This bibliometric study, published in Biomolecules in 2021, provides a comprehensive quantitative analysis of how AI and machine learning research has evolved in the field of diagnostic pathology for lymphoid neoplasms (LN) from 1990 to 2020. The authors retrieved 525 documents from the Clarivate Analytics Web of Science (WoS) core collection database, using a broad query combining AI terms (artificial intelligence, deep learning, machine learning, neural network, support vector, natural language) with lymphoma-related terms (lymphoid neoplasm, lymphoma, lymphoproliferative, lymphocytic, gammopathy, myeloma, histiocytic). The same query run in PubMed yielded 528 documents, confirming consistency across databases.
Why bibliometrics: Bibliometric analysis uses publication metadata, citation counts, co-authorship networks, and keyword co-occurrence patterns to objectively map the structure and evolution of a scientific field. Rather than reviewing the content of individual studies, this approach treats the literature itself as data, revealing which themes are dominant, which institutions lead production, and how research topics have shifted across time. The authors used two specialized software packages for this purpose: SciMAT (Science Mapping Analysis Software Tool) and VOSviewer (Visualizing of Science), both well-validated for biomedical applications.
The 4P medicine context: The introduction frames AI and ML development within the paradigm of 4P medicine (predictive, personalized, preventive, and participative), which requires computational systems capable of integrating clinical, histopathological, and genomic data at scale. Hematopathology is positioned as a pioneer field in translational cancer research, citing molecular targeted therapies in chronic myeloid leukemia and acute promyelocytic leukemia as landmark examples of bench-to-bedside translation. The field of AI in LN pathology sits at the intersection of these traditions and the new data-driven paradigm.
The study's stated goals were to identify global production trends, predict future output volume, determine leading research centers and countries, and perform science mapping analysis (SMA) to elucidate the cognitive (thematic) and social (institutional collaboration) structure of the field across three chronological subperiods: 1990-2005, 2006-2014, and 2015-2020.
The metadata were extracted from the WoS SCI-Expanded Collection in January 2021, restricted to documents published between 1990 and 2020. The authors conducted a structured performance analysis classifying documents by type (original articles, reviews, proceeding papers, meeting abstracts, and other minor types including letters, book chapters, and editorial material). Analysis by research areas, source titles, organizations, and countries was performed on the four primary document types, excluding the minor categories.
SciMAT cognitive mapping: SciMAT constructs a co-occurrence matrix (CM) from the author keywords of each article. Each significant concept or theme is defined as the cluster of author keywords used together across documents during a given period. The simple center algorithm builds thematic networks from this CM. The resulting themes are plotted in a two-dimensional strategic diagram using two parameters originally conceived by Callon et al.: centrality (the degree of interaction of a theme with other themes, indicating its importance to the network) and density (the internal cohesion of the theme, indicating its development). This produces four quadrants: Motor Themes (high centrality, high density), Basic and Transversal Themes (high centrality, low density), Emerging or Declining Themes (low centrality, low density), and Highly Developed or Isolated Themes (low centrality, high density).
VOSviewer social mapping: VOSviewer was used to map the social framework by analyzing bibliometric coupling among institutions and countries. Bibliometric coupling is defined by the presence of common cited references in two documents' reference lists, indicating shared intellectual foundations. The software applies the SMACOF algorithm to position items in a distance map where proximity reflects coupling strength, and label size reflects the number of documents or citations. The analysis was performed for three distinct subperiods (1990-2005, 2006-2014, 2015-2020) to capture longitudinal evolution.
Bradford's law check: To validate the source distribution, the authors applied Bradford's law of bibliographic scattering, which predicts that a small core of journals will account for a disproportionate share of publication volume. This served as an internal consistency check on their database retrieval strategy.
Of the 525 retrieved documents, original journal articles were the dominant type at 359 documents (68.38%), followed by proceeding papers at 106 (20.19%), reviews at 35 (6.66%), and meeting abstracts at 33 (6.28%). The growth of the field was particularly notable from 2017 onward, and cumulative production fits well to both an exponential model (R2 = 0.9112) and a potential model (R2 = 0.9733). These high R2 values indicate that the growth trajectory has been mathematically consistent rather than driven by isolated publication spikes.
