Research hotspots and emerging trends of deep learning applications in orthopedics: A bibliometric and visualized study

Background As a research hotspot, deep learning has been continuously combined with various research fields in medicine. Recently, there is a growing amount of deep learning-based researches in orthopedics. This bibliometric analysis aimed to identify the hotspots of deep learning applications in orthopedics in recent years and infer future research trends. Methods We screened global publication on deep learning applications in orthopedics by accessing the Web of Science Core Collection. The articles and reviews were collected without language and time restrictions. Citespace was applied to conduct the bibliometric analysis of the publications. Results A total of 822 articles and reviews were finally retrieved. The analysis showed that the application of deep learning in orthopedics has great prospects for development based on the annual publications. The most prolific country is the USA, followed by China. University of California San Francisco, and Skeletal Radiology are the most prolific institution and journal, respectively. LeCun Y is the most frequently cited author, and Nature has the highest impact factor in the cited journals. The current hot keywords are convolutional neural network, classification, segmentation, diagnosis, image, fracture, and osteoarthritis. The burst keywords are risk factor, identification, localization, and surgery. The timeline viewer showed two recent research directions for bone tumors and osteoporosis. Conclusion Publications on deep learning applications in orthopedics have increased in recent years, with the USA being the most prolific. The current research mainly focused on classifying, diagnosing and risk predicting in osteoarthritis and fractures from medical images. Future research directions may put emphasis on reducing intraoperative risk, predicting the occurrence of postoperative complications, screening for osteoporosis, and identification and classification of bone tumors from conventional imaging.


Introduction
As a subset of machine learning, deep learning has broken the limitations of traditional machine learning, and can implement more accurate classification and segmentation of images to extract feature elements (1, 2). It is known that deep learning surpasses the performance of other machine learning techniques in several aspects, such as predicting the potential activity of drug molecules (3) and the alternative splicing patterns in tissues (4). Convolutional neural networks (CNNs), consisting of multiple neural layers, are powerful learning structures in deep learning. They are effectively used to solve related problems in computer vision and images. By varying the depth and breadth of the neural layers, the capacity of the CNNs can be expanded to accommodate tens of thousands of images and allow them to interpret correctly (1). Currently, deep learning has been demonstrated to have the potential to identify diseases such as skin lesions and diabetic retinopathy, and it has achieved greater accuracy that comparable to experienced physicians (5,6).
Most musculoskeletal diseases in orthopedics require a large degree of help from images. Given the advantages of deep learning in image processing, there has been a growing number of researches on the application of deep learning in orthopedics in recent years (7). Bibliometrics is a statistical and quantitative method to analyze the academic characteristics of a body of literatures in certain scientific fields. It can help researchers to grasp the priorities and trends, and predict its prospects. Therefore, it has been applied in multiple research areas, including immunology (8), oncology (9,10), nursing (11), vaccine (12), and orthopedics (13). The software Citespace can quantitatively analyze the literature database to evaluate the distribution of countries (or regions) and institutions, as well as excavate the literature that plays a central position in the research field, and visualize the research hotspots and development trends (14).
Currently, no bibliometric analysis has been conducted to quantitatively analyze the progress and current status of deep learning in this emerging field. Herein, this study aimed to elucidate the research hotspots, key fields, and trends of deep learning applications in orthopedics in recent years by using Citespace, Moreover, the research direction and references for further exploration were shown as well.

Data source and collection
Since the details of the documents in the Web of Science are more accurate than other databases, such as Scopus, PubMed, Embase, etc. We retrieved all literature from the Web of Science Core Collection. The retrieval was completed within 1 day on March 29, 2022 to reduce changes due to frequent updates to the bibliographic database. The searching strategy was as follows: #1 and #2 (#1:WC = (orthopedics) OR TS = (orthopedic * OR orthopedic * OR "sports medicine") OR TS = ("bone disease * " OR TS = "bone age" OR bone fracture * OR "joint disease * " OR "joint disorder * " OR TS = "arthritis" OR TS = "joint dislocations" OR TS = "joint dislocations" OR TS = "musculoskeletal system" OR "musculoskeletal disease * " OR "musculoskeletal disorder * " OR "musculoskeletal trauma * " OR TS = "bone tumor * " OR TS = "bone cancer" OR TS = "bone metastasis" OR TS = "bone cyst * " OR TS = "soft tissue tumor * " OR TS = "soft tissue neoplasm * " OR TS = "joint replacement * " OR TS = "arthroplasty" OR TS = "arthroscope * "), #2: TS ="deep learning" OR TS = "convolutional neural network * "). There is no limit to the publication year. The type of literature was selected as articles and reviews without language restriction. Literature not relevant to this topic was excluded, and duplicate literature was removed by Using Citespace (5.8 R3). Finally, a total of 822 articles were retrieved and exported for records in the format of plain text files. Meanwhile, we also obtained the number of annual publications and the amount of publication of the journal.

