Global research trends and foci of artificial intelligence-based tumor pathology: a scientometric study

The Science Citation Index Expanded (SCI-Expanded 1999- present) of Clarivate Analytics’S Web of Science Core Collection (WoSCC) is one of the most comprehensive and influential databases in interdisciplinary fields, containing extensive academic journals and literature, which is wildly used as the data source for bibliometric study. In this study, all the publications we obtained were retrieved and downloaded from WoSCC database on February 24, 2022. The literature searching was performed by two independent researchers to ensure the reliability and authenticity of results. The searching strategy was formulated with reference to previous researches and the searching strategy was shown as follows: topic = (“artificial intelligence” OR “robotic*” OR “expert* system*” OR “intelligent learning” OR “feature* extraction” OR “feature* mining” OR “feature* learning” OR “machine learning” OR “feature* selection” OR “unsupervised clustering” OR “image* segmentation” OR “supervised learning” OR “semantic segmentation” OR “deep network*” OR “bayes* network” OR “deep learning” OR “neural network*” OR “neural learning” OR “neural nets model” OR “artificial neural network” OR “data mining” OR “graph mining” OR “data clustering” OR “big data” OR “knowledge graph”) [25] AND topic = (cancer* OR tumor* OR tumour* OR oncology OR neoplasm* OR carcinoma*) AND topic = (“Pathology” OR “histopathology” OR “Digital pathology” OR “Whole slide imag*” OR “Virtual microscopy” OR “Digital microscopy” OR “Digital slide*” OR “Virtual slide*” OR “Telepathology” OR “Telemicroscopy” OR “Computational pathology” OR “Computer-aided pathology” OR “Digital imag* analysis” OR “Pathology imag*” OR “pathomic*” OR “urine cytology” OR “Histopathological image*” OR “pathological image*”) AND publication year = (1999 to 2021) AND Language = (English). The document types were limited to original articles or reviews. The information including titles, authors, keywords, citations, journals, institutions and references of the publications were saved in plain text format.

First, the documents were imported to Citespace V (Version 5.8.R3, Drexel University, United States) to remove duplicates. Then the targeted documents were collected and imported to GraphPad Prism 8.0 or Microsoft Excel 2019 by two independent authors for quantitative analysis of top-cited/productive countries, journals, authors, institutions, articles and keywords. In addition, H-index, a metric for evaluating the scientific impact of authors’ scholarly output and performance, was also collected from WoSCC [26].

In this study, bibliometric analysis and visualization were performed by Citespace V (Version 5.8.R3), VOSviewer (Version 1.6.16) and a free online platform (http://​bibliometric.​com). VOSviewer is a wildly applied bibliometric analysis tool, which provided three kinds of visualization maps including the network visualization, the overlay visualization and the density visualization [27]. In this research, we mainly adopted VOSviewer to conduct author-keywords co-occurrence analysis, co-authorship analysis of countries/regions, authors, institutions, and co-citation analysis of journals or references. The options and settings of VOSviewer are summarized in Additional file 1: Table S1. Citespace V is another citation visualization analysis software developed by Chen et al. [28, 29]. Compared with VOSviewer, Citespace pays special attention to the connections between knowledge fields. By detecting and monitoring the development and changes of knowledge, it can visually comprehend the research frontiers and hotspots in these fields, and then predict the future development prospects and potential research directions of these fields [30]. In our study, Citespace was utilized to accomplish co-citation analysis of authors and references, dual-map overlay of journals and citation burst of keywords or references.

The number of research papers published in different periods demonstrates the popularity and development tendency of research in a particular field. According to the search strategy and screening process, we collected 2753 literatures related to AI-based tumor pathology in recent 23 years from WoSCC (Fig. 1). As shown in Fig. 2, the number of research articles on AI-based tumor pathology research had been growing. Before 2007, there were only a few articles published. Then from 2008 to 2016, the amount of publications had increased, and after 2016, the number of publications began to grow exponentially, even reached 837 papers in 2021. As of the search date, the total citations of all literatures reached 62,182 times, and the average citations per paper and H-index were 22.59 and 101 respectively.

