Digital false tooth mounting and positioning device based on three-dimensional scanning and method thereof

Abstract

The invention relates to the technical field of oral implantation guidance, and provides a digital false tooth mounting and positioning device based on three-dimensional scanning and a method thereof. The method comprises the following steps: scanning the oral cavity of a patient through an oral cavity scanner to generate a three-dimensional oral cavity digital model; generating a virtual implant of the patient and corresponding initial implantation parameters based on the three-dimensional oral cavity digital model through a SegFormer segmentation network; based on historical operation data, initial implantation parameters, bone density, patient attributes and occlusal force parameters, a long sequence prediction network is adopted to predict the position offset and bone resorption degree of the virtual implant after osseointegration through a segmented attention mechanism and a network memory cache mechanism, and optimal implantation parameters are obtained; and pushing a visual dynamic model containing the candidate positioning strategy to the user based on the optimal implantation parameter, driving an oral cavity model printer to obtain an implantation guide plate model and an oral cavity environment model, assisting the user in verifying the installation effect of the candidate positioning strategy, and improving the accuracy of false tooth installation and positioning.

Description

Technical Field

[0001] This application relates to the field of dental implant guidance technology, and more specifically, to a digital denture installation and positioning device and method based on three-dimensional scanning. Background Technology

[0002] Implant guides are key auxiliary devices used to position implants during dental implant surgery.

[0003] In related technologies, physical impressions of the edentulous area and adjacent teeth are obtained using impression materials such as silicone rubber. A plaster model is then poured to create a physical oral cavity model. Finally, based on clinical experience, the dentist manually marks the implant location on the plaster model and manually cuts and drills resin sheets to create the implant guide. This method relies heavily on manual operation, and errors are easily introduced during manual marking, cutting, and drilling, leading to implant angle deviations and inaccurate depths. Postoperative problems may include malocclusion and implant loosening. Furthermore, the process is lengthy, involving multiple steps from impression taking to plaster model curing and manual guide fabrication. It also makes it difficult to accurately replicate the subtle morphology of the alveolar bone, resulting in poor fit between the guide and oral tissues and a high risk of displacement during surgery. In addition, manual fabrication makes it difficult to adjust the guide structure based on individual data such as the patient's alveolar bone density and the stress on adjacent teeth, only meeting basic positioning requirements and unable to adapt to complex edentulous situations.

[0004] Therefore, there is an urgent need to design a technical solution to address at least one of the technical problems in the above solutions, namely, inaccurate planting positioning, complex and long production process, and inability to flexibly adapt to complex planting scenarios. Summary of the Invention

[0005] In this context, the embodiments of this application aim to provide a digital denture installation and positioning device and method based on three-dimensional scanning, which can solve the technical problems of inaccurate implant positioning, complex manufacturing process, and inability to flexibly adapt to complex implant scenarios in related technologies, improve the adaptability of implants to the oral environment, and enhance the accuracy of denture installation and positioning.

[0006] In a first aspect of the embodiments of this application, a digital denture fitting and positioning method based on three-dimensional scanning is provided, comprising: Three-dimensional anatomical data of the patient's oral cavity and intraoral surface data are obtained through an oral scanner; The three-dimensional anatomical data and the intraoral surface data are elastically aligned to generate an initial model; a three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and the three-dimensional deformation field is superimposed and fused into the initial model to generate a three-dimensional digital oral cavity model. The SegFormer segmentation network is loaded and adapted using 3SHAPE software, and the 3D digital oral model is input into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters. The SegFormer segmentation network is a semantic segmentation network that combines the Transformer architecture and a lightweight MLP decoder. Based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, a long sequence prediction network is used to predict the positional displacement and degree of bone resorption of the virtual implant after osseointegration through a segmented attention mechanism matched with the oral monitoring cycle and a network memory caching mechanism associated with the context. The long sequence prediction network is a Transformer-XL network that integrates a medical constraint model. Based on the positional offset of the virtual implant after osseointegration and the degree of bone resorption, the initial implantation parameters of the virtual implant are dynamically adjusted to obtain the optimal implantation parameters. Based on the optimal implantation parameters, push a visual dynamic model containing candidate positioning strategies to the user; After the user confirms the candidate positioning strategy in the visualized dynamic model, the dental model printer is driven to print the implant guide model using resin liquid and the oral environment model using plaster.

