Tooth position prediction method fusing multi-scale feature extraction and conditional diffusion model

Abstract

The application discloses a tooth position prediction method fusing multi-scale feature extraction and a conditional diffusion model, and has the following process: a tooth position prediction model is built, a three-dimensional dental arch model is input into the tooth position prediction model, the three-dimensional dental arch model is segmented, a tooth point cloud set of each tooth is obtained, data preprocessing and enhancement are performed on each point cloud in the tooth point cloud set, and a tooth array point cloud set of each tooth is obtained; multi-scale fusion conditional features of each tooth are obtained by performing feature extraction and fusion on each tooth point cloud in the tooth array point cloud set; the multi-scale fusion conditional features are input into a conditional diffusion pose regression module, a target rigid transformation matrix of each tooth from a current state to a target state is predicted, and a target tooth array structure is generated according to the target rigid transformation matrix. The method solves the problems of insufficient extraction of local geometric details, poor multi-scale feature fusion effect and limited stability of complex tooth array target pose prediction in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of digital orthodontic technology and relates to a tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model. Background Technology

[0002] Digital orthodontics typically involves several stages, including 3D jaw acquisition, tooth segmentation, tooth position prediction, orthodontic pathway planning, and treatment plan generation. Among these, tooth position prediction, based on the patient's current malocclusion, predicts the spatial position and orientation of each tooth in an ideal occlusal state, serving as a crucial foundation for subsequent orthodontic pathway planning and clear aligner design. During orthodontic treatment, teeth can generally be approximated as rigid bodies, with their movement primarily manifested as translation and rotation in three-dimensional space. Therefore, the tooth position prediction problem can be modeled as a multi-rigid-body pose estimation problem, i.e., predicting the rigid body transformation matrix of each tooth from its current state to its ideal state based on the current dentition structure. The accuracy of the prediction results directly impacts subsequent pathway planning, aligner design, and clinical treatment plan evaluation.

[0003] Existing tooth position prediction methods mostly employ point cloud networks, graph-based networks, or Transformer networks for feature extraction and output rotation and translation parameters through direct regression. While these methods can achieve end-to-end prediction, they still have the following problems in complex malocclusions and edentulous cases: the extraction and fusion of multi-scale tooth structural information, such as cusps, fissures, crown contours, and the overall arch structure, is insufficient, making it difficult to simultaneously consider local geometric details and global dentition spatial relationships; furthermore, the direct regression method has limited ability to model the uncertainty of the target tooth pose, which can easily lead to unstable prediction of the target rigid body transformation matrix, thus affecting the accuracy of subsequent orthodontic path planning and treatment plan generation.

[0004] To address the problems existing in the prior art, this application proposes a tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model. Summary of the Invention

[0005] The purpose of this invention is to provide a tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model, thereby solving the problems of insufficient multi-scale tooth structure feature extraction and fusion and insufficient stability of target tooth pose prediction in existing tooth position prediction methods.

[0006] The technical solution adopted in this invention is a tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model. The prediction process is as follows: a tooth position prediction model is built, a three-dimensional jaw model is input into the tooth position prediction model, the three-dimensional jaw model is segmented to obtain a set of tooth point clouds for each tooth, data preprocessing and enhancement are performed on each point cloud in the set of tooth point clouds to obtain a set of tooth row point clouds for each tooth; feature extraction and fusion are performed on each tooth point cloud in the set of tooth row point clouds to obtain multi-scale fused conditional features for each tooth; the multi-scale fused conditional features are input into the conditional diffusion pose regression module to predict the target rigid body transformation matrix of each tooth from the current state to the target state, and the target tooth row structure is generated based on the target rigid body transformation matrix.

[0007] The invention is further characterized by: The tooth position prediction model includes a data preprocessing and enhancement module, a multi-scale structural perception feature extraction module, a dynamic context feature fusion module, and a conditional diffusion pose regression module. The data preprocessing and enhancement module is used to receive three-dimensional dental jaws. The data preprocessing and enhancement module includes a mesh point cloud transformation unit, a coordinate system unit, a data enhancement unit, a curvature-aware adaptive sampling unit, and a missing tooth interpolation and filling unit. The multi-scale structure-aware feature extraction module includes a parallel multi-scale point cloud feature encoder, a tooth keypoint semantic feature extraction branch, and a tooth position feature encoding branch. The multi-scale point cloud feature encoder extracts the geometric features of each tooth point cloud in the dental arch point cloud set. The tooth keypoint semantic feature extraction branch includes a keypoint extraction network and a keypoint encoder, used to extract and encode a set of tooth keypoints based on each tooth point cloud and its corresponding tooth category conditions in the dental arch point cloud set. The tooth position feature encoding branch obtains tooth position features based on the set of tooth centroid coordinates in a unified oral world coordinate system. The dynamic context feature fusion module receives geometric features of tooth point clouds, semantic features of key points, and tooth position features. It then concatenates these three types of features, models contextual relationships, and performs nonlinear mapping to obtain multi-scale fused conditional features. The dynamic context feature fusion module includes a dynamic normalization unit, a multi-head self-attention unit, a feedforward network unit, a hierarchical scaling unit, a random path discarding unit, and a residual connection unit. The dynamic normalization unit, multi-head self-attention unit, and feedforward network unit are connected sequentially according to the feature processing direction. The hierarchical scaling unit, random path discarding unit, and residual connection unit are set in the feature transmission path of the multi-head self-attention unit and the feedforward network unit, and are used to fuse and model the spatial dependencies between different teeth and the semantic associations between multiple types of features. The conditional diffusion pose regression module receives multi-scale fused conditional features and predicts the rigid body transformation matrix of each tooth from its current state to the target state based on these features. The module includes a diffusion space construction unit, a conditional adaptive diffusion denoising network, a cross-layer conditional injection unit, and a pose recovery unit. The diffusion space construction unit constructs the noise pose, the conditional adaptive diffusion denoising network predicts noise disturbances based on the noise pose, diffusion time step, and multi-scale fused conditional features, the cross-layer conditional injection unit injects the multi-scale fused conditional features into each layer of the conditional adaptive diffusion denoising network, and the pose recovery unit recovers the target rigid body transformation matrix of each tooth based on the predicted noise disturbances.

[0008] The specific steps for tooth position prediction are as follows: Step 1: Obtain the patient's 3D dental model, segment the 3D dental model to obtain a 3D mesh model, convert the 3D mesh model of each tooth into a tooth point cloud set; determine the tooth category conditions corresponding to each tooth based on the segmentation results of the 3D dental model; establish a unified oral world coordinate system based on the overall geometric structure of the dental arch, establish a local coordinate system for each tooth, calculate the centroid coordinates of each existing tooth under the unified oral world coordinate system, and obtain a set of tooth centroid coordinates. Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth. Step 3: Using the multi-scale structure perception feature extraction module, extract features from the tooth row point cloud set, the tooth category conditions corresponding to each point cloud, and the tooth centroid coordinate set, respectively, to obtain the tooth point cloud geometric features, tooth key point semantic features, and tooth position features for each tooth. Step 4: The dynamic context feature fusion module is used to fuse the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fusion conditional features of each tooth. Step 5: Input the multi-scale fused conditional features as conditional guidance information into the conditional guidance-based conditional diffusion pose regression module, and predict the rigid body transformation matrix of each tooth from the current state to the target state through the conditional adaptive diffusion denoising network. Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

[0009] Step 2 is detailed below: Step 2.1: Enhance each point cloud in the tooth point cloud set of each tooth by using the mesh point cloud transformation unit, coordinate system unit and data augmentation unit to obtain the enhanced tooth point cloud set of each tooth. Step 2.2: Perform curvature-aware adaptive point cloud sampling on the enhanced tooth point cloud set to preserve the high curvature key structures of each point cloud in the sampled point cloud, and obtain the tooth point cloud set after sampling for each tooth. Step 2.3: For cases where there are missing teeth in the sampled tooth point cloud set, perform missing tooth filling based on the fitted dental arch line interpolation to generate the estimated centroid, site cloud, or positional features of the missing teeth, and obtain a complete set of tooth row point clouds of the whole set of teeth with complete structure and preserved geometric details.

[0010] Step 3 is detailed below: Step 3.1: Input the point clouds of each tooth in the dental arch point cloud set into a multi-scale point cloud feature encoder to extract the geometric features of each tooth point cloud. ; Step 3.2: Input the tooth point cloud and its corresponding tooth category conditions from the tooth row point cloud set into the tooth keypoint semantic feature extraction branch, extract the tooth keypoint set of each tooth point cloud and encode it to obtain the keypoint semantic features. ; Step 3.3: Input each coordinate from the set of tooth centroid coordinates in the unified oral world coordinate system into the tooth position feature encoding branch, and map each coordinate into a high-dimensional position feature, i.e., tooth position feature. ; (1) in, Features of tooth position The set of centroid coordinates of teeth includes the centroid coordinates of existing teeth and the estimated centroid coordinates of missing teeth obtained through tooth filling.

[0011] Geometric features of tooth point cloud in step 3.1 The extraction process is as follows: Step 3.1.1: For each tooth point cloud in the dentition point cloud set, construct multi-scale neighborhoods using local K-nearest neighbor search at different neighborhood scales, and encode the local spatial relationships of the point clouds within each scale neighborhood to generate multi-scale neighborhood features for each tooth point cloud; among them, small-scale neighborhoods are used to extract local detail features such as cusps, incisal edges, and pits and fissures, while large-scale neighborhoods are used to extract the overall crown contour and overall dentition structural features; Step 3.1.2: Perform residual connections and channel splicing fusion on the neighborhood features of different scales in the multi-scale neighborhood features of each tooth point cloud to obtain the geometric features of each tooth point cloud. ; (2) in, In the set of tooth row point clouds, the first... Dot cloud of a tooth This represents a multi-scale point cloud feature encoder. Indicates the first Geometric features of the point cloud of a tooth; Keypoint semantic features in step 3.2 The extraction process is as follows: Step 3.2.1: Input the point cloud of each tooth in the dental arch point cloud set and its corresponding tooth category conditions into the key point extraction network. Enhance the spatial structure representation of the point cloud through geometric perception position encoding, and output the response probability of each point in each tooth point cloud belonging to the key point region; wherein, the tooth category conditions are used to characterize the category or tooth position information of the corresponding tooth. Step 3.2.2: Based on the response probability of each point in the point cloud of each tooth belonging to the key point region, filter the set of key points in the cusp, incisal edge, pit and fissure, or occlusal contact area. ; and by the keypoint encoder The semantic features of key points are obtained by performing linear mapping, normalization, nonlinear activation, pooling, and multilayer perceptron processing. ; (3) in, Indicates the first The key points of each tooth Indicates the first The first tooth Coordinates of key points Indicates the first Number of key points of each tooth Indicates the first Key semantic features of each tooth.

