Automatic tooth arrangement method and system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent tooth arrangement, particularly discloses an automatic tooth arrangement method and system based on artificial intelligence, and aims to solve the problems that existing AI automatic tooth arrangement lacks biomechanics and dynamic function constraints, and efficiency and precision are difficult to consider at the same time. According to the method, the enhanced condition feature vector is generated by fusing the three-dimensional scanning data of the oral cavity of the patient, the jaw position relation record and biomechanics and mandibular motion parameters, the initial tooth arrangement scheme is output through the pre-training condition generation model, and the differentiable static mechanics and dynamic function agent model is called to obtain the quantitative evaluation index. And a multi-objective total objective function is constructed, and through gradient descent iterative optimization, overall coordination after man-machine cooperation is maintained in combination with an interactive local re-optimization mechanism. According to the method, the tooth arrangement scheme can be upgraded from reasonable form to mechanical reliability and dynamic coordination, the design efficiency and clinical safety are greatly improved, the AI decision interpretability is enhanced, and deep integration of man-machine intelligence can be helped.

Description

Technical Field

[0001] This invention relates to the field of intelligent tooth alignment technology, and in particular to an automatic tooth alignment method and system based on artificial intelligence. Background Technology

[0002] With the rapid development of digital dental technology, AI-based automated tooth alignment methods have gradually become an important technical means in clinical dental restoration. Existing automated tooth alignment methods mostly rely on deep learning models, such as convolutional neural networks and generative adversarial networks. By learning from a large amount of tooth alignment case data from professional technicians, they can automatically generate tooth arrangement plans for edentulous patients. This type of method significantly reduces the reliance on clinicians' manual tooth alignment experience, improves the efficiency of denture design, and shortens the design cycle.

[0003] However, existing AI-based automated tooth arrangement methods still have significant technical limitations. The core issue lies in their purely data-driven nature, which leads to insufficient integration of physical and biomechanical principles, making it difficult to balance the long-term clinical safety and dynamic functionality of the tooth arrangement plan. On the one hand, existing methods only generate tooth arrangement plans that meet visual rationality in terms of geometric shape, without rigorous occlusal mechanics analysis. They cannot accurately assess the distribution of occlusal forces on the alveolar ridge mucosa, easily resulting in stress concentration areas. Long-term use of such dentures can accelerate local alveolar bone resorption and increase the risk of denture base fracture. If high-fidelity finite element analysis is incorporated to ensure mechanical rationality, the extremely high computational cost will significantly extend the design cycle, failing to meet the needs of immediate clinical design. Furthermore, existing AI methods mostly optimize designs under static jaw relationships, neglecting the dynamic functional balance requirements when the mandible performs protrusion, lateralization, and other masticatory movements. The generated tooth arrangement plan is prone to functional interference during dynamic occlusion, thereby affecting the patient's masticatory efficiency and potentially harming the health of the temporomandibular joint.

[0004] Therefore, there is an urgent need for automated tooth alignment methods and systems based on artificial intelligence to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic tooth alignment method based on artificial intelligence, comprising the following steps: The patient's oral cavity three-dimensional scan data and jaw position relationship record are obtained. The geometric features and neutral zone information of the alveolar ridge are extracted through a feature extraction network. The geometric features are then fused with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. The enhanced conditional feature vector is input into a pre-trained conditional generation model, which outputs an initial three-dimensional tooth arrangement scheme. The system invokes a pre-trained, differentiable static mechanical proxy model and a dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the dental arch and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. Based on the initial three-dimensional tooth arrangement scheme, the translation and rotation parameters of each tooth are used as optimization variables to construct the overall objective function. The gradient of the overall objective function with respect to the tooth position parameters is calculated by utilizing the differentiability of the static mechanical proxy model and the dynamic functional proxy model. The tooth position parameters are then iteratively updated through the optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. In response to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, the parameters of the adjusted tooth are locked, and based on the updated overall constraints, the parameters of the unlocked tooth are automatically re-optimized locally to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

[0006] Furthermore, the step of fusing the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector includes: A three-dimensional scanning mesh model of the patient's oral cavity in an edentulous state was obtained, along with jaw position relationship data of the relative positions of the upper and lower jaws. The three-dimensional scanning mesh model was then cleaned and repaired to obtain a standardized digital model of the dentition. The standardized digital model of the teeth and jaws is processed by a three-dimensional convolutional neural network to automatically identify and locate key anatomical landmarks. Based on the identified anatomical landmarks, geometric feature vectors describing the shape of the dental arch and the condition of the alveolar bone are calculated. At the same time, by analyzing the surface curvature of the model and combining it with biomechanical principles, the mechanical equilibrium region where the teeth should be arranged, i.e., the neutral zone information, is estimated. From a pre-set clinical knowledge base, a set of biomechanical material parameters adapted to the characteristics of the patient's oral tissues is matched. The set of parameters includes parameters characterizing mucosal elasticity and grade coefficients characterizing bone density. Based on the individual patient condition or general standards, the motion characteristic parameters of the mandible during protrusion and lateral movements are determined. The calculated geometric feature vector, the neutral zone information, the biomechanical material parameter set, and the mandibular motion feature parameters are concatenated and combined, and then processed by a neural network layer for feature fusion and dimensionality reduction. Finally, a fixed-dimensional enhanced conditional feature vector containing multi-source constraint information is output.

[0007] Furthermore, the step of inputting the enhanced conditional feature vector into a pre-trained conditional generation model to output an initial three-dimensional tooth alignment scheme includes: The conditional generation model is a neural network model built on conditional generative adversarial networks, configured to generate the spatial position and pose of all teeth based on the input enhanced conditional feature vector. The enhanced conditional feature vector is used as a generation condition and input into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. Based on a pre-set standardized artificial tooth 3D model database, the digital model of each standard tooth is spatially transformed according to the position and posture parameters corresponding to the initial parameter matrix, and placed in the virtual denture base coordinate system to assemble a visualized 3D initial tooth arrangement plan containing a complete dentition.

[0008] Furthermore, the step of outputting the evaluation index of dynamic occlusion status includes: The static mechanics surrogate model is a differentiable feedforward deep neural network that is trained under supervision using a large amount of offline finite element analysis simulation data. This static mechanics surrogate model is used to simulate the mechanical effects of tooth alignment on supporting tissues and to predict the concentration and uniformity of stress distribution. The dynamic functional surrogate model is a differentiable graph neural network that is trained under supervision using a large amount of offline mandibular motion dynamics simulation data. The dynamic functional surrogate model is used to quickly evaluate whether there is dynamic occlusal interference in the tooth arrangement scheme and whether there is sufficient balanced contact. The current tooth alignment parameters are input into the static mechanical proxy model in real time to obtain the predicted peak alveolar ridge stress and stress distribution uniformity evaluation value under this scheme. The current three-dimensional relationship of the dentition is input into the dynamic functional proxy model to obtain the predicted number of dynamic occlusal interference points and the number of effective balance contact points under this scheme.

