Transfer Learning via Articial Intelligence for Guiding Implant Placement in the Posterior Mandible: an in vitro Study

Background: To explore the capacity of a single shot multibox detector (SSD) and Voxel-to-voxel prediction network for pose estimation (V2V-PoseNet) based articial intelligence (AI) system in automatically designing implant plan. Methods: 2500 and 67 cases were used to develop and pre-train the AI system. After that, 12 patients who missed the mandibular left rst molars were selected to test the capacity of the AI in automatically designing implant plan. There were three algorithms-based implant positions. They are Group A, B and C (8, 9 and 10 points dependent implant position, respectively). The AI system was then used to detect the characteristic annotators and determine the implant position. For every group, the actual implant position was compared with the algorithm-determined ideal position. And global, angular, depth and lateral deviation were calculate. One-way ANOVA followed by Tukey’s test was performed for statistical comparisons. The signicance value was set at P< 0.05. Results: Group C represented the least coronal (0.6638±0.2651, range: 0.2060 to 1.109 mm) and apical (1.157±0.3350, range: 0.5840 to 1.654 mm) deviation, the same trend was observed in the angular deviation (5.307 ±2.891°, range: 2.049 to 10.90°), and the results are similar with the traditional statistic guide. Conclusion: It can be concluded that the AI system has the capacity of deep learning. And as more characteristic annotators be involved in the algorithm, the AI system can gure out the anatomy of the object region better, then generate the ideal implant plan via deep learning algorithm.


Background
Optimal three-dimensional (3D) position of dental implants is the key for functional and aesthetic outcomes 1, 2 .
Virtual implantation improves the accuracy of implant placement. Subsequently, guided surgery opens up possibilities to transfer the 3D pre-surgical planning to implant surgery, to achieve ideal 3D implant placement [1][2][3][4] .
As surgical templates can reduce human error, guided surgery has come to play an important role in precise bone drilling and implant placement [4][5][6] . However, traditional surgical templates still have several disadvantages, including extensive manual work for preoperative planning and increased cost of template fabrication for each case. Besides that, the accuracy of guided implant surgery depends on the dentist and technician, while there are learning curves and expenses related to training 7 . Our approach to improve the traditional surgical template is the use of arti cial intelligence (AI).
AI refers to a serious of technologies that allow computers and machines to imitate human intelligence 8 . It has come to play an important role in healthcare, including analyzing a diverse array of patient data and simulating human logic to perform some tasks [9][10][11][12] . AI iteratively learn the intrinsic statistics underlying the pairing data and algorithm, to make plan on unseen data 8, 10 . Recently, there has been a signi cantly increasing tendency in the number of AI studies in medicine, mostly for disease detection and classi cation purposes 8, 10,[13][14][15] . And the use of AI for radiological images reading has been widely explored 12,13,16 .
AI system would enable all clinicians to design and practice the complicate cases at the same level of practical expertise as the very best clinicians. Since database could be shared between physicians without the privacy risks of leaking patient data, there is almost limitless potential to renew the system learned from the diverse physicians and diverse patients 11,17 .
Recently, AI technique has also been used in dentistry for various purposes 8,18,19 . For instance, Shintaro et al. 20 used AI to accurately classify dental implant brands from various panoramic X-ray images, Joel et al. 21 presented a deep learning system for mandibular canal segmentation, Jin-Ju et al. 22 used AI for automatic 3D analysis of bone alteration after maxillary sinus augmentation. Jun et al. 16 developed a deep neural network to determine the periimplant marginal bone loss. The accuracies for most of the tasks are promising.
The usage of AI for dental diagnostics has been developed quickly in the recent years 8, 18, 23 . It worth further focusing on the possibilities of automatically generated implant surgery plans. To our knowledge, there were rare studies in this eld. This study explored a deep learning model, named SSD 24 , with V2V-PoseNet 25 as its backbone architecture, to automatically recognize the target region, detect the bone contour and capture the characteristic annotators, then generate the implant plan. Then we assessed the accuracy of the automatically designed templates, which satis es the well-established algorithm.
The purpose of this in vitro study was to explore the capacity of the AI system in automatically designing implant plan, through deep learning. And compare the accuracy of 8, 9 and 10 characteristic annotators based algorithm guided surgery in posterior mandible.