Statistical analysis and visualization
The number of annual publications was imported into Microsoft Office Excel 2021 and the trend was further analyzed. CiteSpace is a visual bibliometrics tool based on Java language for quantitative analysis of literature in specific research areas. The conception of "co-citation analysis theory" is as follows: when two documents appeared in the reference list of a third cited document, these two documents form a co-citation relationship. By mining the "co-introduction relationship", it is possible to reveal important turning points, the evolution of keywords, and the frontiers in related research fields. In this study, the time spans were set to 2015 to 2022 based on the publication date of all 822 literature. Time slices were set to 1 year per slice, and selection criteria was selected as g-index. Authors, countries, and institutions were selected to perform the cooperative network analysis. Keywords were selected for co-occurrence, burst, cluster analysis, and timeline viewer. Reference, cited-authors, and cited journals were chosen for citation analysis. Cosine was select for the link strength. Besides, pathfinder, pruning sliced networks, and pruning the merged network in pruning algorithm were selected. In the visual network maps, node colors change from cold to warm from the inside to the outside, representing the year from the original to the most recent. The purple circle on the outside represented high centrality, and the node connection indicated cooperation, co-occurrence, and co-citation.

Trends of annual publications
Overall, a total of 822 publications (759 articles and 63 reviews) were identified. The number of studies published in each year can help us understand the general trends in the relevant research. Since 2022 has just begun, the number of documents in 2022 cannot show the overall publication situation. Therefore, the analysis only included the publication from 2015 to 2021. As shown in Figure 1, the number of studies concerning the application of deep learning in orthopedics has increased year by year since 2017, indicating that deep learning has received high attention from orthopedics in this period. Compared with publications in 2019, the total number of articles published in 2020 were 227, showing an explosive growth. To further understand the trend of annual publications in this field between 2015 and 2021, a trend line of publications in this period was plotted, and the results showed the exponential function Y = 0.2941e1.0998x (R2 = 0.8983, Y is the annual publication, and X is the year). It is clear that the application of deep learning in orthopedics has great potential, and research trends are likely to continue.

Distribution of countries/ regions and institutions
We used Citespace to conduct co-citation analysis of the countries (or regions) and institutions. The country or region distribution map consisted of 56 nodes and 53 links. As demonstrated in Figure 2A and Table 1 Table 2). Analyzing nodes reveals that they have more connections to other countries and continents. England, Australia, and Canada had high centrality and publications. The institution distribution map consisted of 225 nodes and 239 links. As shown in Figure 2B and Table 2

Analysis of authors and cited authors
A total of 250 authors were involved in the cooperative map ( Figure 3A). As shown in Table 3, VALENTINA PEDOIA was the author with the most published literature. Interestingly, among the top ten authors, we found three major collaborative networks. They were VALENTINA PEDOIA (University of California System, USA) and SHARMILA MAJUMDAR (University of California System, USA), PAUL H YI (Johns . /fpubh. .    Hopkins University, USA) and JAN FRITZ (New York University, USA), JARET M KARNUTA (Cleveland Clinic Foundation, USA), PREM N RAMKUM AR (Cleveland Clinic Foundation, USA) and HEATHER S HAEBERLE (Cleveland Clinic Foundation, USA), separately. Co-cited authors are two (or more authors) who are cited in one or more subsequent papers at the same time, the two or more authors constitute a co-cited relationship. The co-cited author map consisted of 487 notes and 747 links ( Figure 3B). As displayed in Table 3, LeCun Y was the most frequently co-cited author. Among the top ten co-cited authors, Litjens G had the most centrality.

