Deep learning-based bite registration jaw position compensation method and storable medium

Abstract

This invention discloses a deep learning-based method for jaw position compensation in occlusal recordings and a storage medium, comprising the following steps: S1, obtaining the initial relative pose matrix of the maxilla and mandible; S2, extracting the patient's corresponding dental arch morphological features, occlusal recording material parameters, occlusal recording geometric information, and occlusal force vector, inputting the above multi-source features into a deep learning deformation prediction model, and predicting a point-by-point three-dimensional displacement deformation field; S3, extracting the displacement vector corresponding to each material point from the point-by-point three-dimensional displacement deformation field, and solving for the optimal rigidity compensation transformation matrix; S4, compensating the initial relative pose matrix of the maxilla and mandible based on the optimal rigidity compensation transformation matrix. This invention addresses the problem of inaccurate jaw position relationships caused by in vitro elastic deformation of occlusal recordings, applying deep learning technology to the accurate prediction of the three-dimensional elastic deformation field of occlusal recordings, fundamentally overcoming the long-standing industry accuracy bottleneck of inconsistency between the actual intraoral occlusal position and the jaw position registered in extraoral scanning.

Description

Technical Field

[0001] This invention relates to the field of oral medicine technology, and in particular to a deep learning-based method for occlusal recording and jaw position compensation, as well as a storage medium. Background Technology

[0002] In digital dentistry, after obtaining digital models of the upper and lower jaws (usually through intraoral scanning), it is necessary to establish their three-dimensional spatial relationship in the target jaw position (such as centric relation or intercuspal position), i.e., to perform "jaw position transfer". One of the current mainstream methods is the "occlusal recording scanning method": an uncured occlusal recording material (such as polypolymer silicone rubber, polyether, etc.) is placed in the patient's mouth and the patient bites into the target jaw position. After the material cures, it is removed, and the occlusal recording block with occlusal imprints and the upper and lower jaws are scanned separately. Then, the digital models of the upper and lower jaws are registered to the corresponding imprints of the occlusal recording block in the software, thereby indirectly obtaining the relative pose of the upper and lower jaws.

[0003] However, occlusal recording materials are essentially viscoelastic polymers, which undergo compression deformation and solidify under intraoral occlusal pressure. When the occlusal recording block is removed from the mouth, the external pressure is removed, and the material elastically recovers. Furthermore, the temperature and humidity differences between the ex vivo and intraoral environments cause additional dimensional deformation. This elastic deformation leads to a shift in the position of the occlusal impression obtained from the ex vivo scan compared to the position of the impression under actual intraoral occlusion. Ultimately, this results in a systematic deviation between the registered upper and lower jaw relative positions and the patient's actual intraoral jaw position. This deviation can reach tens to hundreds of micrometers in both the vertical and anteroposterior directions. When the deviation exceeds a certain amount, it significantly increases the occlusal adjustment time after prosthesis placement and can even cause occlusal trauma, prosthesis chipping, and other problems. The impact of this error is particularly pronounced in high-precision scenarios such as full-mouth occlusal reconstruction and multi-unit fixed restorations.

[0004] In existing technologies, optimization for the aforementioned deformation error problem mainly focuses on two directions: material improvement and manual adjustment. Material improvement involves increasing the rigidity of the occlusal recording material and using hard wax or light-cured resin to reduce deformation. However, increased material rigidity increases occlusal resistance, preventing the masticatory muscles from achieving a physiologically relaxed state, thus altering the actual target jaw position and introducing greater clinical errors. Manual adjustment involves physicians manually adjusting the relative positions of the upper and lower jaw digital models based on clinical experience to roughly compensate for deformation. However, this method lacks quantitative mechanical basis, and the compensation accuracy is greatly affected by the physician's subjective experience, resulting in poor repeatability and an inability to achieve standardized precision control. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a deep learning-based method for occlusal recording and jaw position compensation, as well as a storage medium.