Methods
All patients consented to the use of their data for research purposes and signed informed consent forms accordingly.
This study protocol was approved by the ethical committees of the Guanghua School of Stomatology, Hospital of Stomatology, Sun Yat-sen University (KQEC-2020-063-01).

Sample
The datasets were divided into pre-training, training and test sets. The pre-training set, including 2500 various edentulous sites undergoing CBCT image (2018.10-2020.10), were used to train and develop the model parameters.
After that, the model hyper parameters and architecture selected using the training set, which includes 67 cases who undergo CBCT and intraoral scanning. The selected 67 patients should meet the following conditions: 1) just miss the mandibular left rst molar; 2) with the normal occlusion; 3) without any metallic prosthesis; 4) ≥18 years old. Finally, 12 patients undergoing CBCT (NewTom VGi CBCT imaging unit, Verona, Italy) and intraoral scanning (SWEDEN & MARTINA, Veneto, Italy) who met the above criteria were collected for the following automatically designed implant surgery guide generation. To avoid over-optimistic results because of over-tting, that is, memorizing speci c features of edentulous sites, each of the three sets had an independent set of patients.

Software work ow
The AI software work ow is presented in Fig. 1. A model named SSD 24 has been proposed for edentulous site and related key points detection.

Evaluation of edentulous area condition
The image ow chart of automatically implant position design is shown in Figure 2. SSD model, a convolutional neural network (CNN), was used to construct the AI system. All les were uploaded to SSD for evaluation of edentulous area condition. First, the software automatically generated a panoramic image and detect the missing tooth position through CBCT (DICOM data) reading. Second, the region of interest (ROI) was constructed. All slices were extracted from ROI of the panoramic image ribbon and a small box area was cut out from the DICOM data for processing. Third, detect the bone contour and capture the designed characteristic annotators based on V2V-PoseNet. Then the implant axis was generated. Finally, the implant position was created automatically.

haracteristic annotators capture
The characteristic annotators were recorded based on the condition of the alveolar bone 3D features, the neighboring teeth and the opposite teeth (

Implantation plan generation
All of the characteristic points were detected using V2V-PoseNet method. The characteristic points provided information about the implant position. In brief, there are three measurements. The rst measurement showed the central spot of the implant neck (P a ), the second measurement showed the axial feature spot (P b ), and the third measurement showed the implant axis (Fig. 4). The linkage between P a and P b represented the implant axis. Besides that, the inferior alveolar nerve were marked by the operator. All the planed implants were bone level implants (NobelReplace CC; Nobelbiocare) of 4.3 mm × 10 mm or 4.3 mm × 11.5 mm. The implant length was determined by the operator, to keep a safe distance (≥ 2mm) from the inferior alveolar nerve.
There are three kinds of calculation methods for the three measurements (P a , P b and implant axis). Accordingly, for every case, there are three groups of implantation plans. They are (1) Group A: 8 points (P1-P8) dependent implant position; (2) Group B: 9 points (P1-P9) dependent implant position; (3) Group C: 10 points (P1-P10) dependent implant position. In each group, P a and P b follow speci c algorithm (Table 1). Based on the plan, implant surgical templates were fabricated from ultraviolet sensitive liquid resin E-Dent (EnvisionTEC). Each surgical guide was 3mm thick. All the templates were teeth-supported and full guided. And all the templates were equipped with metal sleeves to serve as the guidance for drill key. According to the 12 partially edentulous cases, physical surgical resin models (B9R-1-RED; B9Creations) were 3D printed. For each guide and surgical model, one implant was inserted on the basis of the fully guided surgical protocol.
The implant site preparation and implant placement were performed using commercial surgical guide kits (Nobel Surgical Kit). Once the templates were properly tted on the in vitro models, all the drilling steps and implant placement took place without removing the guide. The manufacturer's recommendations were followed through the process of implant site preparation and implant placement, to avoid unpredictable deviation. After the osteotomy, a planned NobelReplace CC implant was placed with the guided portable adaptor. And all the surgeries were accomplished by the same clinician, to estimate the intra-observer variability.

