A personalized tooth morphology restoration and reconstruction method based on deep learning

Abstract

The application provides a personalized tooth morphology recovery and reconstruction method based on deep learning, and belongs to the technical field of artificial intelligence graphics processing and tooth morphology. The method comprises the following steps: obtaining a natural tooth anatomical crown model to form a data set; constructing a ToothDIT network model comprising a DeformNet subnetwork and an SDFDecoder subnetwork; automatically sampling data samples by the model, and converting the sampling points into spatial coordinates to calculate the SDF values of the coordinates to the corresponding tooth surface, and uniformly representing the tooth morphology; using the data set to complete the training and verification of the model, predicting the unique identification of the remaining tooth morphology of the defective tooth, driving the conversion from the tooth template to the corresponding tooth morphology, recovering the original morphology of the defective tooth, and performing subsequent oral repair; the application can customize an accurate digital oral repair reference morphology, not only considering the individual difference characteristics, but also greatly improving the efficiency of oral repair morphology design.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence graphics processing and tooth morphology technology, specifically a method for personalized tooth morphology restoration and reconstruction based on deep learning. Background Technology

[0002] With the continuous advancement and innovation of digital technology, the practice of tooth morphology restoration in the field of dental prosthodontics has undergone significant changes. The traditional manual stage, including the precise fabrication of models and prototypes, time-consuming manual molding, and complex firing processes, is gradually being replaced by advanced computer-aided design (CAD) and computer-aided manufacturing (CAM) technologies. This change has greatly improved the efficiency of restoration fabrication and has also resulted in a qualitative leap in accuracy and aesthetic effects.

[0003] In modern dental restoration, the application of tooth morphology templates represents a significant breakthrough. With pre-set digital template libraries, technicians no longer need to design each restoration from scratch. Instead, they can select existing templates similar to the target tooth and make detailed adjustments, including size matching, positional calibration, and personalized refinement of morphological details. This not only shortens the design cycle but also ensures more accurate and natural tooth morphology reconstruction. However, the template resources built into current digital design software have limitations. For example, the difference between the template and the patient's actual tooth morphology may be significant, the selection range is limited, and it heavily relies on the operator's subjective judgment. These factors often lead to a lengthy design process, requiring doctors to repeatedly try on and finely adjust restorations in clinical practice, thus prolonging the patient's treatment cycle and reducing patient satisfaction.

[0004] Meanwhile, artificial intelligence technologies, especially generative adversarial networks (GANs) in deep learning, have begun to permeate the design process of dental restorations, aiming to address issues of efficiency and accuracy. Nevertheless, when dealing with anterior teeth regions with complex aesthetic requirements, particularly in cases of missing prominent cusps, current AI technologies still struggle to achieve the level of naturalness and personalized design comparable to that of professional technicians. Furthermore, when facing complex oral defects such as occlusal reconstruction, current AI algorithms lack sufficient interpretability and robustness; their decision-making processes in complex situations are difficult to understand, and they are not robust enough to changes in input data. These challenges urgently require further research and refinement. Summary of the Invention

[0005] Based on the current state of the art, the purpose of this invention is to deeply integrate and innovatively apply deep learning technology with implicit template construction strategies to solve the problem of precise restoration of damaged teeth in the field of oral medicine. The method of this invention can tailor highly personalized and detailed restoration templates for each patient. These templates not only fully consider individual differences but also provide a solid scientific basis and precise reference coordinates for subsequent tooth restoration design. This method significantly improves the refinement, accuracy, and long-term stability of tooth morphology restoration plans, thereby effectively reducing the uncertainties traditionally reliant on technicians' personal experience and subjective judgment.

[0006] The method of this invention has been validated, and the final prosthesis design completed under the guidance of the new template demonstrates outstanding performance in both aesthetics and functional matching. It also verifies the significant advantage of this template-based design process in improving overall design efficiency and further explores the potential revolutionary impact of artificial intelligence on the dental prosthesis design industry.