[0012] Step 4 is detailed below: Step 4.1: Geometric features of the point cloud for each tooth. Key semantic features and tooth position features The initial spliced ​​features are obtained by concatenating the features along the channel dimension. ; (4) in, This is the initial splicing feature. Geometric features of tooth point cloud, For key point semantic features, For location features, Indicates splicing; Step 4.2: Initial splicing features of each tooth Inputting dynamic normalization units, multi-head self-attention units, and feedforward network units, context relationship modeling and nonlinear mapping are performed to obtain fused features. ; (5) in, Features of fusion; Step 4.3: Analyze the fusion features of each tooth. With initial splicing features The data is then stitched together again to obtain multi-scale fusion conditional features. .

[0013] Step 5 is detailed below: Step 5.1: During the training phase, conditional features are fused across multiple scales. As the conditional guidance information input to the conditional diffusion pose regression module, the true rigid body transformation matrix of each tooth from its current state to its target state under the constraints of the unified oral world coordinate system and local coordinate system is used as the true target pose in the diffusion modeling space. The current state is the initial malocclusion pose of the teeth in the input 3D dental model, and the target state is the target pose of the teeth under ideal occlusion. Step 5.2: Set the actual target pose The input diffusion space construction unit, based on the diffusion time step t, noise scheduling parameters, and Gaussian noise, constructs the true target pose. Forward diffusion noise processing is performed to obtain the noise pose at time step t. The reparameterized form of the forward diffusion process is: ,in, Indicates Gaussian noise. This represents the coefficient obtained by accumulating noise scheduling parameters. For the true target pose, The noise pose at time step t; Step 5.3: The noise pose obtained in step 5.2 Diffusion time step t and multi-scale fusion condition features An input-conditional adaptive diffusion denoising network is used; simultaneously, multi-scale fusion of conditional features is employed. The diffusion time step t is input to the cross-layer conditional injection unit as the conditional modulation parameter, and the conditional modulation parameter is injected into each network layer of the conditional adaptive diffusion denoising network to guide the conditional adaptive diffusion denoising network to predict noise perturbations. Its expression is: ;in, This represents the noise perturbation predicted by the conditional adaptive diffusion denoising network. The parameter is Conditional adaptive diffusion denoising network, Indicates time step Noise pose at the location Indicates the diffusion time step. Indicates the conditional features of multi-scale fusion; Step 5.4: Apply the predicted noise perturbation obtained in Step 5.3. and noise pose Input pose recovery unit, pose recovery unit according to predicted noise disturbance Noise pose Inverse recovery is performed to obtain the predicted rigid body transformation matrix; during the training phase, a multi-constraint joint loss function is used to supervise and optimize the predicted rigid body transformation matrix in order to update the parameters of the conditional adaptive diffusion denoising network. Step 5.5: In the inference stage, the multi-scale fusion conditional features obtained in Step 4 are... As conditional guidance information, the pose representation is initialized from random noise, and multi-step reverse denoising is performed through a conditional adaptive diffusion denoising network to gradually generate the target rigid body transformation matrix of each tooth from the current state to the target state.

[0014] Step 5.3's multi-constraint joint loss function includes the transformation matrix regression loss. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss The specific process of using a multi-constraint joint loss function to perform supervised optimization of the predicted rigid body transformation matrix is ​​as follows: Step 5.3.1: Calculate the transformation matrix regression loss based on the predicted rigid body transformation matrix and the actual rigid body transformation matrix. This is used to constrain the consistency between the predicted rotation matrix and translation vector and the true annotation; the transformation matrix regression loss. The expression is as follows: (6) in, Indicates the first The predicted rigid body transformation matrix for each tooth. , Indicates the first The true rigid body transformation matrix of each tooth. , and They represent the first The predicted rotation matrix and the actual rotation matrix of each tooth. and They represent the first The predicted translation vector and the actual translation vector of each tooth. Indicates a collection of teeth. For the number of teeth, Represents the logarithmic mapping of the rotation matrix. Represents the L2 norm; Step 5.3.2: Apply the predicted transformation matrix and the true transformation matrix to each point cloud in the tooth point cloud set, respectively, to obtain the predicted transformed point cloud set and the true transformed point cloud set. Calculate the point cloud geometric consistency loss using the chamfer distance. Geometric consistency loss is used to constrain the consistency of prediction results in geometric space. The expression is as follows: (7) in, This represents the set of teeth involved in the calculation. Indicates the number of teeth involved in the calculation. Represents a collection of teeth The first in One tooth, Indicates the first The original tooth dot cloud of a tooth. Indicates the first Predicted transformation point cloud of individual teeth, Indicates the first The target point cloud of a tooth, Indicates the chamfer distance; Step 5.3.3: Calculate the collision avoidance loss based on the nearest distance between adjacent teeth in the predicted transformed point cloud set. When the nearest distance between adjacent teeth is less than a preset safety threshold, a squared penalty is applied to the degree of insufficient distance to avoid collision damage. The expression is as follows: (8) in, Indicates the first Teeth and the first The teeth are topologically adjacent. Represents the first element in the predicted transformed point cloud set. Teeth and the first The nearest neighbor distance between teeth This indicates the preset safe distance threshold; Step 5.3.4: Calculate the motion constraint loss based on the distance between the centroids of adjacent teeth in the predicted transformed point cloud set. When the distance between adjacent teeth exceeds a preset maximum distance threshold, a penalty is imposed on the degree of excessive separation, resulting in motion constraint loss. The expression is as follows: (9) in, This represents the set of teeth involved in the calculation. and Represents a collection of teeth The two teeth in the middle, Indicates the first Teeth and the first The teeth are topologically adjacent. Represents the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth Represents the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth This represents the Euclidean distance between the centroids of two adjacent teeth. This represents the maximum permissible physiological distance threshold between adjacent teeth. This represents the squared penalty applied when the centroid distance between two adjacent teeth exceeds the maximum physiological distance threshold. Step 5.3.5: Regress the loss using the transformation matrix. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss By performing weighted summation, we obtain the joint loss function with multiple constraints. And update the parameters of the conditional adaptive diffusion denoising network according to the multi-constraint joint loss function; (10) in, , , and These are the weighting coefficients corresponding to each loss term. For the transformation matrix regression loss, For point cloud geometric consistency loss, To avoid damage from a collision, This is the loss due to motion constraint.

[0015] The beneficial effects of this invention are as follows: This invention extracts the geometric features of tooth point clouds, the semantic features of key tooth points, and the positional features of teeth through a multi-scale structural perception feature extraction module. It then fuses tooth structural information of different levels and types through a dynamic contextual feature fusion module. This simultaneously enhances the expressive ability of local geometric details such as cusps, incisal edges, and fissures, as well as the overall spatial relationship of the dental arch, solving the problem of insufficient multi-scale tooth structural feature extraction and fusion in existing tooth position prediction methods. By using the multi-scale fusion conditional features as conditional guidance information input into the conditional diffusion pose regression module, the target rigid body transformation matrix is ​​gradually recovered from the noisy pose through a conditional adaptive diffusion denoising network. Furthermore, the pose prediction results are supervised and optimized using a multi-constraint joint loss function. This improves the stability and accuracy of tooth target pose prediction in complex malocclusion and edentulous cases, providing a more reliable target dentition structure for subsequent orthodontic path planning and treatment plan generation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the unified oral world coordinate system in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 2 This is a schematic diagram of the local coordinate system of teeth in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model of the present invention; Figure 3 This is a schematic diagram illustrating the prediction principle of the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model of the present invention. Figure 4 This is a diagram illustrating the overall architecture of the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model according to the present invention. Figure 5 This is an unsampled image of the tooth point cloud set in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 6 This is an image showing the effect of FPS sampling on the tooth point cloud set in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 7 This is a diagram showing the effect of sampling the tooth point cloud set using the AFPS+NPDU sampling method in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention. Figure 8 This is a diagram showing the effect of sampling the tooth point cloud set using the CurvFPS sampling method in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention. Figure 9 This is a schematic diagram of the missing tooth interpolation filling principle in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 10This is a network structure diagram of the multi-scale structure-aware feature extraction module in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 11 This is a branch structure diagram of semantic feature extraction of key points of teeth in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 12 This is an image showing the extraction effect of key points in the tooth position prediction method without wisdom teeth in the tooth position prediction method of this invention, which integrates multi-scale feature extraction and conditional diffusion model. Figure 13 This is an image showing the extraction effect of key points of wisdom tooth malocclusion in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model of this invention; Figure 14 This is a structural diagram of the dynamic context feature fusion module in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 15 This is a structural diagram of the conditional diffusion pose regression module in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 16 This is a diagram of the conditional adaptive diffusion denoising network structure in the tooth position prediction method of this invention, which integrates multi-scale feature extraction and conditional diffusion model. Figure 17 This is the average point-to-point error rate curve for tooth position prediction in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model in this invention; Figure 18 This is a visualization of the complete dentition tooth position prediction results in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model of this invention; Figure 19 This is a visualization of the tooth position prediction results for missing dentition rows in the tooth position prediction method that integrates multi-scale feature extraction and conditional diffusion model of this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] The tooth position prediction model used in the following embodiments includes a data preprocessing and enhancement module, a multi-scale structural perception feature extraction module, a dynamic context feature fusion module, and a conditional diffusion pose regression module; the data preprocessing and enhancement module is used to receive three-dimensional teeth and jaws, and includes a mesh point cloud conversion unit, a coordinate system unit, a data enhancement unit, a curvature-aware adaptive sampling unit, and a missing tooth interpolation filling unit; The multi-scale structure-aware feature extraction module includes a parallel multi-scale point cloud feature encoder, a tooth keypoint semantic feature extraction branch, and a tooth position feature encoding branch. The multi-scale point cloud feature encoder extracts the geometric features of each tooth point cloud in the dental arch point cloud set. The tooth keypoint semantic feature extraction branch includes a keypoint extraction network and a keypoint encoder, used to extract and encode a set of tooth keypoints based on each tooth point cloud and its corresponding tooth category conditions in the dental arch point cloud set. The tooth position feature encoding branch obtains tooth position features based on the set of tooth centroid coordinates in a unified oral world coordinate system. The dynamic context feature fusion module receives geometric features of tooth point clouds, semantic features of key points, and tooth position features. It then concatenates these three types of features, models contextual relationships, and performs nonlinear mapping to obtain multi-scale fused conditional features. The dynamic context feature fusion module includes a dynamic normalization unit, a multi-head self-attention unit, a feedforward network unit, a hierarchical scaling unit, a random path discarding unit, and a residual connection unit. The dynamic normalization unit, multi-head self-attention unit, and feedforward network unit are connected sequentially according to the feature processing direction. The hierarchical scaling unit, random path discarding unit, and residual connection unit are set in the feature transmission path of the multi-head self-attention unit and the feedforward network unit, and are used to fuse and model the spatial dependencies between different teeth and the semantic associations between multiple types of features. The conditional diffusion pose regression module receives multi-scale fused conditional features and predicts the rigid body transformation matrix of each tooth from its current state to the target state based on these features. The module includes a diffusion space construction unit, a conditional adaptive diffusion denoising network, a cross-layer conditional injection unit, and a pose recovery unit. The diffusion space construction unit constructs the noise pose, the conditional adaptive diffusion denoising network predicts noise disturbances based on the noise pose, diffusion time step, and multi-scale fused conditional features, the cross-layer conditional injection unit injects the multi-scale fused conditional features into each layer of the conditional adaptive diffusion denoising network, and the pose recovery unit recovers the target rigid body transformation matrix of each tooth based on the predicted noise disturbances.