[0006] The objective of this invention is achieved through the following technical solution: a jaw position compensation method based on deep learning for occlusal recording, comprising the following steps:

[0007] S1. Obtain the digital model of the patient's maxilla, mandible, and occlusion record respectively. Based on the occlusion imprint features on the occlusion record digital model, register the digital model of the maxilla and mandible with the occlusion record digital model respectively to obtain the initial relative pose matrix of the maxilla and mandible. S2. Extract the dental arch morphology features, occlusal recording material parameters, occlusal recording geometric information, and occlusal force vector corresponding to the patient. Input the above multi-source features into the pre-trained deep learning deformation prediction model. Predict the point-by-point three-dimensional displacement deformation field of the occlusal recording from the intraoral force-fixed state to the off-body free state through the deep learning deformation prediction model. S3. Extract the displacement vectors corresponding to each material point in the mandibular tooth imprint region from the point-by-point three-dimensional displacement deformation field, and use the rigid body motion estimation algorithm to solve the optimal rigid compensation transformation matrix for offsetting the elastic deformation of the occlusal record. S4. The initial relative pose matrix of the upper and lower jaws is compensated based on the optimal rigidity compensation transformation matrix to obtain the corrected relative pose matrix of the upper and lower jaws.

[0008] Preferably, in step S1, the ICP algorithm is used to register the maxillary digital model, the mandibular digital model, and the occlusal record digital model; the initial relative pose matrix of the maxilla and mandible is the initial transformation matrix of the mandibular digital model relative to the maxillary digital model.

[0009] Preferably, in step S2, the dental arch morphological features are composed of global morphological codes extracted from the point clouds of the maxillary and mandibular dentitions by a pre-trained point cloud autoencoder, or are composed of geometric measurement parameters including dental arch width, occlusion, and overjet; the material physical parameters are the viscoelastic constitutive model parameters corresponding to the current occlusal recording material; the occlusal recording geometric information are the spatial geometric features of the occlusal recording digital model, including the thickness distribution field of the occlusal recording digital model; and the occlusal force vector includes the magnitude, direction, and area of ​​action of the force applied during the occlusal recording process.

[0010] Preferably, in step S2, the deep learning deformation prediction model adopts a network architecture based on point cloud or sparse three-dimensional convolution. The deep learning deformation prediction model is equipped with a conditional injection module, which is used to fuse multi-source features into the geometric coding branch of the deep learning deformation prediction model to output a point-by-point three-dimensional displacement deformation field.

[0011] Preferably, in step S2, the loss function of the deep learning deformation prediction model consists of a data fidelity term, a physical residual term, and a material consistency constraint term; the physical residual term is constructed based on the elasticity equilibrium equation and the material constitutive relation; the material consistency constraint term is used to constrain the physical consistency of the deformation prediction results under the same material parameters.

[0012] As a preferred embodiment, the specific method for step S3 is as follows: S31: Segment the mandibular tooth imprint region from the occlusal record digital model, obtain the point set of each material point in the region, and extract the displacement vector corresponding to each point from the point-by-point three-dimensional displacement deformation field. Assign a corresponding weight to each material point. ; S32. Construct an objective optimization function for rotation matrix R and translation vector matrix t; the objective optimization function is: ; S33. Based on the objective function, the optimal rotation matrix is ​​solved by weighted centroid calculation, centroid-free coordinate transformation, and singular value decomposition of the covariance matrix. and the optimal translation vector matrix ; S34, Based on the optimal rotation matrix and the optimal translation vector matrix Obtain the optimal rigid compensation transformation matrix ; .

[0013] Preferably, in step S4, the optimal rigidity compensation transformation matrix is ​​multiplied by the initial relative pose matrix of the mandible and mandible to obtain the corrected relative pose matrix of the mandible and mandible.

[0014] Preferably, the method for constructing the deep learning deformation prediction model is as follows: N1: Constructing a training dataset: Based on finite element simulation, we traverse various working conditions of digital models of the upper and lower jaws with different dental arch shapes, constitutive parameters of different occlusal recording materials, and different occlusal force vectors. We simulate the unloading process of the occlusal recording material from the intraoral stress-cured state to the detached free state under each working condition, and generate the three-dimensional deformation field corresponding to each working condition as the training ground value. N2: Constructing sample input features: For each training sample, extract the corresponding dental arch morphology features, material physical parameters, occlusal record geometric information, and occlusal force vector, and fuse them into network input features; N3: Iterative training of the network: The structured input features are input into the deep learning network, and iterative training is performed with a preset loss function as the optimization objective until the loss function converges, resulting in a trained deep learning deformation prediction model.

[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a deep learning-based occlusal recording jaw position compensation method.

