This study was performed in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [31] and registered in the International Prospective Register of Systematic Reviews (PROSPERO) database (registration no. CRD42022323650). The study was designed based on the population, intervention, comparison, and outcomes question as follows:

The following electronic databases were searched: PubMed, Scopus, Web of Science, Cochrane Library, and Google Scholar. In addition, the references cited in the full-text and relevant articles in several dental journals such as Journal of Clinical Periodontology, Periodontology 2000, Journal of Periodontology, Clinical Oral Implant Research, and Clinical Implant Dentistry and Related Research were manually searched. The search strategy for each database was established using Boolean logic (AND, OR, NOT) to combine search terms. The individual Boolean operators established for searching for each data source are listed in Table 1. The retrieved articles were sorted using reference manager software (EndNote; version x9.2; Clarivate Analytics Inc).

This systematic review considered all original studies published until March 21, 2022 (ie, clinical trials; cohort studies; case-control studies; and preclinical, in vitro, and ex vivo studies) that evaluated the accuracy of AR navigation without any language restrictions. Studies without metric measurements for reporting implant positions and those that were unrelated to dental implant placement were excluded. Similarly, systematic or narrative reviews without implant placement accuracy assessments, letters to the editor, author or editorial opinion articles, and epidemiological studies were excluded. Gray literature and unpublished data were not included in this study to avoid a possible research bias. Only studies that reported deviation values between planned and executed implant positions were included in the meta-analysis. The detailed inclusion and exclusion criteria are listed in Table 2.

For study selection, 2 reviewers (HNM and DHL) independently screened and selected eligible studies based on inclusion and exclusion criteria. The agreement between reviewers was calculated using Cohen κ coefficient. Any disagreement was resolved through a discussion between the reviewers or consultation with an expert (VVD). Relevant information was extracted from the eligible studies, which included study characteristics (ie, authors, year of publication, and study design), participant characteristics (ie, number of participants, type of jaws, number of jaws, type of edentulism, position of the implant in the jaws, and number of implants or drilling channels), AR system characteristics (ie, type of digital image projected on the surgical site, tracking technology, user interface, and image display technology), the outcome of interest (ie, deviation of image-to-registration and positional deviations between planned and placed implants), comparison groups (ie, FH method, CN method, TG method, and traditional dynamic guiding system), and major findings from each included study. In case of missing or ambiguous data, the corresponding authors were contacted for data clarification. Information was extracted independently by 2 reviewers and recorded on an electronic spreadsheet (Microsoft Office Excel 2019; Microsoft Corp). After the initial data extraction, the final data table was re-evaluated by both reviewers, and inconsistencies were corrected either by discussion between the 2 reviewers or by consultation with a third reviewer (VVD).

The risk of bias in the included studies was assessed using the Risk Of Bias In Non-randomized Studies of Interventions tool [32] based on seven domains: (1) bias owing to confounding, (2) bias in the selection of study participants, (3) bias in the classification of interventions, (4) bias owing to deviations from intended interventions, (5) bias owing to missing data, (6) bias in outcome measurement, and (7) bias in the selection of the reported result. The overall risk of bias in the included studies was judged as “low risk” if all the domains were judged to have a low risk of bias, “moderate risk” if all or some domains were judged to have a moderate risk of bias, as “serious risk” if at least one domain was judged to have a serious risk of bias, and as “critical risk” if at least one domain was judged to have a critical risk of bias. The “no information” category was used for studies in which insufficient data were reported to permit judgment. Publication bias in the included studies was assessed using Egger linear regression statistical test [33]. The Duval and Tweedie trim-and-fill method [34] was used to further assess the effects of publication bias and redress asymmetry.

All statistical analyses were performed using R statistical software (version 3.6.0; R Foundation for Statistical Computing), and the significance level was set at .05. Weighted bars and traffic light plots were used to visualize the results of the risk of bias assessments, and funnel plots were generated to visually report publication bias.

The positional deviations between the planned and placed implants were analyzed using 6 outcome variables: lateral coronal deviation (LCD), lateral apical deviation (LAD), global coronal deviation (GCD), global apical deviation (GAD), depth deviation, and angular deviation (Figure 1).

To obtain a general view of the accuracy of implant placement performed using AR navigation, single-arm meta-analyses for continuous data that involved estimating the mean deviations for each deviation variable were used for data analysis. Weighted means with 95% CIs were assessed, and subgroup analyses were performed according to deviation outcome variables (ie, lateral, global, depth, and angular deviations).

