Automatic tooth alignment method, device and equipment based on point cloud understanding and storage medium

Abstract

The embodiment of the application relates to the technical field of tooth correction, in particular to an automatic tooth arrangement method and device based on point cloud understanding, equipment and medium, the method comprises the following steps: acquiring tooth point cloud based on cone beam computed tomography image; the tooth point cloud comprises three-dimensional point cloud data of complete teeth containing tooth roots; tooth features are extracted based on tooth point cloud data and a graph convolution network of an attention mechanism; the tooth features comprise a global feature vector of a dentition and a local feature vector of the teeth; the tooth features are aggregated and updated based on a dynamic graph recommendation method to obtain a pose prediction result; the pose prediction result is optimized based on the aggregated and updated tooth features and an iterative optimization method to obtain a target pose result. The automatic tooth arrangement method based on point cloud understanding can automatically calculate corresponding target poses according to original dentitions of patients with malocclusion, and the accuracy of automatic tooth arrangement results is improved.

Description

Technical Field

[0001] This application relates to the field of orthodontic technology, and in particular to an automatic tooth alignment method, device, equipment, and storage medium based on point cloud understanding. Background Technology

[0002] Orthodontic tooth alignment refers to the planning of the target dentition position for patients with malocclusion, and it is a crucial step in orthodontic treatment. Currently, the clinical orthodontic treatment process, including treatment planning and diagnostic tooth alignment, first requires the dentist to perform an intraoral scan of the patient, sending the results to the orthodontic manufacturer. The manufacturer's technician then completes the tooth alignment according to the dentist's specific instructions. To arrive at the final treatment plan, multiple rounds of communication, feasibility analysis, and trials are necessary between the dentist and the technician. The realization of automated tooth alignment can improve the automation level of digital orthodontic treatment, shortening the above process and saving significant time for both doctors and patients.

[0003] Existing automated tooth alignment methods typically employ deep learning networks to determine the target tooth positions. However, these deep learning-based methods train the network using non-transparent scanned data, which only includes the crowns and lacks information on the roots and alveolar bone beneath the gum line. Consequently, the resulting tooth alignment may not fully meet the aforementioned requirements. Furthermore, improper tooth spacing or missing teeth can lead to discrepancies between the desired dentition and the ideal dentition distribution, resulting in significant prediction errors. Summary of the Invention

[0004] This application provides an automatic tooth alignment method, device, equipment, and storage medium based on point cloud understanding. It automatically calculates the corresponding target pose based on the original dentition of malocclusion patients, thereby improving the accuracy of automatic tooth alignment results.

[0005] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide an automatic tooth alignment method based on point cloud understanding, comprising: acquiring tooth point clouds based on cone-beam computed tomography (CBCT) images; the tooth point clouds including three-dimensional point cloud data of complete teeth containing roots; extracting tooth features based on the tooth point cloud data and a graph convolutional network with an attention mechanism; the tooth features including global feature vectors of the dentition and local feature vectors of the teeth; aggregating and updating the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; and optimizing the pose prediction result based on the aggregated and updated tooth features and an iterative optimization method to obtain a target pose result.

[0006] In some exemplary embodiments, a dynamic graph recommendation method is used to aggregate and update tooth features to obtain pose prediction results. This includes: serializing and encoding tooth features, and constructing a dynamic graph based on the adjacency relationship of teeth; assigning an embedding node capable of network learning to each tooth, performing global pooling on the dynamic graph, inputting it into the dynamic graph recommendation network, aggregating and updating tooth features, and obtaining pose prediction results.

[0007] In some exemplary embodiments, the tooth features are aggregated and updated using equation (1):

[0008]

[0009] in, For each embedded node; This represents information from neighboring nodes, where k is the update time step. Both Update and Aggregate are differentiable functions. Aggregate aggregates information from neighboring nodes, while Update responds to the information and updates the embedded nodes.

[0010] In some exemplary embodiments, the pose prediction result is optimized based on the aggregated and updated tooth features and an iterative optimization method to obtain the target pose result, including: obtaining a pose transformation matrix based on the aggregated and updated tooth features; obtaining a three-dimensional model of the tooth according to the serialized encoding of each tooth, loading the pose transformation onto the corresponding tooth, and adjusting the tooth position; and iteratively reducing the error of the pose prediction result based on the iterative optimization method to obtain the target pose result.

