Knowledge graph-based orthodontic reasoning method
By constructing an orthodontic geometry map and utilizing the U-ViT architecture and graph neural networks, the problem of opaque reasoning in existing digital orthodontic methods is solved, enabling interpretable tooth movement planning and improving the credibility and transparency of treatment plans.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2025-11-18
- Publication Date
- 2026-04-21
AI Technical Summary
Existing digital orthodontic methods lack an interpretable, structured reasoning framework, resulting in opaque and unreliable tooth movement planning outcomes. They fail to provide step-by-step reasoning support, undermining clinicians' confidence and trust.
We construct an orthodontic reasoning method based on geometric knowledge chains. By building an orthodontic geometric atlas containing tooth entities and spatial relationships, we use a diffusion model based on the U-ViT architecture for latent space embedding and perform structured association learning through graph neural networks to generate tooth pose change paths that conform to biomechanical laws.
It enables interpretable and structured orthodontic treatment planning, improves the transparency and clinical credibility of orthodontic plans, and generates intermediate pose sequences that conform to biomechanical principles.
Smart Images
Figure CN121905463A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital treatment planning technology in orthodontics, and more particularly to an orthodontic reasoning method and apparatus based on knowledge graphs. Background Technology
[0002] In the field of orthodontics, the orthodontic process is divided into multiple stages. Doctors adjust the archwire tension through regular follow-up visits, gradually guiding the teeth to their target positions. Traditional orthodontics heavily relies on the doctor's experience, lacking a comprehensive plan for the specific procedures and expected outcomes of each stage in the early stages. Currently, in the field of digital orthodontics, automatically and accurately inferring the final treatment goal and key intermediate stages from the initial dental and jaw state is a complex technical challenge. Existing deep learning-based methods lack an interpretable, structured reasoning framework, making it difficult to understand and model the complex geometric constraints and clinical rules governing tooth movement. This results in an opaque reasoning process, low reliability of results, and an inability to provide step-by-step reasoning evidence.
[0003] In the current field of digital orthodontics, there are schemes that use diffusion processes to describe orthodontic tooth movement planning, or that directly predict adaptive length tooth movement sequences based on Transformers, or that achieve automatic tooth movement planning based on optimization methods and solve using interior point methods. However, these methods fail to meet the needs of clinical practice because the treatment plans they generate lack transparent decision-making basis, thereby undermining clinicians' confidence and trust. Summary of the Invention
[0004] The present invention aims to at least partially solve one of the technical problems in the related art.
[0005] Therefore, the first objective of this invention is to propose a method for orthodontic reasoning based on geometric knowledge chains, comprising: S1, construct an orthodontic geometric atlas containing tooth entities and spatial relationships, transform the patient's three-dimensional dental point cloud data and pose information into structured entities, and define the spatial relationships between teeth as the semantic mode of the atlas. S2, extract entity features from orthodontic geometric atlas, and perform latent space embedding on entity features based on the diffusion model of U-ViT architecture; S3. Obtain the dental and jaw entity embeddings that are in the unified latent space but have not yet fully interacted in step S2. Perform structured association learning through graph neural networks, aggregate the relationship features between entities to generate association feature representations that integrate local and global geometric constraints, and construct an orthodontic geometric perception model. S4. Based on similarity retrieval and chain reasoning mechanism, a customized geometric knowledge chain is generated. The orthodontic geometric perception model is iteratively called to predict the embedded representation of the new patient's dental and jaw state, and the state nodes in the geometric knowledge chain are decoded to serialize and generate a tooth pose change path that conforms to the laws of biomechanics.
[0006] In one embodiment of the present invention, step S1 further includes: S11 defines the entities of the orthodontic geometry atlas. The three-dimensional triangular mesh model of the patient's jaw is preprocessed into high-precision point cloud data of the teeth. Each entity contains FDI tooth position number, centroid coordinates, surface curvature, normal vector, pose matrix and tooth type attributes. S12 constructs an attribute graph model containing semantic relationships between entities by using the contact point position vectors of adjacency relationships, the cusp-fossa correspondence matrix of occlusal relationships, the tilt angle tensor of axial relationships, and the dental arch classification labels of the relationships.
[0007] In one embodiment of the present invention, step S2 further includes: S21, extract the dental and jaw feature vectors of each entity in the orthodontic geometry from the entity attributes; S22 is constructed using a diffusion model based on the U-ViT architecture. Gaussian noise is gradually added to the entity feature vector through a Markov chain forward diffusion process until it reaches a pure noise state. S23: The probability distribution of entity features in the latent space is learned using the inverse denoising process, where the temporal step embedding and the tooth type conditional information are jointly input into the model for feature decoding.
