A comprehensive search was performed across six databases namely PubMed, Scopus, IEEE Xplore, Embase, Medline, and Web of Science using keywords like “robotic-assisted surgery,” “minimally invasive surgery,” “deep learning,” and “computer vision”, for the time period 2017–2024. We limited the analysis to 2017–2024 for two reasons. First, 2017 corresponds to the earliest year in which deep learning–based approaches for surgical instrument analysis in minimally invasive/video-guided surgery were consistently represented in major indexing services, coinciding with the emergence of widely used community benchmarks and challenges (e.g., EndoNet-era laparoscopic video recognition and the EndoVis challenge series (6). Second, 2024 was the most recent complete publication year at the time of our search. We did not include 2025 because the year was ongoing during data collection and indexing and citation accrual would be incomplete.

We included papers that focused on laparoscopic or robotic-assisted surgeries besides utilizing DL models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Generative Adversarial Networks (GANs), or Vision Transformers (ViT). Because instrument-related surgical video studies are not consistently labeled with instrument-specific keywords in titles and abstracts, we intentionally used a high-sensitivity search strategy focused on MIS + deep learning + video/endoscopy terms. Specificity was then enforced during screening using predefined eligibility criteria requiring a direct link to instrument segmentation, detection, or tracking.

The search strategy is shown in Table 1 below:

A thorough process was implemented to ensure the inclusion of eligible studies. Firstly, the search results were exported in RIS format and imported into Endnote 21.3 (Clarivate Analytics, Philadelphia, PA, USA). Duplicate records were then removed, after which all retrieved studies underwent independent title and abstract screening by two reviewers, followed by full-text assessment by two reviewers to determine eligibility. Discrepancies were resolved by consensus. This strategy ensured high sensitivity at the search stage while maintaining strict specificity through subsequent screening based on predefined eligibility criteria.

Our exclusion criteria included research focusing solely on conventional image processing or non-DL machine learning models. Additionally, studies that were applied to ex-vivo testing or simulation kits were not considered. Secondary research like reviews and meta-analyses were excluded to maintain the focus on original studies. Furthermore, studies focusing on workflow recognition without a direct connection to surgical tool detection were also excluded, though those incorporating tool detection as part of surgical phase identification were included.

Data were extracted clearly and systematically from the 217 included papers. The extracted metadata included the title, list of authors, year of publication, citation count, title of the journal, publisher, institutional affiliations of authors, and keywords. Citation counts were extracted from each study's respective indexing database. We report absolute citation counts as primary bibliometric indicators. Because citations accumulate over time and are influenced by publication age and indexing practices, we also report citations per year (total citations divided by years since publication) for the most-cited studies and interpret cross-country and cross-institution comparisons cautiously. Field-normalized metrics such as field-weighted citation impact (FWCI) were not used because they were not consistently available across all retrieved records. Papers with 50 citation counts or above were analyzed critically, by extracting the DL methodology, type of procedure and application. This approach helped us understand the correlation between the popularity of these papers and the direction of research. Later, the metadata of the 217 papers were analyzed using specific bibliometric software application. It allows the identification of important patterns of research, the identification of influential authors, and the study of relationships in the structure.

After duplicates removal, the RIS file containing bibliometric data of the 217 studies was imported into Bibliometrix (22) a bibliometrics analysis tool driven by R programming language (version 4.1.3; R Foundation for Statistical Computing, Vienna, Austria). The data in this file was processed in a systematic manner to create tables and figures representing major trends and emergent patterns in the research landscape. The mapping of the co-authorship and citation patterns was created using the VOSviewer 1.6.20 (23) software package developed at Leiden University by Van Eck. Hence, these visualizations allowed deep insights into the relations among countries, authors, and institutions and thus shed light on the collaborative nature of research in DL in MIS.

The utilization of DL in the segmentation of surgical instruments for MIS has seen substantial advancement over the last eight years (Figure 1). The initial articles on surgical data recognition via deep learning first appeared in 2017, starting with eight papers (6, 24–30). The sector experienced consistent progress, resulting in a linear growth trajectory start in 2021. During this interval, an average of 35 papers were published per year over two consecutive years, indicating that the publication rate had reached a plateau. Nonetheless, this plateau was momentary, as the field resumed its upward trend, ultimately reaching a peak in 2023 and exceeding 50 publications.