Journals and co-cited academic journals
The 822 articles retrieved in this study were published in 308 journals. As shown in Table 4, The journal Skeletal Radiology (33 times) had the highest number of outputs, followed by IEEE Access (28 times), Scientific Reports (25 times), Computer Methods and Programs in Biomedicine (19 times), Applied Sciences Basel (17 times), Diagnostics (17 times), International Journal of Computer Assisted Radiology and Surgery (15 times), Journal of Digital Imaging (14 times), Medical Image Analysis (14 times), and Sensors (14 times). The co-occurrence analysis of cited journals obtained from Citespace was shown in Figure 4A and Table 4, with 532 nodes and 817 links. The greater the node, the higher the cocitation frequency of journals. Through the analysis of cited journals, the distribution of journals that focus on this field can be obtained. Co-citation frequency reflects the quality and influence of journals. Journals with high co-citation frequency are often regarded as mainstream journals. The top ten cited journals were Lecture Notes in Computer Science (345 times),

In Proceedings of The IEEE Conference on Computer Vision
Frontiers in Public Health frontiersin.org . /fpubh. .

Analysis of cited references
A total of 464 nodes and 733 links presenting the cocitation relationship of the references formed the cited reference network in Figure 4B. Co-citation means that two or more articles are cited by one or more papers at the same time,  Deep learning for automated skeletal bone age assessment in X-ray images 2017 Spampinato C Article and important references have higher co-citation frequency. As shown in Table 5, among the top ten most frequently cited references, "Deep learning" had the most frequency. Meanwhile, there were three proceedings papers, namely the "Densely Connected Convolutional Networks, " "Deep Residual Learning for Image Recognition, " and "U-Net: Convolutional Networks for Biomedical Image Segmentation." Most of these articles were preliminary explorations of the neighborhood, and the publication of these articles was related to the gradual increase in annual publications that began in 2017.

Analysis of keywords, burst value, clustering, and time evolution
The high-frequency keywords the article can help us understand the main research hotspots of the filed. Keyword co-occurrence analysis was performed by Citespcace. The cooccurrence network of keywords had a total of 265 nodes and 404 links ( Figure 5A). The high-frequency keywords were listed in Table 6. The top 20 keywords were as follows: classification, segmentation, convolutional neural network, MRI, system, diagnosis, osteoarthritis, artificial intelligence, image, neural network, CT, model, hip, knee, bone, fracture, children, reliability, disease, and prediction. From these keywords, we found that the main research methods were artificial intelligence, neural network, and convolutional neural network. The main research tasks were classification, segmentation, diagnosis, and prediction. The main research content was image, and the main research objects were bone, hip, knee, osteoarthritis, fracture, and disease. To sum up, classification, diagnosis, and risk prediction of various types of fractures and osteoarthritis by image were major research hot-spots.
The strong burst keywords can help us explore the development trends of the filed. We use Citespace's burst function to analyze keywords that have attracted much attention from academia. The burst keywords were shown in Figure 6. Keywords with higher burst values over a period mean that they have received special attention during the corresponding time intervals and may become a new research trend. In the burst keywords after 2020, we paid attention to the following main keywords: risk factor, identification, localization, and surgery. At the same time, a timeline viewer of keywords was plotted, which may help to visualize the phased hotspots and directions from the temporal dimension ( Figure 7).

