On February 15, 2022, all citation data published between January 1, 2012, and December 31, 2021, were retrieved from the Web of Science Core Collection (WoSCC). The data were independently verified by two authors (YL and JZ). The detailed search string is listed in Figure 1. The document type was article. From each publication, we gathered the following basic data: title, abstract, authors, institution, country or region, journal, keywords, and references. The detailed search and analysis processes are depicted in Figure 1.

Unpublished and document types other than articles were excluded. The citation data were downloaded on February 15, 2022, and some 2021 documents included by WoSCC were not published and were thus not included in this study. Some data were excluded because their document types were not articles, such as procedures, papers, review articles, meeting abstracts, early access, editorial materials, book chapters, letters, corrections, data papers, books, and retracted publications.

Collaborative networks of countries, institutions, journals, keywords, and research categories were analyzed and visualized using the bibliometrics online analysis platform Bibliometrc [16], CiteSpace V, and VOSviewer. We collected detailed citation features for analysis, including the number of annual publications, countries, institutes, journals, subject categories, cocited references, and keywords. The H-index represents an estimate of the importance and general impact of the research contributions [17].

We analyzed a total of 2275 papers published between 2012 and 2021. The numbers of published studies on the application of AI technology to retinal illnesses over time are summarized in Figure 2. Since 2017, the number of studies on the use of AI in the treatment of retinal illnesses has skyrocketed.

These citations mentioned a total of 86 nations or territories. In Figure 3, the publications from different countries or regions are represented by blocks of different colors. The size of the colored block area represents the number of citations and the size of the different colored coverage areas represents the intensity of the cooperation. In Figure 4, a larger label area for a given country represents a greater contribution to the related literature. The purple node area indicates the strength of the centrality; the higher the centrality value, the more cooperative the relations it establishes in the country where the node is located. As per Figures 3 and 4, China and the United States have contributed the largest number of documents to this field. The United States, the United Kingdom, and Singapore cooperated more with other countries.

Table 1 lists the top 10 countries cited. China had the largest number of publications, followed by the United States, India, and the United Kingdom. Britain had the strongest centrality, followed by the United States and Singapore.

A total of 197 institutions have published relevant papers, and the clustering of their cooperative relationships is shown in Figure 5. The top 10 institutions regarding the frequency of cited institutions are listed in Table 1, including three Chinese institutions (Sun Yat Sen University, Chinese Academy of Sciences, and Shanghai Jiao Tong University), three US institutions (Johns Hopkins University, Oregon Health and Science University, and Stanford University), two Singapore institutions (Singapore National Eye Centre and National University Singapore), one Austrian institution (Medical University of Vienna), and one UK institution (University College London). Among them, the number of citations with authors from University College London ranked in 10th position; however, their H-index was the highest. In addition, Johns Hopkins University and University College London, which were the two highest-ranked institutions in the center, appeared in the same cluster.

Citing journals represents the research frontier and cited journals represents the research foundation. The top 10 citing journals and cited journals are shown in Table 2. The most frequently cited journals in the included citations were Translational Vision Science Technology, Scientific Reports, and IEEE Access. The journals that appeared most frequently among the cited journals were Ophthalmology, British Journal of Ophthalmology, and IEEE Transactions on Medical Imaging. PLoS One appeared in both the top-ranked citing and cited journals lists. The dual map of the journals is shown in Figure 6. Red represents the discipline field with the greatest influence. The research influence in the field of mathematics/systems/mathematical subject ranked first among the citing journals.

Figure 7 and Table 3 present the research areas of the citations. The most involved research areas were Ophthalmology and Engineering Electrical Electronic. The highest H-index score areas were Engineering Biomedical and Radiology Nuclear Medicine Medical Imaging. This indicates that research on AI in retinal diseases is primarily focused within the fields of computer engineering and medical imaging.

