An orthodontic process monitoring method based on a three-dimensional model prediction of a dentition
By combining a multi-step algorithm with deep learning and simulated annealing, the problems of matching accuracy and computational efficiency in 3D tooth reconstruction were solved, enabling accurate capture and reconstruction of dynamic changes in the dentition, thus improving the precision and effectiveness of dental treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-23
- Publication Date
- 2026-07-24
Smart Images
Figure CN122440339A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of orthodontic technology, and in particular to a method for monitoring the orthodontic process based on a three-dimensional model of the dental arch. Background Technology
[0002] Orthodontic treatment requires tracking the movement of a patient's teeth, typically necessitating multiple visits to the hospital or clinic for diagnosis and treatment evaluation during the treatment cycle. With the continuous development of mobile networks and information technology, telemedicine is increasingly favored by patients and doctors due to its exceptional convenience. In orthodontic treatment, telemedicine helps patients avoid the inconvenience of in-person visits, strengthens communication with their doctors, allows for more flexible and efficient treatment plans based on remote consultation results, and enables patients to track and monitor their treatment progress. Monitoring tooth movement and predicting the dentition using three-dimensional models at different times are the most crucial aspects of telemedicine.
[0003] Currently, Dental Monitoring (DM), a US-based company, has developed a product called Dental Monitoring that enables remote dental clinics through a combination of hardware, software, and AI. It utilizes a tooth movement tracking algorithm to achieve high-precision positioning of moving teeth during clinical orthodontic procedures and to predict 3D models of the dental arch at different times. The entire technical process includes: (i) the orthodontist uploads the scanned STL file of the teeth; (ii) the patient uses the DM application on their smartphone to perform a pre-processed video or photographic examination using a DM-patented cheek retractor; (iii) DM uses the initial scan and pre-processed video or photographs to establish a baseline for tooth position and occlusion, and calculates future movements accordingly; (iv) calculations are performed using the previous 3D model during the orthodontic cycle to generate the next 3D model.
[0004] In step (iii), DM mainly uses two evaluation functions. The entire tooth movement tracking algorithm is divided into four parts:
[0005] 1) Define the reference model to be tested as the initial reference model.
[0006] 2) Using heuristic algorithms, the virtual acquisition conditions (i.e., the various parameters captured during shooting) are made as close as possible to real acquisition conditions. This requires generating a 2D image of a reference model (an initial model obtained through certain transformations) and processing it to obtain a reference image containing the aforementioned discrimination information. The updated reference model's information image is compared with the information image obtained from the actual shooting, so that the value of the first evaluation function is determined based on the difference between the two information images. This value will determine whether to continue or stop searching for virtual acquisition conditions.
[0007] 3) The above results are then processed through a second discriminant evaluation function. Similar to the second step, the value of this function is used to determine whether to continue or stop searching for virtual acquisition conditions.
[0008] 4) For all the tooth models, compare the initial model with the model obtained through the above steps.
[0009] Track the amount of movement at each step.
[0010] DM did not elaborate on the technical details of the patent, but stated that the search could be conducted using meta-search, evolutionary methods, or simulated annealing. According to DM's experimental report, the technology can achieve relatively accurate tooth tracking, but it was only tested on Type I dental models, and its effectiveness in handling special malocclusions cannot be guaranteed.
[0011] Besides DM, Chaohou Company has also proposed a method for predicting 3D digital models of dental arches. This method is based on tooth contour matching, optimizing camera intrinsic and extrinsic parameters to obtain a set of virtual camera intrinsic parameters and a set of virtual camera extrinsic parameters corresponding to each of the N tooth photographs. Based on tooth contour matching, the tooth pose of the first 3D digital model is optimized to obtain an updated 3D digital model. The optimization of both camera parameters and tooth pose is based solely on the contour of each tooth. However, this technology still suffers from poor accuracy and insufficient refinement of the reconstructed dental model.