Accuracy evaluation
Following implant placement, speci c implant scan body was used for accuracy evaluation. In brief, the scan body was attached on the implant, and intraoral scanning was performed to obtain STL data. Then the reverse engineering software (Geomagic Studio; Raindrop Geomagic) was used to analyze the deviation between the implant position and the algorithm bases ideal implant position.
For every group, the virtual implant position according to the various algorithm was marked for reference. Through aligning the data sets, the actual implant position was compared with the algorithm-determined position, and the deviations were performed in 3D 26 (Fig. 5). Besides, perforations of the apical buccal/lingual bone were assessed.
The angular deviation was de ned as the angle between the centerline of the actual and virtual algorithm-determined implant (α). The global deviation was calculated as the 3D distance of apical/coronal center between the placed and algorithm-determined implant. The global deviation was subdivided into the lateral deviation (perpendicular to it) and the depth deviation (along the implant axis). Moreover, the lateral deviation was split into the bucco-lingual deviation (along the bucco-lingual axis) and the mesio-distal deviation (perpendicular to the bucco-lingual axis). Furthermore, to illustrate the deviations in exact directions, a positive value was used when the placed implant was mesial/buccal/apical to the virtual algorithm-determined implant, while a negative value corresponded to distal/lingual/coronal placement compared with the algorithm-determined position. For this study, all the measurements were accomplished by one observer.

Statistics
All quantitative data are presented as means ± SD. One-way analysis of variance (ANOVA) followed by Tukey's test was performed for statistical comparisons. The signi cance value was set at P< 0.05. All data were analyzed with SPSS statistics v25.0. (IBM Corporation, Armonk, NY, USA).

Results
The parameters for the three different groups are shown in Tables 2 and 3. The tables illustrating the differences among the groups are presented in Figure 6A-F. In Table 2, the global (apical and coronal), depth and angular deviations are shown. According to the statistical tests, Group C represented the best in accuracy. Group C represented signi cantly lower apical global deviation (1.157±0.3350) than Group A (1.717±0.3355) (P=0.0045). The same trends were observed in coronal global deviation and depth deviation, with no signi cant difference. And for angle deviation, Group B (4.424±1.505) represented the best, however, no signi cant difference was shown.
The lateral, bucco-lingual (apical and coronal), and mesio-distal (apical and coronal) deviations were shown in Table   3. Group C represented the best in accuracy. The apical bucco-lingual deviations were signi cantly smaller in Group C in both buccal direction (0.6377±0.5287, P = 0.0103) and absolute value (0.7149±0.4069, P =0.0011) when compared with Group A. Group C represent the same trend in the lateral, coronal bucco-lingual, coronal mesio-distal, and apical mesio-distal (mesial direction) deviations, with no signi cant difference. And for the mesio-distal (absolute value) deviation, Group B (4.424±1.505) represented the best, with no signi cant difference.
In regard to the exact direction, for all the groups the placed implants were buccal to the algorithm calculated position. Group C represented the least buccal shift both at the apical (0.6377 ±0.5287 mm) and coronal level (0.2088 ±0.2523 mm). And the difference was signi cant between Group C and Group A at the apical level (P = 0.0103).
Moreover, in this study no buccal or lingual bone perforation was found. For mesio-distal deviation, Group C showed the least shift. And for Group C, no obvious tendency was found either towards mesial or distal, at coronal or apical level. Moreover, there was neither buccal nor lingual perforation in all the groups.