[0007] The design process optimized by deep learning has been significantly simplified, which powerfully demonstrates that CAD technology not only has great potential in improving design efficiency, but also has profound significance for revolutionizing traditional interactive design patterns and promoting the development of digital healthcare.

[0008] The present invention employs the following technical solutions to achieve its objective:

[0009] A personalized tooth morphology restoration and reconstruction method based on deep learning, the method comprising the following steps:

[0010] S1. Based on the task requirement of restoring the morphology of missing teeth, tooth morphology data of each tooth position is obtained and preprocessed to obtain a dataset of natural teeth with complete anatomical crown morphology for model training.

[0011] S2. Based on SDF theory and DeepSDF network structure, an improved DIT algorithm is constructed, forming the ToothDIT network model, which includes the DeformNet sub-network and the SDFDecoder sub-network. The ToothDIT network model is used to deform the trained template tooth and restore the morphology of the defective tooth while maintaining the original scale of the tooth morphology.

[0012] S3. In the ToothDIT network model, multi-point sampling is performed on the preprocessed dataset. The DeformNet sub-network is used to transform the sampling points in the dataset into spatial coordinates. The SDFDecoder sub-network is used to calculate the SDF value from any sampling point to the corresponding tooth surface, thereby achieving a unified representation of tooth morphology.

[0013] S4. Use the dataset to train and validate the ToothDIT network model, enabling it to have model inference capabilities. Obtain the average template of the corresponding tooth position in the dataset and use the inferred unique identifier to complete the conversion from the tooth template to the corresponding tooth shape, thereby generating personalized tooth shapes and applying them to actual tasks of restoring the shape of missing teeth.

[0014] S5. In practical applications, the data of the missing teeth that need to be restored are obtained, and after using the same preprocessing method, they are registered and aligned with the tooth template. Then, the ToothDIT network model is used to complete the inference prediction to obtain the complete tooth shape after restoration, which serves as the basis for the restoration and reconstruction of the missing tooth shape, and the digital restoration is made.

[0015] Preferably, during data preprocessing, the tooth morphology data of each tooth position are processed graphically in a uniform manner based on the centroid and orientation; using the centroid matching method, the tooth morphology data of each tooth position are first roughly aligned according to the same standard, and then the cervical ends of the two are manually aligned to reflect the difference in the incisal edge morphology of the teeth, and this difference is also reflected in the dataset, thus completing the integration of the dataset; the integrated dataset is divided into a training set and a validation set for the training and validation of the ToothDIT network model.

[0016] Furthermore, for steps S2 and S3, based on SDF theory, a unified representation method for tooth morphology is established by constructing an implicit shape space. To represent the shape of all teeth; and at the same time establish a mapping Implicit shape space Any point, i.e., potential code Map to a specific tooth shape in a standard set of tooth shapes; when characterizing tooth morphology, perform multi-point sampling for each tooth shape in the dataset, collect the 3D coordinates (x, y, z) of the sampling points, and then calculate the corresponding SDF value.

[0017] Furthermore, based on the DeepSDF network structure, deep learning techniques are used to generate SDF functions. The specific process is as follows: based on the autodecoder, a DeepSDF network structure is established, and mappings are constructed. The latent code of the shape The signed distance from any point p in space to the shape. Furthermore, based on the DeepSDF network structure, the mapping f is represented as... This allows for the reconstruction and completion of a type of shape.

[0018] Furthermore, the DeformNet subnetwork is used to extract the tooth template shape for the corresponding tooth position. Then, based on the shape features of the missing tooth, it infers the potential unique identifier representing the tooth shape and deforms the tooth template shape to obtain the complete shape. The extracted tooth template shape corresponds to the common features of that type of tooth. This process is achieved through DIT. The DeformNet subnetwork can thus learn the tooth shape variations between individuals.