[0009] Furthermore, the step of iteratively updating the tooth position parameters through an optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions includes: A total objective function is constructed, which is obtained by weighted summation of aesthetic loss terms, biomechanical loss terms, and functional loss terms. The weight coefficients of each term can be preset and adjusted according to the clinical focus. The loss values ​​for each sub-item are calculated. The aesthetic loss item is scored by a trained aesthetic evaluation network on visual elements such as the dental arch morphology and anterior aesthetic line of the current tooth arrangement scheme, and the scores are converted into loss values. The biomechanical loss item is calculated based on the stress peak value and stress distribution uniformity evaluation value. The functional loss item is calculated based on the number of dynamic occlusal interference points and the number of effective balance contact points. Using neural network automatic differentiation technology, the gradient of the total objective function with respect to the position parameters of each degree of freedom of each tooth is calculated through backpropagation. The gradient descent optimization algorithm is then used to iteratively update the position parameters of all teeth based on the calculated gradient direction and magnitude. Based on the updated tooth positions, the various losses are reassessed, and the parameters are updated again. When the value of the overall objective function drops below a preset threshold, the corresponding tooth arrangement scheme is output as the final optimized tooth arrangement scheme.

[0010] Furthermore, the steps of responding to a user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, locking the parameters of the adjusted teeth, and automatically performing local re-optimization on the parameters of the unlocked teeth based on the updated overall constraints include: The optimized tooth arrangement scheme is displayed in a three-dimensional interactive interface. The user's interactive operation on the tooth model in the scheme through the input device is continuously monitored. When the user executes the translation or rotation adjustment command on one or more teeth, the final position and angle parameters of these adjusted teeth are immediately recorded and fixed. Set the user-fixed tooth parameters as unchangeable variables, and set the remaining tooth parameters in the scheme that have not been adjusted by the user as optimizable variables. Keep the overall objective function unchanged, but limit its calculation scope to the new tooth arrangement scheme composed of the current optimizable variables and fixed variables. Starting with the user-adjusted overall solution, a local optimization process with a limited number of iterations is initiated. This process only performs gradient calculations and updates on the tooth parameters of the optimizable variables to fine-tune their positions and automatically finds a new, overall coordinated solution that makes the overall objective function better. Generate and output the final tooth alignment plan that integrates user preferences and system optimization.

[0011] Furthermore, the present invention also discloses an artificial intelligence-based automatic tooth alignment system, comprising: The acquisition module is used to acquire the patient's oral cavity three-dimensional scan data and jaw position relationship record. It extracts the geometric features and neutral zone information of the alveolar ridge through the feature extraction network, and fuses the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. The first output module is used to input the enhanced conditional feature vector into a pre-trained conditional generation model and output an initial three-dimensional tooth arrangement scheme. The second output module is used to call the pre-trained differentiable static mechanical proxy model and dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the dental arch and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. The optimization module is used to construct a total objective function based on the initial three-dimensional tooth arrangement scheme, with the translation and rotation parameters of each tooth as optimization variables. It uses the differentiability of the static mechanical proxy model and the dynamic functional proxy model to calculate the gradient of the total objective function with respect to the tooth position parameters, and iteratively updates the tooth position parameters through the optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. The update module is used to respond to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, lock the parameters of the adjusted teeth, and automatically perform local re-optimization on the parameters of the unlocked teeth based on the updated overall constraints, so as to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

[0012] Furthermore, the first output module includes: The construction unit is configured to generate the spatial position and orientation of all teeth based on the input enhanced conditional feature vector, since the conditional generation model is a neural network model built on a conditional generative adversarial network. A generation unit is used to input the enhanced conditional feature vector as a generation condition into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. The assembly unit is used to spatially transform the digital model of each standard tooth according to the position and orientation parameters in the initial parameter matrix based on the pre-set standardized artificial tooth three-dimensional model database, and place it in the virtual denture base coordinate system to assemble a visualized three-dimensional initial tooth arrangement plan containing a complete dentition.

[0013] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described artificial intelligence-based automatic tooth alignment method.

[0014] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described artificial intelligence-based automatic tooth alignment method.

[0015] The beneficial effects of this application are as follows: Firstly, this application addresses the lack of mechanical simulation guidance in existing technologies by constructing a differentiable static mechanical proxy model and embedding it into the AI ​​optimization process. This allows biomechanical constraints to be directly incorporated into the generation of the tooth arrangement plan, enabling accurate prediction of alveolar ridge stress distribution, reducing the risk of stress concentration, improving the physical reliability of denture design from the algorithmic source, reducing clinical risks such as alveolar bone resorption and denture base fracture caused by long-term use, and enhancing the scientific rationality of the plan.

[0016] Secondly, this application can overcome the shortcomings of existing technologies that neglect dynamic functional balance. By training a differentiable dynamic functional surrogate model, the mandibular motion trajectory parameters are incorporated into the optimization system to ensure that the tooth arrangement scheme has no harmful interference during the simulation of dynamic occlusion such as protrusion and lateral movement, and has sufficient balance contact points. This significantly improves the chewing efficiency of dentures and the comfort of patients, and protects the health of the temporomandibular joint.

[0017] Third, this application addresses the contradiction between mechanical simulation and design efficiency in existing technologies by using a lightweight, differentiable surrogate model trained offline to replace online high-fidelity finite element analysis and dynamic simulation. This reduces the physical calculations that originally took several hours to seconds, achieving a balance between physical law-guided optimization and real-time clinical design, and meeting the actual clinical needs for design efficiency. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a method flow proposed in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the system structure proposed in an embodiment of the present invention.

[0020] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] like Figure 1 As shown, this application provides an artificial intelligence-based automatic tooth alignment method, including the following steps: S1. Obtain the patient's oral cavity three-dimensional scan data and jaw position relationship record, extract the geometric features and neutral zone information of the alveolar ridge through the feature extraction network, and fuse the geometric features with the preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. S2, The enhanced conditional feature vector is input into a pre-trained conditional generation model to output an initial three-dimensional tooth arrangement scheme; S3, invoke the pre-trained differentiable static mechanical proxy model and dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the tooth row and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. S4. Based on the initial three-dimensional tooth arrangement scheme, and using the translation and rotation parameters of each tooth as optimization variables, a total objective function integrating aesthetic evaluation, biomechanical evaluation, and functional evaluation is constructed. Utilizing the differentiability of the static mechanical proxy model and the dynamic functional proxy model, the gradient of the total objective function with respect to the tooth position parameters is calculated. The tooth position parameters are iteratively updated through an optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. S5, responding to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, locking the parameters of the adjusted tooth, and automatically performing local re-optimization on the parameters of the unlocked tooth based on the updated overall constraints, so as to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

[0023] As described in steps S1-S5 above, in clinical dental restoration scenarios, the long-term safety and chewing efficiency of dentures depend on the biomechanical rationality of tooth arrangement and the dynamic occlusal coordination. In existing technologies, purely data-driven AI-based tooth arrangement methods only generate plans by learning from technician tooth arrangement data, without considering the stress distribution patterns of the alveolar ridge mucosa. This can easily lead to stress concentration, accelerating alveolar bone resorption or denture base fracture. Furthermore, it neglects occlusal balance during dynamic movements such as mandibular protrusion and lateralization, potentially causing chewing discomfort or temporomandibular joint problems. If real-time high-fidelity finite element analysis and dynamic simulation optimization are used, the computational load would be too large to meet the immediate clinical design requirements. Therefore, a tooth arrangement method that balances efficiency and multi-dimensional performance is urgently needed.

[0024] This application integrates multi-source oral data features, pre-trained condition generation models, differentiable physical and functional simulation proxy models, and multi-objective iterative optimization mechanisms, combined with interactive local re-optimization functions, to achieve intelligent automatic tooth arrangement that combines aesthetic harmony, biomechanical reliability, and dynamic functional balance. It solves the core problem of traditional AI tooth arrangement that only focuses on morphological rationality while neglecting physical laws and dynamic functions, and provides an efficient and scientific technical solution for the design of complete dentures and partial removable dentures.