[0011] In some exemplary embodiments, the pose transformation matrix is ​​calculated by using a multilayer perceptron to embed the features of the nodes into Lie group rotations and three-dimensional translations.

[0012] In some exemplary embodiments, the target pose result is obtained by iteratively reducing the error of the pose prediction result based on an iterative optimization method, including: treating the pose prediction value in the previous iteration as an approximate estimate of the target pose, obtaining the error using the iterative nearest point method, feeding the error back into the network as part of the input for the next iteration, and iteratively optimizing the parameters of the network.

[0013] Secondly, this application provides an automatic tooth alignment device, characterized in that it comprises: a data preprocessing module, a feature extraction module, a feature propagation module, and a pose prediction module connected in sequence; the data preprocessing module is used to segment cone-beam computed tomography (CBCT) images to obtain binary segmentation results; and to reconstruct the binary segmentation results to obtain a three-dimensional tooth model; and then to obtain a tooth point cloud through the three-dimensional tooth model; the feature extraction module includes a convolutional network with an attention mechanism for extracting tooth features; the tooth features include global feature vectors of the dentition and local feature vectors of the teeth; the feature propagation module aggregates and updates the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; the pose prediction module includes an iterative optimization module for iteratively reducing the error of the pose prediction result to obtain a target pose result.

[0014] In some exemplary embodiments, the pose prediction module further includes a feature transformation module; the feature transformation module embeds and maps the tooth features into Lie group rotations and three-dimensional translations, calculates the tooth pose transformation matrix; then reads the three-dimensional model of the tooth according to the serialized encoding of each tooth, loads the pose transformation onto the corresponding tooth, and moves and optimizes the tooth pose.

[0015] In addition, this application also provides an electronic device, including: 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, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described automatic tooth alignment method based on point cloud understanding.

[0016] In addition, this application also provides a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the above-mentioned automatic tooth alignment method based on point cloud understanding.

[0017] The technical solution provided in this application has at least the following advantages:

[0018] This application provides an automatic tooth alignment method, apparatus, device, and storage medium based on point cloud understanding. The method includes: acquiring tooth point clouds based on cone-beam computed tomography (CBCT) images; the tooth point clouds include 3D point cloud data of complete teeth containing roots; extracting tooth features based on the tooth point cloud data and a graph convolutional network with an attention mechanism; the tooth features include global feature vectors of the dentition and local feature vectors of the teeth; aggregating and updating the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; and optimizing the pose prediction result based on the aggregated and updated tooth features and an iterative optimization method to obtain a target pose result. The automatic tooth alignment method provided in this application improves the accuracy of the automatic tooth alignment results by extracting tooth features through a graph convolutional network with an attention mechanism, aggregating and updating the tooth features through a dynamic graph recommendation method, and optimizing the pose prediction result through an iterative optimization method. In addition, for cases of missing teeth that require subsequent dentures or implants, the automatic tooth arrangement method provided in this application can pre-insert a three-dimensional model of a virtual denture into the original input, and then delete it from the output dentition after the tooth arrangement is completed, thereby reserving space for dentures. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0020] Figure 1 A flowchart illustrating an automatic tooth alignment method based on point cloud understanding, provided as an embodiment of this application;

[0021] Figure 2 A deep learning network structure diagram for automatic tooth alignment based on point cloud understanding provided in an embodiment of this application;

[0022] Figure 3 This is a schematic diagram illustrating the process of aggregation and updating of adjacent teeth by a feature propagation module provided in an embodiment of this application.

[0023] Figure 4 A schematic diagram of an automatic tooth-aligning device based on point cloud understanding is provided for an embodiment of this application;

[0024] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] As can be seen from the background technology, the existing automatic tooth alignment methods have a large prediction error in the tooth alignment results.

[0026] With the rapid development of artificial intelligence in recent years, automatic tooth alignment methods based on deep learning have emerged. One proposed deep learning network for automatic tooth alignment, TANet, consists of five modules: (1) a preprocessing module: performing tooth segmentation and normalization calculations on oral scan images; (2) a feature extraction module: using a point cloud processing network (PointNet) to extract features from the segmented teeth; (3) a feature propagation module: using a gated graph neural network (GGSNN) to propagate tooth features to other teeth with prior positional relationships; and (4) an assembler module: after splicing the features, using a multilayer perceptron to map them into the amount of tooth movement, thereby providing the target position of the tooth. Another proposed deep learning network for automatic tooth alignment, PSTN, is similar to TANet, using PointNet and PointNet++ networks as encoders to extract tooth features and a multilayer perceptron as a decoder. The network outputs the target position of the tooth based on the learned target position.