[0019] Furthermore, in step S4, when training the ToothDIT network model, it is instructed to generate implicit template teeth for each tooth position. The implicit template teeth contain the basic features common to teeth in the dataset, and through training, the ToothDIT network model is enabled to predict latent codes. The latent codes are used as unique identifiers for storing conversion information to realize the conversion from tooth templates to corresponding tooth morphologies, thereby generating personalized tooth morphologies.

[0020] When performing model inference in the ToothDIT network model, the parameters of the trained network model are fixed in advance. Then, based on the incomplete shape point cloud data of the missing teeth, the potential code of the complete shape is inferred by backpropagation iteration, the SDF value of each sampling point coordinate is generated, and the corresponding zero isosurface is extracted into a grid of arbitrary precision using the Marching Cubes algorithm, thereby achieving the subsequent tooth morphology restoration and reconstruction.

[0021] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows:

[0022] Compared to traditional computer-aided design methods, this invention utilizes the ToothDIT network model for morphological restoration of tooth defects, demonstrating significant advantages. This method not only greatly improves the efficiency of the restoration process and significantly reduces the time required from diagnosis to achieving the target morphology, but also achieves a breakthrough in accuracy.

[0023] Through precise comparison of three-dimensional spatial morphology, it can be found that the ToothDIT network model achieves significantly better tooth morphology reconstruction accuracy than traditional design methods. The deviation between the generated restoration and the ideal three-dimensional morphology of real teeth is significantly reduced, which means that it can more accurately simulate and reproduce the complex three-dimensional structural features of natural teeth.

[0024] When focusing on the functional and aesthetic performance at the level of two-dimensional details, the ToothDIT network model also demonstrates its outstanding capabilities. Whether it is the subtle occlusal texture and proximal relationships, or the important aesthetic lines and shapes of the anterior teeth, the ToothDIT network model can restore them with higher accuracy, ensuring that the two-dimensional morphological difference between the final restoration and the natural teeth is minimal, thereby providing patients with a more natural and harmonious oral function and aesthetic effect.

[0025] Furthermore, as the extent and complexity of tooth defects increase, the ToothDIT network model maintains excellent stability, consistently demonstrating high design efficiency in restorations of various sizes of defects. This is particularly valuable clinically for handling complex and large-area tooth defect cases. Overall, the application of the ToothDIT network model not only accelerates the restoration of tooth defect morphology but also comprehensively surpasses traditional CAD techniques in terms of accuracy, stability, and detail reproduction. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0027] Figure 2 A schematic diagram of obtaining a tooth model from CBCT data;

[0028] Figure 3 This is a schematic diagram illustrating the process of obtaining and registering anatomical crowns.

[0029] Figure 4 A schematic diagram for preliminary matching of missing teeth;

[0030] Figure 5 This is a diagram illustrating the manual apex alignment of a missing tooth.

[0031] Figure 6 This is a schematic diagram illustrating the effect of the SDF sampling process;

[0032] Figure 7 This is a schematic diagram illustrating the process of gradually sampling points around the teeth.

[0033] Figure 8 This is a schematic diagram of the ToothDIT network model during the training phase.

[0034] Figure 9 This is a schematic diagram of the ToothDIT network model during the inference and prediction phase.

[0035] Figure 10 This is a schematic diagram of the Marching Cubes algorithm. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0037] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0038] Example

[0039] A personalized tooth morphology restoration and reconstruction method based on deep learning. Figure 1 The flowchart of this method is shown and can be viewed simultaneously. The steps of this method can be summarized as follows:

[0040] S1. Based on the task requirement of restoring the morphology of missing teeth, tooth morphology data of each tooth position is obtained and preprocessed to obtain a dataset of natural teeth with complete anatomical crown morphology for model training.

[0041] S2. Based on SDF theory and DeepSDF network structure, an improved DIT algorithm is constructed, forming the ToothDIT network model, which includes the DeformNet sub-network and the SDFDecoder sub-network. The ToothDIT network model is used to deform the trained template tooth and restore the morphology of the defective tooth while maintaining the original scale of the tooth morphology.