[0012] Current 3D tooth reconstruction technology can track changes in the dentition and determine the movement of each tooth in the upper and lower jaws. However, it still suffers from problems such as low matching accuracy, inaccurate pose estimation, slow computation speed, and large memory consumption. These difficulties mainly stem from the challenges of coordinating hardware and software, variations in human tissue, and conditions such as imaging and lighting. Finding more accurate and stable matching features for the algorithm and optimizing its logic are key technical issues. Summary of the Invention
[0013] To address the aforementioned technical problems, this invention proposes a method for monitoring orthodontic processes based on a three-dimensional dental arch model. The aim is to achieve three-dimensional model prediction of the dental arch at different time points by accurately analyzing and processing two-dimensional camera images and three-dimensional dental model data. This technical solution consists of four main steps: segmentation of multiple two-dimensional camera images and the three-dimensional dental model, reference point detection, camera pose estimation, and tooth pose estimation. Through the organic combination of these steps, the difficulty of accurately capturing and reconstructing dynamic changes in teeth in existing technologies can be effectively solved, thereby improving the precision and effectiveness of dental treatment. The technical solution of this invention is implemented as follows:
[0014] A method for monitoring orthodontic processes based on a three-dimensional dental arch model includes the following steps:
[0015] S1, obtain T n The three-dimensional dental model data at any given time is segmented and processed, and the mask of a single tooth in the three-dimensional dental model and the corresponding tooth number are output.
[0016] S2, obtain N T cards n+m The two-dimensional image data of the teeth at any given time is processed and segmented to output a mask of a single two-dimensional tooth and its corresponding tooth number.
[0017] S3, Reference Point Detection: Includes prediction via immobile teeth and manual marking;
[0018] When T n and T n+m When there are immobile teeth at any given time, reference points are detected by predicting immobile teeth.
[0019] When T n and T n+m When all teeth have undergone translational or rotational transformations at any given time, reference points are detected by manual marking.
[0020] Immovable tooth prediction: First, for each tooth, using single-view camera pose estimation technology, the position of the tooth in the T-wave region is calculated. n+m The camera pose corresponding to each moment; then, by performing cluster analysis on the camera poses corresponding to multiple different teeth, teeth whose geometric positions remain unchanged can be identified.
[0021] S4, Camera pose estimation:
[0022] By matching the regions of the reference points, extracting their time and frequency domain features, and selecting an appropriate loss function, the similarity between the predicted results and the actual observation results is measured.
[0023] S5, Tooth pose estimation:
[0024] Calculate each tooth at T n+m The three-dimensional coordinates and rotation angle at any given moment;
[0025] S6, three-dimensional reconstruction of the dental arch;
[0026] Based on the S5 computing architecture, T n Each tooth in the 3D dental model at time T is translated and rotated to reconstruct the T-shaped model. n+m A three-dimensional dental model at any given time.
[0027] Preferably, in S1, the segmentation process of the three-dimensional dental model is as follows: the input three-dimensional dental model data is processed by a pre-trained three-dimensional convolutional neural network to generate a mask for a single tooth and the corresponding tooth number.
[0028] The segmentation process for two-dimensional image data is as follows: segmentation is performed using a pre-trained two-dimensional convolutional neural network.
[0029] Preferably, in step S3, the specific process of predicting immobile teeth is as follows:
[0030] S3,1, Using single-view camera pose estimation technology, referencing T n Using 3D dental model data at time T, the T value of each tooth was calculated. n+m The camera pose at any given moment;
[0031] S3,2, perform cluster analysis on the camera poses of multiple different teeth to identify teeth whose geometric positions remain unchanged.
[0032] Preferably, step S3.2 is implemented using the DBSCAN clustering algorithm.
[0033] Preferably, S3 further includes a multi-view voting strategy: statistically analyzing the tooth poses under different views to filter out teeth that are judged to be immobile in most views.
[0034] Preferably, S4 includes coarse mixing and fine mixing;
[0035] Rough calculation: By using geometric relationships and traversal calculations, the approximate starting point for the camera's six-dimensional parameters is obtained;
[0036] Fine-tuning: The simulated annealing algorithm is used to perform multiple iterations under the guidance of the loss function, and finally the accurate camera pose parameters are obtained.
[0037] Preferably, the loss function of S4 is the Dice coefficient.
[0038] Preferably, step S5 integrates several occlusal and labial tooth views, performs comprehensive analysis of multi-angle tooth information, adopts a simulated annealing algorithm, and selects an appropriate loss function (including but not limited to the Dice coefficient) to perform pose optimization calculation based on the multi-view information.