[0027] Besides deep learning-based methods, there are also tooth arrangement methods based on computer-aided design / computer-aided manufacturing technologies, which mainly rely on manual tooth arrangement. Although these methods can predict tooth positions, they do not take into account issues such as root position, improper tooth spacing, and missing teeth.

[0028] In orthodontics, the ideal orthodontic target position places specific requirements on the position and orientation of the tooth roots, including the parallelism between the roots and the distance between the roots and the alveolar bone. However, existing deep learning-based automatic tooth alignment methods use non-transparent data obtained through scanning to train the network, which only includes the crowns and lacks information on the roots and alveolar bone under the gingiva. Therefore, the tooth alignment results output by the trained network do not fully meet the above requirements.

[0029] Furthermore, in an ideal dental arch, teeth need to be in close contact with each other. Because deep learning networks have some error compared to human gold standards during inference, their automatic tooth alignment results in excessively large gaps between teeth, leading to poor aesthetics, a poor chewing experience, or a degree of tooth overlap, making it impractical for actual treatment. Although TANet attempted to address this issue through a conditionally weighted algorithm, the results were unsatisfactory.

[0030] In addition, patients with malocclusion in clinical practice may have problems such as bone fractures or loss of permanent teeth, or they may need to extract teeth to create enough space for movement due to crowding. This can lead to a certain difference between the expected dentition and the ideal dentition with the same number of teeth before and after treatment. In addition, the training set samples may be unbalanced. Therefore, the situation of missing teeth needs to be given special consideration.

[0031] To address the aforementioned technical problems, this application provides an automatic tooth alignment method based on point cloud understanding, comprising: acquiring tooth point clouds based on cone-beam computed tomography (CBCT) images; the tooth point clouds including 3D point cloud data of complete teeth containing roots; extracting tooth features based on the tooth point cloud data and a graph convolutional network with an attention mechanism; the tooth features including global feature vectors of the dentition and local feature vectors of the teeth; aggregating and updating the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; and optimizing the pose prediction result based on the aggregated and updated tooth features and an iterative optimization method to obtain a target pose result. The automatic tooth alignment method based on point cloud understanding provided in this application can automatically calculate the corresponding target pose based on the original dentition of malocclusion patients, improving the accuracy of the automatic tooth alignment results.

[0032] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0033] See Figure 1 This application provides an automatic tooth alignment method based on point cloud understanding, including the following steps:

[0034] Step S1: Obtain tooth point cloud based on cone-beam computed tomography (CBCT) images; the tooth point cloud includes three-dimensional point cloud data of a complete tooth containing the tooth root.

[0035] Step S2: Extract tooth features based on tooth point cloud data and a graph convolutional network with an attention mechanism; tooth features include global feature vectors of the dental arch and local feature vectors of the teeth.

[0036] Step S3: Based on the dynamic graph recommendation method, the tooth features are aggregated and updated to obtain the pose prediction results.

[0037] Step S4: Based on the aggregated and updated tooth features and the iterative optimization method, optimize the pose prediction results to obtain the target pose result.

[0038] In step S1, as Figure 2As shown, teeth are segmented from cone-beam computed tomography (CBCT) images and converted into 3D point clouds, which are then used as input to the network after data augmentation. Since CBCT images provide complete teeth including roots compared to oral scans containing only crowns, CBCT images are segmented, and the binary segmentation results are reconstructed into 3D tooth models. Then, the vertices of the 3D model are downsampled to ensure consistent output point counts to obtain the tooth point cloud. Random translations and rotations of the tooth point cloud are performed to expand the dataset size.

[0039] Step S2 mainly involves the extraction of tooth features. Graph convolutional networks have demonstrated a reliable ability to learn knowledge from irregularly structured graph data structures on various tasks, including 3D shape recognition and segmentation algorithms based on frequency domain or spatial domain graph convolutional networks. These algorithms define the spatial relationships of points or other basic units as graphs and extract and aggregate their topological features. Some studies have also used attention mechanisms to extract more fine-grained geometric features.