[0019] The method of the present invention will be further illustrated below through several embodiments: Example 1: This embodiment integrates multi-scale feature extraction and conditional diffusion model for tooth position prediction. The specific steps for tooth position prediction are as follows: Step 1: Obtain the patient's 3D dental model, segment the 3D dental model to obtain a 3D mesh model, convert the 3D mesh model of each tooth into a tooth point cloud set; determine the tooth category conditions corresponding to each tooth based on the segmentation results of the 3D dental model; establish a unified oral world coordinate system based on the overall geometric structure of the dental arch, establish a local coordinate system for each tooth, calculate the centroid coordinates of each existing tooth under the unified oral world coordinate system, and obtain the set of tooth centroid coordinates.

[0020] Among them, the tooth category condition is used to characterize the category or position information of the corresponding tooth. For example... Figure 1 As shown, the unified oral world coordinate system is used to unify the spatial reference datum for different tooth point clouds, tooth centroid coordinates, and rigid body transformation matrices. Its X-axis represents the left-right direction of the dental arch, the Y-axis represents the front-back direction of the dental arch, and the Z-axis represents the vertical direction of the teeth; as shown... Figure 2 As shown, the local coordinate system is used to describe the pose changes of a single tooth. It takes the tooth centroid as the origin and determines the local coordinate axes according to the principal direction of the tooth point cloud.

[0021] Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth.

[0022] The tooth row point cloud set is used as input for the subsequent multi-scale structure-aware feature extraction module. For example... Figures 5 to 8 As shown, a curvature-aware adaptive sampling unit is used to preserve key structures with high curvature, such as cusps, incisal edges, and pits, while reducing the point cloud size; for example... Figure 9 As shown, the missing tooth interpolation filling unit is used to estimate the centroid, occupancy cloud, or location features of the missing tooth based on the centroid distribution of existing teeth, in order to maintain the continuity of the dentition structure in cases of missing teeth.

[0023] Step 3: Using the multi-scale structure perception feature extraction module, feature extraction is performed on the tooth row point cloud set, the tooth category conditions corresponding to each point cloud, and the tooth centroid coordinate set, respectively, to obtain the tooth point cloud geometric features, tooth key point semantic features, and tooth position features for each tooth.

[0024] Among them, such as Figure 10 As shown, the multi-scale structure-aware feature extraction module is used to extract tooth structural information of different levels and types from the tooth row point cloud set; such as Figure 11 As shown, the tooth keypoint semantic feature extraction branch is used to extract a set of tooth keypoints and encode them based on the point cloud of each tooth in the tooth row and its corresponding tooth category conditions; for example... Figure 12 and Figure 13 As shown, the set of key points for teeth includes key points in the cusps, incisal edges, pits and fissures, or occlusal contact areas. Geometric features of the tooth point cloud are used to characterize the local geometric details and overall morphology of the tooth point cloud; semantic features of the key points of the teeth are used to characterize key anatomical structural information of the teeth; and positional features of the teeth are used to characterize the spatial positional relationship of the teeth in a unified oral world coordinate system.

[0025] Step 4: The dynamic context feature fusion module is used to fuse the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fusion conditional features of each tooth.

[0026] Among them, such as Figure 14 As shown, the dynamic context feature fusion module is used to concatenate, model contextual relationships, and perform nonlinear mapping on the geometric features of tooth point clouds, semantic features of key points, and tooth position features to obtain multi-scale fused conditional features. These multi-scale fused conditional features simultaneously express local geometric information of teeth, semantic information of key points, and spatial positional relationships of the tooth row, and serve as conditional guidance information for the conditional diffusion pose regression module.

[0027] Step 5: Input the multi-scale fused conditional features as conditional guidance information into the conditional guidance-based conditional diffusion pose regression module, and predict the rigid body transformation matrix of each tooth from the current state to the target state through the conditional adaptive diffusion denoising network.

[0028] Among them, such as Figure 15 As shown, the conditional diffusion pose regression module is used to predict the target rigid body transformation matrix of each tooth from its current state to the target state based on multi-scale fused conditional features; as shown... Figure 16 As shown, the conditional adaptive diffusion denoising network is used to predict noise disturbances based on noise pose, diffusion time step, and multi-scale fusion conditional features, and to progressively recover the target rigid body transformation matrix. The target rigid body transformation matrix describes the rotation and translation transformation relationship required for the corresponding tooth to move from the current state to the target state.

[0029] Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

[0030] Among them, such as Figure 18 As shown, in a complete dentition scenario, the target dentition structure can reflect the predicted result of each tooth being arranged from its initial state to the target state; for example... Figure 19 As shown, in the scenario of missing teeth, the target dentition structure can reflect the predicted results after spatial alignment of the missing tooth area and adjacent teeth.

[0031] Through the above steps, this embodiment can convert the patient's current three-dimensional dental model into a set of dental point clouds, extract and fuse the geometric features of the tooth point clouds, the semantic features of the key points of the teeth, and the positional features of the teeth, and then use the conditional diffusion pose regression module to output the target rigid body transformation matrix of each tooth, thereby generating the target dental structure.

[0032] Example 2: This embodiment integrates a tooth position prediction method that combines multi-scale feature extraction and a conditional diffusion model, such as... Figure 3 and Figure 4 As shown, the specific steps for tooth position prediction are as follows: Step 1: Obtain the patient's 3D dental model, segment the 3D dental model to obtain a 3D mesh model, convert the 3D mesh model of each tooth into a tooth point cloud set; determine the tooth category conditions corresponding to each tooth based on the segmentation results of the 3D dental model; establish a unified oral world coordinate system based on the overall geometric structure of the dental arch, establish a local coordinate system for each tooth, calculate the centroid coordinates of each existing tooth under the unified oral world coordinate system, and obtain a set of tooth centroid coordinates. Tooth category criteria include tooth type information or tooth position number information; such as Figure 1 and Figure 2 As shown, the unified oral world coordinate system is used to unify the spatial reference of different tooth point clouds, tooth centroid coordinates, and rigid body transformation matrices. Its X-axis represents the left-right direction of the dental arch, Y-axis represents the front-back direction of the dental arch, and Z-axis represents the vertical direction of the tooth. The local coordinate system is used to describe the pose changes of a single tooth. It takes the tooth centroid as the origin and determines the local coordinate axes according to the principal direction of the tooth point cloud.

[0033] Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth. The specific steps are as follows: Step 2.1: Enhance each point cloud in the tooth point cloud set of each tooth by using the mesh point cloud transformation unit, coordinate system unit and data augmentation unit to obtain the enhanced tooth point cloud set of each tooth. Step 2.2: Perform curvature-aware adaptive point cloud sampling on the enhanced tooth point cloud set to preserve the high curvature key structures of each point cloud in the sampled point cloud, and obtain the tooth point cloud set after sampling for each tooth. Step 2.3: For cases where missing teeth exist in the sampled tooth point cloud set, perform missing tooth filling based on the fitted dental arch line interpolation. This generates the estimated centroid, site cloud, or location features of the missing teeth, resulting in a complete set of tooth row point clouds with intact structure and preserved geometric details. The specific principle is as follows: Figure 9 As shown.

[0034] Step 3: The multi-scale structure-aware feature extraction module extracts features from the tooth row point cloud set, the tooth category conditions corresponding to each point cloud, and the tooth centroid coordinate set, respectively. This yields the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features for each tooth. The network structure of the multi-scale structure-aware feature extraction module used in this step is as follows: Figure 10 As shown; The specific steps are as follows: Step 3.1: Input the point clouds of each tooth in the dental arch point cloud set into a multi-scale point cloud feature encoder to extract the geometric features of each tooth point cloud. ; Step 3.2: Input the tooth point cloud and its corresponding tooth category conditions from the tooth row point cloud set into the tooth keypoint semantic feature extraction branch, extract the tooth keypoint set of each tooth point cloud and encode it to obtain the keypoint semantic features. The branch structure diagram of the semantic feature extraction of key teeth points used is as follows: Figure 11 As shown; Step 3.3: Input each coordinate from the set of tooth centroid coordinates in the unified oral world coordinate system into the tooth position feature encoding branch, and map each coordinate into a high-dimensional position feature, i.e., tooth position feature. ; (1) in, Features of tooth position The set of centroid coordinates of teeth includes the centroid coordinates of existing teeth and the estimated centroid coordinates of missing teeth obtained through tooth filling.

[0035] Step 4: The dynamic context feature fusion module fuses the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fused conditional features of each tooth. The structure of the dynamic context feature fusion module used in this step is as follows: Figure 14 As shown; Step 4 is detailed below: Step 4.1: Geometric features of the point cloud for each tooth. Key semantic features and tooth position features The initial spliced ​​features are obtained by concatenating the features along the channel dimension. ; (4) in, This is the initial splicing feature. Geometric features of tooth point cloud, For key point semantic features, For location features, Indicates splicing; Step 4.2: Initial splicing features of each tooth Inputting dynamic normalization units, multi-head self-attention units, and feedforward network units, context relationship modeling and nonlinear mapping are performed to obtain fused features. ; (5) in, Features of fusion; Step 4.3: Analyze the fusion features of each tooth. With initial splicing features The data is then stitched together again to obtain multi-scale fusion conditional features. .

