Dental orthodontic intermediate motion process prediction method, system and device based on conditional diffusion model

Abstract

The invention relates to the field of orthodontics, in particular to a tooth orthodontics intermediate motion prediction method based on a conditional diffusion model. The invention aims to solve the problem that in the existing digital orthodontic technology, the motion trail generation in the correction intermediate process depends on manual design, and the efficiency is low. According to the method, a conditional diffusion model is constructed, and supervised training is carried out by using data sets of tooth movement tracks before, after and during tooth treatment. In the reasoning stage, the model only needs to input tooth model data before and after treatment, and a complete intermediate movement track meeting biomechanical constraints and clinical requirements can be automatically and accurately generated. Through a data driving mode, expert experience is learned and reproduced, tooth collision can be effectively avoided, the occlusion relation is ensured, and the automation level and prediction precision of orthodontic scheme design are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of orthodontic technology, and in particular to a method, system, and device for predicting intermediate movement processes in orthodontics based on a conditional diffusion model. Background Technology

[0002] Digital orthodontics is a significant development in modern dentistry, its core being the precise formulation and simulation of orthodontic treatment plans. In designing these plans, predicting and generating the intermediate movement sequence from the initial uncorrected state to the final ideal state is crucial. This sequence must not only ensure the accuracy of the final result but also guarantee that the entire process conforms to oral biomechanical principles, such as avoiding unreasonable collisions between teeth, maintaining a healthy occlusion, and considering the stability of teeth as anchorages.

[0003] Currently, most mainstream methods for tooth motion path planning rely on geometrically based algorithms or simplified physical simulations. For example, some methods involve setting a series of discrete keyframes, which are then manually fine-tuned by the operator. This process is extremely time-consuming and highly dependent on the dentist's personal experience, making it difficult to standardize. Other methods attempt to incorporate machine learning, but the models are often relatively simple and struggle to capture the complex, nonlinear collaborative motion relationships between multiple teeth. These traditional methods generally suffer from the following drawbacks:

[0004] (1) Low level of automation and high labor costs;

[0005] (2) It is difficult to fully consider complex clinical constraints, and the generated pathways may not conform to biological reality;

[0006] (3) The high dependence on doctors' experience leads to inconsistent quality of treatment plans.

[0007] Therefore, there is an urgent need for an innovative method that can automatically learn and generate high-precision, clinically compatible tooth movement trajectories to drive digital orthodontics toward a more intelligent and precise direction. Summary of the Invention

[0008] The technical problem to be solved by this invention is to design a method, system, and device for predicting intermediate motion processes in orthodontics based on a conditional diffusion model. By constructing a conditional diffusion model and using a dataset of tooth motion trajectories before, after, and during the intermediate process of tooth treatment for supervised training, the model only needs to input tooth model data before and after treatment to automatically and accurately generate complete intermediate motion trajectories that meet biomechanical constraints and clinical requirements. Through a data-driven approach, it learns and reproduces expert experience, effectively avoids tooth collisions, ensures occlusal relationships, and significantly improves the automation level and prediction accuracy of orthodontic treatment plan design, thereby solving existing technical problems.

[0009] To address the aforementioned technical problems, this invention provides a method for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, comprising the following steps:

[0010] Step S1: Dataset Construction and Preprocessing: Collect and construct a dataset containing a large number of patients' pre-treatment dental conditions. Post-treatment condition And intermediate states, meticulously marked by hand by a professional orthodontist. The triplet dataset was used, and uniform position and pose registration was performed on all model data.

[0011] Step S2: Conditional Diffusion Model Design: Construct a conditional diffusion model based on a 3D U-Net architecture. During training, this model receives the pre-treatment model... Post-treatment model As conditional information, and as an intermediate state for doctors. The model performs a forward diffusion process to add noise; in the reverse denoising process, the model learns to predict and remove noise based on conditional information and the current noise level in order to reconstruct the original intermediate state.

[0012] Step S3: Differentiated Model Training: In view of the significant differences in the dental physiological structure between adults and children, separate datasets for adult and pediatric patients are constructed. Using these two independent datasets, two sets of conditional diffusion models with independent parameters are trained, which focus on learning the movement patterns of the fixed dental arch in adults and the dynamic changes in the dentition during the growth period in children, respectively.