[0042] S3. In the ToothDIT network model, multi-point sampling is performed on the preprocessed dataset. The DeformNet sub-network is used to transform the sampling points in the dataset into spatial coordinates. The SDFDecoder sub-network is used to calculate the SDF value from any sampling point to the corresponding tooth surface, thereby achieving a unified representation of tooth morphology.

[0043] S4. Use the dataset to train and validate the ToothDIT network model, enabling it to have model inference capabilities. Obtain the average template of the corresponding tooth position in the dataset and use the inferred unique identifier to complete the conversion from the tooth template to the corresponding tooth shape, thereby generating personalized tooth shapes and applying them to actual tasks of restoring the shape of missing teeth.

[0044] S5. In practical applications, the data of the missing teeth that need to be restored are obtained, and after using the same preprocessing method, they are registered and aligned with the tooth template. Then, the ToothDIT network model is used to complete the inference prediction to obtain the complete tooth shape after restoration, which serves as the basis for the restoration and reconstruction of the missing tooth shape, and the digital restoration is made.

[0045] This embodiment will describe the details of each step in the above process sequence.

[0046] First, in step S1, data collection and processing are performed. Tooth morphology data for each tooth position is acquired using CBCT scanning and intraoral scanning, and saved in DICOM format. Subsequently, in Mimics software, the anatomical morphology of various teeth (such as the right maxillary central incisor) is extracted, reconstructed, and exported as OBJ format. (See attached image.) Figure 2 The illustration. Then use the Geomagic software, such as... Figure 3 As shown, based on the center of gravity and direction, the morphological data of teeth in each position are processed in a unified manner using graphics, and a standard database of three-dimensional morphological models of natural tooth crowns in each position is established. Taking the 28 teeth that a person has as an example, 28 databases can be established, corresponding to the teeth in these 28 positions.

[0047] This data collection and processing process is to ensure the consistency of shape representation in subsequent SDF-related content, thereby ensuring the accuracy of size restoration during morphological restoration. However, in clinical applications, more attention needs to be paid to the morphological changes of the incisal edge of the tooth. Therefore, the preprocessing process in this embodiment can preferably be as follows: using the centroid matching method, the morphological data of each tooth position are first roughly aligned according to the same standard (in this embodiment, the 001 natural crown model is used as the benchmark), and then the cervical ends of the two are manually aligned to reflect the difference in the morphology of the incisal edge of the tooth, and this difference is also reflected in the dataset, thus completing the integration of the dataset.

[0048] In this embodiment, the integrated dataset is divided into a training set and a validation set. 80 cases can be randomly selected from the overall data pool as the training set and 10 cases as the validation set for subsequent training and validation of the ToothDIT network model.

[0049] Steps S2 and S3 form a coherent, progressive explanation to clearly illustrate the characteristics of the DeformNet and SDFDecoder subnetworks in this embodiment. To establish a unified representation method for tooth shape, this embodiment is based on SDF theory and constructs an implicit shape space. To represent the shape of all teeth; and at the same time establish a mapping Implicit shape space Any point, i.e., potential code Mapped to a specific tooth shape in a standard set of tooth profiles.

[0050] Figure 6 The SDF sampling process is shown, where the white outline area represents the shape boundary of the tooth morphology; Figure 7 This also shows the process of gradually sampling points around the teeth.

[0051] To achieve the above method, this embodiment samples 500,000 points for each tooth shape and collects the 3D coordinates (x, y, z) of the sampled points, and then calculates the corresponding SDF value, i.e., the signed distance field value.

[0052] The SDF value is defined as the minimum Euclidean distance from each sampling point to the tooth surface; if a point is located outside the tooth, its distance is considered positive, while points located inside the tooth are assigned a negative distance.

[0053] The SDF value calculation process is as follows:

[0054]

[0055]

[0056] In the formula, The zero isosurface is represented by {p|SDF(p)=0}; inf represents the lower bound; p represents any point in space; Represents the set of real numbers; in this embodiment, the 3D coordinates and corresponding SDF values ​​of each sampling point in the dataset are used as uniform inputs during the training of the ToothDIT network model.