Polynomial regression forecast: The authors fit a third-degree polynomial equation (y = 0.0518x3 - 1.511x2 + 17.345x - 35.972, R2 = 0.9701) to predict future output. This model projects that global production will double by 2027 and reach three times the 2020 level by approximately 2031. The authors explicitly compare this trajectory to Price's law of the growth of science, which postulates that publications double every 10-15 years, and note that the predicted growth rate in this field significantly outpaces Price's expectation. They also caution that polynomial regression has inherent limitations: as a form of linear regression, its predictive accuracy depends on the number of independent variables, and the model must be updated as new data accumulate.
Research area distribution: Computer science (CS) led with 128 documents (24.38%), followed by engineering (EN) with 89 (16.95%), radiology and nuclear medicine (RNM) with 61 (11.61%), biochemistry and molecular biology (BM) with 60 (11.43%), and oncology (ON) with 54 (10.29%). From 2014 onward, a general growth trend was evident across both biomedical research areas (oncology, radiology, biochemistry) and bioinformatics areas (computer science, engineering), indicating convergence of these disciplines around AI applications in LN pathology.
Gompertzian model interpretation: The authors also place the field within a Gompertzian model of scientific evolution, which defines three stages: an initial seminal phase, an exponential growth phase, and a consolidation phase marked by an increase in reviews. Based on the increasing rate of review publications from 2015 onward alongside continued exponential article production, they conclude the field is transitioning between the second and third stages, meaning it is no longer emerging but is in an active consolidation process.
For the entire 1990-2020 period, the USA dominated with 190 documents (36.19% of the total corpus). The People's Republic of China ranked second with 72 documents (13.71%), followed by Germany with 44 (8.38%), India with 35 (6.66%), and France with 31 (5.90%). Notably, both China and India showed dramatic growth during the study period: China increased from 1.56% of total production in 1990-2005 to 13.71% overall, and India from 1.56% to 6.67%, reflecting significant investment in AI research infrastructure in both countries.
Global heterogeneity trends: The number of countries producing more than five documents increased markedly across subperiods: 3 countries (4.84% of all nations analyzed) in 1990-2005, 7 (11.29%) in 2006-2014, and 19 (30.65%) in 2015-2020. This expanding geographic distribution demonstrates that the field has shifted from a concentration in a handful of Western nations to a genuinely global research enterprise, even if the USA and China still account for nearly half of all output.
Institutional rankings: The University of Texas System led overall institutional production (3.62%), followed by INSERM of France (2.86%) and Harvard University (2.48%). Thirteen of the twenty most productive institutions globally were located in the USA. Among the top five centers worldwide, the Chinese Academy of Sciences (2.29%) was the only non-USA institution, demonstrating the strength of Chinese AI research in this domain. European leaders included the Assistance Publique Hopitaux Paris (APHP, 1.71%), the Centre National de la Recherche Scientifique (CNRS, 1.71%), the Technical University of Munich (1.52%), and the Goethe University of Frankfurt (1.33%).
Subperiod analysis: During the 2015-2020 subperiod, the Chinese Academy of Sciences tied with Emory University at 10 documents (2.87% each) as the most productive institutions, signaling China's rapid ascent within this most recent and most active period. The Memorial Sloan Kettering Cancer Center, the University of Pennsylvania, and the Icahn School of Medicine at Mount Sinai also featured prominently in recent production, consistent with the clinical focus of increasingly translational AI research.
A total of 397 scientific journals published at least one of the 525 documents retrieved. However, 388 of these journals (96.47%) published fewer than five documents each, classifying them as secondary sources. This distribution is consistent with Bradford's law of bibliographic scattering, which predicts that the bulk of literature in any field will be concentrated in a small nucleus of highly productive journals, making it inefficient to extend systematic searches far beyond that core.