[0040] Existing deep learning-based automatic tooth alignment methods use PointNet, which has limited local fine-grained feature extraction capabilities, to extract tooth features. Based on a dual-flow graph convolutional network, this invention constructs a graph attention mechanism module to compute the global feature vector of the dentition and the local feature vectors of the teeth. The input original malocclusion point cloud is uniformly set to N*1000 points, where N is the number of teeth. The output global and local feature vectors have 512 and 32 channels, respectively.

[0041] Step S3 mainly involves aggregating and updating the tooth features. In some embodiments, step S3 uses a dynamic graph recommendation method to aggregate and update the tooth features to obtain pose prediction results. This includes: sequentially encoding the tooth features and constructing a dynamic graph based on the adjacency relationships of the teeth; assigning an embedding node capable of network learning to each tooth; performing global pooling on the dynamic graph; inputting the dynamic graph into the dynamic graph recommendation network; aggregating and updating the tooth features; and obtaining pose prediction results.

[0042] like Figure 3 As shown, this application proposes a graph propagation module based on a dynamic graph recommendation method to aggregate and update the features of adjacent teeth. Since the input teeth have already undergone instance segmentation, they can be serialized and encoded. The method used for tooth serialization and encoding can be any; in this embodiment, based on the FDI general tooth position representation, all teeth are encoded as 11-17 (right maxilla), 21-27 (left maxilla), 31-37 (left mandible), and 41-47 (right mandible). A graph is constructed based on the adjacency relationships of the teeth. A learnable embedding node is assigned to each tooth, and after global pooling, it is input into the dynamic graph recommendation network to update and further aggregate features.

[0043] In some embodiments, the tooth features are aggregated and updated using formula (1):

[0044]

[0045] That is, the expression for feature propagation is shown in equation (1), where, For each embedded node; This represents information from neighboring nodes, where k is the update time step. Both Update and Aggregate are differentiable functions. Aggregate aggregates information from neighboring nodes, while Update responds to the information and updates the embedded nodes.

[0046] In some embodiments, step S4 optimizes the pose prediction result based on the aggregated and updated tooth features and an iterative optimization method to obtain the target pose result, including: obtaining a pose transformation matrix based on the aggregated and updated tooth features; obtaining a three-dimensional model of the tooth according to the serialized encoding of each tooth, loading the pose transformation onto the corresponding tooth, and adjusting the tooth position; and iteratively reducing the error of the pose prediction result based on the iterative optimization method to obtain the target pose result.

[0047] In some embodiments, the pose transformation matrix is ​​calculated by using a multilayer perceptron to embed the features of the nodes into Lie group rotations and three-dimensional translations.

[0048] In some embodiments, based on an iterative optimization method, the error of the pose prediction result is iteratively reduced to obtain the target pose result. This includes: treating the pose prediction value in the previous iteration as an approximate estimate of the target pose; using the iterative nearest-point method to obtain the error; feeding the error back into the network as part of the input for the next iteration; and iteratively optimizing the network parameters. Specifically, the iterative optimization method is mainly based on unsupervised confidence scoring and corresponding parameter optimization methods to iteratively reduce the pose prediction error during network training, thereby improving the tooth pose result.

[0049] See Figure 4This application provides an automatic tooth alignment device, comprising: a data preprocessing module 101, a feature extraction module 102, a feature propagation module 103, and a pose prediction module 104 connected in sequence; the data preprocessing module 101 is used to segment cone-beam computed tomography (CBCT) images to obtain binary segmentation results; and to reconstruct the binary segmentation results to obtain a three-dimensional tooth model; and then to obtain a tooth point cloud through the three-dimensional tooth model; the feature extraction module 102 includes a convolutional network with an attention mechanism for extracting tooth features; the tooth features include global feature vectors of the dentition and local feature vectors of the teeth; the feature propagation module 103 aggregates and updates the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; the pose prediction module 104 includes an iterative optimization module 1041 for iteratively reducing the error of the pose prediction result to obtain a target pose result.

