PTv3-based intelligent automatic tooth arrangement method and equipment

Abstract

The invention provides a PTv3-based intelligent automatic tooth arrangement method and device, and the method comprises the steps: collecting three-dimensional tooth point cloud data after orthodontic treatment, carrying out the noise filtering, outlier removal and triangular meshing processing of the point cloud data, and separating a single tooth model through a deep learning segmentation network; carrying out principal component analysis and principal axis alignment on the segmented tooth model, calculating curvature characteristics in combination with normal covariance, and dynamically allocating voxel retention density by adopting a curvature-driven gradient voxel downsampling strategy; generating data pairs before and after simulation orthodontics, and constructing an orthodontics database in combination with displacement direction probability constraint and an adjacent tooth collision punishment mechanism; a PTv3-based double-branch network architecture is constructed as an intelligent automatic tooth arrangement network model, a multi-objective loss function fusing geometric errors, collision constraints and clinical prior is set, and the model is trained through weighted optimization; and inputting the preprocessed tooth point cloud into the trained model, outputting spatial transformation parameters and generating a final arrangement result.

Description

Technical Field

[0001] The present invention belongs to the field of automatic arrangement of three-dimensional tooth models, and in particular relates to an intelligent automatic tooth arrangement technical solution based on PTv3. Background Art

[0002] With technological advancements and rising living standards, public concern for oral health is growing, and demand for oral healthcare services is expanding. In addition to traditional dental treatments, demand for tooth whitening, cosmetic restorations, and orthodontic treatments is also increasing. my country's oral healthcare market continues to expand, particularly in orthodontics and dental implants, where growth has surpassed the industry average. With advances in digital technology, invisible braces have become widely used in orthodontic treatments as an effective treatment for malocclusion, offering advantages such as convenience, invisibility, and aesthetics. However, designing invisible braces typically requires a significant investment of dentists' time, limiting the efficiency of orthodontic treatment. Therefore, machine learning-based digital orthodontic technology has become an emerging research area. Determining the corrective goal is crucial during the design process. Crown positions must be predicted and set based on the malocclusion crown model to achieve a normal occlusion. Traditionally, this process is a collaborative effort between dentists and orthodontic design technicians. While computer-aided design (CAD) systems can improve efficiency, the quality of the results relies on the dentist's experience, is subject to subjective factors, and lacks consistency. Therefore, a data-driven correction target prediction method is needed to reduce misjudgments caused by lack of experience and automatically generate more robust correction targets.

[0003] Currently, existing tooth arrangement network models rely on post-orthodontic tooth segmentation results to calculate the loss function, which serves as the basis for tooth position after orthodontic treatment. However, in clinical practice, the shape of the tooth crown may change, resulting in inconsistent segmentation results before and after orthodontic treatment, which affects the calculation of the loss function and, in turn, the prediction accuracy of the transformation matrix. In addition, existing tooth arrangement algorithms do not fully consider the problem of collisions between teeth, resulting in unreasonable tooth arrangement results, which limits the application of the algorithm in real medical scenarios. Therefore, how to accurately and automatically predict the tooth correction target and solve the consistency issues in existing methods has become a challenge that needs to be urgently addressed in the field of orthodontic treatment. Summary of the Invention

[0004] In order to solve the above technical problems, the technical solution of the present invention provides an intelligent automatic teeth arrangement method based on PTv3.