A total of 86 countries/regions had contributed in the field of AI-based tumor pathology. Figure 3A showed the exponential publication changes of the top 10 countries from 1999 to 2021. Table 1 listed the top 10 productive countries in this field and it was obvious that the United States ranked first with 1138 (41.34%) publications, far more than twice that of China (541, 19.65%). Moreover, H-index and total citations of the United States were several times ahead of any other country with 85 and 35,539 respectively, indicating that the United States was the most cutting-edge country in this field in the world. The world map in Fig. 3B showed that publications in this field were mainly published by countries from North America, East Asia and Western Europe. Meanwhile, apart from the United States, countries with more than 200 publications also included China and the United Kingdom.

The international cooperation analysis was shown in Fig. 3C. It could be found that the cooperation among productive countries/regions was closely matched. As the most productive country, the United States cooperated closely with China, Germany and the United Kingdom. However, there was relatively little cooperation among other countries, especially the developing countries. As shown in Fig. 3D, 30 countries/regions were included and displayed. Among them, the top three countries/regions with the largest total link strength (TLS) were the United States (TLS = 836), the United Kingdom (TLS = 423), Germany (TLS = 365). In addition, the United States was the first country to start AI-based tumor pathology research, with an average publishing year of 2017.00, while the average publishing year of China was 2019.72.

More than 3600 institutions participated in publishing research papers on AI-based tumor pathology. The polar bar chart in Fig. 4A summarized the counts, TLS and total citations of the top productive 10 institutions in detail. It was evident that all the institutions in the top 10 were from North America, of which 8 were from the United States and 2 were from Canada. Specifically, Harvard Medical School ranked first with 61 papers, followed by the University of Toronto and Stanford University. In terms of TLS, Harvard Medical School, University of Toronto and Memorial Sloan Kettering Cancer Center ranked in the top three with 335, 229, 217 respectively. The top three institutions with the highest total citations were Case Western Reserve University (2413), Case Western Reserve University (1502) and Stanford University (1500). The result of Additional file 1: Fig. S1 showed that no organization had a betweenness centrality (BC) value higher than 0.1 and the density of the network map was low. In addition, Additional file 1: Fig. S2, an overlay visualization map generated by VOSviewer, showed that most institutions from North America or Europe, such as the University of Pennsylvania, Johns Hopkins University and Case Western Reserve University, entered the field earlier, while almost all Chinese institutions participated in this field after 2019 (the nodes were reddish).

Funding agencies played a key role in the conduct of research and the publication of articles. In the sight of that, Fig. 4B summarized the top 15 funding agencies by publications. From the results, there were a total of 7 funding agencies from the United States, of which the National Institutes of Health, United States Department of Health Human Services and NIH National Cancer Institute occupied the top three in this field. This result clearly demonstrated that the United States' leading position in this field was closely related to its strong economic foundation and support.

At present, the research papers related to AI-based tumor pathology had been published in 763 scholarly journals. From the results of Table 2, the journal Scientific Reports had the highest number of publications, with a total citations of 1537 times, followed by IEEE Access and Frontiers in Oncology. The total citations of Medical Image Analysis were 5491 times, which was higher than that of any other journals. Moreover, it was worth noting that although European Urology ranked 10th, its H-index (24) and total citations (2724 times) were much better than most journals listed. According to the 2020 Journal Citation Report (JCR), all the top 10 journals were located in Q1/Q2, and among them, European Urology (IF = 20.096) had the highest impact factor (IF). Journal co-citation is a significant index to reflect the influence of a journal. In this study, 73 journals had been co-cited at least 300 times and we used VOSviewer to generate a journal co-cited network map (Fig. 5A). As shown in Fig. 5A, the top three journals with the highest TLS were Lecture Notes in Computer Science, IEEE Transactions on Medical Imaging, and Scientific Reports. Figure 5B summarized the journals with BC value no less than 0.1, indicating the important role of these journals in this field. Computerized Medical Imaging and Graphics and WMJ had the largest BC value (0.19) and ranked first among all journals, followed by IEEE Transactions on Medical Imaging (0.17) and Telemedicine Journal and e-Health (0.17).

Figure 5C was a dual-map, which was used to represent the discipline distribution of journals involved in AI-based tumor pathology research, and through this method, we could clearly understand the knowledge flows among different disciplines and the frontier or hotspot of each discipline. it could be found that the literature published in Molecular/Biology/Immunology or Medicine/Medical/Clinical journals often cited the literature from Molecular/Biology/Genetics or Health/Nursing/Medicine journals.