Top journals overall: Lecture Notes in Computer Science led with 14 documents (2.67%), followed by Blood with 12 (2.28%) and the European Journal of Nuclear Medicine and Molecular Imaging (EJNMMI) with 10 (1.90%). Scientific Reports had 8 documents (1.52%), and BMC Bioinformatics and PLOS ONE each contributed 7 (1.33%). The presence of Lecture Notes in Computer Science as the top source reflects the historically strong contribution of CS conference proceedings to this research area, particularly in the earlier subperiods.
Temporal shifts in source journals: During 2015-2020, Blood (9 documents), EJNMMI (10), Journal of Nuclear Medicine (6), and Lecture Notes in Computer Science (6) emerged as co-leaders, alongside Frontiers in Oncology (5), IEEE Access (5), and the American Journal of Clinical Pathology (4). The growing prominence of clinical oncology and nuclear medicine journals in the most recent period reflects the maturation of the field toward clinical translation, contrasting with the computer science and bioengineering conference proceedings that dominated earlier output.
The authors note that source title rankings depend on the bibliographic database used and would differ if Scopus or PubMed were the primary source instead of WoS, urging interpretation with caution. WoS covers more than 250 disciplines, 21,000 journals, and 1.6 billion cited references from 1900 to the present, making it the standard database for bibliometric analysis but not the exclusive repository of relevant literature.
The SciMAT strategic diagrams visualize the cognitive structure of the field across all three subperiods. In the first period (1990-2005), the ensemble classifier was the only Motor Theme (MT) identified, meaning it had both high internal development and strong connections to other research themes. Ensemble methods compensate for partial errors by passing the output of one base model as input to the next algorithm, progressively improving average prediction accuracy. This approach was applied as early as the 1990s to predict outcomes after hematopoietic stem cell transplantation (HSCT), demonstrating the early intersection of AI and hematological malignancies.
Period 2 (2006-2014) thematic expansion: Motor Themes in this period expanded to include non-Hodgkin lymphoma, machine learning, support vector machine (SVM), and subgroups, reflecting the growing centrality of NHL as the primary disease focus and the adoption of SVM-based classifiers as methodological workhorses. Basic and Transversal Themes shifted from bioinformatics and support vectors (period 1) to neural networks, mass spectrometry, antitumor drug design, and poor prognosis, indicating a broadening of the field into drug response prediction and survival modeling.
Period 3 (2015-2020) thematic maturation: The most recent period shows Motor Themes including antitumor drug design, resistance, and magnetic resonance imaging (MRI), reflecting the integration of functional imaging and drug evaluation into AI-driven research. Lymphoma classification, chronic lymphocytic leukemia (CLL), Hodgkin lymphoma (HL), and random forest appeared as Basic and Transversal Themes, indicating these are now well-established research foundations. The strategic diagram for this period also shows new clinical entities (CLL, HL) that did not appear in earlier diagrams, alongside a continued emphasis on NHL classification now supported by multi-algorithm approaches.
Specific algorithm applications highlighted: The discussion section provides concrete examples of the algorithms represented in the cognitive framework. A random forest algorithm was trained and validated to discriminate the most frequent B-cell NHL categories among 510 NHL cases using ligation-dependent RT-PCR and next-generation sequencing (NGS). SVM was used to stratify 414 DLBCL patients from gene expression profiling (GEP) data obtained from CHOP/R-CHOP-treated cohorts into two biologically distinct subgroups. Logistic regression and Cox proportional hazards were employed to build a cell-of-origin (COO) classifier in DLBCL using targeted RNA sequencing (RNA-seq). Convolutional neural network (CNN) algorithms were developed from digitized histopathological slides using Aperio ImageScope to discriminate between Burkitt lymphoma (BL) and DLBCL.