[0005] The present invention provides an intelligent automatic tooth arrangement method based on PTv3, which includes the following steps: Collect 3D tooth point cloud data after orthodontic treatment, perform noise filtering, outlier removal, and triangulation on the point cloud data, and separate the single tooth model through a deep learning segmentation network; The segmented tooth model is subjected to principal component analysis and principal axis alignment, and the curvature features are calculated by combining the normal covariance. A curvature-driven gradient voxel downsampling strategy is used to dynamically allocate voxel retention density and optimize the point cloud data. Generate simulated before-and-after orthodontic data pairs, and build an orthodontic database by combining displacement direction probability constraints and adjacent tooth collision penalty mechanism; A dual-branch network architecture based on PTv3 was constructed as an intelligent automatic tooth arrangement network model, including centralized and decentralized branches, which decoupled and learned the translation and rotation parameters of the teeth, respectively, and predicted the 6-DOF rigid transformation matrix through feature fusion and dimensionality reduction modules. Setting a multi-objective loss function that integrates geometric error, collision constraint, and clinical priors, and training the model through weighted optimization, wherein the loss function includes rotation loss, displacement loss, reconstruction loss, angle loss, and collision loss; The preprocessed tooth point cloud is input into the trained model, which outputs the spatial transformation parameters and generates the final arrangement result.

[0006] Moreover, the deep learning segmentation network adopts TSegNet to perform semantic segmentation on the tooth model, optimizes the segmentation results through morphological operations, and uniformly scales them to a standard size range.

[0007] Furthermore, the curvature-driven gradient voxel downsampling strategy is to divide the point cloud into a fixed-size voxel grid and dynamically allocate the number of reserved points according to the average curvature value within the voxel.

[0008] Furthermore, the displacement direction probability constraint includes a displacement amplitude ratio range along the sagittal axis, and the rotation angle is limited to a range of rotation around the vertical axis.

[0009] Moreover, the PTv3-based dual-branch network architecture adopts an improved PTv3 module, including enhanced conditional position encoding, adaptive layer normalization, serialized attention mechanism and global feature aggregation module.

[0010] Furthermore, the collision loss is obtained by calculating the product sum of the voxel occupancy states of adjacent teeth after voxelization.

[0011] Moreover, in the multi-objective loss function, the corresponding weights of rotation loss, displacement loss and collision loss are dynamically adjusted. On the other hand, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the intelligent automatic tooth arrangement method based on PTv3 as described above is implemented.

[0012] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent automatic teeth arrangement method based on PTv3 as described above.

[0013] On the other hand, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the intelligent automatic teeth arrangement method based on PTv3 as described above.

[0014] In summary, the present invention provides a technical solution for intelligent automatic tooth alignment based on Point Transformer v3 (PTv3). Using a 3D scanner, a 3D tooth model of a patient treated under the guidance of a professional orthodontist is captured. A tooth segmentation network model is used to obtain the corresponding individual teeth and their numbers. The orthodontic model is then randomly rotated and translated to construct an orthodontic database. Finally, an intelligent automatic tooth alignment method based on PTv3 is used to achieve tooth alignment. This invention is distinguished by the direct use of post-orthodontic data to construct a simulated dataset and the optimization of the tooth alignment model by combining collision loss with constrained learning weights.

[0015] The present invention has the following positive effects: 1) This paper proposes an intelligent automatic tooth alignment method based on Point Transformer v3 (PTv3). This method uses a handheld 3D oral scanner to capture a 3D tooth model after treatment guided by a professional orthodontist. A tooth segmentation network model is used to identify individual teeth and their corresponding numbers. The orthodontic model is then randomly rotated and translated to construct an orthodontic database. Finally, an intelligent automatic tooth alignment method based on PTv3 is used to achieve this.

[0016] 2) This paper achieves principal axis alignment of tooth point clouds through principal component analysis, combines normal covariance to calculate curvature features, proposes a curvature-driven gradient voxel downsampling strategy, dynamically allocates voxel retention density, and prioritizes the retention of high-curvature key structures such as the crown occlusal surface and inter-tooth contact area, thereby improving the model's ability to capture tooth morphological characteristics.

[0017] 3) Based on the clinical laws of orthodontics, this paper designs a displacement direction probability constraint model and an adjacent tooth collision penalty mechanism. By restricting the random transformation matrix generation space and collision detection algorithm, a simulation data enhancement scheme that conforms to biomechanical characteristics is constructed.