[0016] The beneficial effects of this invention are: 1. This invention is the first to address the problem of inaccurate jaw position relationship caused by in vitro elastic deformation of occlusal records. It applies deep learning technology to the field of accurate prediction of the three-dimensional elastic deformation field of occlusal records, filling the technical gap that existing technologies cannot quantify and reverse compensate for in vitro elastic deformation. It fundamentally overcomes the long-standing industry accuracy bottleneck of inconsistency between the actual intraoral occlusal position and the jaw position registered by extraoral scanning, and provides a brand-new path to improve the accuracy of digital jaw position relationship transfer.

[0017] 2. This invention adopts a multimodal feature fusion architecture, using four types of multi-source information as network inputs: dental arch morphological features, material viscoelastic constitutive parameters, occlusal record spatial geometric features, and occlusal force vectors. This endows the model with general predictive capabilities across patients, materials, and occlusal conditions. In the clinical application stage, only the trained network model needs to be called to complete millisecond-level rapid inference, eliminating the need for time-consuming and costly finite element mechanical simulations for individual cases. This significantly improves clinical diagnostic and treatment efficiency while ensuring accuracy.

[0018] 3. This invention embeds a physical residual term based on the equilibrium equation of elasticity and the constitutive relationship of materials into the network loss function, supplemented by a material consistency constraint term, so that the deformation field output by the network strictly follows the objective laws of mechanics, effectively reducing the model's dependence on the number of training samples; even under the condition of limited training samples, it can still output deformation prediction results with physical rationality, significantly improving the model's prediction reliability, working condition extrapolation ability and prediction accuracy for untrained cases, ensuring the stability and accuracy of clinical applications.

[0019] 4. This invention constructs a fully automated error correction closed loop of "intelligent prediction of deformation field + optimal rigidity compensation transformation solution + precise correction of jaw position relationship". The accuracy of jaw position relationship after compensation can reach the micrometer level, which can effectively improve the occlusal accuracy and clinical adaptability of restorations. This method can be directly embedded into the standard workflow of existing digital dental CAD software without adding additional hardware equipment or changing the routine clinical operation process. It has a low threshold for implementation and is easy to promote and apply on a large scale. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0021] Figure 2 This is a schematic diagram of the digital model of the maxilla of the present invention.

[0022] Figure 3This is a schematic diagram of the digital model of the mandible of the present invention.

[0023] Figure 4 This is a schematic diagram of the bite recording digital model of the present invention.

[0024] Figure 5 This is a schematic diagram illustrating the registration of the maxillary digital model and mandibular digital model with the occlusal recording digital model, respectively, according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0026] Those skilled in the art should understand that, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0027] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0028] like Figures 1 to 5 As shown, the deep learning-based occlusal recording jaw position compensation method includes the following steps: S1. Obtain the digital model of the patient's maxilla, mandible, and occlusion record, respectively. Based on the occlusion imprint features on the occlusion record digital model, register the digital model of the maxilla and mandible with the occlusion record digital model to obtain the initial relative pose matrix of the maxilla and mandible.

[0029] In this step, the ICP algorithm is used to register the maxillary digital model, mandibular digital model and occlusal record digital model; the initial relative pose matrix of the maxilla and mandible is the initial transformation matrix of the mandibular digital model relative to the maxillary digital model.

[0030] The ICP algorithm, or Iterative Nearest Point algorithm, aims to find the optimal rigid body transformation (rotation + translation) between two sets of 3D point sets, minimizing the error between the corresponding positions of the two point sets after registration. The ICP algorithm uses iterative optimization as its core logic: it continuously searches for the nearest neighboring point between the point set to be registered and the reference point set, solves for the optimal rigid body transformation based on the corresponding point, updates the point set pose, and iterates repeatedly until the registration error converges, ultimately obtaining the accurate spatial transformation relationship between the two point sets.

[0031] The maxillary digital model and mandibular digital model are three-dimensional model data obtained by scanning the inside of the patient's mouth using an intraoral scanner (e.g., Figure 2 and Figure 3 As shown in the figure, its format can be STL.

[0032] Occlusal recording digital model refers to a three-dimensional model obtained by scanning occlusal recording materials after occlusal testing (such as...). Figure 4 As shown in the figure, its format can be STL.