Keywords were retrieved and examined from the relevant literature. Table 4 lists the top 20 keywords used. Among them, the keywords cited over 200 times were “diabetic retinopathy,” “classification,” “validation,” and “imaging.” The keywords of the 2267 articles were analyzed and divided into four categories (deep learning, DR, optical coherence tomography, and classification), as shown in Figure 8. The time trend was examined using the hotspot transfer method, which was applied to the first 15 keywords with the highest citation outbreak. As shown in Figure 9, the key words with the greatest outburst intensity were “pattern” and “retinal ganglion cell.” The red grid indicates the emergence of keywords. “Information” and “neuron” were the keywords with the longest use (2012-2019), and “eye disease” and “enhancement” were the most popular keywords from 2018 to 2021.

A reference’s citation frequency represents its influence. We constructed a cocited literature network and analyzed the scientific relevance of the publications. The modularity of the network was measured using the modularity index; the higher the modularity Q score, the better the network’s clustering. Q>0.3 indicates that the network community structure obtained is noteworthy. The better the network’s homogeneity, the closer the Silhouette S value is to 1. The cited documents were grouped into 12 clusters, as shown in Figure 10. The homogeneity of these clusters was characterized by modularity Q=0.838 and weighted mean Silhouette S=0.9474. Cluster markers were created using the index items taken from the literature. The modularity of the network was measured using the modularity index. “Deep learning“ was the cluster with the widest range.

There were 2275 SCI papers reviewed in this study of AI and retinal diseases, which were published from 2012 to 2021, with a major increase in these publications since 2017. This is due to one study [18] being cited 506 times in 2016, which provided an important scientific and technical reference for researchers in this field. China and the United States had the maximum number of publications. The United Kingdom had the most cooperative relations. In addition, the three outcome indicators of the institution indicated that the research conducted by University College London and Johns Hopkins University has had a major influence on this research field. The number of citations in this field published by citing journals and cited journals indicated that the application of AI in retinal diseases has mainly been in the fields of computer engineering and medical imaging based on digital technology. This same pattern also applied to the citation categories.

The important nodes in the clustering demonstrated this research field’s knowledge bases after clustering the cited references. They were labeled #0 deep learning, #1 artificial intelligence, #2 optical coherence tomography, #3 image segmentation, #4 vessel segmentation, #5 Kohonen neural network, #6 Brownian motion, #7 diabetic retinopathy, #8 retina imaging, #9 age-related macular degeneration, #10 image analysis, and #11 glaucoma. The top 10 cocited references in these clusters are listed in Table 5. These studies’ conclusions require a substantial amount of basic work to be performed by ophthalmologists. For example, from the top-ranking article, 54 ophthalmologists or ophthalmic trainees participated in the identification of 5-year image data of three eye hospital patients presenting for image-based DR screening. From the second-ranked article, 494,661 retinal photographs were used to assess the diagnosis performance of a deep-learning system for DR and related eye illnesses. Each retinal image was analyzed by two trained senior, certified nonmedical professional graders. To achieve progress in the application of AI in retinal diseases, more cooperative and collaborative relationships among ophthalmologists, imaging technicians, and computer technology researchers are required in the future.

This bibliometric analysis only included the literature data in WoSCC. Some other databases were not included, such as PubMed, Medline, and Cochrane. In addition, the citation data analyzed were only from the literature published from 2012 to 2021, rather than collecting all articles published in this research field to date. Some 2022 studies are still ongoing and have not yet been published. These criteria may result in publication bias.

This study provides a systematic literature analysis on the use of AI in retinal diseases. Bibliometric analysis enabled obtaining objective and comprehensive results. Judging by the volume of published papers and research subjects, this study area is still popular and a noteworthy topic with major interdisciplinary exploration space. Ophthalmologists, imaging experts, and computer algorithm researchers in developing and developed countries need to make full use of population advantages or core technologies in different regions to strengthen collaboration. This idea has become a research hotspot that uses the existing basic clinical research results and a more advanced algorithm mode to develop a high-quality ophthalmic examination image system and further verify its clinical applicability. At present, an algorithm program with 100% diagnostic accuracy for retinal disease has been developed [42]. In the future, high-quality retinal image–forming AI technology with strong stability and clinical applicability will continue to be encouraged.