[0025] The core logic and principle of this application is to achieve fully automated tooth arrangement from personalized feature extraction to multi-objective optimization and human-machine collaborative adjustment by integrating multi-source oral clinical data and multi-dimensional constraints, and using an intelligent framework that combines data-driven and physical simulation proxy models. First, it acquires the patient's three-dimensional oral cavity scan data and jaw position records, uses a feature extraction network to extract the geometric features of the alveolar ridge and estimate the neutral zone information, and then integrates preset biomechanical material parameters and mandibular motion trajectory parameters to generate an enhanced conditional feature vector containing multi-source constraints. This provides comprehensive personalized input for tooth arrangement plan generation. Subsequently, this feature vector is input into a pre-trained conditional generation model, which quickly outputs a reasonably shaped initial three-dimensional tooth arrangement plan. Then, it calls a pre-trained offline differentiable static mechanical proxy model and a dynamic functional proxy model to quickly... The system acquires alveolar ridge stress distribution and dynamic occlusal status assessment indicators to provide mechanical and functional quantitative basis for subsequent optimization. Then, based on the initial tooth arrangement plan, the system constructs a general objective function that integrates aesthetics, biomechanics, and functionality, using the translation and rotation parameters of each tooth as optimization variables. The gradient is calculated through backpropagation using the differentiability of the surrogate model, and the tooth position parameters are iteratively updated using the gradient descent optimization algorithm until the general objective function meets the preset convergence condition, resulting in a multi-dimensional performance-balanced optimized tooth arrangement plan. Finally, by responding to the user's manual adjustment of specific teeth, the system locks the adjusted tooth parameters and performs local re-optimization on the unlocked parameters. While respecting the user's subjective wishes, the system maintains the overall biomechanical and functional coordination of the plan, ultimately achieving intelligent automatic tooth arrangement that balances personalized adaptation, scientific reliability, and clinical applicability.

[0026] In one embodiment, the step of fusing the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector includes: S11, Obtain a three-dimensional scanning mesh model of the patient's oral cavity in an edentulous state, and record the jaw position relationship data of the relative positions of the upper and lower jaws. Clean and repair the data of the three-dimensional scanning mesh model to obtain a standardized digital model of the dentition. S12, using a three-dimensional convolutional neural network to process the standardized digital model of the teeth and jaws, automatically identifying and locating key anatomical landmarks, including a series of points on the alveolar ridge crest line, the center point of the incisor papilla, and the highest point of the bilateral molar retrocuspid pads. Based on the identified anatomical landmarks, a geometric feature vector describing the dental arch morphology and alveolar bone condition is calculated. The geometric feature vector includes the dental arch width, dental arch length, and relative height of the alveolar ridge. At the same time, by analyzing the surface curvature of the model and combining it with biomechanical principles, the mechanical equilibrium region where the teeth should be arranged, i.e., the neutral zone information, is estimated. S13, from the preset clinical knowledge base, match a set of biomechanical material parameters that are adapted to the characteristics of the patient's oral tissue. The set of parameters includes parameters that characterize mucosal elasticity and grade coefficients that characterize bone density. Based on the individual patient condition or general standards, determine the motion characteristic parameters of the mandible when performing protrusion and lateral movements. S14, the calculated geometric feature vector, the neutral zone information, the biomechanical material parameter set, and the mandibular motion feature parameters are concatenated and combined, and then processed by a neural network layer for feature fusion and dimensionality reduction, finally outputting a fixed-dimensional enhanced conditional feature vector containing multi-source constraint information.

[0027] As described in steps S11-S14 above, the oral cavity three-dimensional scanning data and jaw position relationship data are standardized, anatomical landmarks and geometric features are automatically extracted and neutral zone information is estimated, and appropriate biomechanical material parameters and mandibular movement feature parameters are matched. Then, through feature fusion and dimensionality reduction, an enhanced conditional feature vector containing multi-source constraint information is generated, providing input basis for the generation of the initial tooth arrangement plan.

[0028] The rationality of tooth alignment highly depends on a precise understanding of the patient's oral anatomy, tissue biomechanical properties, and mandibular movement patterns. These factors collectively constitute the core constraints of the tooth arrangement plan. Current technologies often limit feature extraction to single geometric information, and anatomical landmark identification frequently relies on manual annotation. This is not only inefficient and prone to errors, but also lacks a systematic fusion of biomechanical material parameters and mandibular movement trajectory parameters. Consequently, the generated feature vector information is one-sided and cannot provide comprehensive constraints for tooth arrangement generation, thus affecting the adaptability and functionality of the tooth arrangement plan. By employing automated data processing, multi-dimensional feature extraction, precise parameter matching, and efficient fusion and dimensionality reduction, an enhanced conditional feature vector containing multi-source constraints from anatomy, mechanics, and motion is formed, specifically addressing the shortcomings of traditional technologies.

[0029] This process involves acquiring a 3D scanning mesh model of the patient's edentulous jaw and recording the relative positions of the upper and lower jaws. The 3D scanning mesh model is obtained using a 3D oral scanner, and the jaw position data is acquired through a clinical jaw position measurement procedure. The 3D scanning mesh model undergoes data cleaning and repair to remove noise points, isolated points, and surface defects generated during the scanning process, resulting in a standardized digital model of the dentition with a complete topological structure and a smooth surface. This addresses the issue of inconsistent quality in the original scanning data, providing a unified and reliable data foundation for subsequent feature extraction and avoiding feature calculation biases caused by data defects.

[0030] A standardized digital model of the dental jaw was processed using a three-dimensional convolutional neural network (CNN). This CNN consisted of five convolutional layers and two max-pooling layers, with kernel sizes of 3×3×3 and 5×5×5, respectively. Deep spatial features of the model were extracted through layer-by-layer convolutional operations, enabling automatic identification and localization of key anatomical landmarks. These landmarks included a series of points on the alveolar ridge crest, the center point of the incisal papilla, and the highest points of the bilateral retromolar pads. Based on these precisely located anatomical landmarks, geometric feature vectors such as arch width, arch length, and relative height of the alveolar ridge were calculated using spatial distance calculations. These feature vectors objectively quantified the patient's dental arch morphology and alveolar bone condition. Simultaneously, by calculating the curvature values ​​of different regions on the model surface and combining this with the biomechanical principle of tooth force balance, the information of the mechanical equilibrium zone (neutral zone) where the teeth should be positioned was estimated. By using an automated neural network recognition method to replace the traditional method of manually marking anatomical landmarks, the efficiency and accuracy of landmark positioning are greatly improved. It also enables the simultaneous acquisition of geometric features and neutral zone information, providing a basic constraint for tooth arrangement plans that conforms to the patient's anatomical structure and mechanical balance requirements.