[0033] The initial relative pose matrix of the mandible and mandible in this application is a 4×4 homogeneous transformation matrix that describes the three-dimensional spatial position and posture of the mandibular digital model relative to the mandible, with the mandibular digital model as the reference. It is a preliminary jaw position relationship result obtained through the traditional occlusal record registration process. Because the occlusal record undergoes elastic deformation after being removed from the body, the initial relative pose matrix of the mandible and mandible contains systematic pose deviations and is the reference object for subsequent elastic deformation compensation.

[0034] S2. Extract the dental arch morphology features, occlusal recording material parameters, occlusal recording geometric information, and occlusal force vector corresponding to the patient. Input the above multi-source features into the pre-trained deep learning deformation prediction model. Predict the point-by-point three-dimensional displacement deformation field of the occlusal recording from the intraoral force-fixed state to the off-body free state through the deep learning deformation prediction model.

[0035] In this step, the dental arch morphological features are composed of global morphological codes extracted from the point clouds of the maxillary and mandibular dentitions by a pre-trained point cloud autoencoder, or are composed of geometric measurement parameters including dental arch width, occlusion, and overlay.

[0036] The material physical parameters are the viscoelastic constitutive model parameters corresponding to the current bite recording material. Among them, the viscoelastic constitutive model parameters are a set of quantitative parameters describing the elastic deformation and viscous rheological composite mechanical behavior of the bite recording material. They are the core physical basis for generating training ground truth values ​​for finite element simulation, input features for deep learning networks, and construction of physical constraint loss.

[0037] Occlusal record geometry information is the spatial geometric features of the occlusal record digital model, which includes the thickness distribution field of the occlusal record digital model; The bite force vector includes the magnitude, direction, and area of ​​the force applied during bite recording.

[0038] The point-by-point three-dimensional displacement deformation field is a vector field describing the three-dimensional spatial displacement of each point within the geometric region of the occlusal recording material during its elastic recovery from an intraoral stressed-solid state to an in vitro free state. For any point in the occlusal recording, its position vector is... When a material point deforms from its initial configuration (stressed and solidified state inside the mouth) to its free state outside the body, the point moves to a new spatial position x(X).

[0039] Then the three-dimensional displacement vector of the material point Defined as: ; in, , , These represent the displacement components of the material point along the three coordinate axes. The point-by-point three-dimensional displacement deformation field maps each point within the region into a three-dimensional displacement vector.

[0040] In the application scenario of this invention, the physical direction of the displacement field is: from the "intraoral compression position" to the "position after free rebound", corresponding to the elastic recovery direction of the bite record.

[0041] The deep learning deformation prediction model adopts a network architecture based on point cloud or sparse 3D convolution. The deep learning deformation prediction model has a conditional injection module inside, which is used to fuse multi-source features into the geometric coding branch of the deep learning deformation prediction model to output a point-by-point 3D displacement deformation field.

[0042] The loss function of a deep learning deformation prediction model consists of a data fidelity term. Physical residuals Material consistency constraints Among them, the data fidelity term is a conventional supervised loss for deep learning tasks. Its function is to make the point-by-point three-dimensional displacement deformation field predicted by the network match the training true value (the deformation field generated by finite element simulation or measured and calibrated) as closely as possible.

[0043] The physical residual term is constructed based on the equilibrium equations of elasticity and the constitutive relation of materials. It forces the deformation field to satisfy objective physical laws through the basic equations of elasticity, thereby fundamentally limiting the output space of the network.

[0044] The material consistency constraint term is used to constrain the physical consistency of deformation prediction results under the same material parameters. The core is to ensure that the equivalent mechanical properties of the same occlusal recording material remain consistent under different dental arch shapes and different occlusal force conditions, so as to avoid material property disorder when the network tries to fit individual samples.

[0045] The total loss function is a weighted sum of three terms, namely: ; These are preset weight coefficients used to balance the optimization intensity of different constraints, and can be dynamically adjusted according to the training stage and data scale.

[0046] The construction method of deep learning deformation prediction model is as follows: N1: Constructing a training dataset: Based on finite element simulation, we traverse various working conditions of digital models of the upper and lower jaws with different dental arch shapes, constitutive parameters of different occlusal recording materials, and different occlusal force vectors. We simulate the unloading process of the occlusal recording material from the intraoral stress-cured state to the detached free state under each working condition, and generate the three-dimensional deformation field corresponding to each working condition as the training ground value.