Discussion
This study utilized the principles of bibliometrics analysis and the method of Citespace visualization. In the application of deep learning in orthopedic filed, the annual number of articles, countries (or regions) and institutions, authors and cited authors, journals and cited journals, and keywords were extensively analyzed to reveal the current research hot-spots and trends in this field.
Based on an analysis of countries (or regions) and institutions, the USA and China were the two most documented countries, but there was a lack of international cooperation between them. More international cooperation is needed for China (0.00 centrality) to jointly promote development in this field. England, Australia, and Canada had high centrality and publications, suggesting that these countries might play an essential role in bridging research. At the same time, most of   the top ten institutions with the most centrality and publications are from China and the USA, but these institutions basically cooperate with other institutions in their countries. Therefore, from the perspective of cooperation networks, we hope that there will be more cooperation between different countries. Through the analysis of the author cooperation network, we found that the top ten authors mainly constituted three cooperative networks. In the cooperation between the two authors in the deep learning field, VALENTINA PEDOIA (University of California System) and Sharmila MAJUMDAR (University of California System), the most cited literature showed that the improved U-net can automatically segment cartilage and meniscus from MRI (15). PAUL H YI (Johns Hopkins University, USA) and JAN FRITZ (New York University, USA) have reported that deep learning can significantly address the identification of implants after knee replacement surgery and help with subsequent revision surgeries (16). JARET M KARNUTA (Cleveland Clinic Foundation, USA), PREM N RAMKUMAR (Cleveland Clinic Foundation, USA) and HEATHER S HAEBERLE (Cleveland Clinic Foundation, USA) summarized some application of artificial intelligence in orthopedics (17). However, it is worth noting that in the entire network of cooperative map, none of the authors had a centrality <0.1, indicating a lack of international cooperation among these authors. The entire network presented a very weak partnership.
As to the analysis of a network of cited authors, the most frequently cited author-LeCun Y believed that deep learning is good at processing high-dimensional data, and can accurately identify and classify targets in image recognition tasks (18). Similarly, Litjens G et al. with the most centrality in the top ten co-cited authors published "A survey on deep learning in medical image analysis (19), " which plays a key role in the field.
According to the journals and co-cited journals in Table 4. The journal Skeletal Radiology had the most publications. The average impact factor of the top ten journals in 2020 was 4.086, and the impact factor of Medical Image Analysis was the highest (8.545). Furthermore, seventy percent of journals .
/fpubh. . had an impact factor over three, and fifty percent of journals belonged to Q1 (JCR). The journal Nature had the highest impact factor in the cited journals. In addition to the cited journals that cannot obtain impact factor and JCR partitions, seventy-five percent of the cited journals belonged to Q1. These data suggested that this area has attracted academic attention from high-scoring journals. Among top ten co-cited references, these references are usually the basis of relevant knowledge fields and play an important role in knowledge structure. For three proceedings papers, Ronneberger O, He KM, and Huang G built new models for deep learning in 2015, 2016, and 2017, respectively, which strengthened the ability of deep learning to process images. These models are deep residual learning, densely connected convolutional networks, and U-net (20)(21)(22). Article entitled "Deep learning" by LeCun Y et al. indicated that deep learning can precisely solve tasks including image recognition and speech recognition (18). Therefore, these articles laid the foundation of this field. "A survey on deep learning in medical image analysis" by Litjens G et al. summarized the research on the application of deep learning in various medical fields before Feb 1, 2017 in details, and these researches were basically published in 2016 and early 2017. We found that these studies on the application of deep learning in musculoskeletal disease were basically belong to proceedings papers (19). Today, the trend of publication in the journals is increasing year by year (Figure 1), suggesting that research in this field will still be a fertile area in the next few years. Among the top ten most frequently cited reference, a study by Olczak J et al. was the first article about deep learning application in fracture, and the performance of deep learning was like that of senior orthopedic surgeons (23).
Keywords are the cores of a paper that reflecting the concerns of relevant field. By analyzing the high-frequency .

FIGURE
The timeline view of keywords.
keywords and strong-burst keywords, we can explore the hotspots and research trends. The main research hotspots and research trends are as follows: For hot keyword-fracture After the first report of deep learning in fracture by Olczak et al. (23), the number of studies concerning this field increased with years.
For upper extremity fracture, deep learning is commonly applied to distal radius fracture and humeral fracture (24-26). Gan et al. applied two deep learning models to identify distal radius fracture (24). They trained a Faster R-CNN with the ability to automatically annotate the regions of interest (ROIs), and then another CNN model for diagnostic was trained with images that were annotated by Faster R-CNN Model; The average intersection of the union (IOU) value of Faster R-CNN was 0.87, and the AUC of inception-v4 was 0.96. However, their data consisted only of anterior-posterior (AP) image, whereas images in both the AP and lateral directions were often clinically required, and there was no adequate amount of data from multiple centers. Recently, Suzuki et al. addressed these two issues, and the results showed that the CNN model had better accuracy than orthopedic surgeons (25). In addition to identification of fractures, deep learning can also directly classify fracture types. Chung et al. shown that CNN was superior to orthopedic surgeons in the classification of proximal humeral three-and four-part fracture, with an AUC exceeding 90%, while both fractures are easily misjudged in clinical work (27). Besides, Langerhuizen et al. tried to identify scaphoid bone fracture, a common fracture of the wrist, by using a small data set and found that deep learning performed inferior than surgeons in 2020 (28). However, Yoon et al. later established two model, namely the Apparent Fracture Model and the Occult Fracture Model, which excelled in identifying obvious and potential scaphoid fractures, respectively. Their research used more datasets and escalated deep learning algorithms, and both models showed higher sensitivity and specificity than that of Langerhuizen et al. (29).
In lower limb, hip fractures attracted much attention. Cheng et al. proved that deep learning model can automatically identify femoral neck fracture and trochanteric fracture through pelvic X-rays, with 98% sensitivity and 91% accuracy (30).  (38). Usually, compression fractures of the vertebrae are associated with osteoporosis. Improving the recognition of compression fractures is also the direction of the current solution. Now, deep learning can identify vertebral compression fractures on radiography and can help distinguish between fresh and old compression fractures, solving the challenge of identifying fresh compression fractures on radiography (39). Similarly, on MRI images, deep learning has also reached the level of specialists in identifying fresh compressed bones (40). More recently, a study by Suri et al. also adopted deep learning to accurately segment and identify the vertebral body and intervertebral discs on MRI, CT, and X-rays, which can offer substantial help in the clinical practice concerning spine-related diseases (41).