[0013] Step S4: Model Inference and Trajectory Generation: Input the patient's pre-treatment model and post-treatment model Based on the patient type (adult or child), the appropriate conditional diffusion model is selected. Guided by conditional information, a high-fidelity intermediate state model is finally generated through a multi-step iterative reverse denoising process. By adjusting the target time step, a series of intermediate states are generated, forming a complete dynamic motion trajectory, thereby completing the process prediction.

[0014] Furthermore, in step S1, the dataset is obtained by segmenting the STL data of each tooth from the complete maxillary and mandibular STL models of high-quality cases, converting it into point cloud data, recording the position information of each tooth in the dentition space, obtaining the three-dimensional rotation angle of the tooth itself based on the small coordinates of the tooth, and saving it in JSON format.

[0015] Furthermore, in step S1, the data preprocessing method is as follows: the point cloud data of each tooth is simplified by fusing the farthest point sampling and voxel mesh downsampling methods, while preserving the detailed features of the tooth outline as much as possible. All model data are uniformly registered in position and pose, including 3D model normalization, rigid body registration based on iterative nearest point (ICP), and vertex correspondence mapping, so as to ensure the spatial consistency of different data in the world coordinate system.

[0016] Furthermore, in step S1, the PointNet++ model is used in conjunction with tooth position encoding to extract features from the point cloud of each tooth. Through multi-level sampling and grouping operations, the feature information of the point cloud is gradually aggregated to form a 128-dimensional global feature of the point cloud. The position encoder encodes the affine matrix of the initial and target positions of each tooth, and uses a fully connected neural network to map the 4×4 three-dimensional translation and rotation information to a 128-dimensional high-dimensional feature representation.

[0017] Furthermore, based on the important relationships between teeth, such as adjacency, occlusion, and symmetry, a graph attention network is used for feature fusion. When constructing the graph neural network, each tooth is regarded as a node in the graph. The adjacency, occlusion, and symmetry relationships between teeth are represented by edges. In the process of information propagation in the graph, the node (tooth) not only receives its own feature information, but also receives the feature information of the adjacent nodes (related teeth) through the edges. At the same time, the attention weights of different edges (relationships) are calculated, and the influence weights of various relationships on the orthodontic effect are dynamically adjusted.

[0018] Furthermore, in step S2, to optimize model performance and more accurately simulate the orthodontic process, a loss function, Loss, is set, and its calculation formula is as follows:

[0019] ;

[0020] in,

[0021] Represents relative position loss, used to measure the difference between the model-predicted relative positions of teeth and the actual relative positions, accurately grasping the changes in the positional relationship between teeth;

[0022] This represents the absolute position loss, focusing on the error between the predicted absolute position and the true absolute position of each tooth;

[0023] This represents the loss of the dental arch line position, prompting the model to adjust the position of the teeth to conform to the normal dental arch morphology;

[0024] This represents the loss of consistency in movement trends. It is achieved by comparing the direction and magnitude of tooth position changes at adjacent time steps with the movement trends during actual orthodontic treatment.

[0025] Represents tooth collision loss, calculated by modeling whether different teeth collide during movement;

[0026] Represents the average chamfer distance loss of teeth, used to measure the similarity between the tooth surface predicted by the model and the real tooth surface, and to evaluate the effect of point cloud feature extraction;

[0027] , , , , and These are the weighting coefficients for each loss function.

[0028] Furthermore, in step S2, to ensure that tooth movement conforms to medical principles as much as possible, post-processing operations are performed on the model: First, momentum analysis of the teeth is conducted. By monitoring the magnitude and direction of momentum, it is ensured that the force on tooth movement is within a reasonable range to avoid damage to periodontal tissues. Regarding anchorage rules, the force and movement of anchor teeth during orthodontic treatment are analyzed. As the support point for orthodontic forces, the anchor teeth are adjusted according to medical experience and mechanical principles to ensure that their movement is controlled within an acceptable range when subjected to orthodontic forces, guaranteeing the smooth progress of orthodontic treatment and the achievement of expected results. Through post-processing operations such as momentum analysis and anchorage rule adjustment, the orthodontic plan output by the model is made more medically feasible and safe.

[0029] Furthermore, step S5 is included: Model post-processing and clinical adaptation: The intermediate motion trajectory generated by the diffusion model is refined and optimized to avoid tooth collisions, ensure healthy occlusal relationships, and consider the stability of tooth anchorage in order to improve clinical applicability.