[0018] 4) This paper designs a centralized / decentralized dual-branch network architecture, decoupling the learning of tooth translation and rotation parameters through spatial coordinate transformation. Combined with a feature dimensionality reduction module, this architecture enables accurate prediction of 6-DOF rigid transformations. A multi-objective loss function is constructed that integrates geometric error, collision constraints, and clinical priors, balancing anatomical rationality and tooth arrangement effectiveness through weighted optimization.

[0019] By integrating curvature-aware downsampling, biomechanically constrained data enhancement, and a dual-stream decoupled network architecture, this invention overcomes the three core bottlenecks of traditional automatic tooth arrangement technology: insufficient preservation of three-dimensional tooth features, a lack of clinical simulation data, and difficulties in optimizing multi-degree-of-freedom parameter coupling. Its pioneering gradient downsampling technology effectively resolves the contradiction between dimensionality reduction of point cloud data and loss of anatomical details. The simulated orthodontic database based on collision penalty breaks through the limitations of model training caused by the scarcity of real case data, while the dual-stream PTv3 network achieves precise control of the 6-degree-of-freedom rigid transformation of teeth through a decoupled learning mechanism. Ultimately, an intelligent diagnosis and treatment system with both high-precision tooth arrangement and biomechanical rationality is constructed, providing automated, standardized, and quantifiable decision support for orthodontic treatment planning, significantly improving the efficiency and accuracy of clinical diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] To further clarify the technical solution of the present invention, the accompanying drawings are briefly described below. It should be noted that these drawings represent only a few embodiments of the present invention. Those skilled in the relevant art can readily deduce other possible drawings based on these drawings without requiring additional creative effort.

[0021] Figure 1 Flowchart of an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the intelligent automatic teeth arrangement network structure according to an embodiment of the present invention; Figure 3 Schematic diagram of the Point Transformer v3 module structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0024] The present invention proposes an intelligent automatic tooth arrangement and arrangement method based on PTv3, covering data acquisition, preprocessing, deep learning segmentation, model standardization and optimization, database construction, intelligent automatic tooth arrangement network design and training strategy of comprehensive loss function. High-precision tooth point cloud data is collected by a handheld oral 3D scanner, and PCA alignment, curvature-driven downsampling and other technologies are applied to optimize data quality. A dual-branch network based on PTv3 is constructed to achieve accurate prediction of the displacement and rotation parameters of the teeth, and a comprehensive loss function, including rotation, translation, reconstruction and collision loss, is used to ensure the accuracy and rationality of the model. This invention can show broad application potential in many fields such as dental diagnosis, treatment planning, digital design, virtual reality education and patient communication, significantly improving the accuracy and efficiency of dental digital processes and promoting the innovation and development of dental technology.

[0025] See also Figure 1 In an embodiment of the present invention, an intelligent automatic tooth arrangement and arrangement method based on PTv3 is proposed, comprising the following steps: Step 1) Acquisition and segmentation of 3D tooth model data. The preferred implementation recommendation is to ensure the use of a high-precision handheld oral 3D scanner to collect high-quality tooth point cloud data and achieve accurate tooth segmentation. The implementation process includes the following steps: Step 1.1: Use a handheld 3D oral scanner to collect 3D tooth point cloud data after orthodontic treatment: Use a high-precision handheld 3D oral scanner to scan the oral cavity, covering all teeth and gums in the upper and lower jaws. For specific implementation, a handheld 3D oral scanner with an accuracy of 0.01mm to 0.05mm is recommended.

[0026] The recommended scanning order is to scan the outer (buccal) surface of the tooth first, then the inner (lingual) surface, and finally the occlusal surface. Ensure that the entire tooth surface and the required soft tissue are fully covered during the scan to avoid missing any details. For this purpose, products such as the 3DIFY-JMO1 oral scanner can be used.

[0027] Step 1.2: Construct a mesh based on the point cloud data, filter the noise and remove outliers on the collected point cloud, and convert the point cloud into a triangular mesh model.