[0031] The system matches a set of biomechanical material parameters adapted to the patient's oral tissue characteristics from a pre-defined clinical knowledge base. This knowledge base stores a large amount of parameter data corresponding to different oral mucosal elasticity and bone mineral density grades. By inputting the patient's oral mucosal elasticity and bone mineral density test results, the corresponding parameters characterizing mucosal elasticity and grade coefficients characterizing bone mineral density can be obtained. Based on the patient's age, chewing habits, and other individual factors, or using a general standard of a 15° Bennett angle and a 30° protrusion condyle guide angle, the motion characteristic parameters of the mandible during protrusion and lateral movements are determined. This supplements the biomechanical and dynamic motion constraint information missing in traditional feature extraction, ensuring that the subsequently generated tooth arrangement plan not only adapts to anatomical morphology but also meets the requirements of mechanical load-bearing and dynamic movement, thus improving the functionality and applicability of the tooth arrangement plan.

[0032] The calculated geometric feature vectors, neutral zone information, biomechanical material parameter sets, and mandibular motion feature parameters are sequentially concatenated to ensure complete preservation of multi-source constraint information. The concatenated high-dimensional features are then input into a fully connected neural network layer with 128 neurons for feature fusion and dimensionality reduction. This neural network layer optimizes weight parameters through offline training, effectively uncovering the intrinsic relationships between different types of parameters, achieving deep fusion of multi-source information, and compressing the high-dimensional features to 256 dimensions, reducing the computational complexity of subsequent conditional generation models while retaining key constraint information. The final output fixed-dimensional enhanced conditional feature vector ensures both the comprehensiveness and synergy of multi-source constraint information and adaptability to subsequent pre-trained conditional generation models, achieving efficient transmission of constraint information.

[0033] In one embodiment, the step of inputting the enhanced conditional feature vector into a pre-trained conditional generation model to output an initial three-dimensional tooth alignment scheme includes: S21, The conditional generation model is a neural network model built on conditional generative adversarial network, configured to generate the spatial position and pose of all teeth according to the input enhanced conditional feature vector. The model has been pre-trained using a dataset containing a large number of professional tooth arrangement cases and has the ability to generate reasonable tooth arrangement according to oral conditions. S22, The enhanced conditional feature vector is used as the generation condition and input into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. S23, based on the pre-set standardized artificial tooth three-dimensional model database, the digital model of each standard tooth is spatially transformed according to the corresponding position and posture parameters in the initial parameter matrix, and placed in the virtual denture base coordinate system to assemble a visualized three-dimensional initial tooth arrangement plan containing a complete dentition.

[0034] As described in steps S21-S23 above, a pre-trained conditional generative adversarial network is used to quickly generate a three-dimensional initial tooth arrangement scheme that is morphologically reasonable and fits individual conditions, by taking an enhanced conditional feature vector containing multi-source constraints of patient oral anatomy, biomechanics and mandibular movement as input. This provides a reliable starting point for iterative optimization that integrates biomechanics and dynamic functions.

[0035] The initial tooth arrangement plan is the foundation of the entire automated tooth arrangement process. Its morphological rationality and individual adaptability directly determine the efficiency of subsequent optimization and the quality of the final plan. In clinical applications, the tooth arrangement plan must simultaneously adapt to the patient's alveolar ridge geometry, the biomechanical requirements of the neutral zone, and the mandibular movement patterns. If the initial plan deviates from these constraints, it will lead to excessive subsequent optimization iterations, increased computational costs, and even failure to converge to the optimal solution. In existing technologies, conditional generation models mostly rely on training with single geometric features and lack multi-source constraint fusion. The generated initial plans often have the problem of mismatch with the patient's oral conditions, requiring a lot of manual adjustments before entering the optimization stage, which seriously affects design efficiency. By using a conditional generative adversarial network pre-trained based on a large number of professional tooth arrangement cases, combined with enhanced conditional feature vectors containing multi-source constraints, the problems of poor adaptability and low generation efficiency of the initial plan are specifically solved, achieving rapid and accurate generation of the initial plan.

[0036] The conditional generation model is a neural network model built on conditional generative adversarial networks (GANs), with two core components: a generator and a discriminator. The generator consists of 6 convolutional layers and 3 fully connected layers, all with 3×3×3 kernels. The fully connected layers have 1024, 512, and 256 neurons respectively. It extracts features through layer-by-layer convolution and decodes them to output tooth arrangement parameters. The discriminator consists of 4 convolutional layers and 2 fully connected layers, used to compare the generated tooth arrangement scheme with real professional tooth arrangement schemes and output a rationality score. This model can be pre-trained using several sets of professional dental technician case datasets. Each dataset includes patient oral feature vectors, biomechanical parameters, mandibular movement parameters, and corresponding optimal tooth arrangement position and pose parameters. Through training, the model learns the mapping relationship between different oral conditions and reasonable tooth arrangement schemes, enabling it to generate morphologically reasonable tooth arrangements based on individual patient oral conditions. This provides model support with professional tooth arrangement knowledge for initial scheme generation, ensuring that the generated results meet the basic morphological requirements of clinical tooth arrangement.

[0037] The generator network of the conditional generation model uses enhanced conditional feature vectors as input to generate conditions. The generator network first performs dimensionality enhancement on the 256-dimensional enhanced conditional feature vectors, transforming them into high-dimensional feature maps. Subsequent convolutional layers extract deep features, which are then mapped to the tooth parameter space via fully connected layers. An initial parameter matrix is ​​inferred and generated in the implicit space. The initial parameter matrix is ​​a 28x6 matrix, with each row corresponding to a tooth and the six columns corresponding to three translation parameters and three rotation parameters. The translation parameters are in millimeters, and the rotation parameters are in degrees, describing the tooth's position and orientation in three-dimensional space. Through deep decoding of multi-source constraint features, the abstract individual patient constraint information is transformed into specific tooth alignment parameters, ensuring that the initial design conforms to the patient's alveolar ridge geometry, neutral zone biomechanical requirements, and mandibular movement patterns.

[0038] Based on a pre-built standardized 3D artificial tooth model database, containing high-precision 3D mesh models of 28 standard teeth, each tooth model uses the crown center as the origin of the coordinate system to ensure uniformity in spatial transformation. According to the position and attitude parameters corresponding to the initial parameter matrix, a spatial transformation algorithm is used to process the digital model of each standard tooth. Translation parameters are used to move the tooth model along the 3D coordinate axes to a specified position, while rotation parameters are used to adjust the pitch, yaw, and roll angles of the tooth model, ensuring the tooth presents an attitude that meets the parameter requirements. Subsequently, all spatially transformed tooth models are placed within a virtual denture base coordinate system, with the midpoint of the alveolar ridge centerline as the origin, ensuring a reasonable relative position of the teeth to the denture base. Finally, this is assembled to form a visualized 3D initial tooth arrangement plan. This plan can be displayed in real-time using 3D CAD software, clearly presenting the arrangement and spatial relationships of all teeth, providing operable 3D objects for subsequent evaluation of static mechanical and dynamic functional proxy models.

[0039] In one embodiment, the step of outputting an evaluation index for dynamic occlusion includes: S31, the static mechanical surrogate model is a differentiable feedforward deep neural network, which is trained under supervision by a large amount of offline finite element analysis simulation data. During training, the input data are the position and orientation parameters of different tooth arrangement schemes, and the output data are the calculated stress values ​​of key areas on the alveolar ridge surface under the corresponding scheme. After training, the static mechanical surrogate model is used to approximate the mechanical influence of tooth arrangement on supporting tissues and quickly predict the concentration and uniformity of stress distribution. S32, the dynamic functional surrogate model is a differentiable graph neural network. It is trained under supervision using a large amount of offline mandibular motion dynamics simulation data. During training, the input data is the three-dimensional geometric relationship diagram of the upper and lower dentition and the motion trajectory parameters. The output data is the contact state between the lower dentition in specific functional motion positions (such as protrusion position and lateral occlusal position). After training, the dynamic functional surrogate model is used to quickly evaluate whether the tooth arrangement plan has dynamic occlusal interference and whether it has sufficient balanced contact. S33. During the optimization process, the tooth arrangement parameters of the current iteration are input into the static mechanical proxy model in real time to obtain the predicted peak value of alveolar ridge stress and the evaluation value of stress distribution uniformity under this scheme. The current three-dimensional relationship of the dentition is input into the dynamic functional proxy model to obtain the predicted number of dynamic occlusal interference points and the number of effective balance contact points under this scheme. These evaluation values ​​are used as quantitative indicators for loss calculation in subsequent optimization steps.