[0030] The present invention also provides a prediction system for intermediate movement processes in orthodontics based on a conditional diffusion model, including a dataset acquisition module, a data preprocessing module, a feature extraction module, a feature fusion module, a model building module, a model training module, a trajectory generation module, and a control module.

[0031] The dataset acquisition module is used to collect and construct a triplet dataset containing a large number of patients' pre-treatment, post-treatment, and intermediate states manually and meticulously annotated by professional orthodontists.

[0032] The data preprocessing module is used to simplify the point cloud data of each tooth and perform unified position and pose registration on all model data.

[0033] The feature extraction module is used to extract features from the point cloud of each tooth by combining the PointNet++ model with the tooth position encoding.

[0034] The feature fusion module is used to perform feature fusion using a graph attention network and dynamically adjust the influence weights of various relationships on the orthodontic effect.

[0035] The model building module is used to build a conditional diffusion model based on the 3D U-Net architecture.

[0036] The model training module includes an adult patient training unit and a pediatric patient training unit. The adult patient training unit is used to build and train an adult patient conditional diffusion model using an adult dataset, and the pediatric patient training unit is used to build and train a pediatric patient conditional diffusion model using a pediatric dataset.

[0037] The trajectory generation module is used to generate a high-fidelity intermediate state model based on the patient's pre-treatment and post-treatment models, using a trained conditional diffusion model and a multi-step iterative reverse denoising process. By adjusting the target time step, a series of intermediate states are generated to form a complete dynamic motion trajectory.

[0038] The control module is used to control the dataset acquisition module, data preprocessing module, feature extraction module, feature fusion module, model building module, model training module, and trajectory generation module to perform relevant operations according to the aforementioned method for predicting intermediate movement processes in orthodontics based on the conditional diffusion model.

[0039] The present invention also provides a device for predicting intermediate movement processes in orthodontics based on a conditional diffusion model, specifically comprising:

[0040] At least one processor;

[0041] and at least one memory communicatively connected to the processor;

[0042] The memory stores instructions that can be executed by a processor, which are then executed by the processor to cause the device to perform the aforementioned method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model.

[0043] Compared with the prior art, the present invention has the following significant technical effects:

[0044] (1) High accuracy and clinical realism: By learning from a large amount of expert-annotated data, the model can generate motion trajectories that are highly consistent with biomechanical constraints (such as occlusion and collision avoidance), and the prediction results are closer to clinical reality.

[0045] (2) High degree of automation: The reasoning process is fully automated. Only the state before and after treatment needs to be input to generate a complete path, which greatly liberates the doctor's productivity and shortens the treatment plan design cycle.

[0046] (3) Strong generalization ability: Differentiated training strategies are adopted to build dedicated models for adults and children respectively, ensuring the applicability and accuracy of the method in patient groups of different age groups.

[0047] (4) End-to-end solution: It realizes the end-to-end mapping from the boundary state (before / after treatment) to the prediction of the complete process, which simplifies the technical process and has high clinical application value and promotion potential. Attached Figure Description

[0048] The specific embodiments of the present invention will be further explained below with reference to the accompanying drawings.

[0049] Figure 1 This is a flowchart of the method for predicting intermediate movement processes in orthodontics based on the conditional diffusion model in Example 1.

[0050] Figure 2 To apply the pre-treatment dental condition information of the case according to the method in Example 1.

[0051] Figure 3 To apply the tooth status information of intermediate state 1 of the case according to the method in Example 1.

[0052] Figure 4 To apply the tooth status information of intermediate state 2 of the case according to the method in Example 1.

[0053] Figure 5 To apply the tooth status information of intermediate state 3 of the case according to the method in Example 1.

[0054] Figure 6 This is to obtain information on the tooth status of the case after treatment is completed, using the method described in Example 1.

[0055] Figure 7 This is a block diagram of the orthodontic intermediate motion prediction system based on the conditional diffusion model in Example 2. Detailed Implementation

[0056] Example 1

[0057] Combination Figure 1 This embodiment provides a method for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, including the following steps:

[0058] Step S1: Dataset Construction and Preprocessing: Collect and construct a dataset containing a large number of patients' pre-treatment dental conditions. Post-treatment condition And intermediate states, meticulously marked by hand by a professional orthodontist. The triplet dataset was used to perform uniform position and pose registration on all model data.