[0007] In a second aspect of the embodiments of this application, a digital denture fitting and positioning device based on three-dimensional scanning is provided, comprising: An oral scanner is used to acquire three-dimensional anatomical data and intraoral surface data of the patient's oral cavity. A user terminal is used to load the SegFormer segmentation network using 3SHAPE software, and elastically align the three-dimensional anatomical data and intraoral surface data to generate an initial model. A three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and this field is superimposed and fused into the initial model to generate a three-dimensional digital oral model. The SegFormer segmentation network is loaded and adapted using 3SHAPE software, and the three-dimensional digital oral model is input into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters. The SegFormer segmentation network is a semantic segmentation network that combines a Transformer architecture and a lightweight MLP decoder. The user terminal is also used to predict the positional displacement and degree of bone resorption of the virtual implant after osseointegration based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters. This is achieved using a long-sequence prediction network with a segmented attention mechanism matched to the oral monitoring cycle and a network memory caching mechanism associated with the context. Based on the positional displacement and degree of bone resorption of the virtual implant after osseointegration, the initial implantation parameters of the virtual implant are dynamically adjusted to obtain the optimal implantation parameters. The long-sequence prediction network is a Transformer-XL network that integrates a medical constraint model. Based on the optimal implantation parameters, a visual dynamic model containing candidate positioning strategies is pushed to the user. The dental model printer is used to print implant guide models using resin liquid and dental environment models using plaster after the user confirms the candidate positioning strategy in the visualized dynamic model.

[0008] In a third aspect of the embodiments of this application, a terminal device is provided, the terminal device comprising: at least one processor, a memory, and an input / output unit; wherein the memory is used to store a computer program, and the processor is used to invoke the computer program stored in the memory to execute the digital denture installation and positioning method based on three-dimensional scanning as described in the first aspect.

[0009] In a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, which includes instructions that, when executed on a computer, cause the computer to perform the digital denture installation and positioning method based on three-dimensional scanning as described in the first aspect.

[0010] In a fifth aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the digital denture installation and positioning method based on three-dimensional scanning as described in the first aspect.

[0011] According to an embodiment of this application, a digital denture mounting and positioning device and method based on three-dimensional scanning is disclosed. First, three-dimensional anatomical data and intraoral surface data of the patient's oral cavity are acquired using an oral scanner. The three-dimensional anatomical data includes bone density and nerve distribution, while the intraoral surface data includes the shape of adjacent teeth, occlusal relationship, and gingival morphology. Then, the three-dimensional anatomical data and the intraoral surface data are elastically aligned to generate an initial model. A three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and this field is superimposed and fused into the initial model to generate a three-dimensional digital oral cavity model. Next, the SegFormer segmentation network is loaded and adapted using 3SHAPE software, and the three-dimensional digital oral cavity model is input into the SegFormer segmentation network to generate a virtual implant for the patient and corresponding initial implantation parameters. The SegFormer segmentation network is a semantic segmentation network combining a Transformer architecture and a lightweight MLP decoder. The initial implantation parameters include the initial implantation position, angle, and depth of the virtual implant. Based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, a long-sequence prediction network is employed. This network utilizes a segmented attention mechanism matched to the oral monitoring cycle and a network memory caching mechanism with associated context to predict the positional displacement and bone resorption degree of the virtual implant after osseointegration. The long-sequence prediction network is a Transformer-XL network incorporating a medical constraint model. The initial implantation parameters of the virtual implant are dynamically adjusted based on the positional displacement and bone resorption degree after osseointegration to obtain optimal implantation parameters. Finally, a visual dynamic model containing candidate positioning strategies is pushed to the user based on the optimal implantation parameters. After the user confirms the candidate positioning strategies in the visual dynamic model, the oral model printer is activated to print an implant guide model using resin and an oral environment model using plaster, assisting the user in verifying the installation effect of the candidate positioning strategies. This application's implementation method utilizes the collaborative operation of oral 3D scanning technology, SegFormer segmentation network, long sequence prediction network, and digital model printing technology to achieve the construction process of an individualized denture installation and positioning model for the patient's oral cavity. This assists users in verifying the installation effect of candidate positioning strategies, greatly improving denture installation and positioning efficiency, and enhancing the intelligence level, positioning accuracy, and safety redundancy of denture installation in clinical operations. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a digital denture installation and positioning method based on three-dimensional scanning, as shown in this application. Figure 2 This application shows a schematic diagram of a digital denture mounting and positioning device based on three-dimensional scanning. Figure 3The present invention provides a schematic diagram of the structure of a medium according to an embodiment of the present application. Detailed Implementation