[0036] Step 5: Input the multi-scale fused conditional features as conditional guidance information into the conditional diffusion pose regression module based on conditional guidance. Predict the rigid body transformation matrix of each tooth from the current state to the target state through the conditional adaptive diffusion denoising network. The structure of the conditional diffusion pose regression module used in this step is as follows: Figure 15 As shown, the conditional adaptive diffusion denoising network structure adopted is as follows: Figure 16 As shown; Step 5 is detailed below: Step 5.1: During the training phase, conditional features are fused across multiple scales. As the conditional guidance information input to the conditional diffusion pose regression module, the true rigid body transformation matrix of each tooth from its current state to its target state under the constraints of the unified oral world coordinate system and local coordinate system is used as the true target pose in the diffusion modeling space. The current state is the initial malocclusion pose of the teeth in the input 3D dental model, and the target state is the target pose of the teeth under ideal occlusion. Step 5.2: Set the actual target pose The input diffusion space construction unit, based on the diffusion time step t, noise scheduling parameters, and Gaussian noise, constructs the true target pose. Forward diffusion noise processing is performed to obtain the noise pose at time step t. The reparameterized form of the forward diffusion process is: ,in, Indicates Gaussian noise. This represents the coefficient obtained by accumulating noise scheduling parameters. For the true target pose, The noise pose at time step t; Step 5.3: The noise pose obtained in step 5.2 Diffusion time step t and multi-scale fusion condition features An input-conditional adaptive diffusion denoising network is used; simultaneously, multi-scale fusion of conditional features is employed. The diffusion time step t is input to the cross-layer conditional injection unit as the conditional modulation parameter, and the conditional modulation parameter is injected into each network layer of the conditional adaptive diffusion denoising network to guide the conditional adaptive diffusion denoising network to predict noise perturbations. Its expression is: ;in, This represents the noise perturbation predicted by the conditional adaptive diffusion denoising network. The parameter is Conditional adaptive diffusion denoising network, Indicates time step Noise pose at the location Indicates the diffusion time step. Indicates the conditional features of multi-scale fusion; Step 5.4: Apply the predicted noise perturbation obtained in Step 5.3. and noise pose Input pose recovery unit, pose recovery unit according to predicted noise disturbance Noise pose Inverse recovery is performed to obtain the predicted rigid body transformation matrix; during the training phase, a multi-constraint joint loss function is used to supervise and optimize the predicted rigid body transformation matrix in order to update the parameters of the conditional adaptive diffusion denoising network. The multi-constraint joint loss function includes the transformation matrix regression loss. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss ; Step 5.5: In the inference stage, the multi-scale fusion conditional features obtained in Step 4 are... As conditional guidance information, the pose representation is initialized from random noise, and multi-step reverse denoising is performed through a conditional adaptive diffusion denoising network to gradually generate the target rigid body transformation matrix of each tooth from the current state to the target state.

[0037] Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

[0038] Example 3: This embodiment integrates a tooth position prediction method that combines multi-scale feature extraction and a conditional diffusion model, such as... Figure 3 and Figure 4 As shown, the specific steps for tooth position prediction are as follows: Step 1: Obtain the patient's 3D dental model, segment the 3D dental model to obtain a 3D mesh model, convert the 3D mesh model of each tooth into a tooth point cloud set; determine the tooth category conditions corresponding to each tooth based on the segmentation results of the 3D dental model; establish a unified oral world coordinate system based on the overall geometric structure of the dental arch, establish a local coordinate system for each tooth, calculate the centroid coordinates of each existing tooth under the unified oral world coordinate system, and obtain a set of tooth centroid coordinates. Tooth category criteria include tooth type information or tooth position number information; such as Figure 1 and Figure 2 As shown, the unified oral world coordinate system is used to unify the spatial reference of different tooth point clouds, tooth centroid coordinates, and rigid body transformation matrices. Its X-axis represents the left-right direction of the dental arch, Y-axis represents the front-back direction of the dental arch, and Z-axis represents the vertical direction of the tooth. The local coordinate system is used to describe the pose changes of a single tooth. It takes the tooth centroid as the origin and determines the local coordinate axes according to the principal direction of the tooth point cloud.

[0039] Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth. The specific steps are as follows: Step 2.1: Enhance each point cloud in the tooth point cloud set of each tooth by using the mesh point cloud transformation unit, coordinate system unit and data augmentation unit to obtain the enhanced tooth point cloud set of each tooth. Step 2.2: Perform curvature-aware adaptive point cloud sampling on the enhanced tooth point cloud set to preserve the high curvature key structures of each point cloud in the sampled point cloud, and obtain the tooth point cloud set after sampling for each tooth. Step 2.3: For cases where missing teeth exist in the sampled tooth point cloud set, perform missing tooth filling based on the fitted dental arch line interpolation. This generates the estimated centroid, site cloud, or location features of the missing teeth, resulting in a complete set of tooth row point clouds with intact structure and preserved geometric details. The specific principle is as follows: Figure 9 As shown.

[0040] Step 3: The multi-scale structure-aware feature extraction module extracts features from the tooth row point cloud set, the tooth category conditions corresponding to each point cloud, and the tooth centroid coordinate set, respectively. This yields the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features for each tooth. The network structure of the multi-scale structure-aware feature extraction module used in this step is as follows: Figure 10 As shown; The specific steps are as follows: Step 3.1: Input the point clouds of each tooth in the dental arch point cloud set into a multi-scale point cloud feature encoder to extract the geometric features of each tooth point cloud. ; Geometric features of tooth point cloud The extraction process is as follows: Step 3.1.1: For each tooth point cloud in the dentition point cloud set, construct multi-scale neighborhoods using local K-nearest neighbor search at different neighborhood scales, and encode the local spatial relationships of the point clouds within each scale neighborhood to generate multi-scale neighborhood features for each tooth point cloud; among them, small-scale neighborhoods are used to extract local detail features such as cusps, incisal edges, and pits and fissures, while large-scale neighborhoods are used to extract the overall crown contour and overall dentition structural features; Step 3.1.2: Perform residual connections and channel splicing fusion on the neighborhood features of different scales in the multi-scale neighborhood features of each tooth point cloud to obtain the geometric features of each tooth point cloud. ; (2) in, In the set of tooth row point clouds, the first... Dot cloud of a tooth This represents a multi-scale point cloud feature encoder. Indicates the first The point cloud geometric features of each tooth are used to comprehensively encode the local details and overall morphological structure of the tooth surface in a high-dimensional space. Through the above multi-scale point cloud feature extraction process, the complex geometric structure of the tooth surface can be effectively captured. In the anterior tooth region, the model can identify subtle changes in the incisal edge and crown morphology. In the molar region, the spatial distribution of cusps and pits and fissures can be obtained.

[0041] Step 3.2: Input the tooth point cloud and its corresponding tooth category conditions from the tooth row point cloud set into the tooth keypoint semantic feature extraction branch, extract the tooth keypoint set of each tooth point cloud and encode it to obtain the keypoint semantic features. The branch structure diagram of the semantic feature extraction of key teeth points used is as follows: Figure 11 As shown; Keypoint semantic features The extraction process is as follows: Step 3.2.1: Input the point cloud of each tooth in the dental arch point cloud set and its corresponding tooth category conditions into the key point extraction network. Enhance the spatial structure representation of the point cloud through geometric perception position encoding, and output the response probability of each point in each tooth point cloud belonging to the key point region; wherein, the tooth category conditions are used to characterize the category or tooth position information of the corresponding tooth. Step 3.2.2: Based on the response probability of each point in the point cloud of each tooth belonging to the key point region, filter the set of key points in the cusp, incisal edge, pit and fissure, or occlusal contact area. ; and by the keypoint encoder The semantic features of key points are obtained by performing linear mapping, normalization, nonlinear activation, pooling, and multilayer perceptron processing. ; (3) in, Indicates the first The key points of each tooth Indicates the first The first tooth Coordinates of key points Indicates the first Number of key points of each tooth Indicates the first Key semantic features of each tooth.

[0042] Step 3.3: Input each coordinate from the set of tooth centroid coordinates in the unified oral world coordinate system into the tooth position feature encoding branch, and map each coordinate into a high-dimensional position feature, i.e., tooth position feature. ; (1) in, Features of tooth position The set of centroid coordinates of teeth includes the centroid coordinates of existing teeth and the estimated centroid coordinates of missing teeth obtained through tooth filling.

[0043] Step 4: The dynamic context feature fusion module fuses the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fused conditional features of each tooth. The structure of the dynamic context feature fusion module used in this step is as follows: Figure 14 As shown; Step 4 is detailed below: Step 4.1: Geometric features of the point cloud for each tooth. Key semantic features and tooth position features The initial spliced ​​features are obtained by concatenating the features along the channel dimension. ; (4) in, This is the initial splicing feature. Geometric features of tooth point cloud, For key point semantic features, For location features, Indicates splicing; Step 4.2: Initial splicing features of each tooth Inputting dynamic normalization units, multi-head self-attention units, and feedforward network units, context relationship modeling and nonlinear mapping are performed to obtain fused features. ; (5) in, Features of fusion; Step 4.3: Analyze the fusion features of each tooth. With initial splicing features The data is then stitched together again to obtain multi-scale fusion conditional features. .