To compare the accuracy of AR navigation (experimental group) with other implant placement methods (control groups), 2-arm meta-analyses for continuous data were performed. The standardized mean difference (SMD) with a 95% CI between the experimental and reference data sets was analyzed to estimate the effect size [35]:

where d_mean is the difference in mean values between the experimental and control groups and SD of the measurements. Subgroup analyses were performed to compare the AR navigation group with the FH, TG, and CN navigation groups.

Meta-analyses were performed using either a random effects model or a fixed-effects model based on the heterogeneity among studies evaluated by the Higgins I² statistic [36]:

where Q is the chi-square statistic and df is the degree of freedom. A higher I² value indicates stronger heterogeneity across studies. In cases of I²≥50%, the data were considered statistically heterogeneous, and a random-effects model was used for data analysis.

The search results are shown in the PRISMA flowchart (Figure 2). In summary, 425 articles were selected after the initial search. The remaining 300 records were screened after excluding 125 duplicate articles. After screening the titles and abstracts, 270 irrelevant articles were excluded. Therefore, 30 articles were subjected to full-text assessment. Among them, 15 articles were considered eligible for narrative review, 8 articles were considered suitable for single-arm meta-analysis, and 4 articles were included in the 2-arm meta-analysis. Cohen κ coefficient showed a high degree of agreement between the 2 reviewers when selecting the studies (κ=0.88).

The characteristics of the included studies are summarized in Multimedia Appendix 1 [20-23,37-47]. Among the 15 eligible studies, 5 studies [22,23,37,38] involved volunteers or patients, 2 studies were performed on cadavers [39-41], 5 studies [20,42-45] involved plastic models, and 3 studies [21,46,47] used actual patient casts. Regarding the type of edentulous jaw, the number of mandibles (98/157, 62.4%) was higher than that of maxillary jaws (59/157, 37.6%), and the amount of partial edentulism (9/14, 64%) was higher than that of complete edentulism (5/14, 36%). The number and type of edentulism were not reported in a previous study [23]. Regarding implant placements, 7 studies involved [22,37-41,45] dental implants, 2 studies [21,47] used pins, and 4 studies [42-44,46] exploited drill channels to measure implant positional deviations. Two studies [20,23] did not clarify the number of implant placements.

Regarding AR technologies, most AR-assisted implant navigation systems allow the appearance of virtually planned implants. Some systems provide additional guidance such as the surgical path or drilling axis of the drills and the alveolar nerve or mandibular canal. Some systems included a warning function to alert surgeons and aid adjustments to the surgical strategy. In particular, one study [46] adopted the AR technology to facilitate the use of static implant-guiding systems. Most studies applied marker-based tracking technology; only 3 studies [23,38,47] used marker-free tracking technology for real-time image displays during implantation. To display the AR image, a head-mounted display (HMD) was used in 7 studies [22,39,40,42,44-46], an integral videography (IV) overlay device was used in 5 studies [21,23,38,43,47], and a smartphone was used in 1 study [41]. The general structure of the AR image display devices is summarized in Figure 3, and detailed information on AR-assisted implant navigation systems reported in the included studies is provided in Figure 4 and Table 3. Among the studies that included a control group, 2 [21,45] included the FH group, 2 [41,45] included the TG group, and 4 [40,41,44,47] included the CN group as the control group.

The risk of bias assessment is shown in Figure 5. An overall moderate risk of bias was found in 4 studies [20,40-42]. Three studies [23,37,38] reported insufficient data to permit judgment. The main limitation of the included studies was the limited sample size, which might have affected the results, and some articles did not specify the blinding of the outcome assessment. Therefore, though there was no serious and critical risk of bias among the assessed domains, domain 6 (ie, bias in outcome measurement) and domain 7 (bias in the selection of the reported result) exhibited higher chances of a moderate risk of bias than the others.

The funnel plot (Figure 6) and the result of Egger test indicated the presence of asymmetry (95% CI −4.78 to −1.93; df=−4.604; P<.001). Visual inspection of the contour-enhanced funnel plot was performed to assess potential publication bias. The asymmetry in the plots, which might be owing to the increased missing data present in the contours of the statistically significant region (gray shaded regions) than in the areas of nonstatistical significance (white region), suggests that the cause of the asymmetry might be because of factors other than publication bias such as variable study quality [48]. After adjusting for missing data using the Duval and Tweedie trim-and-fill method [34], 11 data points were added to adjust the asymmetry of the funnel plot and reduce the effects of existing publication bias (95% CI −2.89 to −.82; df=−1.097; P=.28).