Since 2017, 217 papers on the application of DL for surgical instrument recognition in MIS have been published, with an average citation around 20 per paper. These include 128 journal articles and 89 conference papers (Table 2). Different journals and conferences showed interest in this topic, resulting in 93 various sources. The average number of authors per article was 6.27, with 25.11% of them involving international co-authorships.

Geographical analysis reveals that many countries have made significant contributions to the field (Figure 2). The United States leads in the total number of publications (35) (26, 31–64), followed by China (25) (44–68) and Germany (24) (24, 65–87) Notably, France also stands out with 19 publications (6, 29, 88–103), and the United Kingdom (104–121) and Japan (8, 102–117) make considerable contributions with 18 and 17 publications, respectively. The data indicates that a large portion of publications, totaling 88, originate from Europe, with substantially fewer contributions from regions such as Latin America and Africa. Countries in East Asia, including Hong Kong, South Korea, Japan, Vietnam, and China, represent a significant proportion of contributions from that region.

A total of 370 institutions contributed to at least one article relevant to the production of the 217 included papers (Figure 3). The distribution of contributing institutions spans several countries, with notable contributions from France, Germany, the United States, the United Kingdom, and China. France's contributions were primarily driven by institutions like the University of Strasbourg (86 articles). German institutions like Furtwangen University (42 articles) and the German Cancer Research Center (DKFZ) (9 articles) also featured prominently. In the United States, Stanford University (17 articles) and Johns Hopkins University (16 articles), among others, were notable contributors. Additionally, it is relevant to note that two institutions with 11 articles were not included in the figure, namely the University of Texas Southwestern Medical Center and the Shanghai Jiao Tong University.

A total of 974 authors have contributed to articles in this field, however, a few prominent authors emerge (Table 3). Padoy (6, 8, 53, 69, 73, 88–91, 93, 95, 96, 98, 99, 101, 103, 122) has the most at 17 publications from the University of Strasbourg, followed closely by Stoyanov (53, 91, 102, 106, 109, 110, 113, 114, 116–121, 123) from University College of London with 15. Next, Jalal (65, 66, 68, 70, 71, 74–77, 80, 82, 86, 117) from Furtwangen and Mutter (6, 88, 90, 91, 94, 96, 98, 99, 101, 103) from Strasbourg, have 13 and 10 publications, respectively. Heng (25, 63, 106, 124–129). from the Chinese University of Hong Kong contributed 9 publications, while Jin (106, 124, 126, 130–133), Moeller (65, 66, 68, 70, 71, 74, 76, 77), Mascagni, (53, 88, 91, 95, 96, 98, 99, 122), Neumuth (65, 68, 71, 74–78), and Nwoye (69, 88–91, 96, 99, 101) have 8 article contributions each.

The analysis of the published studies showed that the countries with the highest number of citations were also the countries with the largest number of contributions (Figure 4). France has more than 1,000 citations, making it the country with the highest citation count, followed by Hong Kong (424). United Kingdom and China have the same number of citations at 273, followed by Japan at 254. Most of the top 10 most cited countries have over 100 citations, however, South Korea (97) and Canada (60) fall below this.

The top five most occurring keywords were “Laparoscopy” (223), “Deep Learning” (206), “Convolutional Neural Network” (115), “Surgical Equipment” (101), and “Robotic Surgery” (96) (Table 4). Further, “Videorecording” (86) and “Endoscopic Surgery” (62) were frequently mentioned.

The timeline (Figure 5A) clearly shows a growing trend in research activity of the keywords, especially from 2017 onwards, which is consistent with the number of article publications displayed in Figure 1. For example, mentions of “deep learning” increased from just 3 occurrences in 2017 to 74 in 2021, reaching 206 by the end of 2023. Similarly, “laparoscopy” grew significantly during the same period. Figure 5B illustrates the total occurrences of the top keywords. Although “Laparoscopy” dominates in frequency, other specific surgeries categorized under it, such as “Endoscopic Surgery,” “Transplant Surgery,” and “Cataract Surgery,” are also frequently found.