[0057] Next, based on the DeepSDF network structure and related technologies, this embodiment uses deep learning technology to generate SDF functions. The specific process is as follows:

[0058] Based on the automatic decoder, a DeepSDF network structure is established, and a mapping is constructed. The latent code of the shape The signed distance from any point p in space to the shape. Furthermore, based on the core concepts of the DeepSDF network structure, the mapping f is represented as... This allows for the reconstruction and completion of a type of shape.

[0059] In this embodiment, based on SDF theory and DeepSDF network structure, the SIREN architecture is used to construct the SDFDecoder subnetwork; the SDFDecoder subnetwork is used as the implementation of the tooth template symbol distance field and is used to reflect the concave and convex details of the tooth surface; the SDFDecoder subnetwork includes multiple hidden nodes distributed in multiple layers, preferably including 256 hidden nodes in 4 layers.

[0060] The DeformNet subnetwork is used to extract the tooth template shape for the corresponding tooth position. Then, based on the characteristic shape of the missing tooth, the tooth template shape is deformed to obtain a more intuitive and complete shape. The extracted tooth template shape corresponds to the common features of that type of tooth, providing good guidance. This process is implemented using the DIT method, as shown in the following equation:

[0061]

[0062] In the formula, Any point p in this space represents the query position p, and it is transformed into a new position p′ in the signed distance domain of the tooth template, representing the latent code of the shape. The correspondence between the template and the tooth; subsequently, the DeformNet subnetwork is used. Let SDF be used to represent the value of any point p′ in the new space of the tooth template, since it is independent of the latent code. That is, it is independent of specific tooth shapes, so it only needs to learn general information about tooth shape, which allows the DeformNet subnetwork to learn tooth shape variations between individuals.

[0063] In this embodiment, the DeformNet subnetwork structure includes a multilayer perceptron and a spatial transform LSTM network, wherein a single layer of the multilayer perceptron has multiple hidden nodes; preferably, a multilayer perceptron with 512 hidden nodes in a single layer can be used as a feature, combined with a spatial transform LSTM network to realize the DeformNet subnetwork.

[0064] The DeformNet sub-network, as a hybrid model that simultaneously includes a multilayer perceptron and a spatial transformation LSTM network, can fully utilize the LSTM's ability to model dynamic sequences, the multilayer perceptron's ability to learn feature representations, and the spatial transformation layer's ability to adaptively correct the input data geometrically. This allows it to achieve better performance when dealing with complex problems in both time and space dimensions, making it suitable for the scenario described in this embodiment.

[0065] When the DeformNet subnetwork transforms any point p in 3D space into a new position p′ in the signed distance domain of the tooth template, it does so by predicting the offset Δp of point p, i.e., p′=p+Δp. In this embodiment, the data is processed at the original scale of the tooth, so only the translation transformation of the shape needs to be implemented.

[0066] The main structure of the DeformNet subnetwork is represented as follows:

[0067]

[0068] In the formula, φ and ψ are both unit states; the DeformNet subnetwork is iterated N times in the above manner, and the point after the i-th iteration is represented as p. i =p i-1 +Δp i The spatial correspondence between the final specific shape and the shape of the tooth template is represented as follows:

[0069] Through the above process, implicit shape space The mapping relationship to specific tooth morphology is as follows:

[0070]

[0071] Finally, the DeformNet subnetwork and the SDFDecoder subnetwork are combined to obtain the ToothDIT network model.

[0072] In step S4 of this embodiment, when training the ToothDIT network model, it generates an implicit tooth template. The tooth template contains the basic features common to teeth in the dataset. Through training, the ToothDIT network model is equipped with the function of predicting latent codes. The latent codes are used as unique identifiers for storing conversion information to realize the conversion from the tooth template to the corresponding tooth shape, thereby generating personalized tooth shapes.