[0059] In this preferred embodiment, in step S1, the dataset is acquired as follows: STL data of each tooth is segmented from the complete maxillary and mandibular STL models of high-quality cases, converted into point cloud data, the position information of each tooth in the dentition space is recorded, and the three-dimensional rotation angle of the tooth itself is obtained based on the small coordinates of the tooth, and saved in JSON format. In addition, the dataset needs to be divided into adult patient datasets and pediatric patient datasets according to age group.

[0060] Specifically, in this embodiment, complete three-dimensional STL models of the maxilla and mandible were collected from 214 independent clinical cases, covering both adult and pediatric patients. An automated segmentation algorithm was used to segment each model into 20–28 individual tooth STL models based on FDI numbers (dynamically adjusted according to the eruption status of children's teeth). Each tooth STL was further converted into a point cloud representation, with the number of points initially downsampled to approximately 1200 points per tooth.

[0061] In this embodiment, in step S1, the data preprocessing method is as follows: the point cloud data of each tooth is simplified by fusing the farthest point sampling and voxel mesh downsampling methods, while preserving the detailed features of the tooth outline as much as possible. All model data are uniformly registered in position and pose, including 3D model normalization, rigid body registration based on iterative nearest point (ICP), and vertex correspondence mapping, so as to ensure the spatial consistency of different data in the world coordinate system.

[0062] Specifically, in this embodiment, the point cloud of a single tooth is sequentially sampled by farthest point sampling (FPS) and then downsampled using a voxel mesh (voxel size = 0.15 mm), and standardized to N=400 points. The point set is normalized with the centroid of each tooth as the origin, and the coordinates are zero-mean and unit variance normalized.

[0063] In this embodiment, all cases are aligned with the mandibular reference plane (Z=0), and global rigid body registration is performed using the ICP algorithm to ensure spatial consistency across cases. The position of each tooth is recorded. (Before treatment) (After treatment) and some Pose at keyframe time (using quaternions) , , , ) and translation ( , , (This is represented as ) and saved as JSON.

[0064] In this preferred embodiment, in step S1, the PointNet++ model is used in conjunction with tooth position encoding to extract features from the point cloud of each tooth. Through multi-level sampling and grouping operations, the feature information of the point cloud is gradually aggregated to form a 128-dimensional global feature of the point cloud. The position encoder encodes the affine matrix of the initial and target positions of each tooth, and uses a fully connected neural network to map the 4×4 three-dimensional translation and rotation information to a 128-dimensional high-dimensional feature representation.

[0065] Specifically, in this embodiment, the point cloud of each tooth is input into an improved PointNet++ feature extraction network. The network adopts a three-layer SA structure, with sampling point numbers of 512, 128, and 1 respectively, and sphere radii of 0.1 / 0.2 / 0.4, 0.2 / 0.4 / 0.8, and global, respectively. The MLP output dimensions are 320, 640, and 1024, respectively. After global pooling, a 128-dimensional global morphological feature of each tooth is obtained by mapping through three fully connected layers (512-256-128).

[0066] In this embodiment, specifically, for each tooth, its clinical framework features (21 dimensions including apex, incisal edge, contact point, etc.) are additionally encoded. Pose information ( / The homogeneous transformation matrix (4×4, flattened to 16 dimensions) is mapped to 128 dimensions through a fully connected layer. Finally, the feature vector of each tooth node is 128 (point cloud) + 128 (pose) + 21 (clinical) = 277 dimensions.

[0067] In this preferred embodiment, based on the important relationships between teeth such as adjacency, occlusion, and symmetry, a graph attention network is used for feature fusion. When constructing the graph neural network, each tooth is regarded as a node in the graph. The adjacency, occlusion, and symmetry relationships between teeth are represented by edges. In the process of information propagation in the graph, the node (tooth) not only receives its own feature information, but also receives the feature information of adjacent nodes (related teeth) through the edges. At the same time, the attention weights of different edges (relationships) are calculated, and the influence weights of various relationships on the orthodontic effect are dynamically adjusted.

[0068] In this embodiment, the tooth relationship graph structure is constructed as follows: for each case, a directed relationship graph is established based on the relationships of adjacent teeth, opposing teeth, mirror symmetry, etc., with nodes representing teeth and edges representing relationship pairs.