Although there were 93 sources in the form of journals and conference proceedings, a few specific journals stood out in terms of total number of publications. The journal with the highest number of articles published was the International Journal of Computer Assisted Radiology and Surgery (24, 36, 38, 41, 54, 56, 67, 73, 78, 83, 90, 98, 101, 109, 110, 117, 119, 126, 134–144), with 27 articles and an impact factor (IF) of 2.3. Lecture Notes in Computer Science (25, 58, 69, 84, 99, 103, 104, 118, 121, 130, 145–153) came next with 19 articles published followed by Medical Image Analysis (88, 89, 91, 93, 96, 102, 113, 124, 127, 154, 155) with a high IF of 10.9. Surgical Endoscopy (31, 40, 42, 50, 53, 95, 156–159) and IEEE Transactions on Medical Imaging (6, 7, 32, 34, 106, 114, 120, 125, 160, 161) both had 10 publications; however, they had IFs of 2.4 and 10.6, respectively. The next 5 journals all had fewer than 10 articles in each and IFs of less than 5.

The ten most cited articles using DL for surgical tool detection in MIS is shown in Table 5. Most of the highly cited articles are published by IEEE, with a few contributions published by Elsevier, JAMA and Springer. The most cited paper is “EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos,” (6) authored by Twinanda (2017, 924 citations) which is published in IEEE. The paper proposed a CNN model performs surgical tools classification and workflow recognition. This paper is one of the first papers in the field of instruments segmentation using DL model. Shvets' “Automatic Instrument Segmentation in Robot-Assisted Surgery using Deep Learning” (60) (2018, 433 citations), which explored binary, semantic and tools’ parts segmentation using CNN blocks in an encoder-decoder fashion. The model was more specific compared to other models as it was developed for robotic-assisted surgeries. Jin's 2018 article, “Tool Detection and Operative Skill Assessment in Surgical Videos Using Region-Based Convolutional Neural Networks” (61), with 327 citations, utilizes instrument segmentation to assess the surgeons' skill level for training purposes.

Moreover, Du's “Articulated Multi-Instrument 2-D Pose Estimation Using Fully Convolutional Networks” (120) (2018, 40 citations), focuses on pose estimation in laparoscopic surgery, relying on segmentation for accurate tool positioning. The accurate positioning was implemented through the detection of instruments joints. Colleoni has also developed a three-dimensional CNN that detects and localize surgical tools joints, published under “Deep Learning Based Robotic Tool Detection and Articulation Estimation With Spatio-Temporal Layers” (2019, 115 citations). Jin's (163) “Multi-task recurrent convolutional network with correlation loss for surgical video analysis” (127) (2020, 179 citations), and Yu's “Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques” (164) (2019, 106 citations) utilized surgical tool detection in workflow recognition. Nwoye's “Weakly supervised convolutional LSTM approach for tool tracking in laparoscopic videos” (101) (2019, 144 citations) utilized instrument segmentation to achieve accurate and reliable tool detection and tracking.

The 3 most cited articles from 2020 onwards in surgical video analysis are shown in Table 6. Out of the 3 articles, 2 were published in ScienceDirect Elsevier and 1 in IEEE. The most highly cited article, Jin, “Multi-task recurrent convolutional network with correlation loss for surgical video analysis,” (127) was published by ScienceDirect (Elsevier), with 179 citations. The paper talks about surgical video analysis, focusing on tool detection and phase recognition. Jin and his co-authors were able to achieve such task through the utilization of CNN and RNNs to maintain the temporal elements. Next, Kitaguchi published the second most cited article during the last five years at 99, titled “Automated laparoscopic colorectal surgery workflow recognition using artificial intelligence” (165). The paper introduces a new dataset “LapSig300” which encompasses 300 intraoperative videos that were collected from 19 high-volume centers. This novel work describes the dataset components such as surgical workflows that were classified into nine phases, 3 actions and 5 tools. Other notable contributions during this recent timeframe include Jin's (2021) “Temporal Memory Relation Network for Workflow Recognition” (125) with 81 citations. The work presents a novel network that detects the surgical phase based on the presence of surgical tools by utilizing LSTM that preserves temporal information.

Additionally, a growing subset of recent papers explores Transformer-based models and multimodal learning approaches, signaling a shift from purely convolutional or recurrent architectures toward methods capable of integrating video, textual, and even sensor data for more robust surgical context understanding.