[0044] Step 5: Input the multi-scale fused conditional features as conditional guidance information into the conditional diffusion pose regression module based on conditional guidance. Predict the rigid body transformation matrix of each tooth from the current state to the target state through the conditional adaptive diffusion denoising network. The structure of the conditional diffusion pose regression module used in this step is as follows: Figure 15 As shown, the conditional adaptive diffusion denoising network structure adopted is as follows: Figure 16 As shown; Step 5 is detailed below: Step 5.1: During the training phase, conditional features are fused across multiple scales. As the conditional guidance information input to the conditional diffusion pose regression module, the true rigid body transformation matrix of each tooth from its current state to its target state under the constraints of the unified oral world coordinate system and local coordinate system is used as the true target pose in the diffusion modeling space. The current state is the initial malocclusion pose of the teeth in the input 3D dental model, and the target state is the target pose of the teeth under ideal occlusion. Step 5.2: Set the actual target pose The input diffusion space construction unit, based on the diffusion time step t, noise scheduling parameters, and Gaussian noise, constructs the true target pose. Forward diffusion noise processing is performed to obtain the noise pose at time step t. The reparameterized form of the forward diffusion process is: ,in, Indicates Gaussian noise. This represents the coefficient obtained by accumulating noise scheduling parameters. For the true target pose, The noise pose at time step t; Step 5.3: The noise pose obtained in step 5.2 Diffusion time step t and multi-scale fusion condition features An input-conditional adaptive diffusion denoising network is used; simultaneously, multi-scale fusion of conditional features is employed. The diffusion time step t is input to the cross-layer conditional injection unit as the conditional modulation parameter, and the conditional modulation parameter is injected into each network layer of the conditional adaptive diffusion denoising network to guide the conditional adaptive diffusion denoising network to predict noise perturbations. Its expression is: ;in, This represents the noise perturbation predicted by the conditional adaptive diffusion denoising network. The parameter is Conditional adaptive diffusion denoising network, Indicates time step Noise pose at the location Indicates the diffusion time step. Indicates the conditional features of multi-scale fusion; Step 5.4: Apply the predicted noise perturbation obtained in Step 5.3. and noise pose Input pose recovery unit, pose recovery unit according to predicted noise disturbance Noise pose Inverse recovery is performed to obtain the predicted rigid body transformation matrix. During the training phase, a multi-constraint joint loss function is used to supervise and optimize the predicted rigid body transformation matrix to update the parameters of the conditional adaptive diffusion denoising network. The multi-constraint joint loss function includes the transformation matrix regression loss. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss ; Step 5.5: In the inference stage, the multi-scale fusion conditional features obtained in Step 4 are... The input conditional diffusion pose regression module uses random noise as the initial input from the diffusion space construction unit to obtain the initial random noise pose; subsequently, the initial random noise pose, diffusion time step t, and multi-scale fused conditional features are used to... The input conditional adaptive diffusion denoising network fuses multi-scale conditional features through cross-layer conditional injection units. Each network layer of the injection conditional adaptive diffusion denoising network is used to perform multi-step inverse denoising, and the pose recovery unit gradually recovers the target rigid body transformation matrix of each tooth from the current state to the target state.

[0045] Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

[0046] Example 4: This embodiment integrates a tooth position prediction method that combines multi-scale feature extraction and a conditional diffusion model, such as... Figure 3 and Figure 4 As shown, the specific steps for tooth position prediction are as follows: Step 1: Obtain the patient's 3D dental model, segment the 3D dental model to obtain a 3D mesh model, convert the 3D mesh model of each tooth into a tooth point cloud set; determine the tooth category conditions corresponding to each tooth based on the segmentation results of the 3D dental model; establish a unified oral world coordinate system based on the overall geometric structure of the dental arch, establish a local coordinate system for each tooth, calculate the centroid coordinates of each existing tooth under the unified oral world coordinate system, and obtain a set of tooth centroid coordinates. Tooth category criteria include tooth type information or tooth position number information; such as Figure 1 and Figure 2 As shown, the unified oral world coordinate system is used to unify the spatial reference of different tooth point clouds, tooth centroid coordinates, and rigid body transformation matrices. Its X-axis represents the left-right direction of the dental arch, Y-axis represents the front-back direction of the dental arch, and Z-axis represents the vertical direction of the tooth. The local coordinate system is used to describe the pose changes of a single tooth. It takes the tooth centroid as the origin and determines the local coordinate axes according to the principal direction of the tooth point cloud.

[0047] Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth. The specific steps are as follows: Step 2.1: Enhance each point cloud in the tooth point cloud set of each tooth by using the mesh point cloud transformation unit, coordinate system unit and data augmentation unit to obtain the enhanced tooth point cloud set of each tooth. Step 2.2: Perform curvature-aware adaptive point cloud sampling on the enhanced tooth point cloud set to preserve the high curvature key structures of each point cloud in the sampled point cloud, and obtain the tooth point cloud set after sampling for each tooth. Step 2.3: For cases where missing teeth exist in the sampled tooth point cloud set, perform missing tooth filling based on the fitted dental arch line interpolation. This generates the estimated centroid, site cloud, or location features of the missing teeth, resulting in a complete set of tooth row point clouds with intact structure and preserved geometric details. The specific principle is as follows: Figure 9 As shown.

[0048] Step 3: The multi-scale structure-aware feature extraction module extracts features from the tooth row point cloud set, the tooth category conditions corresponding to each point cloud, and the tooth centroid coordinate set, respectively. This yields the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features for each tooth. The network structure of the multi-scale structure-aware feature extraction module used in this step is as follows: Figure 10 As shown; The specific steps are as follows: Step 3.1: Input the point clouds of each tooth in the dental arch point cloud set into a multi-scale point cloud feature encoder to extract the geometric features of each tooth point cloud. ; Geometric features of tooth point cloud The extraction process is as follows: Step 3.1.1: For each tooth point cloud in the dentition point cloud set, construct multi-scale neighborhoods using local K-nearest neighbor search at different neighborhood scales, and encode the local spatial relationships of the point clouds within each scale neighborhood to generate multi-scale neighborhood features for each tooth point cloud; among them, small-scale neighborhoods are used to extract local detail features such as cusps, incisal edges, and pits and fissures, while large-scale neighborhoods are used to extract the overall crown contour and overall dentition structural features; Step 3.1.2: Perform residual connections and channel splicing fusion on the neighborhood features of different scales in the multi-scale neighborhood features of each tooth point cloud to obtain the geometric features of each tooth point cloud. ; (2) in, In the set of tooth row point clouds, the first... Dot cloud of a tooth This represents a multi-scale point cloud feature encoder. Indicates the first The point cloud geometric features of each tooth are used to comprehensively encode the local details and overall morphological structure of the tooth surface in a high-dimensional space. Through the above multi-scale point cloud feature extraction process, the complex geometric structure of the tooth surface can be effectively captured. In the anterior tooth region, the model can identify subtle changes in the incisal edge and crown morphology. In the molar region, the spatial distribution of cusps and pits and fissures can be obtained.

[0049] Step 3.2: Input the tooth point cloud and its corresponding tooth category conditions from the tooth row point cloud set into the tooth keypoint semantic feature extraction branch, extract the tooth keypoint set of each tooth point cloud and encode it to obtain the keypoint semantic features. The branch structure diagram of the semantic feature extraction of key teeth points used is as follows: Figure 11 As shown; Keypoint semantic features The extraction process is as follows: Step 3.2.1: Input the point cloud of each tooth in the dental arch point cloud set and its corresponding tooth category conditions into the key point extraction network. Enhance the spatial structure representation of the point cloud through geometric perception position encoding, and output the response probability of each point in each tooth point cloud belonging to the key point region; wherein, the tooth category conditions are used to characterize the category or tooth position information of the corresponding tooth. Step 3.2.2: Based on the response probability of each point in the point cloud of each tooth belonging to the key point region, filter the set of key points in the cusp, incisal edge, pit and fissure, or occlusal contact area. ; and by the keypoint encoder The semantic features of key points are obtained by performing linear mapping, normalization, nonlinear activation, pooling, and multilayer perceptron processing. ; (3) in, Indicates the first The key points of each tooth Indicates the first The first tooth Coordinates of key points Indicates the first Number of key points of each tooth Indicates the first Key semantic features of each tooth.

[0050] Step 3.3: Input each coordinate from the set of tooth centroid coordinates in the unified oral world coordinate system into the tooth position feature encoding branch, and map each coordinate into a high-dimensional position feature, i.e., tooth position feature. ; (1) in, Features of tooth position The set of centroid coordinates of teeth includes the centroid coordinates of existing teeth and the estimated centroid coordinates of missing teeth obtained through tooth filling.

[0051] Step 4: The dynamic context feature fusion module fuses the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fused conditional features of each tooth. The structure of the dynamic context feature fusion module used in this step is as follows: Figure 14 As shown; Step 4 is detailed below: Step 4.1: Geometric features of the point cloud for each tooth. Key semantic features and tooth position features The initial spliced ​​features are obtained by concatenating the features along the channel dimension. ; (4) in, This is the initial splicing feature. Geometric features of tooth point cloud, For key point semantic features, For location features, Indicates splicing; Step 4.2: Initial splicing features of each tooth Inputting dynamic normalization units, multi-head self-attention units, and feedforward network units, context relationship modeling and nonlinear mapping are performed to obtain fused features. ; (5) in, Features of fusion; Step 4.3: Analyze the fusion features of each tooth. With initial splicing features The data is then stitched together again to obtain multi-scale fusion conditional features. .