[0008] In one embodiment of the present invention, step S3 further includes: S31, Defining attribute association graphs based on orthodontic geometric atlas ,in, V Let the set of nodes be the set of nodes in the set of nodes. Corresponding entities in the atlas, E This is a set of edges, where each edge corresponds to a defined spatial relationship between entities. S32, the association graph G is input into the multi-layer graph convolutional network. The message passing mechanism is executed through the multi-layer graph convolutional network (GCN). The features of each node are aggregated with the association features of its neighboring nodes in each layer, and finally the association feature representation that integrates local and global geometric constraints is generated, forming a complete orthodontic geometric perception model.
[0009] In one embodiment of the present invention, step S4 further includes: S41, When a new patient's three-dimensional jaw is input, repeat steps S1-S3 to construct the initial orthodontic geometry subgraph; S42: Based on the retrieved optimal geometric knowledge chain template, input the initial state of the new patient into the orthodontic geometric perception model for reasoning to obtain the geometric knowledge chain of the new patient; S43, the geometric knowledge chain of the new patient is input into the diffusion model for decoding to achieve feature reconstruction of each dental entity, and further analyze the three-dimensional pose rectangle of each tooth in the intermediate stage to form a traceable pose sequence of the intermediate process of orthodontic treatment.
[0010] In one embodiment of the present invention, the spatial relationships include adjacency, interlocking, and axial relationships.
[0011] In one embodiment of the present invention, step S11 further includes: preprocessing the three-dimensional dental triangular mesh model through segmentation, denoising and registration.
[0012] In one embodiment of the present invention, the dental feature vector includes geometric morphological features, pose features, and topological relationship features.
[0013] To achieve the above objectives, a second aspect of the present invention provides an apparatus for orthodontic reasoning based on a knowledge graph, comprising: The geometric atlas construction module is used to convert the patient's three-dimensional dental point cloud data and pose information into structured entities, and to define the adjacency, occlusion, axis and belonging relationship between teeth as the semantic mode of the atlas. The latent space embedding module is used to perform latent space embedding on the entity features of the orthogonal geometric map based on the diffusion model of the U-ViT architecture. It gradually adds Gaussian noise to the entity feature vector through the forward diffusion process until it becomes pure noise, and uses the reverse denoising process to learn the probability distribution of entity features in the latent space. The structured association learning module is used to perform structured association learning on the latent space embedding vector through a graph neural network, and aggregate the relationship features between entities to generate an association feature representation that integrates local and global geometric constraints. The geometric knowledge chain generation module is used to generate customized geometric knowledge chains based on similarity retrieval and chain reasoning mechanisms. It predicts the embedded representation of the new patient's dental and jaw state by iteratively calling the orthodontic geometric perception model, and decodes the state nodes in the geometric knowledge chain to serialize and generate a tooth pose change path that conforms to the laws of biomechanics.
[0014] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.
[0015] The methods, systems, and storage media of this invention enable interpretable and structured orthodontic treatment planning. By progressively reasoning through geometric knowledge chains, intermediate pose sequences that conform to biomechanical principles are generated, significantly improving the transparency and clinical credibility of orthodontic plans.
[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a knowledge graph-based orthodontic reasoning method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the module structure of a knowledge graph-based orthodontic reasoning method according to an embodiment of the present invention. Detailed Implementation
[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0020] The following description, with reference to the accompanying drawings, describes an oral orthodontic reasoning method and apparatus based on a knowledge graph, according to an embodiment of the present invention.
[0021] Example 1 Figure 1 This is a flowchart of an embodiment of the orthodontic reasoning method based on geometric knowledge chains according to the present invention.
[0022] like Figure 1 As shown, the orthodontic reasoning method based on geometric knowledge chains includes the following steps: S1. Construct an orthodontic geometric atlas containing tooth entities and spatial relationships. Transform the patient's three-dimensional dental point cloud data and pose information into structured entities and define the spatial relationships between teeth as the semantic pattern of the atlas.
[0023] Specifically, in step S1, constructing an orthodontic geometric atlas containing tooth entities and spatial relationships is the core foundation for the interpretable orthodontic reasoning of this patented technology. This step transforms the patient's three-dimensional jaw point cloud data and tooth pose information into structured entities, and defines the adjacency, occlusion, axis, and belonging relationships between teeth as semantic patterns of the atlas, thereby providing a computable and interpretable knowledge carrier for subsequent geometric perception modeling and reasoning.