[0028] In step 1.3, the segmented model is fed into a tooth segmentation network to generate semantic labels for the teeth and gums. A deep learning-based semantic segmentation network, such as TSegNet, is used to segment the tooth model. The segmented mesh model is normalized, the bounding box of the model is calculated, and the entire mesh is scaled to a standard size (recommended to be a cube in the range [-1, 1]). Based on the network's prediction results, each mesh vertex is assigned to a corresponding semantic category, thereby generating the corresponding semantic label (teeth, gums).

[0029] Step 1.4, separate the individual tooth models and optimize each tooth model: Based on the semantic labels, use logical operations to filter the vertices corresponding to each type label, thereby separating each tooth from the overall model, and smoothing and optimizing the segmentation results of each tooth through morphological operations such as dilation and erosion, and further smoothing the results using median filtering.

[0030] Step 2) Standardization and downsampling of the 3D tooth model. Prioritized implementation recommendations include: implementing a principal component analysis alignment algorithm to standardize the model, and developing a curvature-driven downsampling strategy to optimize data quality and processing efficiency. The specific implementation process is as follows: Step 2.1: Extract the vertex and face data of the model and apply principal component analysis (PCA) to align the principal axes of the point cloud data so that the principal axes of the tooth model are aligned with the real-world coordinate axes to achieve model standardization: Extract vertex and face data from the segmented mesh model, and perform PCA using the scikit-learn library. By calculating the covariance matrix and solving its eigenvalues and orthogonal eigenvectors, determine the three-dimensional coordinate system that represents the main directions of the model's geometric distribution. The first three principal component axes correspond to the dental arch extension direction, the buccal-lingual distribution direction, and the direction perpendicular to the occlusal plane. By calculating the transformation relationship between the eigenvectors and the original coordinate system, the rotation matrix required to align the model's main axes with the target coordinate axes is obtained:

[0031] in, The direction of dental arch extension, The distribution direction is buccal and lingual. The rotation matrix is applied to all vertices of the model to complete the normalization.

[0032] Step 2.2, estimate the normal for each point, and calculate the curvature through covariance analysis of the neighborhood normal to reflect the degree of curvature of the surface around the point, so as to determine whether it is a junction point cloud: The embodiment constructs a local neighborhood for each point based on the standardized point cloud data, and determines the normal direction by the vector corresponding to the minimum eigenvalue of the neighborhood point coordinate covariance matrix. A covariance analysis is further performed on the neighborhood normal, and the eigenvalue distribution of the covariance matrix is calculated. The ratio of the minimum eigenvalue to the total eigenvalue is defined as the curvature value of the current point, thereby quantifying the degree of local surface curvature. Areas of high curvature are marked as key feature points (i.e., the point cloud at the junction).

[0033] Step 2.3, divide the point cloud into a voxel grid of a specified size and calculate the average curvature of the points within each voxel. Sort the voxels according to the curvature value, and assign more reserved points to high-curvature voxels to retain more junction point clouds during the downsampling process: This embodiment further evenly divides the standardized 3D point cloud into a fixed-size voxel grid (recommended range: 1-3mm) and calculates the average curvature value of all points within each voxel. Dynamic retention rules are set based on the curvature distribution: 80% of the original points are retained for the top 20% of high-curvature voxels, 50% for medium-curvature voxels, and only 20% for low-curvature areas. Through non-uniform sampling, the density of key point clouds for high-value anatomical structures such as the crown occlusal surface and the tooth-cervical junction is increased, enhancing the subsequent model's ability to perceive detailed features.

[0034] Step 3) Build a 3D tooth model orthodontic database. The priority implementation recommendation is to generate diverse pre-orthodontic point cloud data through translation and rotation, and ensure structured storage and reasonable division of the data. The specific implementation process is as follows: In step 3.1, considering that there may be large morphological differences between the real tooth models before and after orthodontics, which will affect the loss optimization during the training process, a certain number of tooth models are randomly selected to generate a simulated orthodontic database.