[0013] The following is for reference. Figure 1 , Figure 1 This is a flowchart illustrating a digital denture installation and positioning method based on three-dimensional scanning, provided as an embodiment of this application.

[0014] To address at least one of the aforementioned technical problems, this application provides a digital denture installation and positioning device and method based on three-dimensional scanning. Specifically, multimodal data (including bone density, nerve distribution, adjacent tooth shape, occlusal relationship, etc.) acquired by an oral scanner are used to elastically align the three-dimensional anatomical data and the intraoral surface data to generate an initial model. A three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and this field is superimposed and fused into the initial model to obtain a high-fidelity three-dimensional digital oral model. The SegFormer segmentation network is a semantic segmentation network combining a Transformer architecture and a lightweight MLP decoder. This application further optimizes the virtual implant generation process through the SegFormer segmentation network, enabling accurate identification of anatomical structures and calculation of initial implantation parameters (including position, angle, and depth), avoiding errors caused by reliance on traditional experience. On the other hand, a long-sequence prediction network, combining patient attributes and occlusal force parameters, predicts the positional displacement and bone resorption of the virtual implant after osseointegration through a segmented attention mechanism matched with the oral monitoring cycle and a network memory caching mechanism associated with the context. This dynamically simulates the biomechanical changes after osseointegration (including positional displacement and bone resorption), making implant parameter adjustments more scientifically based and reducing the risk of implant failure due to long-term poor biomechanical adaptation. The long-sequence prediction network, based on historical surgical data and multi-parameter fusion, enables the prediction of long-term implant stability. Furthermore, the implantation parameters can be dynamically adjusted based on the positional displacement and bone resorption of the virtual implant after osseointegration, overcoming the technical problem of not being able to predict the bioremodeling process in traditional static planning, further improving the accuracy of implant positioning. By pushing candidate strategies through a visualized dynamic model, the surgeon's decision-making is assisted, improving the predictability of the surgery and reducing over-reliance on the surgeon's personal experience. Finally, through the visualization of the dynamic model and the printing of physical models (i.e., the implant guide model and the oral environment model), the preoperative plan is presented in three dimensions and verified physically, improving its adaptability to complex implant scenarios. Users (such as doctors) can visually assess the compatibility of implants with the oral environment and simulate occlusal relationships on resin guides and plaster models, further optimizing the treatment plan before surgery, significantly reducing the risk of intraoperative complications. This dual-verification mode is applicable to complex cases (such as areas with dense nerves and blood vessels or multiple missing teeth), improving the error tolerance and safety of the procedure. Through end-to-end digital technologies, such as scanning, segmentation, prediction, and printing, the cumbersome steps of traditional impression taking and wax model fitting are replaced, greatly shortening the production cycle, reducing the complexity of the production process, and decreasing the number of patient visits and intraoperative operation time. Oral scanning is non-contact and discomfort-free, while the printed guide precisely matches the patient's anatomical features, avoiding the hassle of repeated adjustments, especially suitable for individuals with limited mouth opening or sensitive individuals.