[0069] In this embodiment, the graph neural network fusion method is as follows: a graph attention network (GAT) is used, with initial node features of 277 dimensions, and neighborhood information is fused through adaptive attention weighting. The information transmission steps K=3, and after fusion, each tooth outputs context-related features.

[0070] Step S2: Conditional Diffusion Model Design: Construct a conditional diffusion model based on a 3D U-Net architecture. During training, this model receives the pre-treatment model... Post-treatment model As conditional information, and as an intermediate state for doctors. The model performs a forward diffusion process to add noise; in the reverse denoising process, the model learns to predict and remove noise based on conditional information and the current noise level in order to reconstruct the original intermediate state.

[0071] In this embodiment, a diffusion model based on a 3D U-Net architecture is specifically adopted. The input is a 7-dimensional tooth motion sequence (quaternion + translation) of L=20 frames, and the conditional input is... and The model features are described. In the forward diffusion process, Gaussian noise is added to the actual intermediate motion trajectory for T=100 steps, and cosine scheduling is used for noise scheduling. The reverse denoising network receives the noise and conditional features of the current frame, and predicts the residual through a 3D U-Net backbone and a three-layer perceptron.

[0072] In this preferred embodiment, in step S2, in order to optimize model performance and make it more accurately simulate the orthodontic process, a loss function Loss is set, and the calculation formula is as follows:

[0073] ;

[0074] in,

[0075] Relative position loss represents the difference between the model-predicted relative positions of teeth and their actual relative positions. It is calculated by measuring the deviation of the relative positions of each pair of related teeth (such as adjacent teeth or teeth with occlusal relationships) between the predicted and actual states, and is quantified using methods such as mean squared error (MSE). The calculation formula is as follows:

[0076] ;

[0077] in, and Let represent the predicted and actual relative position vectors of the i-th pair of teeth, respectively, and n be the number of tooth pairs.

[0078] Relative position loss helps the model accurately grasp changes in the positional relationship between teeth, ensuring the correctness of the relative position of teeth during orthodontic treatment.

[0079] This represents the absolute position loss, focusing on the error between the predicted absolute position and the true absolute position of each tooth. It is also calculated using the mean squared error formula:

[0080] ;

[0081] in, and ...

[0082] This represents the loss of position along the dental arch line, a crucial reference point in orthodontic treatment. By defining an ideal dental arch line shape, the deviation between the model's predicted tooth positions and this ideal arch line is calculated as the loss. The distance from each tooth to the arch line is calculated, and then these distances are summed and averaged to obtain the final dental arch line position loss. This loss function prompts the model to adjust the position of the teeth to conform to the normal dental arch morphology, thereby improving the aesthetics and functionality of the dental arch after orthodontic treatment.

[0083] Representing the loss of consistency in motion trend, the motion trend of teeth during orthodontic treatment should be consistent. The loss of consistency in motion trend is calculated by comparing the direction and magnitude of tooth position changes at adjacent time steps with the actual motion trend during orthodontic treatment. Methods such as the cosine of the vector angle are used to measure the consistency between the predicted and actual motion trends. The calculation formula is as follows:

[0084] ;

[0085] in, It is the angle between the predicted and actual motion trend vectors at the l-th time step, and k is the total number of time steps.

[0086] This represents tooth collision loss. The calculation model predicts whether different teeth will collide during movement. If a collision occurs, the collision loss is calculated based on the overlap range of the tooth point clouds. A collision distance of less than 0.05 mm is considered acceptable.

[0087] The average chamfer distance loss represents the tooth surface area, which measures the similarity between the model's predicted tooth surface and the actual tooth surface, evaluating the effectiveness of point cloud feature extraction. By calculating the average chamfer distance of the tooth surface point cloud, it reflects the accuracy of changes in tooth morphology during orthodontic treatment. The smaller the average chamfer distance, the closer the predicted tooth morphology is to the actual morphology. The specific calculation formula is as follows:

[0088] ;

[0089] Where P and Q are the point cloud sets of the predicted and actual tooth surfaces, respectively.

[0090] , , , , and These are the weighting coefficients for each loss function. These weights can be fine-tuned according to the dataset and clinical needs.

[0091] In this embodiment, the weight coefficients of each loss function are specifically set as follows: , , , , and .