[0040] As described in steps S31-S33 above, by constructing and training differentiable static mechanical surrogate models and dynamic functional surrogate models offline, a rapid quantitative evaluation of the biomechanical performance and dynamic occlusal function of the tooth arrangement scheme can be achieved. This provides accurate and efficient constraints for subsequent multi-objective iterative optimization, solving the core problems of time-consuming mechanical analysis and lack of dynamic functional considerations in traditional tooth arrangement evaluation.

[0041] The clinical suitability of dentures depends not only on static aesthetics but also on biomechanical rationality and dynamic occlusal coordination. Uneven stress distribution in the alveolar ridge mucosa accelerates alveolar bone resorption, while dynamic occlusal interference reduces masticatory efficiency and affects temporomandibular joint health. Current technologies, such as high-fidelity finite element analysis for mechanical assessment, require several hours for a single calculation, failing to meet real-time optimization needs. Dynamic functional assessments, on the other hand, rely heavily on subjective human judgment, lacking quantitative standards and being inefficient. Purely data-driven AI-based tooth arrangement methods directly ignore these two critical assessment types, leading to potential clinical risks. Therefore, a quantitative assessment scheme that balances efficiency and accuracy is needed. This scheme, through offline training of lightweight, differentiable surrogate models, specifically addresses the inefficiencies and insufficient quantification of traditional assessment methods, achieving sub-second responses for mechanical and dynamic functional assessments and providing real-time support for the optimization process.

[0042] The static mechanics surrogate model is a differentiable feedforward deep neural network with five fully connected layers. Each layer contains 128, 256, 256, 128, and 2 neurons respectively. The activation function is ReLU, and the loss function is mean squared error. The model's training data comes from 10,000 sets of offline finite element analysis simulation data. Each set of data was obtained by applying standard occlusal forces to different tooth alignment schemes using high-fidelity finite element simulation software, calculating the Von Mises stress values ​​of key areas on the alveolar ridge surface. During training, the input data consists of 72-dimensional parameters of tooth position (6 degrees of freedom for each of the 28 teeth), and the output data consists of the stress values ​​at five key points on the alveolar ridge crest line under the corresponding scheme. These stress values ​​are further used to calculate the peak stress and stress distribution uniformity evaluation values. After training, the model can approximately simulate the mechanical effects of tooth alignment on supporting tissues, replacing traditional real-time finite element analysis. It reduces the mechanical evaluation time from several hours to seconds, while maintaining differentiability to ensure that subsequent gradient optimization can backpropagate mechanical constraints, making the optimization process precisely target the direction of better stress distribution.

[0043] The dynamic functional surrogate model is a differentiable GraphSAGE graph neural network. Its network structure consists of 3 graph convolutional layers and 2 fully connected layers, with a node feature dimension of 10 and an edge feature dimension of 5. The training data for this model comes from 20,000 sets of offline mandibular motion dynamics simulation data. The simulation is based on standard mandibular motion trajectory parameters, namely a Bennett angle of 15° and a protruding condyle inclination of 30°. The contact state of the dentition under different tooth arrangement schemes in the protruding and lateral occlusal positions is obtained by calculating the dentition contact state using dynamic simulation software. During training, the input data is a three-dimensional geometric relationship diagram of the upper and lower dentition and motion trajectory parameters, where teeth are simplified as nodes, and features include position and type. Occlusal contact relationships are represented as edges. The output data is the number of abnormal contact points and the number of effective balanced contact points under specific functional motion positions. After training, the dynamic functional proxy model can quickly assess whether there is dynamic occlusal interference and whether there is sufficient balanced contact in the tooth arrangement scheme, filling the gap of traditional AI tooth arrangement lacking dynamic functional quantitative evaluation. Its differentiable properties also support subsequent gradient optimization, enabling dynamic occlusal constraints to be deeply integrated into the tooth arrangement parameter adjustment process.

[0044] During the optimization process, the current tooth alignment parameters are input into the static mechanical surrogate model in real time. This model quickly outputs the predicted peak alveolar ridge stress and stress distribution uniformity evaluation value under this scheme through forward propagation. The peak stress directly reflects the degree of stress concentration, and the stress distribution uniformity evaluation value is obtained by calculating the variance of stress at each key point; the smaller the value, the more uniform the distribution. Simultaneously, the current three-dimensional relationship of the dentition and the preset mandibular motion trajectory parameters are input into the dynamic functional surrogate model. This model quickly outputs the predicted number of dynamic occlusal interference points and the number of effective balance contact points through graph convolution operations. The fewer the interference points and the more effective balance contact points, the better the dynamic occlusal coordination. These evaluation values, as quantitative indicators, are directly input into the overall objective function, transforming into biomechanical loss terms and functional loss terms. This provides a clear adjustment direction for the optimization algorithm, ensuring that each parameter update specifically improves the mechanical performance and dynamic occlusal function of the tooth alignment scheme.

[0045] In one embodiment, the step of iteratively updating the tooth position parameters through an optimization algorithm until an optimized tooth arrangement scheme that meets a preset convergence condition includes: S41, construct an overall objective function, which is obtained by weighted summation of aesthetic loss terms, biomechanical loss terms, and functional loss terms. The weight coefficients of each term can be preset and adjusted according to the clinical focus. S42, calculate the loss values ​​for each sub-item. The aesthetic loss item is scored by a trained aesthetic evaluation network on visual elements such as the dental arch morphology and anterior aesthetic line of the current tooth arrangement scheme, and the score is converted into a loss value. The biomechanical loss item is calculated based on the stress peak value and stress distribution uniformity evaluation value. The calculation principle is to reduce the stress peak value as much as possible and make the stress distribution as uniform as possible. The functional loss item is calculated based on the number of dynamic occlusal interference points and the number of effective balance contact points. The calculation principle is to reduce the number of interference points as much as possible and increase the effective balance contact as much as possible. S43, using the automatic differentiation technique of neural networks, the gradient of the total objective function with respect to the position parameters of each degree of freedom of each tooth is calculated through backpropagation. The gradient descent optimization algorithm is used to iteratively update the position parameters of all teeth according to the calculated gradient direction and magnitude. S44, based on the updated tooth position, re-evaluate each loss and update the parameters again. When the value of the total objective function drops below the preset threshold, output the corresponding tooth arrangement scheme as the final optimized tooth arrangement scheme.

[0046] As described in steps S41-S44 above, by constructing a multi-objective overall objective function that integrates aesthetics, biomechanics and functionality, using neural network automatic differentiation technology to calculate the gradient, and combining gradient descent optimization algorithm to iteratively update tooth position parameters, an optimized tooth arrangement scheme that takes into account multi-dimensional performance and meets preset convergence conditions is finally obtained, realizing the upgrade of the initial tooth arrangement scheme from morphologically reasonable to multi-dimensional optimal.