[0024] At the technical implementation level, this step first uses a 3D dental triangular mesh model obtained from intraoral or CT scans as the input data source. Point cloud segmentation algorithms (such as deep learning-based semantic segmentation models) are used to extract the point cloud data of each tooth, and noise reduction processing (such as using Poisson reconstruction or statistical filtering methods) is performed to improve the accuracy of the geometric data. Subsequently, a registration algorithm (such as the ICP iterative nearest point algorithm) is used to unify all tooth point clouds into the dental coordinate system to obtain their initial pose matrix. This includes translation vectors. With rotation matrix It is used to describe the position and orientation of teeth in three-dimensional space.
[0025] At the parametric level, each tooth entity is assigned a unique identifier (such as the FDI tooth position number) and includes three types of attributes: geometric attributes (such as centroid) Surface curvature Normal vector ), pose attributes (such as The relationship between the teeth is defined based on the principles of oral anatomy and orthodontics. For example, the adjacency relationship describes the position and tightness of the contact points of the teeth in the dental arch, the occlusal relationship describes the cusp-fossa correspondence of the upper and lower jaw teeth, and the axial relationship describes the inclination angle of the long axis of the teeth (such as mesiodistal inclination). Lips and tongue tilted The affiliation is used to distinguish whether a tooth belongs to the maxillary arch or the mandibular arch.
[0026] At the application level, this graph construction process is suitable for the initial modeling stage of digital orthodontic treatment, providing structured input for subsequent tooth movement prediction and treatment path planning. By transforming unstructured scan data into a graph structure with clear semantic relationships, the system can more accurately simulate the geometric changes and mechanical constraints of teeth during treatment.
[0027] In terms of technical effectiveness, this step achieves an explicit representation of tooth entities and relationships within the dentition system through structured modeling, providing an interpretable and computable knowledge base for subsequent training and inference of the geometric perception model. The semantic patterns of the graph not only enhance the model's understanding of spatial dependencies between teeth but also provide clinicians with intuitive visualization and verification methods, thereby increasing the credibility and transparency of treatment plans.
[0028] Furthermore, S1 includes: S11 defines the entities in the orthodontic geometry atlas. The patient's 3D dentistry triangular mesh model is preprocessed into high-precision point cloud data of 32 teeth. Each entity contains FDI tooth position number, centroid coordinates, surface curvature, normal vector, pose matrix, and tooth type attributes. Specifically, in step S11, defining the entities of the orthodontic geometric atlas is a fundamental step in constructing the entire orthodontic knowledge graph. This step uses a 3D dental triangular mesh model obtained from a patient's intraoral scan or CT scan as the input data source. Preprocessing operations are performed first, including tooth segmentation, surface denoising, and coordinate registration. Tooth segmentation can employ a deep learning-based semantic segmentation model (such as 3D U-Net or PointNet++) to separate each of the 32 teeth into independent point cloud datasets, each containing approximately 10,000 to 20,000 points to ensure high fidelity of geometric details. The registration process uses the ICP (Iterative Closest Point) algorithm or the RANSAC method based on feature matching to align the point cloud data of each tooth to a unified dental coordinate system, ensuring its pose matrix... The accuracy.
[0029] Each tooth entity is assigned a unique FDI tooth position number in the atlas (e.g., 11 represents the right maxillary central incisor) and a set of structured attributes. Geometric attributes include centroid coordinates. Surface curvature (Calculated using Gaussian curvature and mean curvature), and the normal vector. This is used to describe the local geometric features of the tooth surface. The pose attribute is determined by the homogeneous transformation matrix. This indicates that it includes translation vectors. With rotation matrix Morphological attributes are used to characterize the position and orientation of teeth in three-dimensional space. Morphological attributes are classified by tooth type (such as incisors, canines, premolars, and molars), providing semantic support for subsequent geometric perception modeling.
[0030] The technical value of this step lies in transforming unstructured 3D dental data into a structured set of entities and attributes, providing a standardized data foundation for subsequent map construction, feature embedding, and inference mechanisms. Simultaneously, the extraction of high-precision point clouds and pose matrices ensures that the model possesses sufficient spatial awareness when dealing with tooth geometric constraints, thereby improving the accuracy and interpretability of overall orthodontic inference.
[0031] S12, Relationship Definition and Graph Pattern Construction: By using the contact point position vectors of adjacency relationships, the cusp-fossa correspondence matrix of occlusal relationships, the tilt angle tensor of axial relationships, and the dental arch classification labels of the relationships, an attribute graph model containing semantic relationships between entities is constructed.