[0073] Figure 8 This diagram illustrates the ToothDIT network model during the training phase. During training, a complete tooth is sampled using SDF and then fed into the ToothDIT network model. Guided by the loss function, the sampled data passes through the DeformNet sub-network and the SDFDecoder sub-network to generate a corresponding implicit template, obtaining the latent code for that tooth and updating the network model's parameters. When the next tooth is input, both the implicit template and the network model parameters are updated again until the later stages of training, when they gradually stabilize.

[0074] When performing model inference in the ToothDIT network model, the parameters of the trained network model are fixed in advance. Then, based on the SDF values ​​of the sampling points of the incomplete shape of the missing tooth, the latent code of the complete shape is constructed in a backpropagation iterative manner. The Marching Cubes algorithm is used to extract the corresponding zero isosurface into a grid of arbitrary precision, thereby achieving the subsequent tooth morphology restoration and reconstruction.

[0075] Figure 9 This diagram illustrates the ToothDIT network model during the inference and prediction phase. In this phase, the missing tooth is sampled via SDF and input into the pre-trained ToothDIT network model. The SDFDecoder sub-network predicts the latent code that characterizes the tooth's morphological features, thereby driving the implicit template to deform into the desired complete tooth shape. This embodiment employs the Marching Cubes algorithm to extract the implicitly expressed morphology into an explicit mesh, facilitating the subsequent morphological reconstruction process.

[0076] like Figure 10 As shown, the Marching Cubes algorithm is an algorithm for extracting isosurfaces from 3D data fields. It provides a standardized method for extracting meaningful surface structures from 3D datasets, which is of great value for understanding and visualizing 3D data.

[0077] In this embodiment, when training the ToothDIT network model, the optimal mapping f of the model is found by optimizing the loss function; the loss function of the ToothDIT network model includes the DeformNet loss. SDFDecoder loss and the regularization loss of potential code loss function for:

[0078]

[0079] Among them, DeformNet loss for:

[0080]

[0081] In the formula, This represents pointwise loss, used to constrain the DeformNet subnetwork to minimize the displacement of each point and avoid incorrect correspondence between a specific shape and the template shape. The point-to-point loss is used to constrain the DeformNet subnetwork to create a smooth deformation of the symbolic distance field from a specific shape to a template shape, preventing extreme distortion of the symbolic distance field; λ pw With λ ppThese are the weighting coefficients for the two losses;

[0082] SDFDecoder loss in the i-th iteration for:

[0083]

[0084] In the formula, K represents the total number of categories; Ω k This represents the set of samples belonging to the k-th category.

[0085] because The limitation on the network's predicted SDF to closely approximate reality means that this embodiment uses Curriculum here. The method controls the effective symbolic distance field within the strip space centered on the isosurface;

[0086] Regularization loss of potential code for:

[0087]

[0088] In the formula, Represents the scaling factor; Potential code representing the k-th category Corresponding vector The square of the L2 norm.

[0089] In this embodiment, during the final practical application of the fully trained ToothDIT network model, the patient's CBCT scan and / or intraoral scan data are first acquired. These high-precision data sources are the foundation for achieving tooth restoration in practical applications. Subsequently, the damaged teeth to be restored are meticulously segmented and extracted to ensure independent imaging, and a three-dimensional mesh is constructed and optimized for subsequent morphological restoration operations.

[0090] Next, the damaged teeth that have undergone the same preprocessing method are precisely matched and aligned with the standard template teeth (or the No. 001 natural tooth in the database) generated by training the ToothDIT network model to ensure that the repaired part can perfectly match the original tooth structure.

[0091] Furthermore, the trained ToothDIT network model is used to predictively restore missing teeth, generating complete tooth morphologies that meet biomechanical and aesthetic requirements. The ideal morphology inferred from the model will be directly applied to the specific morphological restoration of the patient's missing teeth and the design and fabrication of the restoration.