[0035] In the embodiment, several teeth are randomly selected from the standardized tooth model as the target teeth to be arranged, and a controllable rigid body transformation is applied to them. According to the characteristic that teeth mainly move along the buccal and lingual directions in orthodontic clinics, a probability constraint on the displacement direction is designed: in the three-dimensional coordinate system, the displacement amplitude along the sagittal axis accounts for 60%-80%, and the remaining directions are randomly distributed. At the same time, in order to simulate the natural rotation amplitude of teeth during arrangement, the rotation angle range is limited (within ±15° around the vertical axis), and a uniform distribution is used to generate random rotation matrices and translation vectors to ensure that the parameters are equally covered within the constraint range, avoid generating excessively extreme deformations, and construct simulated before and after orthodontic data pairs with controllable morphological differences.

[0036] In step 3.2, considering that most teeth in real orthodontic patients are convex or concave, meaning that the primary displacement direction is forward and backward, constraints are imposed on the rotation and displacement matrices. Furthermore, an optimized simulation constraint scheme is proposed based on the collision penalty for adjacent teeth.

[0037] For each generated simulated displacement scheme, the embodiment calculates the minimum Euclidean distance between adjacent tooth point clouds as a collision detection indicator. A collision threshold is set. The normal distance between adjacent tooth surfaces is approximately 0.1-0.3 mm. Combined with the point cloud resolution, the collision threshold is set to 0.5 mm. If the distance between any adjacent teeth is less than the threshold, a collision penalty mechanism is triggered: the displacement vector is dynamically adjusted through gradient descent to reverse the displacement direction of the collision area. This includes locating the collision point, calculating the gradient direction, and iteratively updating the displacement until a safe distance is met.

[0038] In step 3.3, store the pre- and post-orthodontic data as an npy file. The default number of teeth is 32, and missing teeth are replaced with zero vectors. The orthodontic dataset is divided into a training set, a validation set, and a test set according to the preset ratio.

[0039] In this embodiment, the tooth point clouds before and after orthodontic treatment are arranged in a preset order into a fixed sequence of 32 teeth, with vacant tooth positions filled with zero vectors to form a data matrix of uniform dimensionality. Labels are pre-classified based on the degree of tooth crowding, malocclusion type, overbite, overprotrusion, and dentition type. Finally, the training, validation, and test sets are randomly divided in an 8:1:1 ratio to ensure a balanced distribution of different arch morphologies, tooth loss patterns, and displacement intensities within each set, providing diverse data support for model training.

[0040] Step 4) Build an intelligent automatic tooth arrangement network model based on PTv3 This paper proposes to design and implement a dual-branch PTv3 network to process displacement and rotation parameters separately to ensure efficient feature extraction and fusion: A dual-branch PTv3 network architecture is designed to process displacement and rotation parameters separately. In one branch, each tooth point cloud is centralized to focus on learning displacement parameters. In the other branch, the geometric structure of the point cloud is preserved and the rotation parameters are learned, i.e., decentralized. Using PTv3 as the feature extraction backbone, feature extraction is performed on the point cloud data from each branch, capturing complex spatial relationships and local geometric features. While ensuring that the two branches are relatively independent during the feature extraction process, each branch focuses on different parameter prediction tasks. The feature vectors extracted from the two branches are concatenated to form a comprehensive feature representation that preserves displacement and rotation information. Specific weights are applied to the features of the two branches to perform a weighted sum to highlight important features. The PTv3 backbone network and multiple fully connected layers are used to reduce the dimensionality of the comprehensive features and extract key parameter information. Finally, a dedicated output layer is designed to predict the rotation parameters, expressed as 6-DoF transformation parameters. The predicted rotation parameters are converted into a rotation matrix. This rotation matrix and translation vector are applied to the pre-corrected tooth point cloud to complete the rotation and translation of the tooth model, generating the predicted tooth model after alignment.