[0039] Preferably, in step S5, the loss function is the Dice coefficient.
[0040] The advantages of this invention are as follows:
[0041] 1. The reference point detection method is more automated and accurate, improving the accuracy of 3D dental arch model prediction.
[0042] Current orthodontic procedures often rely on treatment plans to predict tooth movement. However, in actual clinical practice, tooth movement often does not perfectly match the expected outcome of the treatment plan, especially as various unforeseen factors may arise during treatment, causing the tooth movement trajectory to deviate from the anticipated path. Traditional three-dimensional prediction methods, if based on the treatment plan, may result in reconstructions that do not accurately reflect the patient's current dental condition, thus affecting treatment decisions.
[0043] The reference point detection method proposed in this invention effectively avoids problems that may arise from using the treatment plan as a prerequisite. By detecting and matching reference points, this algorithm can independently monitor the actual tooth movement during orthodontic treatment in real time and identify potential abnormalities. This method not only improves the accuracy of predicting the three-dimensional model of the dentition at different times during orthodontic treatment but also provides doctors with more flexible treatment monitoring tools, ensuring accurate three-dimensional prediction of the dentition under various circumstances.
[0044] 2. The combination of coarse and fine matching in camera pose estimation improves the efficiency of the camera pose estimation algorithm.
[0045] In camera pose estimation, existing methods often face a trade-off between time cost and accuracy. While coarse-matching methods based on geometric relationships and simple traversal can quickly generate preliminary estimates of camera pose, their time complexity increases exponentially to meet accuracy requirements, making them computationally inefficient in practical applications. On the other hand, if only simulated annealing algorithms from fine-matching are used, their global search in a multi-dimensional parameter space can easily lead to local optima when the initial parameters are far from the global optimum, resulting in suboptimal pose estimation accuracy.
[0046] This invention successfully solves this problem by innovatively combining coarse and fine modeling. The coarse modeling stage utilizes geometric relationships to quickly locate a starting point close to the global optimum, significantly reducing the search space and thus lowering time costs. Next, based on this starting point, the fine modeling stage uses simulated annealing to perform fine adjustments, approximating the global optimum. In this way, the algorithm not only shortens the running time but also significantly improves the accuracy of camera pose estimation, ensuring that the final reconstruction result has high accuracy and reliability.
[0047] 3. The application of region matching improves the robustness of the 3D dental arch model prediction algorithm.
[0048] In existing pose estimation algorithms, contour matching is typically used as the loss function. However, contour matching is highly sensitive to edge noise and occlusion, and the presence of noise or poor image quality can easily lead to increased matching errors, thus affecting the reconstruction accuracy. This invention employs region matching in both camera pose estimation and tooth pose estimation to improve the algorithm's robustness, enabling it to maintain high matching accuracy even in the presence of some noise.
[0049] By using region matching, this algorithm can more effectively estimate pose, thereby improving the accuracy and stability of reconstruction. Compared with traditional contour matching methods, region matching performs better in complex scenes, especially in cases of poor image quality or complex tooth structures, ensuring the accuracy of the algorithm results.
[0050] 4. The simulated annealing algorithm employed provides better global optimization capabilities.
[0051] Simulated annealing algorithm is widely used in this invention, especially in the fine-tuning stage of camera pose estimation and tooth pose estimation. A significant feature of this algorithm is its global optimization capability, enabling it to perform extensive searches in the multi-dimensional parameter space and avoid getting trapped in local optima. Through multiple iterative optimizations using simulated annealing, the algorithm gradually approaches the global optimum, thereby improving the accuracy of the final pose estimation.
[0052] Compared with other local search algorithms, simulated annealing exhibits stronger robustness, especially in complex parameter spaces. Through multiple iterative optimizations, the final pose estimation results are not only accurate but also highly reliable, providing a solid technical guarantee for the prediction process of 3D dental arch models. Attached Figure Description
[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0054] Figure 1 The flowchart is for an embodiment of the present invention; the accompanying drawings are for illustrating the technical solution of the present invention most clearly. Figure 1 It was also designated as an abstract figure;
[0055] Figure 2 for Figure 1 The diagram shown illustrates the effect of region matching in the embodiment. Detailed Implementation
[0056] The technical solutions of the present invention will now be clearly and completely described with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0057] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used in the detailed description is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.