[0047] Traditional clinical practice cannot accurately measure point-by-point 3D deformation in vivo from patients' intraoral occlusion records. Purely measured data is extremely small in scale and highly noisy, making it unsuitable for supervised training of deep learning models. Finite element simulation can generate spatially continuous, sub-micron level annotations of point-by-point 3D deformation fields in batches at low cost, solving the core problem of scarce training samples at the data source and providing a fundamental guarantee for the accuracy of deep learning models. The dataset covers three core variable dimensions: different dental arch morphologies, different material constitutive parameters, and different occlusal force vectors. It fully matches the characteristics of real clinical scenarios, such as large differences in patient dentition, diverse occlusal recording materials, and significant individual differences in occlusal forces. This diverse range of training conditions allows the model to learn the influence of different factors on deformation.

[0048] N2: Constructing Sample Input Features: For each training sample, extract the corresponding dental arch morphology features, material physical parameters, occlusal record geometric information, and occlusal force vector, and fuse them into network input features.

[0049] Elastic deformation recorded in occlusion is a mechanical result of the combined effects of geometry, material properties, and stress conditions. Four types of features—architectural features, material physical parameters, occlusion record geometric information, and occlusal force vectors—completely cover the core variables determining deformation distribution. Models using these as inputs can learn the intrinsic generation laws of deformation, rather than simply interpolating and fitting specific samples; the prediction logic is highly consistent with the physical essence. The structured multi-source input design allows the model to output the deformation field under corresponding working conditions by switching the input material type, occlusal force parameters, and arch morphology encoding.

[0050] N3: Iterative training of the network: The structured input features are input into the deep learning network, and iterative training is performed with a preset loss function as the optimization objective until the loss function converges, resulting in a trained deep learning deformation prediction model.

[0051] This application employs a deep learning deformation prediction model to predict point-by-point three-dimensional displacement deformation fields, significantly improving computational efficiency and enabling real-time clinically applicable quantitative prediction of deformations, thus providing a foundation for subsequent compensation of the relative pose matrix of the maxilla and mandible.

[0052] S3. Extract the displacement vectors corresponding to each material point in the mandibular tooth imprint region from the point-by-point three-dimensional displacement deformation field, and use a rigid body motion estimation algorithm to solve for the optimal rigid compensation transformation matrix used to offset the elastic deformation of the occlusal record. The specific method for this step is as follows: S31: Segment the mandibular tooth imprint region from the occlusal record digital model to obtain the point set of each material point in the region. And extract the displacement vector corresponding to each material point from the point-by-point three-dimensional displacement deformation field. Assign a corresponding weight to each material point. ; S32. Construct an objective optimization function for rotation matrix R and translation vector matrix t; the objective optimization function is: .

[0053] In the formula, N represents the number of material points in the mandibular tooth imprint region. It is the square of the second norm of the vector, that is, the square of the Euclidean distance between the two points. The rotation matrix R is an orthogonal matrix, and its physical meaning is that rotation transformation only changes the orientation of the object, without changing the shape, size, or relative distance between points, thus satisfying the basic properties of a rigid body. The coordinates of the material points on the surface of the mandibular imprint during the off-site scanning state (free state outside the mouth) of the occlusal recording material are given. The position of these material points carries the spatial deviation caused by the elastic deformation of the material.

[0054] The determinant of the rotation matrix is ​​+1, which means that the transformation is a normal three-dimensional rotation (keeping the chirality of the coordinate system unchanged), eliminating invalid mathematical solutions of mirror reflection, and ensuring that the result conforms to the physical motion laws of real three-dimensional space.

[0055] The objective optimization function described above is a weighted least squares optimization problem with orthogonal constraints. This objective optimization function fits a globally optimal global rigid body transformation from the point-by-point non-uniform elastic displacement field, allowing the mandibular dentition to undergo a single global rotation and translation transformation to maximize the offset of the pose error caused by the elastic deformation of the occlusal record, and finally restore the true maxillary and mandibular positional relationship within the exit.

[0056] S33. Based on the objective function, the optimal rotation matrix is ​​solved by weighted centroid calculation, centroid-free coordinate transformation, and singular value decomposition of the covariance matrix. and the optimal translation vector matrix .