[0092] As can be seen, the method and process of this embodiment can be seamlessly integrated into the existing computer-aided design and manufacturing system, realizing the successful application of deep learning in personalized and precise tooth morphology restoration and restoration fabrication, thereby improving the restoration effect and promoting the intelligent process of oral restoration technology.

Claims

1. A personalized tooth morphology restoration and reconstruction method based on deep learning, characterized in that, The method includes the following steps: S1. Based on the task requirement of restoring the morphology of missing teeth, tooth morphology data of each tooth position is obtained and preprocessed to obtain a dataset of natural teeth with complete anatomical crown morphology for model training. S2. Based on SDF theory and DeepSDF network structure, an improved DIT algorithm is constructed, forming the ToothDIT network model, which includes the DeformNet sub-network and the SDFDecoder sub-network. The ToothDIT network model is used to deform the trained template tooth and restore the morphology of the defective tooth while maintaining the original scale of the tooth morphology. S3. In the ToothDIT network model, multi-point sampling is performed on the preprocessed dataset. The DeformNet sub-network is used to transform the sampling points in the dataset into spatial coordinates. The SDFDecoder sub-network is used to calculate the SDF value from any sampling point to the corresponding tooth surface, thereby achieving a unified representation of tooth morphology. S4. Use the dataset to train and validate the ToothDIT network model, enabling it to have model inference capabilities. Obtain the average template of the corresponding tooth position in the dataset and use the inferred unique identifier to complete the conversion from the tooth template to the corresponding tooth shape, thereby generating personalized tooth shapes and applying them to actual tasks of restoring the shape of missing teeth. S5. In practical applications, the data of the missing teeth that need to be restored are obtained, and after using the same preprocessing method, they are registered and aligned with the tooth template. Then, the ToothDIT network model is used to complete the inference prediction to obtain the complete tooth shape after restoration, which serves as the basis for the restoration and reconstruction of the missing tooth shape, and the digital restoration is made. In step S4, when training the ToothDIT network model, it is instructed to generate implicit template teeth for each tooth position. The implicit template teeth contain the basic features common to teeth in the dataset. Through training, the ToothDIT network model is enabled to predict latent codes. The latent codes are used as unique identifiers for storing conversion information to realize the conversion from tooth templates to corresponding tooth morphologies, thereby generating personalized tooth morphologies. When performing model inference in the ToothDIT network model, the parameters of the trained network model are fixed in advance. Then, based on the SDF values ​​of the sampling points of the incomplete shape of the missing tooth, the potential code of the complete shape is inferred by backpropagation iteration, the SDF values ​​of the coordinates of each sampling point are generated, and the corresponding zero isosurface is extracted into a grid of arbitrary precision using the Marching Cubes algorithm, thereby achieving the subsequent tooth morphology restoration and reconstruction.

2. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 1, characterized in that: In step S1, three-dimensional reconstruction is performed using CBCT scanning and intraoral scanning to obtain tooth morphology data for each tooth position, and the anatomical morphology of the natural teeth in each tooth position is extracted. Then, based on the center of gravity and direction, the tooth morphology data for each tooth position is preprocessed and integrated to obtain a dataset.

3. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 2, characterized in that: During data preprocessing, the tooth morphology data of each tooth position are processed graphically in a uniform manner according to the centroid and orientation. Using the centroid matching method, the tooth morphology data of each tooth position are first roughly aligned according to the same standard, and then the cervical ends of the two are manually aligned to reflect the difference in the incisal edge morphology of the teeth. This difference is also reflected in the dataset, thus completing the integration of the dataset. The integrated dataset is divided into training set and validation set for training and validation of the ToothDIT network model.

4. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 1, characterized in that: In steps S2 and S3, based on SDF theory, a unified representation method for tooth morphology is established by constructing an implicit shape space. To represent the shape of all teeth; and simultaneously establish a mapping. implicit shape space Any point, i.e., potential code Mapped to a specific tooth shape within a standard set of tooth profiles; When characterizing tooth morphology, multiple points are sampled for each tooth morphology in the dataset, and the 3D coordinates of the sampling points are collected. Then calculate the corresponding SDF value, and define the SDF value as the minimum Euclidean distance from each sampling point to the tooth surface; The SDF value calculation process is as follows: In the formula, Represents the zero isosurface. ; Represents the lower limit; Represents any point in space; Represents the set of real numbers; the 3D coordinates and corresponding SDF values ​​of each sampling point in the dataset are used as uniform inputs during the training of the ToothDIT network model.

5. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 4, characterized in that, Based on the DeepSDF network structure, deep learning techniques are used to generate SDF functions. The specific process is as follows: Based on the automatic decoder, a DeepSDF network structure is established, and mappings are constructed. The latent code of the shape With any point in space Symbolic distance mapped to the shape Furthermore, based on the DeepSDF network structure, the mapping is... Represented as This allows for the reconstruction and completion of a type of shape.

6. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 5, characterized in that: Based on SDF theory and DeepSDF network structure, the SIREN architecture is used to construct the SDFDecoder subnetwork. The SDFDecoder subnetwork is used as the implementation of the tooth template symbol distance field and is used to reflect the concave and convex details of the tooth surface. The SDFDecoder subnetwork includes multiple hidden nodes distributed in multiple layers.

7. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 6, characterized in that: The DeformNet subnetwork is used to extract the tooth template shape for the corresponding tooth position. Then, based on the shape features of the missing tooth, it infers the potential unique identifier representing the tooth shape and deforms the tooth template shape to obtain the complete shape. The extracted tooth template shape corresponds to the common features of that type of tooth. This process is implemented using the DIT method, as shown in the following formula: In the formula, Any point in this space This represents the query location. And convert it into a new position in the signed distance domain of the tooth template. , representing the potential code of the shape Correspondence with the dental template; Subsequently, using the DeformNet subnetwork To represent any point in the new space of the tooth template The SDF value allows the DeformNet subnetwork to learn the tooth shape variations between individuals.

8. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 7, characterized in that: The DeformNet subnetwork consists of a multilayer perceptron and a spatial transformation LSTM network, where a single layer of the multilayer perceptron has multiple hidden nodes; the DeformNet subnetwork transforms any point in 3D space... Convert to a new position in the signed distance domain of the tooth template At that time, through the predicted point offset Implementation, that is ; The structure of the DeformNet subnetwork is represented as follows: In the formula, and All are cell states; the DeformNet subnetwork is processed in the manner described above. The iteration, the... The point after the nth iteration is represented as The spatial correspondence between the final specific shape and the shape of the tooth template is represented as follows: ; Through the above process, implicit shape space The mapping relationship to specific tooth morphology is as follows: Finally, the DeformNet subnetwork and the SDFDecoder subnetwork are combined to obtain the ToothDIT network model.

9. The method for personalized tooth morphology restoration and reconstruction based on deep learning according to claim 1, characterized in that: During the training of the ToothDIT network model, the optimal mapping of the model is found by optimizing the loss function. The loss function of the ToothDIT network model includes the DeformNet loss. SDFDecoder loss and the regularization loss of potential code loss function for: Among them, DeformNet loss for: In the formula, This represents pointwise loss, used to constrain the DeformNet subnetwork to minimize the displacement of each point and avoid incorrect correspondence between a specific shape and the template shape. The point-to-point loss is used to constrain the DeformNet subnetwork to create a smooth deformation of the symbolic distance field from a specific shape to a template shape, preventing extreme distortion of the symbolic distance field. and These are the weighting coefficients for the two losses; No. SDFDecoder loss in the next iteration for: In the formula, Represents the total number of categories; Representing the A set of samples for each category; Curriculum DeepSDF is used here. The method controls the effective symbolic distance field within the strip space centered on the isosurface; Regularization loss of potential code for: In the formula, Represents the scaling factor; Representing the Potential code of the category Corresponding vector The square of the L2 norm.

Images