[0058] In the description of specific embodiments of the present invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present invention, "multiple" means two or more, unless otherwise explicitly defined.
[0059] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.
[0060] In the description of the embodiments of this invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this invention, the character " / " generally indicates that the preceding and following associated objects have an "or" relationship.
[0061] The embodiments of the present invention will be described in more detail below through examples. It should be noted that the embodiments of the present invention are not limited to these examples.
[0062] In one specific embodiment, such as Figure 1 As shown, a three-dimensional reconstruction method for dental arches based on a multi-step algorithm includes the following steps:
[0063] S1, obtain T nThe three-dimensional dental model data at time n (where n is a non-negative integer) is segmented and processed, and the mask of a single tooth in the three-dimensional dental model and the corresponding tooth number are output.
[0064] S2, obtain N T cards n+m The two-dimensional image data of the teeth at time m (where m is a non-negative integer) is processed and segmented, and the mask of a single two-dimensional tooth and its corresponding tooth number are output.
[0065] S3, Reference Point Detection: Includes prediction via immobile teeth and manual marking;
[0066] When T n and T n+m When there are immobile teeth at any given time, reference points are detected by predicting immobile teeth.
[0067] When T n and T n+m When all teeth have undergone translational or rotational transformations at any given time, reference points are detected by manual marking.
[0068] Immovable tooth prediction: First, for each tooth, using single-view camera pose estimation technology, the position of the tooth in the T-wave region is calculated. n+m The camera pose corresponding to each moment; then, by performing cluster analysis on the camera poses corresponding to multiple different teeth, teeth whose geometric positions remain unchanged can be identified.
[0069] S4, Camera pose estimation:
[0070] By matching the regions of the reference points, extracting their time and frequency domain features, and selecting an appropriate loss function, the similarity between the predicted results and the actual observation results is measured.
[0071] S5, Tooth pose estimation:
[0072] Calculate each tooth at T n+m The three-dimensional coordinates and rotation angle at any given moment;
[0073] S6, three-dimensional reconstruction of the dental arch;
[0074] Based on the S5 computing architecture, T n Each tooth in the 3D dental model at time T is translated and rotated to reconstruct the T-shaped model. n+m A three-dimensional dental model at any given time.
[0075] This embodiment proposes a multi-step algorithm-based 3D reconstruction method for dental arches, aiming to achieve 3D reconstruction of dental arches at different time points by accurately analyzing and processing 2D camera images and 3D dental model data. The technical solution consists of four main steps: segmentation of multiple 2D camera images and 3D dental model, reference point detection, camera pose estimation, and tooth pose estimation. Through the organic combination of these steps, the challenge of accurately capturing and reconstructing dynamic changes in teeth in existing technologies can be effectively solved, thereby improving the accuracy and effectiveness of dental treatment.
[0076] The specific implementation process of this embodiment is as follows:
[0077] Segmentation of multiple 2D camera images and segmentation of 3D dental models:
[0078] The first step in this embodiment is to T n 3D dental model data and T at time point n+m Segmentation is performed on multiple 2D camera image data at different times. n Three-dimensional dental model data at any given time can be generated from the patient's initial intraoral scan of the dentition, while T... n+m Multiple two-dimensional camera images at different times are acquired by the hardware-based camera. These two types of data represent the state of the patient's teeth at different points in time. Processing this data provides accurate input for subsequent steps.
[0079] The segmentation process of a 3D dental model relies on a 3D segmentation network using deep learning technology. A pre-trained 3D convolutional neural network is used to segment the input T... n The 3D dental model data at any given time is processed to generate a mask for each tooth and its corresponding tooth number. The key to this step is ensuring the accuracy and stability of the segmentation to guarantee the correct geometry and positional information of each tooth.