[0032] Specifically, in step S12, relationship definition and graph schema construction are the core steps in orthodontic geometric graph structured modeling. This step constructs a graph schema with clinical semantics by defining various semantic relationships between tooth entities, thereby providing structured input for subsequent graph neural network modeling and knowledge chain reasoning.
[0033] At the technical implementation level, this step first defines four key inter-entity relationships based on the principles of oral anatomy and orthodontics: adjacency, occlusion, axial relationship, and possession. Adjacency describes the contact state between adjacent teeth in the dental arch, and its attributes include the three-dimensional position vector of the contact point. and contact tightness ,in Identifiers representing adjacent teeth. Occlusal relationships are used to model the cusp-fossa correspondences of the upper and lower jaw teeth in an occlusal state; their attributes are represented by the cusp-fossa correspondence matrix. Indicates matrix elements Indicates the first maxillary teeth and the first The first mandibular tooth in The first apex and the first The matching strength between individual pits. Axial relationship describes the tilt state of the tooth's long axis in space, and its properties are quantized by the tilt angle tensor. This indicates the inclination angles corresponding to the mesial, labiolingual, and vertical directions, respectively. The affiliation is determined by dental arch classification labels. This indicates the type of dental arch to which each tooth belongs.
[0034] In practice, these relationships are encoded using a graph structure, with each edge corresponding to a relationship type, and attribute vectors can be attached to enhance semantic expression. Once the graph pattern is constructed, it can be further instantiated into an attribute graph model and stored in a graph database, such as Neo4j or Apache Jena, to support efficient query and inference operations.
[0035] The technical value of this step lies in its ability to explicitly model the complex geometric and functional constraints of the dentition system by structurally defining the semantic relationships between teeth. This provides an interpretable and computable graph foundation for subsequent graph neural network training and knowledge chain reasoning. In practical applications, this graph model can be widely used for generating orthodontic treatment plans, predicting tooth movement paths, and optimizing personalized treatment plans.
[0036] S2 extracts entity features from the orthodontic geometric atlas and performs latent space embedding on the entity features based on the diffusion model of the U-ViT architecture.
[0037] Specifically, in some implementations, step S2 uses a diffusion model based on the U-ViT architecture to perform latent space embedding on the entity features of the orthodontic geometry map. The underlying principle is based on the forward diffusion and backward denoising process of the Denoising Diffusion Probabilistic Model (DDPM). The specific operation is as follows: First, multimodal feature vectors for each dental entity are extracted from the orthodontic geometry map, including geometric morphological features, pose features, and topological relationship features. Geometric morphological features are extracted from point cloud data using a PointNet network. Pose features are encoded into dense vectors from the tooth pose matrix (containing position and rotation information). Topological relationship features are aggregated using a graph convolutional network (GCN) to aggregate contextual information such as adjacency and occlusion relationships.
[0038] In terms of parameter settings, the U-ViT model adopts a hybrid architecture of UNet and Vision Transformer, where UNet is responsible for multi-scale feature extraction, while the Transformer module is used to model global dependencies. The diffusion process is typically set as follows: At each time step, the forward diffusion process gradually adds Gaussian noise, progressively destroying the entity feature vector from its initial state to a pure noise state. The noise addition process follows a Markov chain, and the noise covariance matrix at each step can be set as follows: ,in For time steps The corresponding noise standard deviation.
[0039] In the reverse denoising process, the model learns to progressively recover the original features from the noise. Its inputs include the noisy data at the current time step, the time step embeddings, and optional entity conditional information (such as tooth type). The model output is the added real noise, thus modeling the probability distribution of entity features in the latent space. By sampling from this distribution and performing the reverse process, a unified embedding vector in the latent space can be generated, providing a structured and semantically consistent feature representation for subsequent graph-based association learning.
[0040] This step plays a crucial role in the overall technical solution. Through latent space embedding, it preserves the geometric and pose information of the tooth entity, providing context-aware node features for subsequent Graph Neural Networks (GCNs). Simultaneously, its decoding capability verifies the completeness of the latent space representation, providing fundamental support for the reasoning and reconstruction of geometric knowledge chains. Technically, this method significantly improves the distinguishability and structural fidelity of dental features in the latent space, providing a solid data representation foundation for interpretable orthodontic reasoning.
[0041] Furthermore, S2 includes: S21, extract the dental and jaw feature vectors of each entity in the orthodontic geometry from the entity attributes; S22 is constructed using a diffusion model based on the U-ViT architecture. Gaussian noise is gradually added to the entity feature vector through a Markov chain forward diffusion process until it reaches a pure noise state. S23: The probability distribution of entity features in the latent space is learned using the inverse denoising process, where the temporal step embedding and the tooth type conditional information are jointly input into the model for feature decoding.
