchers who combine technical AI expertise with deep clinical and imaging biology knowledge. Governments and institutions are encouraged to create incentive structures that reward cross-disciplinary collaboration, to accelerate translation from algorithmic innovation to improved patient outcomes worldwide.
Three dominant research hotspots. The bibliometric analysis identified three major areas at the intersection of AI and pulmonary nodule research: detection and classification of nodules using deep learning, malignancy risk prediction and longitudinal growth modeling, and the emerging application of AI to drug efficacy evaluation for non-surgical treatment.
Priorities for future model development. The analysis recommends developing hybrid models that combine high predictive performance with interpretability through visual explanation outputs, constructing multimodal and multi-center data ecosystems to improve generalizability, and exploring AI-guided personalized treatment planning for patients who cannot undergo surgery.
Limitations of the bibliometric approach. The study's reliance on the Web of Science Core Collection may exclude important non-English publications, and recent high-quality papers may appear underrepresented in citation metrics due to the time required for citations to accumulate. Merging multiple databases would have provided a more complete picture.
A roadmap for stakeholders. For clinicians, AI trends in diagnostic accuracy and malignancy prediction can inform early cancer screening protocols. For AI developers, identified gaps in annotated datasets and interpretability set clear research priorities. For policymakers, the global research distribution highlights where investment and regulatory frameworks for clinical AI should be directed.