[0047] While the initial tooth alignment scheme already possesses basic morphological adaptability, there is still room for optimization in terms of biomechanical performance and dynamic occlusal function. Furthermore, aesthetics, mechanical reliability, and functional coordination may be mutually constrained, making it difficult to achieve global optimization through simple manual adjustments. Current technologies often focus on a single objective or lack efficient gradient calculation methods, resulting in low optimization efficiency and slow convergence speed, failing to simultaneously meet the comprehensive clinical needs for aesthetics, safety, and function. Therefore, a multi-objective collaborative optimization mechanism is needed to achieve balanced optimization of various objectives through precise gradient calculation and efficient iteration. By constructing a weighted and fused overall objective function, leveraging the differentiability of the surrogate model to achieve rapid gradient calculation, and employing a gradient descent algorithm to iteratively update parameters, this approach specifically addresses the problems of single-objective and low-efficiency in traditional optimization, ensuring that the optimization scheme achieves balance across multiple dimensions.

[0048] A total objective function is constructed, which is obtained by weighted summation of aesthetic, biomechanical, and functional loss terms. The expression is: Total objective function equals 1.0 multiplied by the aesthetic loss term plus 2.5 multiplied by the biomechanical loss term plus 2.0 multiplied by the functional loss term. Here, 1.0, 2.5, and 2.0 are preset weight coefficients that can be adjusted according to clinical needs. For example, when emphasizing biomechanical performance, the weight of the biomechanical loss term can be increased; when emphasizing aesthetics, the weight of the aesthetic loss term can be increased. The core function of the total objective function is to transform the performance requirements of the three dimensions into quantifiable and optimizable mathematical indicators. By balancing the importance of each objective through weight allocation, it ensures that the optimization process progresses towards multi-dimensional synergistic optimization.

[0049] The loss values ​​for each component are calculated. The aesthetic loss is obtained through a trained aesthetic evaluation network, which contains four convolutional layers and two fully connected layers. The convolutional kernel size is 3×3×3, and the number of neurons in the fully connected layers is 256 and 128, respectively. This network can be trained using several sets of clinical aesthetic tooth arrangement cases. The network scores visual elements such as the arch morphology and anterior aesthetic lines of the current tooth arrangement plan, with a score range of 0 to 100. The scores are then normalized to convert them into loss values ​​of 0 to 1, with higher scores resulting in lower loss values. The biomechanical loss is calculated based on the peak stress and stress distribution uniformity evaluation values ​​output by the static mechanical surrogate model. Higher peak stress results in higher loss values, and lower stress distribution uniformity evaluation values ​​result in higher loss values. Specifically, it is obtained by weighted summing the proportion of the difference between the peak stress and a preset safety threshold, combined with the reciprocal of the uniformity evaluation value. The formula for calculating the biomechanical loss is as follows: ; Among them, the Represents the biomechanical loss term. Indicates the peak stress at the alveolar ridge. This represents the alveolar ridge's ultimate stress threshold, which is fixed at 5.0 MPa (the upper limit of clinical tolerance). This indicates the safe stress threshold of the alveolar ridge, which is fixed at 2.5 MPa (clinically recognized safe value). This represents the adjustment coefficient, fixed at 0.8, used to optimize the mapping relationship of uniformity evaluation values. This represents the evaluation value for the uniformity of stress distribution, ranging from 0 to 1. The larger the value, the more uniform the distribution.

[0050] The functional loss term is calculated based on the number of dynamic engagement interference points and the number of effective balancing contact points output by the dynamic functional proxy model. A higher number of interference points results in a larger loss value, as does a lower number of effective balancing contact points. The loss is calculated by summing the ratio of the number of interference points to the preset maximum allowable value, combined with the percentage difference between the number of effective contact points and the preset optimal value. The calculation of each sub-item's loss value ensures accurate quantification of each performance dimension, providing reliable input for the overall objective function.

[0051] Utilizing neural network automatic differentiation technology, the gradient of the overall objective function with respect to the positional parameters of each degree of freedom for each tooth is calculated via backpropagation. Automatic differentiation technology tracks the calculation process of the overall objective function, accurately determining the influence of each parameter on the function value. The sign of the gradient indicates the direction of parameter adjustment, and the magnitude of the gradient indicates the priority of adjustment. The Adam optimizer is used as the gradient descent optimization algorithm, with a learning rate of 0.001 and a momentum parameter of 0.9. Based on the calculated gradient direction and magnitude, the three translational parameters and three rotational parameters of each tooth are iteratively updated. For example, if the stress peak of a posterior tooth is too high, the corresponding biomechanical loss term has a positive gradient. The optimization algorithm will adjust the translational parameters of that tooth along the opposite direction of the gradient, moving it towards a position with a more uniform stress distribution. This achieves precise and efficient parameter adjustment, avoiding the inefficiency caused by blind iteration.

[0052] Based on the updated tooth positions, the static mechanical surrogate model, dynamic functional surrogate model, and aesthetic evaluation network are re-invoked to calculate new loss values ​​for each component and the overall objective function value. Gradient calculation and parameter updates are then performed again. The convergence threshold for the overall objective function is preset to 0.01. Iteration stops when the overall objective function value drops below 0.01 or the number of iterations reaches 300. At this point, the corresponding tooth alignment scheme meets the preset requirements in terms of aesthetics, biomechanical performance, and dynamic occlusal function, and is output as the final optimized tooth alignment scheme. This iterative convergence mechanism ensures that the optimization process achieves the desired effect while avoiding the waste of computational resources caused by excessive iteration.

[0053] In one embodiment, the steps of responding to a user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, locking the parameters of the adjusted tooth, and automatically performing local re-optimization on the parameters of the unlocked tooth based on the updated overall constraints include: S51, The optimized tooth arrangement scheme is displayed in the three-dimensional interactive interface. The user's interactive operation on the tooth model in the scheme through the input device is continuously monitored. When the user executes the translation or rotation adjustment command on one or more teeth, the final position and angle parameters of these adjusted teeth are immediately recorded and fixed. S52, set the user-fixed tooth parameters as unchangeable variables, set the remaining tooth parameters in the scheme that have not been adjusted by the user as optimizable variables, keep the total objective function unchanged, but limit its calculation scope to the new tooth arrangement scheme composed of the current optimizable variables and fixed variables; S53, starting with the overall scheme adjusted by the user, initiates a local optimization process with a limited number of iterations. This process only calculates and updates the gradient of the tooth parameters of the optimizable variables to fine-tune their positions, thereby automatically finding a new scheme that makes the overall objective function better and is more coordinated while respecting the user's subjective adjustments. S54 generates and outputs a final tooth alignment plan that integrates user preferences and system optimization. It also provides brief visual feedback on the biomechanical and functional evaluation report of the final plan for confirmation.

[0054] As described in steps S51-S54 above, by allowing users to manually adjust specific teeth in the optimized tooth arrangement scheme through a three-dimensional interactive interface, locking the parameters of the adjusted teeth and setting the remaining tooth parameters as optimizable variables, and performing local re-optimization with a limited number of iterations starting from the adjusted scheme, the final output is a tooth arrangement scheme that integrates user preferences and system optimization and provides visual evaluation feedback. While respecting the user's subjective adjustments, the overall biomechanics and functional coordination of the scheme are maintained.