A comparative analysis of the total number of citations and the average citations per article offers valuable information into the balance between research volume and individual article influence in each country. Figure 6 highlights France as an outlier, with the highest citation count despite not leading in the number of publications. This indicates that France produces a high volume of influential research as the average citations per article is high (Figure 7). Hong Kong follows with fewer total citations compared to France, but still significant in number. Importantly, Hong Kong stands out with the highest average citations per article, indicating that while the number of publications may be lower, the impact of each publication is very high. This shows that the research from institutions in Hong Kong is highly targeted and impactful within its academic field, possibly reflecting a focus on quality over quantity. The UK shows a moderate number of total citations, only slightly behind Hong Kong. Its average citations per article are similarly high, reflecting that although the influence may be less, it is consistent throughout publications. China and Japan display similar trends with moderate total citations. Interestingly, their average citations per article are lower, implying that while these countries may publish frequently, the individual impact of each publication tends to be smaller. Despite the large quantities of research, the overall influence of each article may not be as pronounced.

Despite the reputation of the US in global research, the total number of citations is somewhat lower than expected, considering it has the highest number of publications per country. It has a modest average citation rate per article, showing that the country produces a steady stream of research, but its per-article influence is not as strong as other countries. This may warrant a deeper look into the quality or visibility of US publications in this research area. With both a low total citation count and a modest average citation rate, Germany's overall contribution appears less prominent. However, its consistent output may indicate steady participation in the field without significant high-impact publications. The citations on publications from Italy portray an interesting pattern. Though not among the highest in total citations, Italy's publications have a strong average citation per article, suggesting that its research, while less frequent, tends to be highly impactful. This suggests a focus on producing fewer but higher-quality or more widely recognized studies. Finally, both South Korea and Canada are on the lower end in terms of total citations and average citations per article. This shows that their contributions are either newer or less widely recognized in this field. This could also reflect limited participation or a developing research presence in the field.

However, on an interpretative note, we need to be cognizant of the fact that citation-based indicators reflect visibility and uptake of publications and are influenced by publication age, venue, and citation practices. Accordingly, differences in total citations and average citations per article should not be interpreted as direct measures of methodological quality or clinical effectiveness.

In terms of geographical distribution, the United States occupies a central position in the co-authorship network, characterized by a high number of international collaboration links. Many countries demonstrate frequent co-authorship with the United States, reflecting its central role in the network. France, despite a lower publication volume, shows a comparatively higher average citation rate per article, suggesting a strong impact of its research output. China has also shown rapid growth in both publications and citations, indicating its increasing influence in the field.

Further, institutions such as the University of Strasbourg, Furtwangen University, and Johns Hopkins University are at the forefront of DL research in MIS. This highlights how these institutions can drive the collection of information. Institutions from Europe, North America, and Asia account for the majority of publications, indicating broad geographic representation across these regions. However, collaboration involving institutions from underrepresented regions appears less frequent in the co-authorship network. Additionally, author and institutional collaboration networks demonstrate frequent connections within certain clusters, particularly between European, American, and East Asian countries. Authors like Stoyanov, Padoy, and Heng are central figures in this collaborative environment, driving forward impactful research. There is no doubt still room for growth in terms of incorporating more diverse contributors from underrepresented regions, which could lead to new perspectives and novel advancements.

The collaboration network in these publications, which focuses on several countries, displays a thorough, but unevenly distributed, structure of global cooperation (Figure 8). The United States demonstrates the highest number of co-authorship links with countries in Europe and East Asia. Several European countries, including the United Kingdom, Germany, and France, show frequent co-authorship links with institutions based in the United States. Similarly, East Asian countries like China, Japan, and South Korea also exhibit strong collaborative research both within the region and with the US. Furthermore, while other countries, like India, Australia, and Canada make significant contributions, their collaborative networks are not as extensive or dense as those seen in Europe and East Asia, indicating that there is room for growth and integration.