A total of 15,182 authors participated in the publication of papers in the field of AI-based tumor pathology. Table 3 summarized the top 10 most productive authors and the top 10 co-cited authors. The top 10 most productive authors were mostly from the United States and European countries. Madabhushi Anant, Rajpoot Nasir M and Yang Lin ranked in the top three with 40, 25 and 20 papers respectively. While Van Der Laak Jeroen A. W. M. and Litjens Geert from the Netherlands published fewer papers, their total citations were as high as 5230 and 5117, respectively. Figure 6A was a visualization map of author co-authorship analysis generated by VOSviewer. As shown in Fig. 6A, Van Der Laak Jeroen A. W. M. and Litjens Geert were key authors connecting multiple research clusters. However, overall, there was little collaboration and communication among various research clusters. Through co-citation analysis of authors, we found that the total citations of Bejnordi BE, Litjens Geert and Szegedy C ranked in the top three, and the total citations of the top 10 co-cited authors exceeded 240 times (Table 3). In this study, the BC values of 10 authors exceeded 0.1 (Fig. 6B). Among them, Jemal Ahmedin, Madabhushi Anant and Ficarra Vincenzo were the top three, up to 0.25, 0.21 and 0.21 respectively, showing their centrality in this field. Figure 6C was the map of author co-citation analysis produced by Citespcae. The results also demonstrated that Madabhushi Anant et al. were at the center of the research in this field.

A total of 2753 papers were included in this study, of which 102 publications were cited more than 100 times. Table 4 listed the top 10 cited papers and the most cited article was the review published by Litjens Geert in 2017, with a total of 3777 times, followed by Rhodes [31] and Tothill Richard W. (2008), with 2425 and 929 times respectively. As shown in Fig. 7A, [32‐34] were the three most co-cited references, with a total of 221, 215, 214 citations respectively. The timeline view of co-citation reference is a visual diagram that can reflect the temporal characteristics of the research hotspots in this field. According to the results of Fig. 7B, the Modularity Q was 0.9545, and the mean Silhouette S was also as high as 0.9839, showing the excellent clustering effect and network homogeneity. Among the 12 clusters, #6 cancer detection was the earliest research hotspot in this field, and had been studied until recent years. To date, the most popular research hotspot is #8 breast cancer histopathology, and more researchers may pay more attention to these research foci. Top 25 references with the strongest Citation bursts were summarized in Fig. 7C. The reference citation burst in this study began in 2011 due to the paper published by Breiman Leo in 2001 [35]. The latest reference citation burst was detected in 2019 and last until now. Among them, the paper on new deep residual nets published by Kaiming He et al. [36] in 2016 had the strongest strength value. Currently, most important articles are still cited frequently and it can be speculated that AI-based tumor pathology research will still be a research hotspot in the next few years.

A total of 5279 author keywords were included in this study. Figure 8A summarized the top 20 keywords with the highest frequency. The top three most common keywords were “deep learning”, “machine learning” and “artificial intelligence”, which was consistent with our study topic. In addition, “breast cancer” and “prostate cancer” were currently the most studied tumors in this field. In addition, the top 20 commonly investigated cancers/tumors were also listed in Additional file 1: Table S2. As shown in Fig. 8B, All keywords were marked with different colors depending on the temporal sequence of keyword appearance. It was noticeable that “deep learning” located in the central position of the visualization map. Moreover, except for “robotics”, “laparoscopy” and other surgery-related keywords, most of the keywords, such as “nuclei segmentation”, “computational pathology” or “transfer learning” appeared after 2019, indicating that the research in this field was rising in recent years. In other words, AI-based tumor pathology study is of high research value and will become a hot topic for a long time in the future.

The top 25 keywords with the strongest citation burst were listed in Fig. 8C. “Artificial neural network”, the earliest keyword burst, was detected in 2002. Later, researches related to tumor treatment such as “radiotherapy” or “robotic surgery” became hot topics. The latest keyword burst occurred in 2019 and had last until now. The major keywords were “convolutional neural network“, “magnetic resonance imaging” and “histopathological image”, suggesting that these research topics had received extensive attention in recent years and might become new research foci in the next few years.