[0042] Specifically, in some implementations, step S22 involves constructing a diffusion embedding model based on the U-ViT architecture, using a denoised diffusion probability model (DDPM) to model and reconstruct the latent space of the dental and jaw entity feature vectors. The core technical principle of this step is to gradually add Gaussian noise to the dental and jaw features through the forward diffusion process of a Markov chain, ultimately transforming them into a pure noise state, thereby achieving latent space encoding of the features. Subsequently, through a reverse denoising process, the model learns to gradually recover the original dental and jaw features from the noise, completing the learning and modeling of the probability distribution in the latent space.
[0043] By combining the input time-step embedding with tooth type conditional information, the U-ViT model can more accurately model the feature evolution of teeth at different time steps, thereby learning a semantically structured distribution of tooth features in the latent space. This step provides a high-quality feature representation foundation for subsequent graph association learning and geometric knowledge chain reasoning, and is a key step in realizing interpretable and structured orthodontic reasoning.
[0044] S3. Obtain the dental and jaw entity embeddings that are in the unified latent space but have not yet fully interacted in step S2. Perform structured association learning through graph neural networks, aggregate the relationship features between entities to generate association feature representations that integrate local and global geometric constraints, and construct an orthodontic geometric perception model.
[0045] Specifically, in step S3, the structured association learning of the latent space embedding vectors through a graph convolutional network (GCN) is a key step in realizing the global geometric constraint modeling of the dentition system in this patented technology. This step aims to aggregate the relational features between entities (i.e., between teeth) to generate a relational feature representation that integrates local and global geometric constraints, thereby improving the model's predictive ability and interpretability for tooth movement sequences.
[0046] At the technical implementation level, this step first defines an attribute association graph as the input to the GCN based on the orthodontic geometric atlas constructed in S1. Each node in the graph corresponds to a tooth entity, and its node features are the latent space embedding vectors generated by the diffusion model of the U-ViT architecture in S2. Edges represent the spatial relationships between teeth, such as adjacency, occlusion, and axis, and the edge type can be used as semantic information input to the GCN model. In each layer of GCN, through the message passing mechanism, the feature vector of each node is aggregated with the information of its neighboring nodes, thereby gradually fusing the local geometric features of teeth with the dependencies of the global dental arch structure. Through the stacking of multiple layers of GCN, the model can learn the context-aware representation of teeth in the complex dentofacial system, realizing the structured modeling of geometric constraints between teeth.
[0047] In terms of parameter metrics, GCN models typically employ a multi-layered stacked structure, with 2 to 4 layers to balance model complexity and expressive power. The node feature dimension of each GCN layer is recommended to be 256 to 512 dimensions to ensure sufficient representation of the geometric and pose information of the teeth. Edge weights can be initialized based on the strength of the relationship; for example, the weight of adjacency relationships can be set to 1.0, while the weight of occlusion relationships can be set to 0.8 to reflect the difference in importance between different relationships. Furthermore, the graph regularization loss is used during training to constrain entities with strong relationships to maintain similar embedding representations in the latent space; its weight coefficient is typically set to 0.1 to 0.5 to avoid overfitting.
[0048] In terms of application scenarios, this step is widely used in the intermediate stages of orthodontic treatment prediction and motion planning. By using GCN to perform structured relational learning on the latent space embedding, the model can understand the relative position of teeth in the dental arch, occlusal relationships, and axial change trends, thus providing a structured and interpretable basis for decision-making in subsequent geometric knowledge chain reasoning. For example, when predicting the direction of movement of a tooth, the model not only considers the geometric features of the tooth itself, but also refers to the movement trends and changes in occlusal relationships of its adjacent teeth, ensuring that the generated treatment path conforms to biomechanical and clinical rules.
[0049] From a technical perspective, this step utilizes graph neural networks to model the complex geometric relationships between teeth, significantly enhancing the model's ability to understand the overall structure of the dentofacial system. Its output relational feature representation not only preserves the geometric and pose information of individual teeth but also integrates their topological dependencies within the dental arch, providing a structured and semantically rich feature foundation for subsequent chain-like reasoning and treatment path generation. This process effectively improves the interpretability and prediction accuracy of the orthodontic reasoning model and is a core step in achieving the transition from "black-box prediction" to "structured reasoning."