Discussion
The technology underlying AI applications was machine learning, which iteratively learned the inherent statistical patterns from speci c datasets and algorithms 5 . However, as the most effective number of characteristic annotators was unclear, the data used to train the machine was small, and the result generation process was invisible, there were doubts about the accuracy and replicability 8, 27 .
This study developed a SSD and V2V-PoseNet based system for image recognition and characteristic points capture, then implant position design. Based on the deep learning frame, the datasets were trained using the designed SSD algorithm to generate image identi cation and location models. Besides that, V2V-PoseNet method was used to detect the feature of the ROI. After training the image set, 12 independent test images were used to evaluate the trained model. Our AI system showed good performance for the implant position design and the accuracy is similar with the recently reported statistic guide 28 .
Moreover, we found that the number of characteristic annotators in uenced the performance of our AI system. A trend of more characteristic annotators induced more accuracy was observed. In brief, Group C performed better than Group A and B, for global (apical and coronal), angular, depth, lateral, bucco-lingual and mesia-distal deviations. In another word, Group C was the most accurate, and it was used to further improve the AI system. This phenomenon reminded us that our AI system was full of learning ability, and more information contributed to the accuracy.
In Group C, the average deviations were 0.6638±0.2651 (range: 0.2060 to 1.109 mm) for the coronal deviation, 1.157±0.3350 (range: 0.5840 to 1.654 mm) for the apical deviation, and 5.307 ±2.891° (range: 2.049 to 10.90°) for the angular deviation (  Fig. 6C). And for all the three groups, the implants were in a more coronal direction compared to the virtual position. This phenomenon also has been reported by Vercruyssen et al 29 and Zhaozhao et al 26 . However, as more characteristic annotators been calculated, the deviation decreased.
Regarding the bucco-lingual deviation, the implants tended to move buccally both at coronal and apical level, and a greater deviation was found at the apical level. The coronal/apical deviation may occur during the processes of either drilling or implant placement 26 .However, the bucco-lingual deviation was small in Group C (Table 3, Fig. 6E). As to the mesia-distal deviation, for Group C, there were nearly no tendency neither toward mesial nor distal, both at the coronal and apical level ( Table 3, Fig. 6F). For Group C, tiny bucco-lingual and mesia-distal deviations would bring mechanical equilibrium, which be bene cial to the long-term success of dental implant.
The high accuracy of our AI system mostly depended on SSD, a deep learning technology based superior convolutional neural network (CNN) algorithm 30 . SSD is one of the best AI models in image identi cation and location, which consist of at least 16 layers 31,32 . It is an effective object detector for multiple targets recognition within just one stage 10  To sum up, SSD and V2V-PoseNet were used to train the datasets 32,35 , and the deep learning neural network, especially the 10 characteristic annotators based algorithm group, was fairly accurate and clinically reliable. The core of SSD model is to explore objects and corresponding locations in given bounding boxes, realizing object location prediction 36 . It relatively simpli ed the processing, while increasing the identi cation.
However, due to the limited test size, the results in this study should be interpreted with caution, and more training images were needed to improve the performance of our system. Besides that, there are some real clinical elements, such as limited view and inter-occlusal distance 37 that the present in vitro study cannot re ect. Moreover, based on the pronounced learning abilities of the AI system, in the future more key parameters would be incorporated into the algorithm to gure out the best implant plan.
It can be concluded that the SSD and V2V-PoseNet based AI system has the capacity of deep learning from the data and speci c algorithms. And as more characteristic annotators be involved in the algorithm, the AI system can gure out the anatomy of the object region better, then generate the ideal implant plan via speci c algorithm.

Declarations
Ethics approval and consent to participate This study was approved by the ethical committees of the Guanghua School of Stomatology, Hospital of Stomatology, Sun Yat-sen University (KQEC-2020-063-01).

Consent for publication
All patients consented to the use of their data for research purposes and signed informed consent forms accordingly.

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Figure 1 AI software work ow.   Diagram of three measurements (Pa, Pb and Implant axis).