[0080] The segmentation process for 2D camera images is similar to that for 3D models, employing a pre-trained 2D convolutional neural network. The network input is T... n+m The system continuously captures 2D camera images, outputting a mask for each tooth and its corresponding tooth number. The key to this process is ensuring clear segmentation edges and accurate mask generation, so that the 2D image can be precisely matched with the 3D model later.
[0081] Reference point detection:
[0082] After obtaining T n 3D dental model data and T at time point n+m After obtaining multiple 2D camera image data at each time point, the second step in this embodiment is to perform reference point detection. A reference point refers to a point between two time points (i.e., T). n and Tn+m (At any time), maintain a stable three-dimensional point cloud.
[0083] When T n and T n+m When there is a stationary tooth between two time points, this embodiment uses stationary tooth prediction for reference point detection. n and T n+m At any given moment, the teeth that did not undergo translation or rotation. These stationary teeth provide stable reference points for subsequent camera pose estimation.
[0084] This embodiment employs a method based on machine learning and cluster analysis to predict immobile teeth. First, for each tooth, single-view camera pose estimation technology is used, referencing T... n Using three-dimensional dental model data at time T, the tooth was calculated. n+m The camera pose corresponding to each moment is then determined. By clustering the camera poses of multiple different teeth, teeth whose geometric positions remain unchanged can be identified. The core technology in this step is the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm, which can classify data points based on density without relying on specific cluster shapes, thus effectively identifying immobile teeth.
[0085] To further improve the accuracy of immobile tooth prediction, this embodiment introduces a multi-view voting strategy. Specifically, by statistically analyzing the tooth poses under different views, teeth that are determined to be immobile under multiple views are selected. This strategy can effectively reduce inaccurate predictions caused by single-view errors, thereby ensuring that the final immobile tooth sequence has high reliability.
[0086] When T n and T n+m When all teeth undergo translational and rotational transformations at any given time, this embodiment uses manual marking to detect reference points. During orthodontic treatment, there are some stable tissues in the oral cavity, such as maxillary folds, which can serve as reference points for subsequent matching.
[0087] Camera pose estimation:
[0088] After the reference point detection is completed, this embodiment proceeds to the camera pose estimation step. Accurate estimation of the camera pose is crucial for achieving 3D tooth reconstruction; it defines T... n+m The position and orientation of the camera in the virtual world coordinate system are specifically represented by six-dimensional parameters: three-dimensional coordinates and rotation angle.
[0089] like Figure 2As shown, camera pose estimation is based on a reference point (T). n and T n+m Region matching is performed using a fixed reference point (the "stationary tooth") between two time points. Since the spatial location of the reference point remains unchanged between two time points, it can serve as a stable reference. By performing matching calculations on the regions of the reference points, their time-domain and frequency-domain features are extracted, and an appropriate loss function, such as the Dice coefficient, is selected to measure the similarity between the predicted and actual observation results. The Dice coefficient is a commonly used index to measure the similarity of overlapping regions, ranging from 0 to 1; a higher value indicates a higher degree of matching between the two regions.
[0090] Camera pose estimation is divided into two stages: coarse matching and fine matching. In the coarse matching stage, the algorithm obtains a rough starting point for the camera's six-dimensional parameters through geometric relationships and simple traversal calculations. The goal of coarse matching is to quickly determine a relatively accurate initial pose, providing a good foundation for fine-tuning in the fine matching stage. The fine matching stage employs simulated annealing, a stochastic search optimization method that can gradually approach the global optimum while avoiding local optima. Through multiple iterations guided by a loss function, the precise camera pose parameters are finally obtained.
[0091] Tooth position estimation:
[0092] After camera pose estimation is completed, this embodiment proceeds to the tooth pose estimation step, which is a crucial step in achieving the final 3D reconstruction of the dentition. Tooth pose estimation also involves the calculation of six-dimensional parameters, namely, the position of each tooth in T... n+m The three-dimensional coordinates and rotation angle at any given moment.
[0093] To improve the accuracy of tooth pose estimation, this embodiment uses T... n+m Multiple occlusal and labial views acquired at any given time are integrated. Each view provides tooth information observed from different angles. By comprehensively analyzing this information, errors that may occur from a single viewpoint can be effectively reduced. Specifically, simulated annealing algorithm is also used, and an appropriate loss function is selected to perform pose optimization calculations based on the multi-view information.