[0057] The formula for calculating the weighted centroid is as follows: ; in, A point set of matter points The weighted centroid, Let be the weighted centroid of the target point set, where the target point is the coordinate of a material point on the surface of the interlocking recording material under interlocking force, and the target point set is the collection of target points.

[0058] The calculation method for centroid-free coordinate transformation is as follows: ; In the formula, To find the centroid coordinates of the matter point, Find the centroid coordinates of the target point.

[0059] The singular value decomposition (SVD) of the covariance matrix is ​​calculated as follows: ; In the formula, Let covariance matrix be the variance matrix. Covariance matrix The right singular vector matrix of singular value decomposition. Covariance matrix The diagonal matrix of singular value decomposition. Covariance matrix The left singular vector matrix of singular value decomposition.

[0060] Finally, calculate the optimal rotation matrix. and the optimal translation vector matrix Optimal rotation matrix and the optimal translation vector matrix The formula is as follows: ; .

[0061] S34, Based on the optimal rotation matrix and the optimal translation vector matrix Obtain the optimal rigid compensation transformation matrix ; .

[0062] S4. Based on the optimal rigidity compensation transformation matrix, the initial relative pose matrix of the mandible and jaw is compensated to obtain the corrected relative pose matrix of the mandible and jaw. .

[0063] In this step, the optimal rigidity compensation transformation matrix is ​​multiplied by the initial relative pose matrix of the mandible and jaw to obtain the corrected relative pose matrix of the mandible and jaw, i.e.: .

[0064] In the formula, This is the initial relative pose matrix of the upper and lower jaws.

[0065] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a deep learning-based occlusal recording jaw position compensation method.

[0066] This invention has the following advantages: 1. This invention is the first to address the problem of inaccurate jaw position relationship caused by in vitro elastic deformation of occlusal records. It applies deep learning technology to the field of accurate prediction of the three-dimensional elastic deformation field of occlusal records, filling the technical gap that existing technologies cannot quantify and reverse compensate for in vitro elastic deformation. It fundamentally overcomes the long-standing industry accuracy bottleneck of inconsistency between the actual intraoral occlusal position and the jaw position registered by extraoral scanning, and provides a brand-new path to improve the accuracy of digital jaw position relationship transfer.

[0067] 2. This invention adopts a multimodal feature fusion architecture, using four types of multi-source information as network inputs: dental arch morphological features, material viscoelastic constitutive parameters, occlusal record spatial geometric features, and occlusal force vectors. This endows the model with general predictive capabilities across patients, materials, and occlusal conditions. In the clinical application stage, only the trained network model needs to be called to complete millisecond-level rapid inference, eliminating the need for time-consuming and costly finite element mechanical simulations for individual cases. This significantly improves clinical diagnostic and treatment efficiency while ensuring accuracy.

[0068] 3. This invention embeds a physical residual term based on the equilibrium equation of elasticity and the constitutive relationship of materials into the network loss function, supplemented by a material consistency constraint term, so that the deformation field output by the network strictly follows the objective laws of mechanics, effectively reducing the model's dependence on the number of training samples; even under the condition of limited training samples, it can still output deformation prediction results with physical rationality, significantly improving the model's prediction reliability, working condition extrapolation ability and prediction accuracy for untrained cases, ensuring the stability and accuracy of clinical applications.

[0069] 4. This invention constructs a fully automated error correction closed loop of "intelligent prediction of deformation field + optimal rigidity compensation transformation solution + precise correction of jaw position relationship". The accuracy of jaw position relationship after compensation can reach the micrometer level, which can effectively improve the occlusal accuracy and clinical adaptability of restorations. This method can be directly embedded into the standard workflow of existing digital dental CAD software without adding additional hardware equipment or changing the routine clinical operation process. It has a low threshold for implementation and is easy to promote and apply on a large scale.

[0070] This invention is not limited to the preferred embodiments described above. Anyone can derive other products in various forms under the guidance of this invention. However, regardless of any changes in shape or structure, any technical solution that is the same as or similar to this application falls within the protection scope of this invention.