[0041] The intelligent automatic tooth arrangement network based on PTv3 constructed in the embodiment is shown in the following diagram: Figure 2 As shown, it includes dual-branch feature extraction, comprehensive feature refinement and transformation matrix prediction. Specifically: In step 4.1, the input full-mouth dental point cloud data is first structured. A normalized sampling operation is performed on the complete dentition point cloud, consisting of T individual teeth, with each tooth uniformly sampled to N three-dimensional coordinate points. This results in a regularized tensor representation of dimensions T × N × 3 (T is preferably 32, N is 513, and the feature dimension C = 3). To achieve precise decoupled learning of the spatial transformation parameters of the teeth, the system employs a two-branch parallel processing architecture: the first branch geometrically centers the point cloud of a single tooth through global pooling and fully connected layers, outputting a T × 3-dimensional vector representing the spatial displacement characteristics of each tooth. The second branch decentralizes the point cloud, extracting the morphological features of each tooth through local feature encoding while preserving the spatial distribution characteristics of the original point cloud, generating a T × N × 3-dimensional feature tensor that maintains the original spatial structure.

[0042] In step 4.2, each feature branch performs feature enhancement using a deep neural network. During the feature dimensionality increase phase, the original 3D features are first expanded to a higher-dimensional 256-dimensional space using fully connected layers, significantly improving the model's nonlinear representation capabilities. Subsequently, a deep feature extraction network is constructed using five cascaded Point Transformer v3 modules. This module utilizes a self-attention mechanism and a local neighborhood aggregation strategy to fine-tune the geometric features of the point cloud at multiple scales. By cascading Point Transformer v3 modules, the model is able to progressively extract richer spatial features, particularly by fine-tuning the geometric features of the point cloud at different scales. This hierarchical processing helps capture feature information at different levels and improves the ability to model complex geometric shapes. Each cascaded Point Transformer module not only strengthens the feature representation of the previous level but also enhances the effectiveness of lower-level feature information through cross-layer feature reuse. This enables the model to better utilize the feature representations of each layer, ultimately achieving a more accurate and efficient feature representation through continuously enhanced feature fusion. After completing dual-branch feature extraction, the system achieves feature fusion (Cat) through channel-wise splicing, forming a joint representation that combines global displacement information with local morphological features. To enhance the ability to capture broad contextual information, the fused features are further processed through four cascaded Point Transformer v3 modules. The introduction of a residual connection mechanism effectively mitigates the vanishing gradient problem, while cross-layer feature reuse enhances the network's ability to model complex spatial relationships.

[0043] In step 4.3, at the end of the feature processing flow, the network achieves the final spatial transformation prediction through multi-level feature fusion and parameter decoding. First, the high-dimensional features are mapped to the fully connected layer for dimensionality reduction, and then dedicated parameter output heads are constructed separately: for rotation matrix prediction, a convolutional layer is used with the ReLU activation function to construct a rotation parameter regressor, which outputs 9-dimensional rotation matrix parameters that conform to the SO(3) group characteristics; for displacement vector prediction, a 3-dimensional translation vector is generated through a fully connected layer with constraints. In particular, the present invention recommends introducing an orthogonalization post-processing module in the parameter output stage to ensure that the predicted rotation matrix meets the mathematical constraints of the rigid transformation.

[0044] The present invention preferably recommends that the feature extraction module adopts the Point Transformer v3 module. For the convenience of implementation reference, the present invention improves the Point Transformer v3 module. The improved architecture is as follows Figure 3As shown in Figure 2, its core design combines multi-dimensional feature enhancement with dynamic attention modeling mechanisms. First, an enhanced conditional position encoding is introduced at the input. By tensor-concatenating and nonlinearly mapping geometric features such as the point cloud's 3D coordinates and local curvature with semantic features, a highly discriminative spatial position representation is constructed. Furthermore, an adaptive layer normalization module (AdaLN) is employed to normalize the feature distribution. Compared to traditional LN layer normalization, AdaLN dynamically adjusts the normalization parameters based on input changes, enabling better response to varying point cloud densities and noise, and improving the network's adaptability and robustness.