[0055] In clinical denture design, while optimized tooth arrangement plans meet multi-dimensional objective requirements, dentists or technicians may still request subjective adjustments based on individual aesthetic judgments and clinical experience. Manually adjusting one or more teeth can disrupt the original biomechanical balance and dynamic occlusal coordination. Without subsequent optimization, this can lead to localized stress concentration or dynamic occlusal interference. Current technologies require restarting global optimization after user adjustments, which is time-consuming and may overwrite the user's intended adjustments, or lacks targeted local optimization mechanisms, failing to quickly restore overall coordination. Therefore, a local re-optimization mechanism is needed to precisely optimize unadjusted tooth parameters while preserving the user's adjustment results. By locking adjustment parameters, limiting the optimization range, and performing local iterative optimization, this addresses the problem that traditional optimization mechanisms cannot simultaneously consider user subjective intentions and overall coordination, achieving highly efficient human-machine collaborative optimization.

[0056] The optimized tooth alignment plan is displayed in a 3D interactive interface, integrated into a 3D CAD software plugin, supporting real-time rendering and 3D manipulation of the tooth model. User actions are continuously monitored via input device interfaces, including mice, styluses, or 3D control handles. When a user performs translation or rotation on one or more teeth, position and angle sensors record the final parameters after the operation. These parameters include three translation parameters and three rotation parameters, consistent with the defined tooth position parameters. These parameters are immediately locked to remain unchanged in subsequent optimizations, ensuring the user's adjustment intentions are fully preserved. For example, after a dentist adjusts the labial inclination of the anterior teeth based on the patient's facial contours, the system locks the six degrees of freedom parameters of that anterior tooth to prevent subsequent optimizations from overriding the adjustment.

[0057] The user's fixed tooth parameters are set as invariants, while the remaining unadjusted tooth parameters in the plan are set as optimizable variables. The overall objective function remains unchanged, still a weighted sum of aesthetic, biomechanical, and functional loss terms, with weight coefficients remaining constant at 1.0, 2.5, and 2.0. However, the calculation scope is limited to the new tooth arrangement plan composed of both optimizable and invariant variables. This setting ensures that the optimization process always revolves around the original multi-dimensional objectives, while clearly defining the optimization boundaries, avoiding interference with the user's adjusted teeth, and providing a clear mathematical framework for local re-optimization.

[0058] Starting with the user-adjusted overall solution, a local optimization process is initiated, with iterations limited to 50 steps, far fewer than the 300 steps of global optimization, ensuring rapid response. During optimization, gradient calculations and updates are performed only on the optimizable tooth parameters. Gradient calculations are achieved through neural network automatic differentiation technology, leveraging the differentiability of static mechanical surrogate models and dynamic functional surrogate models to backpropagate the gradient of the overall objective function. For example, after the user adjusts the position of the anterior teeth, the posterior tooth parameters become optimizable variables. The system adjusts the translation and rotation parameters of the posterior teeth through gradient calculations, restoring a uniform distribution of alveolar ridge stress and reducing dynamic occlusal interference points. By limiting the optimization scope and iteration count, overall coordination is quickly restored while avoiding efficiency reduction caused by over-optimization.

[0059] The system generates and outputs a final tooth alignment plan that integrates user preferences with system optimization. This plan includes the final position and orientation parameters of all teeth and supports model export and denture fabrication file generation. It also provides concise visual feedback on biomechanical and functional assessment reports. Biomechanical feedback displays the alveolar ridge stress distribution in the form of a virtual stress cloud map, while functional feedback displays the contact status in protruding and lateral jaw positions in the form of a dynamic contact point report. This feedback information intuitively presents the consistency of the plan, helping users confirm the adjustment effect and ensuring that the final plan meets both subjective preferences and objective clinical requirements.

[0060] like Figure 2 As shown, the present invention also discloses an automatic tooth alignment system based on artificial intelligence, comprising: The acquisition module is used to acquire the patient's oral cavity three-dimensional scan data and jaw position relationship record. It extracts the geometric features and neutral zone information of the alveolar ridge through the feature extraction network, and fuses the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. The first output module is used to input the enhanced conditional feature vector into a pre-trained conditional generation model and output an initial three-dimensional tooth arrangement scheme. The second output module is used to call the pre-trained differentiable static mechanical proxy model and dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the dental arch and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. The optimization module is used to construct a total objective function based on the initial three-dimensional tooth arrangement scheme, with the translation and rotation parameters of each tooth as optimization variables. It uses the differentiability of the static mechanical proxy model and the dynamic functional proxy model to calculate the gradient of the total objective function with respect to the tooth position parameters, and iteratively updates the tooth position parameters through the optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. The update module is used to respond to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, lock the parameters of the adjusted teeth, and automatically perform local re-optimization on the parameters of the unlocked teeth based on the updated overall constraints, so as to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

[0061] In one embodiment, the first output module includes: The construction unit is configured to generate the spatial position and orientation of all teeth based on the input enhanced conditional feature vector, since the conditional generation model is a neural network model built on a conditional generative adversarial network. A generation unit is used to input the enhanced conditional feature vector as a generation condition into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. The assembly unit is used to spatially transform the digital model of each standard tooth according to the position and orientation parameters in the initial parameter matrix based on the pre-set standardized artificial tooth three-dimensional model database, and place it in the virtual denture base coordinate system to assemble a visualized three-dimensional initial tooth arrangement plan containing a complete dentition.

[0062] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described artificial intelligence-based automatic tooth alignment method.

[0063] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described artificial intelligence-based automatic tooth alignment method.

[0064] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0065] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0066] The above description is merely a preferred embodiment of the present invention and does not limit the scope of this application. Any equivalent results or equivalent process transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. An automatic tooth alignment method based on artificial intelligence, characterized in that, Includes the following steps: The patient's oral cavity three-dimensional scan data and jaw position relationship record are obtained. The geometric features and neutral zone information of the alveolar ridge are extracted through a feature extraction network. The geometric features are then fused with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. The enhanced conditional feature vector is input into a pre-trained conditional generation model, which outputs an initial three-dimensional tooth arrangement scheme. The system invokes a pre-trained, differentiable static mechanical proxy model and a dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the dental arch and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. Based on the initial three-dimensional tooth arrangement scheme, the translation and rotation parameters of each tooth are used as optimization variables to construct the overall objective function. The gradient of the overall objective function with respect to the tooth position parameters is calculated by utilizing the differentiability of the static mechanical proxy model and the dynamic functional proxy model. The tooth position parameters are then iteratively updated through the optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. In response to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, the parameters of the adjusted tooth are locked, and based on the updated overall constraints, the parameters of the unlocked tooth are automatically re-optimized locally to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

2. The automatic tooth alignment method based on artificial intelligence according to claim 1, characterized in that, The step of fusing the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector includes: A three-dimensional scanning mesh model of the patient's oral cavity in an edentulous state was obtained, along with jaw position relationship data of the relative positions of the upper and lower jaws. The three-dimensional scanning mesh model was then cleaned and repaired to obtain a standardized digital model of the dentition. The standardized digital model of the teeth and jaws is processed by a three-dimensional convolutional neural network to automatically identify and locate key anatomical landmarks. Based on the identified anatomical landmarks, geometric feature vectors describing the shape of the dental arch and the condition of the alveolar bone are calculated. At the same time, by analyzing the surface curvature of the model and combining it with biomechanical principles, the mechanical equilibrium region where the teeth should be arranged, i.e., the neutral zone information, is estimated. From a pre-set clinical knowledge base, a set of biomechanical material parameters adapted to the characteristics of the patient's oral tissues is matched. The set of parameters includes parameters characterizing mucosal elasticity and grade coefficients characterizing bone density. Based on the individual patient condition or general standards, the motion characteristic parameters of the mandible during protrusion and lateral movements are determined. The calculated geometric feature vector, the neutral zone information, the biomechanical material parameter set, and the mandibular motion feature parameters are concatenated and combined, and then processed by a neural network layer for feature fusion and dimensionality reduction. Finally, a fixed-dimensional enhanced conditional feature vector containing multi-source constraint information is output.