Despite active research clusters in Europe, North America, and East Asia, participation from regions such as Africa and Latin America remains limited, highlighting the need for more inclusive global collaborations and funding initiatives to broaden the field's impact. There are a variety of reasons why certain authors or countries may be isolated. It is possible that the author's work might not be widely disseminated or accessible to the broader research community. This can be the result of publication in less prominent journals, or even limited participation in international conferences and networks. Additionally, if the author's work spans across multiple disciplines, it might not fit neatly into the established research networks of any single field, leading to isolation within both publication and citation networks. Further, geographical or institutional factors might limit the author's opportunities for collaboration. Over time, as the field develops, there may be an increase in both collaborations and citations across authors from different countries and institutions over time.

The co-citation network represents the relationships between authors based on how often they are cited together in these scientific publications (Figure 9). Padoy appears to be a central figure with a lot of connections, which suggests that he is highly influential and widely cited across different studies. Additionally, Stoyanov and Jannin also appear to be very important in their clusters, demonstrating substantial impact and regular co-citation with other authors. Some authors like Jalal, who were found to be isolated in publications networks, were also found to be isolated in the citation networks. Inter-cluster connections suggest that influential authors are not only leaders within their specific areas but also contribute to interdisciplinary research, bridging gaps between different fields and fostering collaboration.

Stoyanov plays a central role in collaborations, followed by Padoy and Heng, who act as bridges between various researchers (Figure 10). Author collaborations indicate high-impact research, with broadening opportunities for further collaboration. Given his extensive connections with other scholars, Stoyanov appears to have a crucial role in promoting cooperation both inside and outside of his cluster. Moreover, Padoy and Heng exhibit a noteworthy collaborative influence, serving as a bridge between various researchers. Certain researchers, like Loukas and Ito, seem to be more isolated with fewer connections.

The frequent occurrence of terms like “Deep Learning,” “Laparoscopy,” and “Convolutional Neural Networks” highlights the focus on laparoscopic surgeries as a key area for DL (Table 4). From these keywords, it is evident that the use of video data as opposed to images for training DL models is most common. Additionally, most DL models utilize CNNs in their architectures, given that it is the most frequent algorithm in the keyword analysis. This indicates that they have the DL algorithm with the most extensive application in image recognition and segmentation tasks. Additionally, Tables 5, 7 illustrate CNNs' popularity as they are widely utilized by the top cited papers in the field. The growing trend in the use of video-based data for DL training, as opposed to static images, is indicative of the shift towards its applications in surgery. This suggests that the future of DL in MIS may lie in improving intraoperative decision-making by providing real-time analysis and feedback to surgeons. This significant increase also aligns with the rapid technological advancements in DL algorithms and their increasing use in medical imaging and surgical robotics. It is clear that this is a developing field, as indicated by the rapid growth of publications and keyword use. By analyzing keywords in these research papers, important information can be acquired about research trends and focus areas in the field of DL use in MIS.

The keyword co-occurrence network in Figure 11 provides insights into key technical terms and their relationships in DL. The central term, “DL,” is closely linked with “image segmentation,” “computer vision,” “CNN,” and “AI,” reflecting the focus on DL techniques for image analysis in surgery. The green cluster highlights terms such as “image segmentation,” “image processing,” and “semantics,” which highlights some of the most commonly used techniques by DL algorithms in the processing of surgical videos. These models are critical for accurately tracking surgical instruments and automating workflow recognition. The blue cluster features “CNN” and “LSTM” (long short-term memory networks), suggesting that both spatial and temporal features are handled through a combination of neural networks. These tools enhance real-time surgical decision-making by identifying various phases of surgery. Lastly, the purple cluster emphasizes “object recognition” and “instrument detection,” addressing the importance of these tasks in developing a function DL algorithm. Recent works are using transformers, however, as it is till new to the field of surgical video analysis it less frequent given that the timeframe is since 2017, where CNNs were dominating the field.

The keyword co-occurrence network in Figure 12 is more focused on surgical and medical terminology, illustrating their relationships to MIS. The central node, “surgical equipment,” is closely connected to multiple terms such as “robotic surgery,” “endoscopic surgery,” and “laparoscopic cholecystectomy,” indicating the DL algorithms used had a particular focus on instrumentation in all these types of surgeries. The green cluster, which includes terms like “surgical instrument” and “surgical tools,” highlights the focus on instrument development and refinement, which is essential for enhancing precision and outcomes in MIS procedures. The connection to “transplant surgery” shows that advancements in surgical tools are not just limited to one type of surgery but also seem to have broad applications across different fields. In the red cluster, terms such as “medical imaging” and “neurosurgery” emphasize the integration of imaging techniques in surgeries, especially for complex procedures like brain surgery. This cluster demonstrates the critical role of imaging in guiding surgical procedures and improving surgical accuracy. The link to “surgical phase recognition” suggests the growing importance of recognizing and automating phases of surgery using advanced tools and DL models.