[0092] In this preferred embodiment, in step S2, to ensure that tooth movement conforms to medical principles as much as possible, post-processing operations are performed on the model: First, momentum analysis of the teeth is conducted. By monitoring the magnitude and direction of momentum, it is ensured that the force on tooth movement is within a reasonable range, avoiding damage to periodontal tissues. Regarding anchorage rules, the force and movement of anchor teeth during orthodontic treatment are analyzed. As the support point for orthodontic forces, the anchor teeth are adjusted according to medical experience and mechanical principles to ensure that the movement of anchor teeth under orthodontic forces is controlled within an acceptable range, guaranteeing the smooth progress of orthodontic treatment and the achievement of expected results. Through post-processing operations such as momentum analysis and anchorage rule adjustment, the orthodontic plan output by the model is made more medically feasible and safe.

[0093] Step S3: Differentiated Model Training: In view of the significant differences in the dental physiological structure between adults and children, datasets of adult and child patients are constructed according to age group. Using these two independent datasets, two sets of conditional diffusion models with independent parameters are trained, so that they focus on learning the movement law of the fixed dental arch in adults and the dynamic change characteristics of the dentition during the growth period in children, respectively.

[0094] In this embodiment, the training process preferably employs a data augmentation strategy to improve the model's generalization ability.

[0095] Step S4: Model Inference and Trajectory Generation: Input the patient's pre-treatment model and post-treatment model Based on the patient type (adult or child), the appropriate conditional diffusion model is selected. Guided by conditional information, a high-fidelity intermediate state model is finally generated through a multi-step iterative reverse denoising process. By adjusting the target time step, a series of high-fidelity intermediate state models are generated, forming a complete dynamic motion trajectory, thereby completing the process prediction.

[0096] In this preferred embodiment, step S5 is also included: model post-processing and clinical adaptation: the intermediate motion trajectory generated by the diffusion model is refined and optimized to avoid tooth collision, ensure a healthy occlusal relationship, and consider the stability of tooth anchorage, so as to improve clinical applicability.

[0097] The method for predicting intermediate movement processes in orthodontics based on the conditional diffusion model in this embodiment was tested:

[0098] Input the pre- and post-treatment status information of a tooth in a case, and the model infers intermediate motion data such as... Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, Figure 2 Information on the condition of the teeth before treatment. Figure 3 This refers to the tooth state information in intermediate state 1. Figure 4 This refers to the tooth state information in intermediate state 2. Figure 5 This refers to the tooth status information in intermediate state 3. Figure 6 Information on the condition of the teeth after treatment is completed.

[0099] After testing on 214 clinical cases, the final crown centroid error was 0.18±0.06 mm. The arch curve fitting R²=0.92, and the motion trend consistency rate (cosine similarity>0.9) was 93%. Blind evaluation by doctors showed that 87% of the intermediate states could be directly used for appliance design, significantly improving the level of clinical automation.

[0100] Example 2

[0101] Combination Figure 7 As shown, this embodiment provides a prediction system for intermediate movement processes in orthodontics based on a conditional diffusion model, including a dataset acquisition module, a data preprocessing module, a feature extraction module, a feature fusion module, a model building module, a model training module, a trajectory generation module, and a control module.

[0102] The dataset acquisition module is used to collect and construct a triplet dataset containing a large number of patients' pre-treatment, post-treatment, and intermediate states manually and meticulously annotated by professional orthodontists.

[0103] The data preprocessing module is used to simplify the point cloud data of each tooth and perform unified position and pose registration on all model data.

[0104] The feature extraction module is used to extract features from the point cloud of each tooth by combining the PointNet++ model with the tooth position encoding.

[0105] The feature fusion module is used to perform feature fusion using a graph attention network and dynamically adjust the influence weights of various relationships on the orthodontic effect.

[0106] The model building module is used to build a conditional diffusion model based on the 3D U-Net architecture.

[0107] The model training module includes an adult patient training unit and a pediatric patient training unit. The adult patient training unit is used to build and train an adult patient conditional diffusion model using an adult dataset, and the pediatric patient training unit is used to build and train a pediatric patient conditional diffusion model using a pediatric dataset.

[0108] The trajectory generation module is used to generate a high-fidelity intermediate state model based on the patient's pre-treatment and post-treatment models, using a trained conditional diffusion model and a multi-step iterative reverse denoising process. By adjusting the target time step, a series of intermediate states are generated to form a complete dynamic motion trajectory.