[0050] Furthermore, S3 includes: S31, Defining attribute association graphs based on orthodontic geometric atlas ,in, V Let the set of nodes be the set of nodes in the set of nodes. Corresponding entities in the atlas, E This is a set of edges, where each edge corresponds to a defined spatial relationship between entities. S32, the association graph G is input into the multi-layer graph convolutional network. The message passing mechanism is executed through the multi-layer graph convolutional network (GCN). The features of each node are aggregated with the association features of its neighboring nodes in each layer, and finally the association feature representation that integrates local and global geometric constraints is generated, forming a complete orthodontic geometric perception model.
[0051] Specifically, in step S32, this invention employs a multi-layer Graph Convolutional Network (GCN) to learn structured relational features of entities in the orthodontic geometric graph. The core of this step lies in aggregating neighbor information of nodes in each layer of graph convolution through a message-passing mechanism, thereby generating a relational feature representation that integrates local and global geometric constraints. This method not only preserves the inherent features of tooth entities but also enhances the model's understanding of the overall structure of the dentition system by modeling the spatial dependencies between teeth through graph structure.
[0052] This step plays a crucial role in the overall technical solution, bridging the gap between previous and subsequent steps. On one hand, it receives the latent space embeddings generated by the diffusion model in S2 as node features; on the other hand, it provides structured and semantically relevant feature representations for geometric knowledge chain reasoning in S4. Through iterative aggregation of multi-layer GCNs, the model can gradually fuse local tooth features with global dental arch structure information, thereby constructing an orthodontic reasoning model with geometric awareness capabilities, significantly improving the interpretability and accuracy of treatment path prediction.
[0053] S4, as Figure 2As shown, a customized geometric knowledge chain is generated based on similarity retrieval and chain reasoning mechanism. The orthodontic geometric perception model is iteratively invoked to predict the embedded representation of the new patient's dentition state, and the state nodes in the geometric knowledge chain are decoded to serialize and generate a tooth movement path that conforms to the laws of biomechanics.
[0054] Specifically, generating a customized geometric knowledge chain based on similarity retrieval and chain-based reasoning mechanisms is the core component of this patented technology for achieving interpretable orthodontic treatment path planning. This step involves interacting with the embedded representation of a new patient's dental and jaw features and a trained orthodontic geometric perception model to progressively predict the embedded representation of the patient's intermediate states during treatment, thereby constructing a tooth movement path that conforms to biomechanical principles.
[0055] In some implementations, this step first transforms the new patient's 3D dental point cloud data into a structured orthodontic geometric atlas through processes S1 to S2, and then encodes it into an embedding vector in a unified latent space using a diffusion model based on the U-ViT architecture. Subsequently, the system performs a similarity search between this embedding vector and multiple reference cases stored in a parametric knowledge base. Specifically, measures such as cosine similarity or Euclidean distance can be used to calculate the similarity of the dental feature embeddings between the new patient and the reference cases, selecting the one with the highest similarity. One reference case was used as a candidate template.
[0056] Furthermore, the system extracts the corresponding treatment paths from these reference cases, namely a series of ordered dental and jaw state embeddings, as templates for the geometric knowledge chain. Based on these templates, the model iteratively invokes the orthodontic geometric perception model to predict the optimal next-state embedding for a new patient in the current state. In each step of inference, the model aims to minimize the joint loss function, ensuring that the predicted embeddings are consistent with clinical rules in terms of geometric shape, pose changes, and graph structure relationships.
[0057] The technical value of this step lies in transforming the previously black-box process of predicting tooth movement into an interpretable and verifiable sequential reasoning process by introducing a chain-like reasoning mechanism. Simultaneously, by combining similarity retrieval, the model can draw upon treatment paths from successful historical cases, thereby improving the clinical rationality and personalized adaptability of the predictions. In practical applications, this method can be widely used for the automatic generation and decision support of digital orthodontic treatment plans, providing doctors with intermediate-stage tooth pose sequences based on clear reasoning.
[0058] The orthodontic reasoning method based on geometric knowledge chains in this invention can achieve transparency and structured reasoning in the orthodontic treatment process, automatically generate intermediate tooth pose sequences that conform to biomechanical laws, and improve the interpretability and clinical credibility of the treatment plan.
[0059] Furthermore, S4 includes: S41, When a new patient's three-dimensional jaw is input, repeat steps S1-S3 to construct the initial orthodontic geometry subgraph; S42: Based on the retrieved optimal geometric knowledge chain template, input the initial state of the new patient into the orthodontic geometric perception model for reasoning to obtain the geometric knowledge chain of the new patient; S43, the geometric knowledge chain of the new patient is input into the diffusion model for decoding to achieve feature reconstruction of each dental entity, further analyze the pose rectangle of each tooth in the intermediate stage, and form a traceable pose sequence of the intermediate process of orthodontic treatment.