[0045] The module's core processing unit employs an innovative sequential attention mechanism, transforming unordered point clouds into topological sequence structures through a spatial grid sorting strategy. Specifically, the attention computation utilizes a multi-head parallel architecture, with each attention head focusing on feature interactions at different scales. A deformable attention window mechanism adaptively adjusts the receptive field, significantly improving the modeling capability of complex geometric structures. After the attention feature reconstruction, the system applies adaptive layer normalization (AdaLN) to scale the output features, ensuring the stability of deep network training and improving the model's adaptability to diverse tasks. To further enhance the model's understanding and capture of global features, the improved model adds a global feature aggregation step before the fully connected layers. This step aggregates all point cloud features using a global average pooling strategy, extracting global context and effectively preventing information loss. This global representation of features enables the network to capture the macroscopic structure of the entire point cloud, further improving the comprehensiveness and accuracy of feature representation. Feature transformation is then performed through a fully connected network composed of gated linear units. This network employs a bottleneck architecture to intelligently compress and expand feature dimensions, preserving valuable information while enhancing the model's nonlinear representation capabilities.

[0046] Notably, the enhanced conditional position encoding, AdaLN adaptive layer normalization and sequential attention mechanism, AdaLN layer normalization, global feature aggregation, and skip connections before and after the fully connected layer achieve cross-layer information fusion through element-by-element addition of feature matrices. This design not only effectively alleviates the vanishing gradient problem but also improves the model's accuracy in capturing geometric details through multi-level feature reuse. Based on the output rotation and translation parameters, the pre-corrected tooth model is spatially transformed to generate the predicted tooth model after alignment.

[0047] Step 5) Design a training strategy that combines collision loss and constrained learning weights. The recommended priority is to implement the basic rotation and displacement loss functions, and then gradually integrate reconstruction loss, angle loss, and collision loss to optimize the overall accuracy and rationality of the model. The specific implementation process is as follows: Step 5.1, calculate the rotation loss and displacement loss based on the predicted parameters and the true parameters, that is, calculate the rotation loss and displacement loss between the predicted rotation matrix and the true rotation matrix, where Spin loss The calculation formula is as follows:

[0048] Where i is the tooth number, N is the total number of teeth, is the predicted rotation matrix, is the real rotation matrix, It constrains learning weights, and SmoothL1 is a smooth L1 loss that combines the advantages of L1 loss and L2 loss. It uses L2 loss for small errors and L1 loss for large errors to reduce sensitivity to outliers.

[0049] Displacement loss The calculation formula is as follows:

[0050] in, is the predicted displacement matrix, is the real displacement matrix.

[0051] Step 5.2, calculate the rotation loss and displacement loss based on the predicted parameters and the actual parameters, that is, calculate the reconstruction loss and angle loss between the predicted point cloud and the actual point cloud, evaluate the spatial correspondence between the predicted point cloud and the actual point cloud, and reflect the matching degree of the point cloud distribution. Reconstruction loss The calculation formula is:

[0052] Where n is the tooth number, N is the total number of teeth, is the predicted tooth point cloud, This is a real tooth point cloud.

[0053] Angle loss The calculation formula is:

[0054] Step 5.3, calculate the collision loss for the predicted point cloud teeth to avoid unreasonable tooth arrangement collision, including voxelizing the tooth point cloud results and calculating the collision loss between a single tooth and adjacent teeth. , the calculation formula is:

[0055] in, For the first tooth in voxel The occupancy status (1 or 0), For the second tooth in voxel The occupancy status (1 or 0).

[0056] In step 5.4, we combine several different losses and weight the loss function based on the prior orthodontic decision to ensure that the model focuses on the teeth that are actually deformed: In this embodiment, different loss functions are combined to obtain the total loss:

[0057] Among them, the weight 、 、 Spin loss , displacement loss , collision losses The corresponding weights can be dynamically adjusted based on the specific distribution of training data and prior orthodontic decisions. The Adam optimization algorithm is used to minimize the comprehensive loss function, and the weight coefficients are dynamically adjusted based on the GradNorn method to improve the accuracy and rationality of tooth arrangement.