3. The automatic tooth alignment method based on artificial intelligence according to claim 1, characterized in that, The step of inputting the enhanced conditional feature vector into a pre-trained conditional generation model to output an initial three-dimensional tooth alignment scheme includes: The conditional generation model is a neural network model built on a conditional generative adversarial network, configured to generate the spatial position and pose of all teeth based on the input enhanced conditional feature vector. The enhanced conditional feature vector is used as a generation condition and input into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. Based on a pre-set standardized artificial tooth 3D model database, the digital model of each standard tooth is spatially transformed according to the position and posture parameters corresponding to the initial parameter matrix, and placed in the virtual denture base coordinate system to assemble a visualized 3D initial tooth arrangement plan containing a complete dentition.

4. The automatic tooth alignment method based on artificial intelligence according to claim 1, characterized in that, The steps for outputting the evaluation index of dynamic occlusion status include: The static mechanics surrogate model is a differentiable feedforward deep neural network that is trained under supervision using a large amount of offline finite element analysis simulation data. This static mechanics surrogate model is used to simulate the mechanical effects of tooth alignment on supporting tissues and to predict the concentration and uniformity of stress distribution. The dynamic functional surrogate model is a differentiable graph neural network that is trained under supervision using a large amount of offline mandibular motion dynamics simulation data. The dynamic functional surrogate model is used to quickly evaluate whether there is dynamic occlusal interference in the tooth arrangement scheme and whether there is sufficient balanced contact. The current tooth alignment parameters are input into the static mechanical proxy model in real time to obtain the predicted peak alveolar ridge stress and stress distribution uniformity evaluation value under this scheme. The current three-dimensional relationship of the dentition is input into the dynamic functional proxy model to obtain the predicted number of dynamic occlusal interference points and the number of effective balance contact points under this scheme.

5. The automatic tooth alignment method based on artificial intelligence according to claim 4, characterized in that, The step of iteratively updating the tooth position parameters through an optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence condition includes: A total objective function is constructed, which is obtained by weighted summation of aesthetic loss terms, biomechanical loss terms, and functional loss terms. The weight coefficients of each term can be preset and adjusted according to the clinical focus. The loss values ​​for each sub-item are calculated. The aesthetic loss item is scored by a trained aesthetic evaluation network on visual elements such as the dental arch morphology and anterior aesthetic line of the current tooth arrangement scheme, and the scores are converted into loss values. The biomechanical loss item is calculated based on the stress peak value and stress distribution uniformity evaluation value. The functional loss item is calculated based on the number of dynamic occlusal interference points and the number of effective balance contact points. Using neural network automatic differentiation technology, the gradient of the total objective function with respect to the position parameters of each degree of freedom of each tooth is calculated through backpropagation. The gradient descent optimization algorithm is then used to iteratively update the position parameters of all teeth based on the calculated gradient direction and magnitude. Based on the updated tooth positions, the various losses are reassessed, and the parameters are updated again. When the value of the overall objective function drops below a preset threshold, the corresponding tooth arrangement scheme is output as the final optimized tooth arrangement scheme.

6. The automatic tooth alignment method based on artificial intelligence according to claim 1, wherein, The steps of responding to a user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, locking the parameters of the adjusted teeth, and automatically performing local re-optimization on the parameters of the unlocked teeth based on the updated overall constraints include: The optimized tooth arrangement scheme is displayed in a three-dimensional interactive interface. The user's interactive operation on the tooth model in the scheme through the input device is continuously monitored. When the user executes the translation or rotation adjustment command on one or more teeth, the final position and angle parameters of these adjusted teeth are immediately recorded and fixed. Set the user-fixed tooth parameters as unchangeable variables, and set the remaining tooth parameters in the scheme that have not been adjusted by the user as optimizable variables. Keep the overall objective function unchanged, but limit its calculation scope to the new tooth arrangement scheme composed of the current optimizable variables and fixed variables. Starting with the user-adjusted overall solution, a local optimization process with a limited number of iterations is initiated. This process only performs gradient calculations and updates on the tooth parameters of the optimizable variables to fine-tune their positions and automatically finds a new, overall coordinated solution that makes the overall objective function better. Generate and output the final tooth alignment plan that integrates user preferences and system optimization.

7. An automatic tooth alignment system based on artificial intelligence, characterized by, include: The acquisition module is used to acquire the patient's oral cavity three-dimensional scan data and jaw position relationship record. It extracts the geometric features and neutral zone information of the alveolar ridge through the feature extraction network, and fuses the geometric features with preset biomechanical material parameters and mandibular movement trajectory parameters to generate an enhanced conditional feature vector. The first output module is used to input the enhanced conditional feature vector into a pre-trained conditional generation model and output an initial three-dimensional tooth arrangement scheme. The second output module is used to call the pre-trained differentiable static mechanical proxy model and dynamic functional proxy model. The static mechanical proxy model takes tooth position parameters as input and outputs an evaluation index of the alveolar ridge stress distribution. The dynamic functional proxy model takes the geometric relationship of the dental arch and motion trajectory parameters as input and outputs an evaluation index of the dynamic occlusal state. The optimization module is used to construct a total objective function based on the initial three-dimensional tooth arrangement scheme, with the translation and rotation parameters of each tooth as optimization variables. It uses the differentiability of the static mechanical proxy model and the dynamic functional proxy model to calculate the gradient of the total objective function with respect to the tooth position parameters, and iteratively updates the tooth position parameters through the optimization algorithm until an optimized tooth arrangement scheme that meets the preset convergence conditions is obtained. The update module is used to respond to the user's manual adjustment of the position of a specific tooth in the optimized tooth arrangement plan, lock the parameters of the adjusted teeth, and automatically perform local re-optimization on the parameters of the unlocked teeth based on the updated overall constraints, so as to maintain the overall biomechanical and functional coordination of the tooth arrangement plan.

8. The automatic tooth alignment system based on artificial intelligence according to claim 7, wherein, The first output module includes: The construction unit is configured to generate the spatial position and orientation of all teeth based on the input enhanced conditional feature vector, since the conditional generation model is a neural network model built on a conditional generative adversarial network. The generation unit is used to input the enhanced conditional feature vector as a generation condition into the generator network of the conditional generation model. Based on the generation condition, the generator network infers and generates an initial parameter matrix representing the spatial position and orientation of all teeth in the implicit space. The assembly unit is used to spatially transform the digital model of each standard tooth according to the position and orientation parameters in the initial parameter matrix based on the pre-set standardized artificial tooth three-dimensional model database, and place it in the virtual denture base coordinate system to assemble a visualized three-dimensional initial tooth arrangement plan containing a complete dentition.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images