A total of 346 implant sites were evaluated (181 implants, 59 drilling channels, and 106 parallel pins). Detailed descriptive characteristics of the AR-guided implant placement systems used in the included studies are presented in Table 3. The detailed positional deviations (mm) of the implants placed with AR navigation regarding the coronal, apical, and global deviations are shown in Figure 7. Accordingly, the pooled total weighted mean lateral deviation was 0.90 (95% CI 0.78-1.02) mm, with a deviation of 0.86 (95% CI 0.68-1.05) mm at the coronal region and 0.93 (95% CI.75-1.11) mm at the apex. The pooled total weighted global deviation was 1.18 (95% CI 0.95-1.41) mm, with a deviation of 0.89 (95% CI 0.58-1.21) mm in the coronal region and 1.47 (95% CI 0.85-2.10) mm at the apex. The pooled weighted depth deviations and angular deviations were 0.78 (95% CI 0.48-1.08) mm and 3.96° (95% CI 3.45°-4.48°), respectively.

Forest plots representing the SMD between the implant positional deviations in the AR navigation and control groups are shown in Figure 8. The AR navigation group showed significantly higher accuracy than the FH group (SMD=−1.01; 95% CI −1.47 to −0.55; P<.001) and CN groups (SMD=−0.46; 95% CI −0.64 to −0.29; P<.001), and similar accuracy to the TG group (SMD=0.06; 95% CI −0.62 to 0.74; P=.73). Detailed positional deviation comparisons are shown in Figure 9. For lateral deviation (mm), the AR navigation group exhibited significantly smaller deviations than the CN group (SMD=−0.68; 95% CI −0.92 to −0.43; P<.001; I²=47%). For global deviation (mm), the implants placed in the AR navigation group showed smaller deviations than those in the control group (SMD=−0.37; 95% CI −0.98 to 0.23; P=.18; I²=78%). The AR navigation group exhibited significantly smaller deviations than the FH group (SMD=−1.20; 95% CI −2.03 to −0.37; P=.02; I²=0%) and the TG group (SMD=−0.10; 95% CI −0.10 to −0.09; P<.001; I²=0%) but showed larger deviations than the CN group (SMD=0.15, 95% CI 0.08-0.21; P=.02; I²=0%). Regarding the depth deviations (mm), the AR navigation group exhibited smaller deviations than the CN group (SMD=−0.52; 95% CI −0.89 to −0.16; P=.02; I²=16%). Angular deviations (°) of the implants placed in the AR navigation group were also significantly smaller than those of the implants placed in the control groups (SMD=−0.35; 95% CI −0.66 to −0.05; P=.03; I²=67%). Specifically, the AR navigation group showed smaller deviations of placed implants than the CN group (SMD=−0.38; 95% CI −0.73 to −0.02; P=.04; I²=75%) and the FH group (SMD=−0.71; 95% CI −2.42 to 1.01; P=.12; I²=0%) but exhibited larger deviations than the TG group (SMD=0.38; 95% CI −0.32 to 1.08; P=.29).

This systematic review aimed to evaluate the accuracy of AR navigation and compare it with that of FH, TG, and CN methods. The results of the meta-analysis indicated that the errors of implant placement using AR navigation were within the safety zone, which is recommended to be in the range of 1-2 mm horizontally and vertically [1,49-51] and up to an angle deviation of 5° [49]. The results also indicate a higher accuracy of AR navigation than the FH method. In addition, the accuracy of AR navigation was higher than that of the CN method, in which the operators needed to look away from the surgical sites to follow the guidance displayed on the computer screen. Although the implants placed using AR navigation exhibited global positional deviations similar to those of the implants placed using the TG method, they exhibited significantly larger angular deviations.

The accuracy of TG implant surgery is superior to that of the FH method [52]. In addition, the accuracy of dynamic implant navigation systems is comparable with that of the static guiding method [10,53,54]. In this review, most of the introduced AR navigation methods were developed by integrating innovative AR technology into traditional implant navigation systems, in which AR was used to project the patient parameters, relevant radiographic images, 3D reconstruction of the preoperatively planned implant, or output of the navigation system screen to allow the surgeon to visualize real-time dynamic guidance without being forced to look away from the patient’s mouth [21,22,39,40,42,44,45]. The use of AR navigation provides surgeons with greater flexibility by allowing adjustments of the presurgical plan during real-time operation following the actual surgical site and oral conditions that would not have been possible with static guiding systems. In particular, in several advanced AR-assisted navigation systems, digital guidelines and warning alerts were included to prevent tolerances of the drills from the safety zone and planned implant positions [40,43-47]. Hence, it is expected to significantly improve the accuracy of traditional dynamic navigation. Moreover, AR technology can be combined with a surgical guide template to enhance both the accuracy and visualization of the implant guidance. Lin et al [46] reported a significant reduction in implant positional deviations by integrating surgical guide templates with AR technology.