[0060] Specifically, in step S41, the system retrieves the K reference cases most similar to the new patient's dental and jaw features from the constructed orthodontic geometric knowledge base using a semantic similarity clustering algorithm. This step is a crucial link in the entire orthodontic inference framework for generating personalized treatment paths, and its technical implementation is based on structured dental and jaw maps and feature representations in a unified latent space.
[0061] In some implementations, the retrieval process first preprocesses the 3D dental model of the new patient, including point cloud segmentation, registration, and pose matrix extraction, to construct its corresponding orthodontic geometric subgraph. Subsequently, using a diffusion model based on the U-ViT architecture trained in S2, the dental entity features in this subgraph are encoded into a unified latent space to obtain its embedded representation. This embedded vector preserves the geometric morphology, pose relationships, and topological context information of the teeth, providing a semantically consistent feature space for subsequent similarity calculations.
[0062] Specifically, the system uses cosine similarity or Euclidean distance as the metric for clustering and retrieval, matching the embedding vectors of all encoded dental entities in the knowledge base. In some implementations, weighted graph structural similarity is optionally introduced, which compares not only entity features but also the degree of matching in terms of adjacency, occlusion, and other relationships to improve the clinical relevance of the retrieval. The parameter K is typically set to 5 to 10 to ensure a balance between diversity and representativeness in the retrieval results.
[0063] Furthermore, the retrieved K reference cases will serve as candidate templates for subsequent geometric knowledge chain construction. Each reference case contains its corresponding treatment path, i.e., an embedded representation of the dental and jaw states of a series of intermediate stages. By selecting the reference case that best matches the characteristics of the new patient, the system can provide the new patient with a clinically feasible initial reasoning path, thereby significantly improving the credibility and interpretability of the prediction results.
[0064] The technical value of this step lies in achieving an efficient mapping from unstructured dental data to inferable treatment paths through the combination of a structured knowledge base and a semantic embedding space. Its technical effectiveness is reflected in: a retrieval accuracy (Recall@K) exceeding 85%, and a retrieval response time controlled within 500ms, providing high-quality initial input for subsequent chained reasoning and tooth movement path prediction.
[0065] The orthodontic reasoning method based on geometric knowledge chains in this invention can achieve transparency and structured reasoning in the orthodontic treatment process, automatically generate intermediate tooth pose sequences that conform to biomechanical laws, and improve the interpretability and clinical credibility of the treatment plan.
[0066] Example 2 The device for orthodontic reasoning based on geometric knowledge chains proposed in this invention includes: The geometric atlas construction module is used to transform the patient's three-dimensional dental point cloud data and pose information into structured entities, and to define the adjacency, occlusion, axis and belonging relationship between teeth as semantic patterns of the atlas.
[0067] The latent space embedding module is used to perform latent space embedding of the entity features of the orthodontic geometric atlas based on the diffusion model of the U-ViT architecture. Through the forward diffusion process, Gaussian noise is gradually added to the entity feature vector until it is in a pure noise state. The reverse denoising process is used to learn the probability distribution of entity features in the latent space. The dental feature vector includes geometric morphological features, pose features and topological relationship features.
[0068] The structured association learning module is used to perform structured association learning on the latent space embedding vectors through graph neural networks, and aggregate the relationship features between entities to generate an association feature representation that integrates local and global geometric constraints.
[0069] The geometric knowledge chain generation module is used to generate customized geometric knowledge chains based on similarity retrieval and chain reasoning mechanisms. It predicts the embedded representation of the new patient's dental and jaw state by iteratively calling the orthodontic geometric perception model, and decodes the state nodes in the geometric knowledge chain to serialize and generate a tooth pose change path that conforms to the laws of biomechanics.
[0070] Furthermore, the geometry mapping module preprocesses the 3D dental triangular mesh model through segmentation, denoising, and registration.
[0071] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described orthodontic reasoning method based on geometric knowledge chains.