DL is significantly contributing to the surgical domain, particularly through convolutional neural networks (CNNs). CNNs are highly effective for processing image data, making them popular for the integration with other algorithms, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), which preserve temporal features. This combination is beneficial when maintaining the temporal information across different frames is critical, like surgical workflow recognition.

Recurrent networks are extensively utilized in tracking applications, such as tool tracking, where they maintain continuity by connecting subsequent frames and preventing the loss of recognized or split items. However, tasks that are within the spatial components only can be achieved using CNNs only like classification, localization, or segmentation of surgical instruments. Notably, Encoder-decoder architecture is suitable for generating segmentation masks, including binary, semantic, or tool parts. In this network, the encoder identifies and captures important features, then the decoder reconstructs them into an abstract form.

Three-dimensional CNNs are used to identify surgical tool joints as they harness data from the third dimension. Moreover, numerous articles utilize existing CNN-based models such as YOLOv3 and ResNet, illustrating the adaptability of CNNs across diverse architectural frameworks to tackle unique issues in surgical operations.

Within our corpus, instrument perception research in MIS can be grouped into three closely related tasks: detection (tool presence and coarse localization), segmentation (pixel-level delineation), and tracking/pose estimation (temporal continuity and articulation). Earlier high-impact studies frequently emphasized detection and tool presence classification, often as components of workflow analysis and training pipelines (166) where coarse spatial cues are sufficient, and were commonly evaluated on public laparoscopic video datasets such as Cholec80 (50).

More recent publications increasingly focus on segmentation, driven by the need for precise spatial understanding and by the maturation of benchmark datasets and challenges. Studies published include approaches that leverage transformer backbones and foundation-model paradigms for robust instrument segmentation under occlusion and scene variability, such as for example, adaptations of Segment Anything–style models for robotic instrument segmentation on EndoVis benchmarks [e.g., Surgical-DeSAM (10)]. Segmentation outputs are also being integrated into downstream applications such as objective performance assessment and context-aware assistance (167).

Tracking and pose estimation extend detection and segmentation across video frames and support higher-level scene understanding by enforcing temporal consistency. Recent spatio-temporal architectures have been used to track instrument motion (31) and articulation (163) and to support phase/step recognition, with benchmarking on datasets such as Cholec80 and EndoVis. Together, these complementary task families suggest a continued shift toward unified, real-time models that jointly model spatial detail and temporal dynamics.

The top ten most cited papers show the use of DL algorithms for a wide variety of purposes, including in MIS (Table 5). Twinada's “EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos” (6) introduced a very popular convolutional neural network architecture named EndoNet. Although it focuses on surgical phase recognition in a variety of surgeries, it has only been applied to laparoscopic surgeries in this study. Instrument segmentation is crucial for this process, enabling the accurate tracking of instruments and allowing the algorithm to recognize the current phase of the surgery. Similarly, Jin's 2018 article “SV-RCNet: Workflow Recognition From Surgical Videos Using Recurrent Convolutional Network” (7) also focuses on workflow recognition, specifically in laparoscopic cholecystectomy videos. Jin's approach combines a convolutional neural network and a recurrent neural network, allowing for better utilization of both the temporal and visual features present in surgical videos. Padoy's 2019 (162) and Dergachyova's 2016 articles also focus on similar objective, using neural networks and other DL algorithms to identify the various stages of surgical procedures and the utilization of endoscopic videos for the training of these algorithms.

Six out of the 10 most cited articles were published in IEEE, likely because the field was in its early stages of development and the research was more technical in nature. At that time, the focus was on algorithms and specific technology, which was not clinically applicable yet. As the field progressed and deep learning applications became more integrated into surgical practice, the research gained clinical relevance. This has led to an increase in publications in medical journals rather than the more technical journals like the ones published by IEEE. This shift reflects the growing importance of these technologies to surgeons as they transitioned from experimental to practical, clinical tools.