[0058] In step 6, during the automatic tooth alignment process, the collected data is first preprocessed as mentioned in steps 1 and 2. The processed data is then fed into the trained model, which outputs rotation and translation parameters. Finally, the processed tooth model is transformed into spatial coordinates to generate the final alignment result.

[0059] In practice, those skilled in the art may employ software techniques to automate the above process. Accordingly, providing a PTv3-based intelligent automatic tooth alignment solution, including a computer or server, and executing the above process on the computer or server for PTv3-based intelligent automatic tooth alignment would also fall within the scope of protection of the present invention.

[0060] In another embodiment, it also relates to an electronic device, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned intelligent automatic teeth arrangement method based on PTv3.

[0061] In another embodiment, it also relates to a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the above-mentioned intelligent automatic teeth arrangement method based on PTv3.

[0062] In another embodiment, a computer program product is provided, comprising a computer program, wherein the computer program implements the above-mentioned intelligent automatic teeth arrangement method based on PTv3 when executed by a processor.

[0063] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of these modules may be selected to achieve the objectives of the present embodiments based on practical needs. Those skilled in the art will be able to understand and implement these embodiments without inventive effort. Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a required general-purpose hardware platform, or, of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes instructions for enabling a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or portions thereof. The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An intelligent automatic teeth arrangement method based on PTv3, characterized in that: The following processes are included: Collect 3D tooth point cloud data after orthodontic treatment, perform noise filtering, outlier removal, and triangulation on the point cloud data, and separate the single tooth model through a deep learning segmentation network; The segmented tooth model is subjected to principal component analysis and principal axis alignment, and the curvature features are calculated by combining the normal covariance. A curvature-driven gradient voxel downsampling strategy is used to dynamically allocate voxel retention density and optimize the point cloud data. Generate simulated before-and-after orthodontic data pairs, and build an orthodontic database by combining displacement direction probability constraints and adjacent tooth collision penalty mechanism; A dual-branch network architecture based on PTv3 was constructed as an intelligent automatic tooth arrangement network model, including centralized and decentralized branches, which decoupled and learned the translation and rotation parameters of the teeth, respectively, and predicted the 6-DOF rigid transformation matrix through feature fusion and dimensionality reduction modules. Setting a multi-objective loss function that integrates geometric error, collision constraint, and clinical priors, and training the model through weighted optimization, wherein the loss function includes rotation loss, displacement loss, reconstruction loss, angle loss, and collision loss; The preprocessed tooth point cloud is input into the trained model, which outputs the spatial transformation parameters and generates the final arrangement result.

2. The method according to claim 1, characterized in that The deep learning segmentation network adopts TSegNet to perform semantic segmentation on the tooth model, optimizes the segmentation results through morphological operations, and uniformly scales them to a standard size range.

3. The method according to claim 1, characterized in that The curvature-driven gradient voxel downsampling strategy is as follows: the point cloud is divided into a fixed-size voxel grid, and the number of reserved points is dynamically allocated according to the average curvature value within the voxel.

4. The method according to claim 1, wherein The displacement direction probability constraint includes the displacement amplitude ratio range along the sagittal axis, and the rotation angle is limited to the range of rotation around the vertical axis.

5. The method according to claim 1, wherein The PTv3-based dual-branch network architecture adopts an improved PTv3 module, including enhanced conditional position encoding, adaptive layer normalization, serialized attention mechanism and global feature aggregation module.

6. The method according to claim 1, characterized in that The collision loss is obtained by calculating the product sum of the voxel occupancy states of adjacent teeth after voxelization.

7. The method according to claim 1, characterized in that In the multi-objective loss function, the corresponding weights of rotation loss, displacement loss and collision loss are dynamically adjusted.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the intelligent automatic teeth arrangement method based on PTv3 as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent automatic teeth arrangement method based on PTv3 as claimed in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the intelligent automatic teeth arrangement method based on PTv3 as claimed in any one of claims 1 to 7 is implemented.

Images