Similar to CN systems, AR navigation involves the same phase of preoperative implant planning based on CBCT and intraoral data of patients and requires a navigation system for tracking the real-time position of the surgical instruments. However, it uses innovative imaging technology in the form of AR to overlay the digital presurgical plan over the actual surgical site. Generally, the setting of an AR-assisted dental implant navigation system consists of 3 components: image registration, image tracking, and AR image display devices.

Marker-based and marker-free methods are the 2 main approaches used for image registration and tracking. For marker-based image matching, specific artificial markers that are attached to the reference template applied in the patient’s mouth and implant drilling instruments are used to merge the digital presurgical plan with the real operation site. In contrast, the marker-free approach uses a point cloud–based registration method and tooth shape tracking methods to merge digital images in a real environment. In this review, most reported AR navigation systems relied on marker-based image registration and image tracking techniques [20-22,39-46], and only 3 studies have reported marker-free AR navigation systems [23,41,47]. Compared with the marker-based registration method, marker-free methods do not require additional markers placed on the patient, which accelerates the registration speed, simplifies the surgical guiding process, reduces patient discomfort, and increases the surgeon’s convenience [47]. However, it is more challenging to use teeth as natural landmarks for tracking because they have less texture, are difficult to distinguish from the background noise using either structured light or stereo vision, occupy only a small portion of the camera view, and are easily covered by lips or other anatomical structures during the movement of the patient and surgical instrument [23,38].

As for AR image display, see-through HMD devices are the main tools [20,22,37,39,40,42,44-46]. Based on image display principles, HMD devices can be classified into optical see-through (OST) and video see-through (VST) HMD [55]. For OST HMDs, half-transparent mirrors placed in front of the user’s eyes are used to optically combine real and digital world images in the user’s eyes. HoloLens (Microsoft Corp) is a popular commercialized OST HMD that enables physicians to obtain immediate insight into patient information by overlaying it with the view of the clinical scenario. It helps medical students gain a better understanding of complex human anatomies or a better experience with treatment procedures and assists patients during rehabilitation and treatment [56]. In this review context, effective and successful applications of HoloLens as an image display component of an AR-assisted implant navigation system have been introduced in 2 studies [22,45]. In contrast to the OST HMD, VST HDMs use 2 miniature video cameras mounted on the headgear to capture real-world images and electronically combine them with digital images. In addition, some AR navigation systems allow AR images to be viewed through IV overlay devices with the surgeon’s naked eye [21,23,38,43,47]. The IV overlay display is a new tool for autostereoscopic display that uses a fast image rendering algorithm to project a digital image through a microconvex lens array using multiple rays so that the observer can view animated 3D objects from various directions as if they are fixed in 3D space [57]. In particular, a smartphone-based AR method for intraoperative implant visualization and final verification of implant position via a dedicated smartphone app was introduced in a study that relied on a marker-based tracking method [41].

This systematic review has some limitations, including the relatively small number of included studies owing to the limited research in this new area and some sources of bias. Another limitation is the risk of high heterogeneity among the included studies. The inconsistency in the included studies was owing to the uniformity of implant placement error measurements, as multiple variables (lateral coronal and apical, global coronal and apical, depth, and angular deviations) were considered. The variety in the selection of the control group (FH, TG, and CN groups) and the subjects of the experiment (human, cadaver, artificial model, and patient’s cast) also contributed to the high heterogeneity among the studies. Therefore, in this review, a random effects model was used to perform a meta-analysis and subgroup analyses were performed to reduce heterogeneity. Further studies that consider patient- and surgeon-related factors, such as user satisfaction, convenience, and comfort, should be conducted to further our understanding of the existing AR-assisted implant navigation systems.

This systematic review supports the relevance of AR-assisted implant navigation methods. The meta-analysis showed that the accuracy of AR implant navigation was comparable with that of the highly recommended dental implant–guided surgery method, TG, and even superior to that of the conventional FH and CN methods. It should be noted that although AR implant navigation may be considered an effective immersive surgical guidance for dental implant placement, limited studies regarding the clinical application are available. Further studies that consider patient- and surgeon-related factors, such as user satisfaction, convenience, and comfort, should be conducted to further our understanding of the existing AR-assisted implant navigation systems.