Claims

1. A deep learning-based method for jaw position compensation in occlusion recording, characterized in that, Includes the following steps: S1. Obtain the digital model of the patient's maxilla, mandible, and occlusion record respectively. Based on the occlusion imprint features on the occlusion record digital model, register the digital model of the maxilla and mandible with the occlusion record digital model respectively to obtain the initial relative pose matrix of the maxilla and mandible. S2. Extract the dental arch morphology features, occlusal recording material parameters, occlusal recording geometric information, and occlusal force vector corresponding to the patient. Input the above multi-source features into the pre-trained deep learning deformation prediction model. Predict the point-by-point three-dimensional displacement deformation field of the occlusal recording from the intraoral force-fixed state to the off-body free state through the deep learning deformation prediction model. S3. Extract the displacement vectors corresponding to each material point in the mandibular tooth imprint region from the point-by-point three-dimensional displacement deformation field, and use the rigid body motion estimation algorithm to solve the optimal rigid compensation transformation matrix for offsetting the elastic deformation of the occlusal record. S4. The initial relative pose matrix of the upper and lower jaws is compensated based on the optimal rigidity compensation transformation matrix to obtain the corrected relative pose matrix of the upper and lower jaws.

2. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, In step S1, the ICP algorithm is used to register the maxillary digital model, mandibular digital model and occlusal record digital model; the initial relative pose matrix of the maxilla and mandible is the initial transformation matrix of the mandibular digital model relative to the maxillary digital model.

3. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, In step S2, the dental arch morphological features are composed of global morphological codes extracted from the point clouds of the maxillary and mandibular dentition by a pre-trained point cloud autoencoder, or are composed of geometric measurement parameters including dental arch width, occlusion, and overlay; the material physical parameters are the viscoelastic constitutive model parameters corresponding to the material of the current occlusion record. Occlusal record geometry information is the spatial geometric features of the occlusal record digital model, which includes the thickness distribution field of the occlusal record digital model; The bite force vector includes the magnitude, direction, and area of ​​the force applied during bite recording.

4. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, In step S2, the deep learning deformation prediction model adopts a network architecture based on point cloud or sparse three-dimensional convolution. The deep learning deformation prediction model is equipped with a conditional injection module, which is used to fuse multi-source features into the geometric coding branch of the deep learning deformation prediction model to output a point-by-point three-dimensional displacement deformation field.

5. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, In step S2, the loss function of the deep learning deformation prediction model consists of a data fidelity term, a physical residual term, and a material consistency constraint term. The physical residual term is constructed based on the elasticity equilibrium equation and the material constitutive relation. The material consistency constraint term is used to constrain the physical consistency of the deformation prediction results under the same material parameters.

6. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, The specific method for step S3 is as follows: S31: Segment the mandibular tooth imprint region from the occlusal record digital model, obtain the point set of each material point in the region, and extract the displacement vector corresponding to each point from the point-by-point three-dimensional displacement deformation field. Assign a corresponding weight to each material point. ; S32. Construct the objective optimization function for rotation matrix R and translation vector matrix t; The objective optimization function is: ； S33. Based on the objective function, the optimal rotation matrix is ​​solved by weighted centroid calculation, centroid-free coordinate transformation, and singular value decomposition of the covariance matrix. and the optimal translation vector matrix ; S34, Based on the optimal rotation matrix and the optimal translation vector matrix Obtain the optimal rigid compensation transformation matrix ; 。 7. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, In step S4, the optimal rigidity compensation transformation matrix is ​​multiplied by the initial relative pose matrix of the mandible and jaw to obtain the corrected relative pose matrix of the mandible and jaw.

8. The deep learning-based occlusal recording jaw position compensation method according to claim 1, characterized in that, The method for constructing the deep learning deformation prediction model is as follows: N1: Constructing a training dataset: Based on finite element simulation, we traverse various working conditions of digital models of the upper and lower jaws with different dental arch shapes, constitutive parameters of different occlusal recording materials, and different occlusal force vectors. We simulate the unloading process of the occlusal recording material from the intraoral stress-cured state to the detached free state under each working condition, and generate the three-dimensional deformation field corresponding to each working condition as the training ground value. N2: Constructing sample input features: For each training sample, extract the corresponding dental arch morphology features, material physical parameters, occlusal record geometric information, and occlusal force vector, and fuse them into network input features; N3: Iterative training of the network: The structured input features are input into the deep learning network, and iterative training is performed with a preset loss function as the optimization objective until the loss function converges, resulting in a trained deep learning deformation prediction model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based occlusal recording jaw position compensation method as described in any one of claims 1-8.

Images