[0050] It should be noted that the input to the feature extraction module 102 can be not only tooth point clouds, but also triangular mesh models. The feature extraction module 102 of this application can be extended by combining it with the MeshSegNet network, which has achieved excellent results in existing tooth segmentation tasks. The input to the feature extraction module 102 of this application is a single modality, namely, tooth point clouds. As another embodiment, the input to the feature extraction module 102 can also be changed to a multimodal parallel form, including raw cone-beam computed tomography (CBCT) images, intraoral scan images, and other types of 3D tooth models, such as triangular mesh models. The network can thus learn tooth feature extraction in parallel from multiple branches. Including raw CBCT images and intraoral scan images facilitates end-to-end learning and can improve the workflow for automated tooth alignment in clinical applications.

[0051] In some embodiments, such as Figure 4 As shown, the pose prediction module also includes a feature transformation module 1042; the feature transformation module 1042 embeds and maps tooth features into Lie group rotation and three-dimensional translation, calculates the tooth pose transformation matrix; then reads the three-dimensional model of the tooth according to the serialized encoding of each tooth, loads the pose transformation onto the corresponding tooth, and moves and optimizes the tooth pose.

[0052] The automatic tooth alignment method of this application improves the network's ability to extract and propagate tooth point cloud features by constructing a graph attention mechanism module and a graph propagation module based on a dynamic graph recommendation method. Furthermore, the iterative optimization module 1041 in the pose prediction module 104 can gradually reduce the pose prediction error during network training by using previous pose prediction values, thus improving the algorithm's robustness to imbalanced training set samples. Secondly, for cases of tooth loss requiring subsequent denture wearing or implantation, this invention allows for the pre-insertion of a 3D model of a virtual denture into the original input, which is then removed from the output dentition after tooth alignment, effectively reserving space for dentures.

[0053] The automatic tooth alignment method and deep learning network provided in this application have been trained and tested on a tooth alignment dataset. The accuracy of the output results uses performance metrics consistent with existing deep learning-based automatic tooth alignment methods, including the mean distance error (ADD) between the predicted tooth pose and the gold standard for artificial tooth alignment, and the translation error (ΔT). avg The methods used include rotation error Δθavg and cosine similarity (CSA) to comprehensively represent movement error. Experimental results show that the accuracy of the automatic tooth alignment results output by the deep learning network provided in this application is superior to existing automatic tooth alignment methods.

[0054] Based on the above technical solutions, embodiments of this application provide an automatic tooth alignment method, apparatus, device, and storage medium based on point cloud understanding. The method includes: acquiring tooth point clouds based on cone-beam computed tomography (CBCT) images; the tooth point clouds include 3D point cloud data of complete teeth containing roots; extracting tooth features based on the tooth point cloud data and a graph convolutional network with an attention mechanism; the tooth features include global feature vectors of the dentition and local feature vectors of the teeth; aggregating and updating the tooth features based on a dynamic graph recommendation method to obtain a pose prediction result; and optimizing the pose prediction result based on the aggregated and updated tooth features and an iterative optimization method to obtain a target pose result. The automatic tooth alignment method provided in this application improves the accuracy of the automatic tooth alignment results by extracting tooth features through a graph convolutional network with an attention mechanism, aggregating and updating the tooth features through a dynamic graph recommendation method, and optimizing the pose prediction result through an iterative optimization method. In addition, for cases of missing teeth that require subsequent dentures or implants, the automatic tooth arrangement method provided in this application can pre-insert a three-dimensional model of a virtual denture into the original input, and then delete it from the output dentition after the tooth arrangement is completed, thereby reserving space for dentures.

[0055] This application proposes an automatic tooth alignment method based on point cloud understanding. Compared with existing similar methods, the automatic tooth alignment method of this application improves the network's ability to extract and propagate tooth point cloud features by constructing a graph attention mechanism module and a graph propagation module based on a dynamic graph recommendation method. The network is trained using 3D point cloud data of complete teeth containing tooth roots obtained by cone-beam computed tomography image segmentation, overcoming the defect of similar methods that do not consider tooth root pose. An iterative optimization method is introduced to optimize the tooth pose in the automatic tooth alignment result, improving the algorithm's robustness to imbalanced training set samples.

[0056] refer to Figure 5 Another embodiment of this application provides an electronic device, including: at least one processor 110; and a memory 111 communicatively connected to the at least one processor; wherein the memory 111 stores instructions executable by the at least one processor 110, the instructions being executed by the at least one processor 110 to enable the at least one processor 110 to perform any of the above method embodiments.