[0052] Step 5: Input the multi-scale fused conditional features as conditional guidance information into the conditional diffusion pose regression module based on conditional guidance. Predict the rigid body transformation matrix of each tooth from the current state to the target state through the conditional adaptive diffusion denoising network. The structure of the conditional diffusion pose regression module used in this step is as follows: Figure 15 As shown, the conditional adaptive diffusion denoising network structure adopted is as follows: Figure 16 As shown; Step 5 is detailed below: Step 5.1: During the training phase, conditional features are fused across multiple scales. As the conditional guidance information input to the conditional diffusion pose regression module, the true rigid body transformation matrix of each tooth from its current state to its target state under the constraints of the unified oral world coordinate system and local coordinate system is used as the true target pose in the diffusion modeling space. The current state is the initial malocclusion pose of the teeth in the input 3D dental model, and the target state is the target pose of the teeth under ideal occlusion. Step 5.2: Set the actual target pose The input diffusion space construction unit, based on the diffusion time step t, noise scheduling parameters, and Gaussian noise, constructs the true target pose. Forward diffusion noise processing is performed to obtain the noise pose at time step t. The reparameterized form of the forward diffusion process is: ,in, Indicates Gaussian noise. This represents the coefficient obtained by accumulating noise scheduling parameters. For the true target pose, The noise pose at time step t; Step 5.3: The noise pose obtained in step 5.2 Diffusion time step t and multi-scale fusion condition features An input-conditional adaptive diffusion denoising network is used; simultaneously, multi-scale fusion of conditional features is employed. The diffusion time step t is input to the cross-layer conditional injection unit as the conditional modulation parameter, and the conditional modulation parameter is injected into each network layer of the conditional adaptive diffusion denoising network to guide the conditional adaptive diffusion denoising network to predict noise perturbations. Its expression is: ;in, This represents the noise perturbation predicted by the conditional adaptive diffusion denoising network. The parameter is Conditional adaptive diffusion denoising network, Indicates time step Noise pose at the location Indicates the diffusion time step. Indicates the conditional features of multi-scale fusion; Step 5.4: Apply the predicted noise perturbation obtained in Step 5.3. and noise pose Input pose recovery unit, pose recovery unit according to predicted noise disturbance Noise pose Inverse recovery is performed to obtain the predicted rigid body transformation matrix; during the training phase, a multi-constraint joint loss function is used to supervise and optimize the predicted rigid body transformation matrix in order to update the parameters of the conditional adaptive diffusion denoising network. The multi-constraint joint loss function includes the transformation matrix regression loss. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss The specific process of using a multi-constraint joint loss function to perform supervised optimization of the predicted rigid body transformation matrix is ​​as follows: Step 5.4.1: Calculate the transformation matrix regression loss based on the predicted rigid body transformation matrix and the actual rigid body transformation matrix. This is used to constrain the consistency between the predicted rotation matrix and translation vector and the true annotation; the transformation matrix regression loss. The expression is as follows: (6) in, Indicates the first The predicted rigid body transformation matrix for each tooth. , Indicates the first The true rigid body transformation matrix of each tooth. , and They represent the first The predicted rotation matrix and the actual rotation matrix of each tooth. and They represent the first The predicted translation vector and the actual translation vector of each tooth. Indicates a collection of teeth. For the number of teeth, Represents the logarithmic mapping of the rotation matrix. Represents the L2 norm; Step 5.4.2: Apply the predicted transformation matrix and the true transformation matrix to each point cloud in the tooth point cloud set, respectively, to obtain the predicted transformed point cloud set and the true transformed point cloud set. Calculate the point cloud geometric consistency loss using the chamfer distance. Geometric consistency loss is used to constrain the consistency of prediction results in geometric space. The expression is as follows: (7) in, This represents the set of teeth involved in the calculation. Indicates the number of teeth involved in the calculation. Represents a collection of teeth The first in One tooth, Indicates the first The original tooth dot cloud of a tooth. Indicates the first Predicted transformation point cloud of individual teeth, Indicates the first The target point cloud of a tooth, Indicates the chamfer distance; Step 5.4.3: Calculate the collision avoidance loss based on the nearest distance between adjacent teeth in the predicted transformed point cloud set. When the nearest distance between adjacent teeth is less than a preset safety threshold, a squared penalty is applied to the degree of insufficient distance to avoid collision damage. The expression is as follows: (8) in, Indicates the first Teeth and the first The teeth are topologically adjacent. Represents the first element in the predicted transformed point cloud set. Teeth and the first The nearest neighbor distance between teeth This indicates the preset safe distance threshold; Step 5.4.4: Calculate the motion constraint loss based on the distance between the centroids of adjacent teeth in the predicted transformed point cloud set. When the distance between adjacent teeth exceeds a preset maximum distance threshold, a penalty is imposed on the degree of excessive separation, resulting in motion constraint loss. The expression is as follows: (9) in, This represents the set of teeth involved in the calculation. and Represents a collection of teeth The two teeth in the middle, Indicates the first Teeth and the first The teeth are topologically adjacent. Represents the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth Represents the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth This represents the Euclidean distance between the centroids of two adjacent teeth. This represents the maximum permissible physiological distance threshold between adjacent teeth. This represents the squared penalty applied when the centroid distance between two adjacent teeth exceeds the maximum physiological distance threshold. This loss term penalizes excessively separated adjacent teeth, restricting tooth movement at the overall layout level. Unlike the minimum distance constraint corresponding to the collision avoidance loss, the motion constraint loss establishes a maximum distance constraint, preventing the dentition from becoming loose during prediction. This loss term plays a crucial role in maintaining the continuity of the dental arch and the overall structural stability, and is suitable for tooth position prediction in complex malformation cases and edentulous scenarios.

[0053] Step 5.4.5: Regress the transformation matrix using the loss function. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss By performing weighted summation, we obtain the joint loss function with multiple constraints. And update the parameters of the conditional adaptive diffusion denoising network according to the multi-constraint joint loss function; (10) in, , , and These are the weighting coefficients corresponding to each loss term. For the transformation matrix regression loss, For point cloud geometric consistency loss, To avoid damage from a collision, This is the loss due to motion constraint.

[0054] Point cloud geometric consistency loss ensures the realistic structural effect of the predicted results at the geometric space level. Collision avoidance loss suppresses tooth overlap. Motion constraint loss constrains the compactness of the dental arch. By setting weighting coefficients, collaborative optimization among multiple constraints can be achieved.

[0055] Step 5.5: In the inference stage, the multi-scale fusion conditional features obtained in Step 4 are... The input conditional diffusion pose regression module uses random noise as the initial input from the diffusion space construction unit to obtain the initial random noise pose; subsequently, the initial random noise pose, diffusion time step t, and multi-scale fused conditional features are used to... The input conditional adaptive diffusion denoising network fuses multi-scale conditional features through cross-layer conditional injection units. Each network layer of the injection conditional adaptive diffusion denoising network is used to perform multi-step inverse denoising, and the pose recovery unit gradually recovers the target rigid body transformation matrix of each tooth from the current state to the target state.

[0056] Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

[0057] Example 5: The difference between this embodiment and embodiment 4 is that: In this embodiment, model training and inference are implemented using Python 3.8.0 and PyTorch 1.12.1 frameworks, with CUDA 11.6 for GPU acceleration. The experimental equipment is an NVIDIA RTX 4090 GPU. The optimizer used is AdamW, and the initial learning rate is set to 2.25 × 10⁻⁶. -4 The weight decay coefficient is set to 1×10. -6 During training, a cosine annealing learning rate scheduling strategy is used, with the minimum learning rate set to 1×10. -6 The batch size is set to 8, and the number of training rounds is set to 4000.

[0058] This embodiment uses real clinical 3D oral scan data as training and evaluation data. The data includes data on non-malformed dentition as well as data on malformed cases such as crowded teeth and missing teeth. Each sample is labeled by a clinical orthodontist with the rigid body transformation matrix of the corresponding tooth from its current state to its ideal state, which is used to supervise the training of the tooth position prediction model. In one experimental setting, the real sample size used for training and evaluation is 834 oral scan data, which are divided into training set, validation set and test set. The test set is used to verify the predictive stability of the model in complex dentition scenarios.

[0059] Module parameter verification: In the data preprocessing and enhancement module, a constraint-based random enhancement strategy and a real-transformation-based intermediate enhancement strategy are used to expand the training sample distribution; in the point cloud sampling stage, an adaptive point cloud sampling method based on curvature awareness is used, with the default number of sampling points set to 512; in the missing tooth processing stage, a missing tooth filling method based on fitting dental arch line interpolation is used; in the feature extraction stage, a multi-scale point cloud feature encoder, a keypoint encoder, and a learnable position encoder are used; in the feature fusion stage, a dynamic context feature fusion module is used; in the pose regression stage, a conditional adaptive diffusion denoising network is used; and in the training stage, a multi-constraint joint loss function consisting of transformation matrix regression loss, point cloud geometric consistency loss, collision avoidance loss, and motion constraint loss is used. The above parameters and module configurations are only one optional implementation of this invention.

[0060] In this embodiment, the oral cavity world coordinate system is as follows: Figure 1As shown in the figure, this diagram illustrates how a unified spatial reference frame is established based on the overall geometry of the dental arch. The X-axis represents the left-right direction of the dental arch, the Y-axis represents the front-back direction, and the Z-axis represents the vertical direction of the teeth. This ensures that predictions and transformations of different tooth poses are performed within the same coordinate system. The local coordinate system of the teeth is shown below. Figure 2 As shown in the figure, this diagram illustrates how a single tooth's rotation and translation in three-dimensional space can be described by using the centroid of the tooth as the origin and determining the local coordinate axes based on the tooth's principal directions.

[0061] Tooth position prediction process as follows Figure 3 As shown in the figure, this diagram illustrates the process of generating a target pose from the initial pose of a malformed tooth through tooth position prediction, demonstrating that the target pose output by this invention can serve as the target state for subsequent orthodontic path planning and treatment scheme design.

[0062] The overall framework of the tooth position prediction method is as follows: Figure 4 As shown in the figure, this diagram illustrates the overall technical process from inputting a 3D tooth model, data preprocessing, multi-scale structural awareness feature extraction, dynamic contextual feature fusion, to conditional diffusion pose regression and reconstruction of the target dentition.

[0063] In this embodiment, the adaptive point cloud sampling effect based on curvature perception is as follows: Figure 5 , Figure 6 , Figure 7 and Figure 8 As shown, the comparison of unsampled, FPS sampled, AFPS+NPDU sampled, and CurvFPS sampled results illustrates that the present invention can reduce the point cloud size while more fully preserving key high-curvature regions such as tooth cusps, incisal edges, and pits and fissures through a curvature sensing mechanism.

[0064] The principle of tooth filling based on fitting dental arch line interpolation, such as... Figure 9 As shown in the figure, this diagram illustrates the process of fitting the dental arch curve with the centroids of existing teeth and estimating the centroids of missing teeth based on the positions of adjacent teeth on the dental arch curve, in order to maintain the continuity of the dental arch structure in cases of edentulism.

[0065] The network structure of the multi-scale structure-aware feature extraction module is as follows: Figure 10 As shown in the figure, this diagram illustrates the multi-branch parallel extraction and joint representation process of tooth point cloud geometric features, tooth keypoint semantic features, and tooth position features.

[0066] The branch structure for extracting semantic features of key teeth is as follows: Figure 11 As shown in the figure, this diagram illustrates the key point extraction process based on geometrically sensed location coding and tooth category conditions, used to obtain key structural information such as cusps, incisal edges, pits and fissures, and occlusal contact areas. Figure 12 and Figure 13The results of key point extraction for dentition without wisdom teeth and malformed dentition with wisdom teeth are presented to illustrate that semantic features of key points can enhance the model's ability to perceive the anatomical structure of teeth.