For hot keyword-osteoarthritis
Osteoarthritis is the most common musculoskeletal disorder, which mainly affects the hip and knee joints with large weight bearings, especially the knee joint. Initially, a proceedings paper showed the use of CNN to analyze Kellgren & Lawrence (K&L) grades based on knee radiographs (42 For the burst keywords-risk factor, identification, localization, and surgery Deep learning is often applied to the classification and diagnosis of diseases. However, with the development of deep learning, this technique will also be applied to perioperative management of orthopedic surgery. Nowadays, deep learning is being gradually explored in the study of orthopedic surgery. Since the visual field in arthroscopic surgery is often affected by different angles of the joint, which requires repositioning in different visual fields, automatic localization under the arthroscopic field of view is urgently needed.  (48). These studies demonstrate the potential of deep learning in assisting orthopedic surgery. For the postoperative aspect of orthopedic surgery, the prediction of postoperative complications is important. Zhu WB et al. used deep learning to infer the possibility of bone necrosis after femoral neck fracture internal fixation from postoperative X-rays, which can help clinicians make timely treatment (49).

For the timeline viewer of keywords
From the timeline viewer of keywords, we found that the study of fractures and osteoarthritis gradually began to study in-depth from 2016 to 2019. We also paid special attention to the direction of gradual in-depth research after 2019, such as osteoporosis, and bone tumors (osteosarcoma). Osteoporosis often occurs in older or postmenopausal women, and is usually detected because of fractures, thereby screening for osteoporosis may help prevent osteoporotic fractures in many patients. Deep learning can help predict osteoporosis and the possible fracture risk (50, 51). Although dual energy X-ray absorptiometry is the current gold standard for osteoporosis diagnosis, the lack of universality limits its clinical application. Accordingly, Loffler et al. used conventional CT combined with deep learning to predict osteoporosis (52). In addition, the present studies on bone tumor rely on radiographic imaging and histopathology.
Deep learning can help improve the segmentation of bone tumor in imaging (53), and can also be used in histopathology to quickly distinguish between normal tumor areas and necrotic tumor areas, thereby assisting to evaluate response to neojuvant chemotherapy (54,55). Therefore, the use of deep learning based on conventional imaging and histopathology to accurately screen osteoporosis, identify and classify bone tumors may become future research directions. Several limitations remain in this study. First, as the literature in the Web of Science Core Collection is constantly updated, and there is currently no uniform regulation on keywords related to the literature, the results of this study may differ from the actual number of documents included. Second, deep learning first appeared in seminars or conferences, and then gradually appeared in journals in the form of article. Since we consider that article has more systematic research, so it is strictly regulated in literature inclusion, which may introduce certain bias to this study. However, it is believed that literature-based visual analysis has undoubtedly laid a foundation for investigators to quickly understand the research hot-spots and development trends of deep learning applied in orthopedics.

Conclusion
In conclusion, according to the annual publication curve, the research on the application of deep learning in orthopedics has developed rapidly, and there is a good research prospect. Accordingly, researchers in the field of deep learning and orthopedics may benefit this study. The USA is the country with the largest amounts of articles in this field. University of California San Francisco, the institution with the largest amounts of articles, is also from the USA. However, cooperation between different countries is markedly insufficient. The application of deep learning in orthopedic research has been published in journals of various discipline categories, and multidisciplinary communication conforms to the mainstream of today's world. The current research hot-spots mainly focused on the classification and diagnosis of orthopedic diseases that rely heavily on medical images, such as fractures and osteoarthritis. The application of deep learning to reduce intraoperative risk, predict postoperative risk, screen osteoporosis, and identify, classify, and segment bone tumors may be at the fore-front of the future. At the same time, with the rapid development of deep learning applications in orthopedics, researches in this field will continue to evolve toward multicentric data and more perfect deep learning algorithms.

Data availability statement
The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Author contributions
CF: investigation, data analysis and visualization, and writing-original draft preparation. XZ, HW, YH, and ZL: investigation and editing. CT: conceptualization, supervision, and writing and revision. All authors have read and agreed to the published version of the manuscript.