[0057] The memory 111 and processor 110 are connected via a bus, which may include any number of interconnecting buses and bridges, connecting various circuits of one or more processors 110 and memory 111. The bus may also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 110 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 110.

[0058] Processor 110 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 111 can be used to store data used by processor 110 during operation.

[0059] Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0060] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0061] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. An automatic tooth alignment method based on point cloud understanding, characterized in that, include: Tooth point cloud acquisition based on cone-beam computed tomography images; The tooth point cloud includes three-dimensional point cloud data of a complete tooth containing the tooth root; Based on the tooth point cloud data and a graph convolutional network with an attention mechanism, tooth features are extracted; the tooth features include global feature vectors of the dental arch and local feature vectors of the teeth. Based on the dynamic graph recommendation method, the tooth features are aggregated and updated to obtain the pose prediction result; Based on the aggregated and updated tooth features and the iterative optimization method, the pose prediction results are optimized to obtain the target pose result; The dynamic graph-based recommendation method aggregates and updates the tooth features to obtain pose prediction results, including: The tooth features are serialized and encoded, and a dynamic graph is constructed based on the adjacency relationship of the teeth; Each tooth is assigned an embedding node capable of network learning. After global pooling of the dynamic graph, it is input into the dynamic graph recommendation network to aggregate and update the tooth features, thereby obtaining the pose prediction result. The method of optimizing the pose prediction result based on the aggregated and updated tooth features and the iterative optimization method to obtain the target pose result includes: Based on the aggregated and updated tooth features, the pose transformation matrix is ​​obtained; The three-dimensional model of each tooth is obtained by serializing the code of each tooth, and the pose transformation is applied to the corresponding tooth to adjust the tooth position. Based on the iterative optimization method, the error of the pose prediction result is iteratively reduced to obtain the target pose result; The pose transformation matrix is ​​calculated by using a multilayer perceptron to embed the features of the nodes into Lie group rotations and three-dimensional translations.

2. The automatic tooth alignment method based on point cloud understanding according to claim 1, characterized in that, The tooth features are aggregated and updated using equation (1): （1） in, For each embedded node; This represents information from neighboring nodes, where k is the update time step. Both Update and Aggregate are differentiable functions. Aggregate aggregates information from neighboring nodes, while Update responds to the information and updates the embedded nodes.

3. The automatic tooth alignment method based on point cloud understanding according to claim 1, characterized in that, The method based on iterative optimization, which iteratively reduces the error of the pose prediction result to obtain the target pose result, includes: The pose prediction value in the previous iteration is regarded as an approximate estimate of the target pose. After obtaining the error using the iterative nearest point method, the error is fed back into the network as part of the input for the next iteration, and the network parameters are iteratively optimized.

4. An automatic tooth alignment device, the device being used to implement the automatic tooth alignment method based on point cloud understanding as described in any one of claims 1 to 3, characterized in that, include: The data preprocessing module, feature extraction module, feature propagation module, and pose prediction module are connected in sequence. The data preprocessing module is used to segment the cone-beam computed tomography (CBCT) image to obtain a binary segmentation result; and to reconstruct the binary segmentation result to obtain a three-dimensional tooth model; and then to obtain a tooth point cloud from the three-dimensional tooth model. The feature extraction module includes a convolutional network with an attention mechanism for extracting tooth features; the tooth features include global feature vectors of the dental arch and local feature vectors of the teeth. The feature propagation module aggregates and updates the tooth features based on the dynamic graph recommendation method to obtain the pose prediction result; The pose prediction module includes an iterative optimization module, which iteratively reduces the error of the pose prediction result to obtain the target pose result.

5. The automatic tooth-aligning device according to claim 4, characterized in that, The pose prediction module also includes a feature conversion module; The feature transformation module embeds and maps the tooth features into Lie group rotation and three-dimensional translation, and calculates the tooth pose transformation matrix. Then, based on the serialized encoding of each tooth, the 3D model of the tooth is read, the pose transformation is loaded onto the corresponding tooth, and the tooth pose is moved and optimized.

6. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the automatic tooth alignment method based on point cloud understanding as described in any one of claims 1 to 3.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the automatic tooth alignment method based on point cloud understanding as described in any one of claims 1 to 3.

Images