[0094] After obtaining the precise pose of each tooth, the final step is to transfer the T-shaped image to the center of the tooth. n The three-dimensional dental model at time T is translated and rotated to predict T. n+m A three-dimensional dental model at a given time. This process is based on the tooth pose information obtained in the previous steps, through T... n The three-dimensional model at time T is adjusted according to the corresponding six-dimensional parameters to align with T. n+m The actual position of the teeth at any given moment is consistent, thus achieving alignment of the T-wave during orthodontic treatment. n+mThe system predicts the movement of teeth using a 3D model at any given time. This enables precise prediction and monitoring of tooth movement during orthodontic treatment, effectively facilitating the smooth implementation of the orthodontic plan.
[0095] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for monitoring orthodontic processes based on a three-dimensional model of the dental arch, characterized in that, The steps include the following: S1, obtain T n The three-dimensional dental model data at any given time is segmented and processed, and the mask of a single tooth in the three-dimensional dental model and the corresponding tooth number are output. S2, obtain N T cards n+m The two-dimensional image data of the teeth at any given time is processed and segmented to output a mask of a single two-dimensional tooth and its corresponding tooth number. S3, Reference Point Detection: Includes prediction via immobile teeth and manual marking; When T n and T n+m When there are immobile teeth at any given time, reference points are detected by predicting immobile teeth. When T n and T n+m When all teeth have undergone translational or rotational transformations at any given time, reference points are detected by manual marking. Immovable tooth prediction: First, for each tooth, using single-view camera pose estimation technology, the position of the tooth in the T-wave region is calculated. n+m The camera pose corresponding to each moment; then, by performing cluster analysis on the camera poses corresponding to multiple different teeth, teeth whose geometric positions remain unchanged can be identified. S4, Camera pose estimation: By matching the regions of the reference points, extracting their time and frequency domain features, and selecting an appropriate loss function, the similarity between the predicted results and the actual observation results is measured. S5, Tooth pose estimation: Calculate each tooth at T n+m The three-dimensional coordinates and rotation angle at any given moment; S6, three-dimensional reconstruction of the dental arch; Based on the S5 computing architecture, T n Each tooth in the 3D dental model at time T is translated and rotated to reconstruct the T-shaped model. n+m A three-dimensional dental model at any given time.
2. The method according to claim 1, characterized in that, In S1, the segmentation process of the three-dimensional dental model is as follows: the input three-dimensional dental model data is processed by a pre-trained three-dimensional convolutional neural network to generate a mask for a single tooth and the corresponding tooth number. The segmentation process for two-dimensional image data is as follows: segmentation is performed using a pre-trained two-dimensional convolutional neural network.
3. The method according to claim 2, characterized in that, In S3, the specific process for predicting immobile teeth is as follows: S3,1, Using single-view camera pose estimation technology, referencing T n Using 3D dental model data at time T, the T value of each tooth was calculated. n+m The camera pose at any given moment; S3,2, perform cluster analysis on the camera poses of multiple different teeth to identify teeth whose geometric positions remain unchanged.
4. The method according to claim 3, characterized in that, S3.2 is implemented using the DBSCAN clustering algorithm.
5. The method according to claim 1, characterized in that, The S3 also includes a multi-view voting strategy: statistically analyzing the tooth poses under different views to filter out the teeth that are judged to be stationary in most views.
6. The method according to claim 1, characterized in that, S4 includes coarse matching and fine matching; Rough calculation: By using geometric relationships and traversal calculations, the approximate starting point for the camera's six-dimensional parameters is obtained; Fine-tuning: The simulated annealing algorithm is used to perform multiple iterations under the guidance of the loss function, and finally the accurate camera pose parameters are obtained.
7. The method according to claim 1, characterized in that, The loss function of S4 is the Dice coefficient.
8. The method according to claim 1, characterized in that, The S5 integrates several occlusal and labial tooth views, performs comprehensive analysis of multi-angle tooth information, adopts a simulated annealing algorithm, selects a loss function, and performs pose optimization calculation based on the multi-view information.
9. The method according to claim 1, characterized in that, In S5, the loss function is the Dice coefficient.