The VOSviewer social framework analysis revealed two major structural regions in the global collaboration network. The first and larger cluster consists of institutions primarily in the USA and Europe, which also collaborate with select Asiatic centers including Shanghai Jiao Tong University, Tongji University, Yonsei University, and Sichuan University. The second region is characterized by the Chinese Academy of Sciences, which appeared positioned more peripherally and relatively isolated from the main USA-Europe cluster in terms of bibliometric coupling strength.
Publication-based vs. citation-based maps: When institutions were mapped based on the number of documents published, the structural arrangement showed two main nodes. However, when citation impact was used as the weighting metric instead of raw document count, the map's overall topology was preserved but Harvard University's centrality notably increased, emerging as a major hub of high-impact citation activity. This divergence between production volume and citation influence is a meaningful finding: it identifies institutions that produce not just quantity but high-impact foundational contributions cited by other researchers worldwide.
Country-level collaboration: At the country level, the USA acted as the central node of the global collaboration map, with England, Japan, and a network of European nations (Italy, Spain, Germany, France) connected outward from it. China and the USA stood out most prominently when raw document production was assessed. However, when citation impact was evaluated, the contribution of Asiatic countries decreased relative to the USA and England, suggesting that while Asian researchers are prolific producers, the most-cited foundational work remains concentrated in Western institutions.
Interpretation of collaboration patterns: The authors propose that European and USA institutions tend toward more collaborative research patterns, whereas Asiatic centers, particularly in China, follow a more unified national research strategy. This interpretation aligns with broader observations in the scientometrics literature about differences in academic publishing cultures and funding structures across regions. However, the authors appropriately note this hypothesis requires more in-depth investigation of country-specific science promotion policies before definitive conclusions can be drawn.
The discussion contextualizes the bibliometric findings within the broader landscape of AI in oncology and pathology. The authors highlight parallel developments in other cancer types as reference points for how far AI translation has progressed elsewhere. In colorectal cancer, AI algorithms can automatically discriminate neoplastic from non-tumorous tissue from scanned preparations. In pancreatic neuroendocrine tumors, deep learning can delineate tumor areas from stroma with 97.8% sensitivity and 88.8% specificity, improving Ki67 quantification. In breast cancer, deep learning classifies digital H&E preparations as benign or malignant with an area under the curve (AUC) of 0.962. The first EU-approved deep learning system with CE-IVD marking was for prostate cancer detection, with published AUC values between 0.98 and 0.997 for biopsy classification.
Imaging and molecular integration in LN: The analysis of research areas and source journals points to molecular imaging and AI as a growing convergence zone. Deep CNNs have been applied to PET/CT imaging in 327 NHL patients to discriminate patterns of tumor infiltration. A fully automatic CNN-based segmentation approach was developed for 3D FDG-PET/CT to predict total metabolic tumor volume (TMTV) in DLBCL, using a convolutional neural network. SVM algorithms have been used to discriminate hypermetabolic lymphomatous lesions from noncancerous processes. A transfer learning algorithm predicted sensitivity to Bortezomib in multiple myeloma patients, while a Bayesian network combined with neural networks and RNA-seq identified novel resistance mechanisms across 150 drugs evaluated in DLBCL.
CLL and HL applications: Although NHL dominates the research corpus, the cognitive analysis identified emerging ML approaches targeting CLL and HL as specific entities. ML models have been developed to identify CLL patients at high risk of infection, and artificial neural networks (ANNs) have been applied to optimize CLL diagnosis using gene expression profiling (GEP). For HL, where the complexity of isolating Hodgkin and Reed-Sternberg (HRS) cells within a predominantly non-tumoral microenvironment poses unique challenges, ML algorithms have demonstrated potential to predict prognosis in both adult and pediatric HL patients.
What the field needs: The authors conclude that the full clinical implementation of AI systems in LN pathology will require collaborative programs connecting pathologists, bioinformaticians, and clinicians. This multidisciplinary approach is framed not as a distant ambition but as an active requirement, given that pathology departments will increasingly need professionals with foundational competency in both AI and bioinformatics to operate in the evolving research and clinical landscape predicted by the polynomial growth model.