[0109] The control module is used to control the dataset acquisition module, data preprocessing module, feature extraction module, feature fusion module, model building module, model training module, and trajectory generation module to perform relevant operations according to the orthodontic intermediate motion process prediction method based on the conditional diffusion model in Example 1.

[0110] In this preferred embodiment, it further includes a loss function calculation module, which is used to calculate the relative position loss, absolute position loss, dental arch line position loss, motion trend consistency loss, tooth collision loss, and average tooth chamfer distance loss of the model.

[0111] In this preferred embodiment, a medical rule post-processing module is also included. This module is used for momentum analysis of teeth and analysis of the force and movement of anchorage teeth during orthodontic treatment. By monitoring the magnitude and direction of momentum, it ensures that the force on tooth movement is within a reasonable range, avoiding damage to periodontal tissues. As the support point for orthodontic forces, the anchorage teeth are adjusted according to medical experience and mechanical principles to ensure that their movement is controlled within an acceptable range when subjected to orthodontic forces, guaranteeing the smooth progress of orthodontic treatment and the achievement of expected results. Through post-processing operations such as momentum analysis and anchorage rule adjustment, the orthodontic plan output by the model becomes more medically feasible and safe.

[0112] Example 3

[0113] This embodiment provides a device for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, specifically including:

[0114] At least one processor;

[0115] and at least one memory communicatively connected to the processor;

[0116] The memory stores instructions that can be executed by a processor, which are then executed by the processor to cause the device to perform the method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model as described in Embodiment 1.

[0117] Many specific details have been set forth in the foregoing description to provide a thorough understanding of the present invention. However, the above description is merely a preferred embodiment of the present invention, and the present invention can be implemented in many other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed above. Furthermore, any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, using the methods and techniques disclosed above, without departing from the scope of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, characterized in that: Includes the following steps: Step S1: Dataset Construction and Preprocessing: Collect and construct a dataset containing a large number of patients' pre-treatment dental conditions. Post-treatment condition And intermediate states, meticulously marked by hand by a professional orthodontist. The triplet dataset was used to perform uniform position and pose registration on all model data; Step S2: Conditional Diffusion Model Design: Construct a conditional diffusion model based on a 3D U-Net architecture. During training, this model receives the pre-treatment model... Post-treatment model As conditional information, and as an intermediate state for doctors. Perform a forward diffusion process with added noise; In the reverse denoising process, the model learns to predict and remove noise based on conditional information and the current noise level in order to reconstruct the original intermediate state. Step S3: Differentiated Model Training: Construct datasets for adult and pediatric patients respectively. Using these two independent datasets, train two sets of conditional diffusion models with independent parameters, so that they focus on learning the movement patterns of the fixed dental arch in adults and the dynamic changes of the dentition during the growth period in children, respectively. Step S4: Model Inference and Trajectory Generation: Input the patient's pre-treatment model and post-treatment model Based on the patient type, the corresponding conditional diffusion model is selected. Guided by the conditional information, a high-fidelity intermediate state model is finally generated through a multi-step iterative reverse denoising process. By adjusting the target time step, a series of intermediate states are generated, forming a complete dynamic motion trajectory, thereby completing the process prediction.

2. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 1, characterized in that: In step S1, the dataset is obtained by segmenting the STL data of each tooth from the complete maxillary and mandibular STL models of high-quality cases, converting it into point cloud data, recording the position information of each tooth in the dentition space, obtaining the three-dimensional rotation angle of the tooth itself based on the small coordinates of the tooth, and saving it in JSON format.

3. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 2, characterized in that: In step S1, the data preprocessing method is as follows: the point cloud data of each tooth is simplified by fusing the farthest point sampling and voxel mesh downsampling methods, while retaining the detailed features of the tooth outline. All model data are uniformly registered in position and pose, including 3D model normalization, rigid body registration based on iterative nearest point, and vertex correspondence mapping.

4. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 2, characterized in that: In step S1, the PointNet++ model is used in conjunction with the position encoding of the teeth to extract features from the point cloud of each tooth. Through multi-level sampling and grouping operations, the feature information of the point cloud is gradually aggregated to form a 128-dimensional global feature of the point cloud.

5. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 4, characterized in that: Feature fusion is achieved using graph attention networks. When constructing the graph neural network, each tooth is regarded as a node in the graph. The adjacent relationships, occlusal relationships, and symmetric relationships between teeth are represented by edges. In the process of information propagation in the graph, nodes not only receive their own feature information, but also receive the feature information of adjacent nodes through edges. At the same time, the attention weights of different edges are calculated, and the influence weights of various relationships on the orthodontic effect are dynamically adjusted.

6. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 1, characterized in that: In step S2, to optimize model performance and more accurately simulate the orthodontic process, a loss function, Loss, is set, and its calculation formula is as follows: ； in, Represents relative position loss, used to measure the difference between the model-predicted relative positions of teeth and the actual relative positions, accurately grasping the changes in the positional relationship between teeth; This represents the absolute position loss, focusing on the error between the predicted absolute position and the true absolute position of each tooth; This represents the loss of the dental arch line position, prompting the model to adjust the position of the teeth to conform to the normal dental arch morphology; This represents the loss of consistency in movement trends. It is achieved by comparing the direction and magnitude of tooth position changes at adjacent time steps with the movement trends during actual orthodontic treatment. Represents tooth collision loss, calculated by modeling whether different teeth collide during movement; Represents the average chamfer distance loss of teeth, used to measure the similarity between the tooth surface predicted by the model and the real tooth surface, and to evaluate the effect of point cloud feature extraction. , , , , and These are the weighting coefficients for each loss function.

7. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 1, characterized in that: In step S2, post-processing operations are performed on the model: First, the momentum of the teeth is analyzed. By monitoring the magnitude and direction of the momentum, it is ensured that the force on the tooth movement is within a reasonable range to avoid damage to the periodontal tissues. The predicted movement of the anchorage teeth in the model is adjusted to ensure that the movement of the anchorage teeth is controlled within an acceptable range when subjected to orthodontic forces, so as to ensure the smooth progress of orthodontic treatment and the achievement of the expected results.

8. The method for predicting intermediate movement processes in orthodontics based on a conditional diffusion model according to claim 1, characterized in that: It also includes step S5: Model post-processing and clinical adaptation: fine-tuning the intermediate motion trajectory generated by the diffusion model to avoid tooth collisions, ensure healthy occlusal relationships, and consider the stability of tooth anchorage in order to improve clinical applicability.

9. A system for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, characterized in that: It includes a dataset acquisition module, a data preprocessing module, a feature extraction module, a feature fusion module, a model building module, a model training module, a trajectory generation module, and a control module; The dataset acquisition module is used to collect and construct a triplet dataset containing a large number of patients' pre-treatment, post-treatment, and intermediate states manually and meticulously annotated by professional orthodontists. The data preprocessing module is used to simplify the point cloud data of each tooth and perform unified position and pose registration on all model data. The feature extraction module is used to extract features from the point cloud of each tooth by combining the PointNet++ model with the tooth position encoding. The feature fusion module is used to perform feature fusion using a graph attention network and dynamically adjust the influence weights of various relationships on the orthodontic effect. The model building module is used to build a conditional diffusion model based on the 3D U-Net architecture; The model training module includes an adult patient training unit and a pediatric patient training unit. The adult patient training unit is used to build and train an adult patient conditional diffusion model using an adult dataset, and the pediatric patient training unit is used to build and train a pediatric patient conditional diffusion model using a pediatric dataset. The trajectory generation module is used to generate a high-fidelity intermediate state model based on the patient's pre-treatment and post-treatment models, using a trained conditional diffusion model and a multi-step iterative reverse denoising process. By adjusting the target time step, a series of intermediate states are generated to form a complete dynamic motion trajectory. The control module is used to control the dataset acquisition module, data preprocessing module, feature extraction module, feature fusion module, model building module, model training module, and trajectory generation module to perform relevant operations according to the method for predicting intermediate movement processes in orthodontics based on the conditional diffusion model as described in any one of claims 1-8.

10. A device for predicting intermediate movement processes in orthodontic treatment based on a conditional diffusion model, characterized in that: include: At least one processor; and at least one memory communicatively connected to the processor; The memory stores instructions that can be executed by a processor, which are executed by the processor to cause the device to perform the method for predicting intermediate movement processes in orthodontics based on the conditional diffusion model as described in any one of claims 1-8.

Images