[0067] The structure of the dynamic context feature fusion module is as follows: Figure 14 As shown, this illustrates the process of fusing multi-scale features into conditional features through dynamic normalization, multi-head self-attention, feedforward networks, residual connections, and feature preservation strategies. The structure of the conditional diffusion pose regression module is as follows: Figure 15 As shown, the process of forward denoising based on the true pose, conditional feature-guided backward denoising, and generation of the target rigid body transformation matrix is ​​illustrated. The conditional adaptive diffusion denoising network structure is as follows: Figure 16 The diagram illustrates the structure of the CADiT network, which predicts noise disturbances through conditional adaptive modulation and cross-layer conditional injection mechanisms. The average point-to-point error rate curve for tooth position prediction is shown below. Figure 17 As shown in the figure. This figure is used to compare the correction rate performance of different methods under different error thresholds, reflecting the advantages of this invention in terms of error distribution stability and prediction accuracy. The visualization of the complete dentition tooth position prediction results in this embodiment is shown below. Figure 18 As shown in the figure, this figure illustrates the differences in prediction performance between the initial input, the comparison method, the method presented in this paper, and the ground truth labels in multiple cases with complete dentition. The visualization of the missing dentition tooth position prediction results is shown below. Figure 19 As shown in the figure, this diagram illustrates the target tooth position prediction results of different methods in a case of edentulous teeth, demonstrating the superior stability of this invention in terms of spatial alignment between the edentulous region and adjacent teeth.

[0068] Experimental results description: Under the same data partitioning, input preprocessing procedures, and evaluation metric settings, the method of this invention can be compared with point cloud-based regression, tooth alignment networks, keypoint-guided methods, and existing diffusion model methods. In experiments, cosine similarity (CSA), root mean square error (RMSE), mean rotation error (MRE), and mean translation error (MTE) can be used to evaluate the prediction results.

[0069] In one experimental result, the method of the present invention achieved superior results in four indices: CSA, RMSE, MRE, and MTE, with CSA at 0.6843, RMSE at 1.1422, MRE at 10.4976°, and MTE at 1.7565 mm. Compared with the comparative method, the present invention can more accurately predict the rigid body transformation relationship of teeth from the initial malocclusion state to the ideal pose.

[0070] Ablation experiments show that the fusion modules of multi-scale point cloud features, key point semantic features, tooth position features, and dynamic context features have complementary effects; the multi-constraint joint loss function can synergistically improve the accuracy and structural rationality of the prediction results from aspects such as pose parameters, point cloud geometry, collision avoidance, and motion compactness.

[0071] Example 6 To verify the effectiveness of the tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model of this invention, experimental verification was conducted using real clinical 3D oral scan data. The experimental data included multiple oral scan cases, each containing the patient's current dentition status and the target dentition status annotated by a clinical orthodontist. For each case, a 3D model of the patient's jaw was first acquired, and the teeth were segmented to obtain a 3D mesh model of a single tooth. Subsequently, the 3D mesh model was converted into a tooth point cloud set, and coordinate system 1, data augmentation, curvature-aware adaptive point cloud sampling, and missing tooth filling processing based on fitted dental arch line interpolation were performed to obtain a dentition point cloud set with complete structure and preserved geometric details.

[0072] In this embodiment, the processed set of tooth row point clouds, tooth category conditions, and tooth centroid coordinates are input into the tooth position prediction model. The tooth position prediction model extracts the geometric features of the tooth point cloud, the semantic features of tooth key points, and the tooth position features through a multi-scale structure-aware feature extraction module, and then fuses these features through a dynamic context feature fusion module to obtain multi-scale fused conditional features. Subsequently, the multi-scale fused conditional features are used as conditional guidance information and input into a conditional diffusion pose regression module. A conditional adaptive diffusion denoising network predicts the rigid body transformation matrix of each tooth from the current state to the target state, and performs spatial transformation on the original tooth mesh model or tooth point cloud according to the rigid body transformation matrix to generate the target tooth row structure.

[0073] To verify the effectiveness of each module of this invention, cosine similarity, root mean square error, average rotation error, and average translation error were used as evaluation metrics, and compared with models that did not employ the key modules of this invention. Experimental results show that when the multi-scale structure-aware feature extraction module, the dynamic context feature fusion module, and the conditional diffusion pose regression module are used simultaneously, the model can achieve higher prediction accuracy and more stable dentition structure results in the tooth position prediction task. Furthermore, by adopting a multi-constraint joint loss function consisting of transformation matrix regression loss, point cloud geometric consistency loss, collision avoidance loss, and motion constraint loss, the model can simultaneously constrain pose parameter errors, point cloud geometric deviations, interdental collisions, and excessive separation of dentition, thereby improving the geometric consistency and clinical rationality of the predicted dentition.

[0074] Furthermore, to verify the impact of data preprocessing steps on the prediction results, comparative experiments were conducted on different point cloud sampling methods and edentulism filling methods. Experimental results show that the curvature-aware adaptive point cloud sampling method can preserve the geometric details of key high-curvature areas such as cusps, incisal edges, and fissures while maintaining the uniformity of point cloud spatial coverage; the edentulism filling method based on fitted arch line interpolation can estimate the reasonable spatial position of missing teeth according to the overall arch morphology, maintaining the continuity of the dentition structure. Therefore, this invention can improve the accuracy, stability, and structural rationality of tooth position prediction in complex malocclusion and edentulism cases, providing a reliable target dentition state for subsequent orthodontic pathway planning and treatment plan generation.

Claims

1. A tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model, characterized in that, The prediction process is as follows: a tooth position prediction model is built, a three-dimensional jaw model is input into the tooth position prediction model, the three-dimensional jaw model is segmented to obtain a set of tooth point clouds for each tooth, and data preprocessing and enhancement are performed on each point cloud in the set of tooth point clouds to obtain a set of tooth row point clouds for each tooth. Feature extraction and fusion are performed on the point clouds of each tooth in the tooth row point cloud set to obtain multi-scale fusion conditional features for each tooth; The multi-scale fusion conditional features are input into the conditional diffusion pose regression module to predict the target rigid body transformation matrix of each tooth from the current state to the target state, and the target tooth row structure is generated based on the target rigid body transformation matrix.

2. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 1, characterized in that, The tooth position prediction model includes a data preprocessing and enhancement module, a multi-scale structural perception feature extraction module, a dynamic context feature fusion module, and a conditional diffusion pose regression module. The data preprocessing and enhancement module is used to receive three-dimensional dental jaws. The data preprocessing and enhancement module includes a mesh point cloud conversion unit, a coordinate system unit, a data enhancement unit, a curvature-aware adaptive sampling unit, and a missing tooth interpolation and filling unit. The multi-scale structure-aware feature extraction module includes a parallel multi-scale point cloud feature encoder, a tooth keypoint semantic feature extraction branch, and a tooth position feature encoding branch. The multi-scale point cloud feature encoder extracts the geometric features of each tooth point cloud in the dental arch point cloud set. The tooth keypoint semantic feature extraction branch includes a keypoint extraction network and a keypoint encoder, used to extract a set of tooth keypoints based on each tooth point cloud and its corresponding tooth category conditions, and encode these keypoints to obtain the tooth keypoint semantic features. The tooth position feature encoding branch obtains the tooth position features based on the set of tooth centroid coordinates in a unified oral world coordinate system. The dynamic context feature fusion module receives the geometric features of the tooth point cloud, the semantic features of key points, and the tooth position features. It then performs concatenation, context relationship modeling, and nonlinear mapping on the three types of features to obtain multi-scale fusion conditional features. The dynamic context feature fusion module includes a dynamic normalization unit, a multi-head self-attention unit, a feedforward network unit, a hierarchical scaling unit, a random path discarding unit, and a residual connection unit. The dynamic normalization unit, the multi-head self-attention unit, and the feedforward network unit are connected sequentially according to the feature processing direction. The hierarchical scaling unit, the random path discarding unit, and the residual connection unit are set in the feature transmission path of the multi-head self-attention unit and the feedforward network unit, and are used to fuse and model the spatial dependencies between different teeth and the semantic associations between multiple types of features. The conditional diffusion pose regression module is used to receive the multi-scale fusion conditional features and predict the rigid body transformation matrix of each tooth from the current state to the target state based on the multi-scale fusion conditional features. The conditional diffusion pose regression module includes a diffusion space construction unit, a conditional adaptive diffusion denoising network, a cross-layer conditional injection unit, and a pose recovery unit. The diffusion space construction unit is used to construct the noise pose. The conditional adaptive diffusion denoising network is used to predict the noise disturbance based on the noise pose, the diffusion time step, and the multi-scale fusion conditional features. The cross-layer conditional injection unit is used to inject the multi-scale fusion conditional features into each network layer of the conditional adaptive diffusion denoising network. The pose recovery unit is used to recover the target rigid body transformation matrix of each tooth based on the predicted noise disturbance.

3. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 2, characterized in that, The specific steps for predicting tooth position are as follows: Step 1: Obtain the patient's three-dimensional dental model, segment the three-dimensional dental model to obtain a three-dimensional mesh model, and convert the three-dimensional mesh model of each tooth into a set of tooth point clouds. And based on the segmentation results of the three-dimensional dental model, the tooth category conditions corresponding to each tooth are determined; A unified oral world coordinate system is established based on the overall geometry of the dental arch, and a local coordinate system is established for each tooth. The centroid coordinates of each existing tooth are calculated under the unified oral world coordinate system to obtain the set of tooth centroid coordinates. Step 2: For each tooth, perform preprocessing and enhancement on each point cloud in the tooth point cloud set in sequence using the mesh point cloud transformation unit, coordinate system unit, data augmentation unit, curvature-aware adaptive sampling unit, and missing tooth interpolation filling unit to obtain the tooth row point cloud set of the whole set of teeth. Step 3: Using the multi-scale structure perception feature extraction module, extract features from the set of tooth row point clouds, the set of tooth row point clouds and the tooth category conditions corresponding to each point cloud, and the set of tooth centroid coordinates to obtain the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth. Step 4: The dynamic context feature fusion module is used to fuse the geometric features of the tooth point cloud, the semantic features of the tooth key points, and the tooth position features of each tooth to obtain the multi-scale fusion conditional features of each tooth. Step 5: Input the multi-scale fusion conditional features as conditional guidance information into the conditional diffusion pose regression module based on conditional guidance. Then, perform stepwise denoising prediction on the noise pose through the conditional adaptive diffusion denoising network, and output the target rigid body transformation matrix of each tooth from the current state to the target state. Step 6: Perform spatial rotation and translation transformations on the three-dimensional dentition according to the rigid body transformation matrix to obtain the target pose of each tooth. Reconstruct and arrange all teeth according to the target pose of each tooth to generate the target dentition structure. Use the target dentition structure as the target state for subsequent orthodontic path planning and treatment plan generation.

4. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 3, characterized in that, The specific steps of step 2 are as follows: Step 2.1: Enhance each point cloud in the tooth point cloud set of each tooth by using the mesh point cloud transformation unit, coordinate system unit and data augmentation unit to obtain the enhanced tooth point cloud set of each tooth. Step 2.2: Perform curvature-aware adaptive point cloud sampling on the enhanced tooth point cloud set to preserve the high curvature key structures of each point cloud in the sampled point cloud, and obtain the tooth point cloud set after sampling for each tooth. Step 2.3: For cases where there are missing teeth in the sampled tooth point cloud set, perform missing tooth filling based on the fitted dental arch line interpolation to generate the estimated centroid, site cloud, or positional features of the missing teeth, and obtain a complete set of tooth row point clouds of the whole set of teeth with complete structure and preserved geometric details.

5. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 3, characterized in that, The specific steps of step 3 are as follows: Step 3.1: Input the point clouds of each tooth in the dental arch point cloud set into a multi-scale point cloud feature encoder to extract the geometric features of each tooth point cloud. ; Step 3.2: Input the tooth point cloud and its corresponding tooth category conditions from the tooth row point cloud set into the tooth keypoint semantic feature extraction branch, extract the tooth keypoint set of each tooth point cloud and encode it to obtain the keypoint semantic features. ; Step 3.3: Input each coordinate from the set of tooth centroid coordinates in the unified oral world coordinate system into the tooth position feature encoding branch, and map each coordinate into a high-dimensional position feature, i.e., tooth position feature. ; (1) in, Features of tooth position The set of centroid coordinates of teeth includes the centroid coordinates of existing teeth and the estimated centroid coordinates of missing teeth obtained through tooth filling.

6. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 5, characterized in that, The geometric features of the tooth point cloud in step 3.1 The extraction process is as follows: Step 3.1.1: For each tooth point cloud in the dentition point cloud set, construct multi-scale neighborhoods using local K-nearest neighbor search at different neighborhood scales, and encode the local spatial relationships of the point clouds within each scale neighborhood to generate multi-scale neighborhood features for each tooth point cloud; among them, small-scale neighborhoods are used to extract local detail features such as cusps, incisal edges, and pits and fissures, while large-scale neighborhoods are used to extract the overall crown contour and overall dentition structural features; Step 3.1.2: Perform residual connection and channel splicing fusion on the neighborhood features of different scales in the multi-scale neighborhood features of each tooth point cloud to obtain the geometric features of each tooth point cloud. ; (2) in, The first tooth row point cloud set represents the tooth row point cloud set. Dot cloud of a tooth This represents a multi-scale point cloud feature encoder. Indicates the first Geometric features of the point cloud of a tooth.

7. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 5, characterized in that, Keypoint semantic features in step 3.2 The extraction process is as follows: Step 3.2.1: Input the point cloud of each tooth in the dental arch point cloud set and its corresponding tooth category conditions into the key point extraction network, enhance the spatial structure representation of the point cloud through geometric perception position encoding, and output the response probability of each point in each tooth point cloud belonging to the key point region; wherein, the tooth category conditions are used to characterize the category or tooth position information of the corresponding tooth; Step 3.2.2: Based on the response probability of each point in the point cloud of each tooth belonging to the key point region, filter the set of key points in the cusp, incisal edge, pit and fissure, or occlusal contact area. ; and by the keypoint encoder K The semantic features of key points are obtained by performing linear mapping, normalization, nonlinear activation, pooling, and multilayer perceptron processing. ; (3) in, Indicates the first The key points of each tooth Indicates the first The first tooth Coordinates of key points Indicates the first Number of key points of each tooth Indicates the first Key semantic features of each tooth.

8. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 3, characterized in that, The specific steps of step 4 are as follows: Step 4.1: Extract the point cloud geometric features of each tooth. Key semantic features and tooth position features The initial spliced ​​features are obtained by concatenating the features along the channel dimension. ; (4) in, This is the initial splicing feature. Geometric features of tooth point cloud, For key point semantic features, For location features, Indicates splicing; Step 4.2: Initial splicing features of each tooth Inputting dynamic normalization units, multi-head self-attention units, and feedforward network units, context relationship modeling and nonlinear mapping are performed to obtain fused features. ; (5) in, Features of fusion; Step 4.3: Analyze the fusion features of each tooth. With initial splicing features The data is then stitched together again to obtain multi-scale fusion conditional features. .

9. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 3, characterized in that, The specific steps of step 5 are as follows: Step 5.1: During the training phase, conditional features are fused across multiple scales. As the conditional guidance information input to the conditional diffusion pose regression module, the true rigid body transformation matrix of each tooth from its current state to its target state under the constraints of the unified oral world coordinate system and local coordinate system is used as the true target pose in the diffusion modeling space. Wherein, the current state is the initial malocclusion pose of the teeth in the input three-dimensional dental model, and the target state is the target pose of the teeth under ideal occlusion. Step 5.2: Set the actual target pose The input diffusion space construction unit, which constructs the diffusion space based on the diffusion time step t, noise scheduling parameters, and Gaussian noise, determines the pose of the real target. Perform forward diffusion noise addition to obtain the noise pose at time step t. The reparameterized form of the forward diffusion process is: ,in, Indicates Gaussian noise. This represents the coefficient obtained by accumulating noise scheduling parameters. For the true target pose, The noise pose at time step t; Step 5.3: The noise pose obtained in step 5.2 Diffusion time step t and multi-scale fusion condition features An input conditional adaptive diffusion denoising network is used; simultaneously, the multi-scale fusion conditional features are incorporated. The diffusion time step t is input to the cross-layer conditional injection unit as the conditional modulation parameter, and the conditional modulation parameter is injected into each network layer of the conditional adaptive diffusion denoising network to guide the conditional adaptive diffusion denoising network to predict noise perturbations. Its expression is: ;in, This represents the noise perturbation predicted by the conditional adaptive diffusion denoising network. The parameter is Conditional adaptive diffusion denoising network, Indicates time step Noise pose at the location. Indicates the diffusion time step. Indicates the conditional features of multi-scale fusion; Step 5.4: Apply the predicted noise perturbation obtained in Step 5.

3. and the noise pose Input pose recovery unit, the pose recovery unit according to the predicted noise perturbation Noise pose Inverse recovery is performed to obtain the predicted rigid body transformation matrix; during the training phase, a multi-constraint joint loss function is used to supervise and optimize the predicted rigid body transformation matrix in order to update the parameters of the conditional adaptive diffusion denoising network. Step 5.5: In the inference stage, the multi-scale fusion conditional features obtained in Step 4 are... As conditional guidance information, the pose representation is initialized from random noise, and multi-step reverse denoising is performed through a conditional adaptive diffusion denoising network to gradually generate the target rigid body transformation matrix of each tooth from the current state to the target state.

10. The tooth position prediction method integrating multi-scale feature extraction and conditional diffusion model according to claim 9, characterized in that, The multi-constraint joint loss function in step 5.4 includes the transformation matrix regression loss. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss The specific process of using a multi-constraint joint loss function to perform supervised optimization of the predicted rigid body transformation matrix is ​​as follows: Step 5.4.1: Calculate the transformation matrix regression loss based on the predicted rigid body transformation matrix and the actual rigid body transformation matrix. This is used to constrain the consistency between the predicted rotation matrix and translation vector and the true annotation; the transformation matrix regression loss. The expression is as follows: (6) in, Indicates the first The predicted rigid body transformation matrix for each tooth. , Indicates the first The true rigid body transformation matrix of each tooth. , and They represent the first The predicted rotation matrix and the actual rotation matrix of each tooth. and They represent the first The predicted translation vector and the actual translation vector of each tooth. Indicates a collection of teeth. For the number of teeth, Represents the logarithmic mapping of the rotation matrix. Represents the L2 norm; Step 5.4.2: Apply the predicted transformation matrix and the true transformation matrix to each point cloud in the tooth point cloud set, respectively, to obtain the predicted transformed point cloud set and the true transformed point cloud set. Calculate the point cloud geometric consistency loss using the chamfer distance. Geometric consistency loss is used to constrain the consistency of prediction results in geometric space. The expression is as follows: (7) in, This represents the set of teeth involved in the calculation. Indicates the number of teeth involved in the calculation. Represents a collection of teeth The first in One tooth, Indicates the first The original tooth dot cloud of a tooth. Indicates the first Predicted transformation point cloud of individual teeth, Indicates the first The target point cloud of a tooth, Indicates the chamfer distance; Step 5.4.3: Calculate the collision avoidance loss based on the nearest distance between adjacent teeth in the predicted transformed point cloud set. When the nearest distance between adjacent teeth is less than a preset safety threshold, a squared penalty is applied to the degree of insufficient distance to avoid collision damage. The expression is as follows: (8) in, Indicates the first Teeth and the first The teeth are topologically adjacent. Indicates the first element in the predicted transformed point cloud set. Teeth and the first The distance between the nearest neighbors of the teeth This indicates the preset safe distance threshold; Step 5.4.4: Calculate the motion constraint loss based on the distance between the centroids of adjacent teeth in the predicted transformed point cloud set. When the distance between adjacent teeth exceeds a preset maximum distance threshold, a penalty is imposed on the degree of excessive separation, resulting in motion constraint loss. The expression is as follows: (9) in, This represents the set of teeth involved in the calculation. and Represents a collection of teeth The two teeth in the middle, Indicates the first Teeth and the first The teeth are topologically adjacent. Indicates the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth Indicates the first element in the predicted transformed point cloud set. The centroid coordinates of a tooth This represents the Euclidean distance between the centroids of two adjacent teeth. This represents the maximum permissible physiological distance threshold between adjacent teeth. This represents the squared penalty applied when the centroid distance between two adjacent teeth exceeds the maximum physiological distance threshold. Step 5.4.5: Regress the transformation matrix using the loss function. Point cloud geometric consistency loss Collision avoidance of loss and motion constraint loss By performing weighted summation, we obtain the joint loss function with multiple constraints. The parameters of the conditional adaptive diffusion denoising network are updated according to the multi-constraint joint loss function; (10) in, , , and These are the weighting coefficients corresponding to each loss term. For the transformation matrix regression loss, For point cloud geometric consistency loss, To avoid damage from a collision, This is the loss due to motion constraint.

Images