[0072] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0073] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A method for orthodontic reasoning based on geometric knowledge chains, characterized in that, include: S1, construct an orthodontic geometric atlas containing tooth entities and spatial relationships, transform the patient's three-dimensional dental point cloud data and pose information into structured entities, and define the spatial relationships between teeth as the semantic mode of the atlas. S2 extracts the dental and jaw features of entities in the orthodontic geometry map and performs unified latent space embedding on the entity features based on the diffusion model of the U-ViT architecture. S3. Obtain the dental and jaw entity embeddings that are in the unified latent space but have not yet fully interacted in step S2. Perform structured association learning through graph neural networks, aggregate the relationship features between entities to generate association feature representations that integrate local and global geometric constraints, and construct an orthodontic geometric perception model. S4. Based on similarity retrieval and chain reasoning mechanism, a customized geometric knowledge chain is generated. The orthodontic geometric perception model is iteratively called to predict the embedded representation of the new patient's dental and jaw state, and the state nodes in the geometric knowledge chain are decoded to serialize and generate a tooth pose change path that conforms to the laws of biomechanics.
2. The method as described in claim 1, characterized in that, Step S1 further includes: S11 defines the entities of the orthodontic geometry atlas. The three-dimensional triangular mesh model of the patient's jaw is preprocessed into high-precision point cloud data of the teeth. Each entity contains FDI tooth position number, centroid coordinates, surface curvature, normal vector, pose matrix and tooth type attributes. S12 constructs an attribute graph model containing semantic relationships between entities by using the contact point position vectors of adjacency relationships, the cusp-fossa correspondence matrix of occlusal relationships, the tilt angle tensor of axial relationships, and the dental arch classification labels of the relationships.
3. The method as described in claim 1, characterized in that, Step S2 further includes: S21, extract the dental and jaw feature vectors of each entity in the orthodontic geometry from the entity attributes; S22 is constructed using a diffusion model based on the U-ViT architecture. Gaussian noise is gradually added to the entity feature vector through a Markov chain forward diffusion process until it reaches a pure noise state. S23: The probability distribution of entity features in the latent space is learned using the inverse denoising process, where the temporal step embedding and the tooth type conditional information are jointly input into the model for feature decoding.
4. The method as described in claim 1, characterized in that, Step S3 further includes: S31, Defining attribute association graphs based on orthodontic geometric atlas ,in, V Let the set of nodes be the set of nodes in the set of nodes. Corresponding entities in the atlas, E This is a set of edges, where each edge corresponds to a defined spatial relationship between entities. S32, the association graph G is input into the multi-layer graph convolutional network. The message passing mechanism is executed through the multi-layer graph convolutional network (GCN). The features of each node are aggregated with the association features of its neighboring nodes in each layer, and finally the association feature representation that integrates local and global geometric constraints is generated, forming a complete orthodontic geometric perception model.
5. The method as described in claim 1, characterized in that, Step S4 further includes: S41, When a new patient's three-dimensional jaw is input, repeat steps S1-S3 to construct the initial orthodontic geometry subgraph; S42: Based on the retrieved optimal geometric knowledge chain template, input the initial state of the new patient into the orthodontic geometric perception model for reasoning to obtain the geometric knowledge chain of the new patient; S43 inputs the geometric knowledge chain of the new patient into the diffusion model for decoding to achieve feature reconstruction of each dental entity, further analyzes the three-dimensional pose rectangle of each tooth in the intermediate stage, and forms a traceable pose sequence of the intermediate process of orthodontic treatment.
6. The method as described in claim 1, characterized in that, The spatial relationships include adjacency, interlocking, and axial relationships.
7. The method as described in claim 1, characterized in that, Step S11 further includes: preprocessing the three-dimensional dental triangular mesh model through segmentation, denoising, and registration.
8. The method as described in claim 3, characterized in that, The dental feature vector includes geometric morphological features, pose features, and topological relationship features.
9. A device for orthodontic reasoning based on knowledge graphs, characterized in that, include: The geometric atlas construction module is used to convert the patient's three-dimensional dental point cloud data and pose information into structured entities, and to define the adjacency, occlusion, axis and belonging relationship between teeth as the semantic mode of the atlas. The latent space embedding module is used to perform latent space embedding on the entity features of the orthogonal geometric map based on the diffusion model of the U-ViT architecture. It gradually adds Gaussian noise to the entity feature vector through the forward diffusion process until it becomes pure noise, and uses the reverse denoising process to learn the probability distribution of entity features in the latent space. The structured association learning module is used to perform structured association learning on the latent space embedding vector through a graph neural network, and aggregate the relationship features between entities to generate an association feature representation that integrates local and global geometric constraints. The geometric knowledge chain generation module is used to generate customized geometric knowledge chains based on similarity retrieval and chain reasoning mechanisms. It predicts the embedded representation of the new patient's dental and jaw state by iteratively calling the orthodontic geometric perception model, and decodes the state nodes in the geometric knowledge chain to serialize and generate a tooth pose change path that conforms to the laws of biomechanics.
10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as claimed in any one of claims 1-8.