[0015] Figure 1The flowchart of a digital denture fitting and positioning method based on three-dimensional scanning provided in one embodiment of this application includes: Step S101: Obtain three-dimensional anatomical data and intraoral surface data of the patient's oral cavity using an oral scanner. In this embodiment, the three-dimensional anatomical data includes bone density and nerve distribution, and the intraoral surface data includes the shape of adjacent teeth, occlusal relationship, and gingival morphology.

[0016] An oral scanner is a specialized device used to acquire data on the internal structure of a patient's oral cavity. Utilizing optical scanning principles (such as structured light scanning and laser scanning), it can capture the three-dimensional morphological information of soft and hard tissues within the oral cavity non-contactly. The procedure eliminates the need for traditional impression materials, allowing scanning to be performed directly inside the patient's mouth, reducing patient discomfort while rapidly generating high-precision digital data. This provides the raw data support for subsequent construction of three-dimensional oral models. The device typically features a miniaturized scanning probe that can flexibly reach various areas of the oral cavity, including areas with limited operating space such as the posterior teeth region. This ensures complete scanning of critical areas such as edentulous areas, adjacent teeth, and gums. Furthermore, the scanning results can be transmitted in real-time to a supporting data processing system for subsequent data integration and analysis.

[0017] Three-dimensional anatomical data is digital information reflecting the anatomical characteristics of the hard tissues inside a patient's oral cavity, corely encompassing two key aspects: bone density and nerve distribution. Bone density data primarily reflects the compactness of the alveolar bone; differences in bone density across different regions directly affect the stability of implant placement, with denser areas generally being more suitable as implant sites. Nerve distribution data precisely records the course and location of nerve pathways within the oral cavity, especially the distribution of nerve bundles within the alveolar bone. This data is crucial for determining the implant placement and depth, preventing damage to nerve tissue during the implantation procedure. Three-dimensional anatomical data is generated through scanning and processing of hard tissues such as the alveolar bone using an oral scanner, providing a visual representation of the spatial structural relationships of oral hard tissues and offering an anatomical basis for implant planning.

[0018] Intraoral surface data is digital information reflecting the morphological characteristics of the soft tissues and tooth surfaces inside the patient's oral cavity. It mainly includes three aspects: adjacent tooth shape, occlusal relationship, and gingival morphology. Adjacent tooth shape data records detailed information such as the crown shape, occlusal texture, and marginal contours of teeth surrounding the edentulous area. This information ensures the morphological coordination between the virtual implant and adjacent teeth, as well as the normal contact relationship between the implanted prosthesis and adjacent teeth. Occlusal relationship data captures the contact state, occlusal trajectory, and occlusal force distribution of the upper and lower jaws during occlusion, providing a reference for determining the implant placement angle to adapt to normal occlusal function. Gingival morphology data records the surface undulations, marginal position, and attachment relationship of the gingival tissue to the teeth, helping to ensure proper fit between the implant neck and the gingival tissue, improving the aesthetics and comfort of the post-implantation oral cavity. Intraoral surface data is acquired through scanning the soft tissues and tooth surfaces using an oral scanner and is a crucial foundation for constructing a complete three-dimensional digital oral model and optimizing the implantation plan.

[0019] Step S102: Elastically align the three-dimensional anatomical data and the intraoral surface data to generate an initial model; construct a three-dimensional deformation field using the Swin-UMamba algorithm, and dynamically repair the scanning blind spots in the initial model using the three-dimensional deformation field to generate a three-dimensional digital oral model.

[0020] As an optional embodiment, in step S102, the three-dimensional anatomical data and the intraoral surface data are first preprocessed to extract alveolar bone feature points from the three-dimensional anatomical data and oral soft tissue feature points from the intraoral surface data. Here, points with stable intraoral anatomical structures, high recognizability, and spatial correspondence are mainly selected as registration benchmarks. For example, alveolar bone feature points include alveolar bone ridge apex and key points of nerve pathways, while oral soft tissue feature points include apex of the occlusal surface of adjacent teeth and gingival margin feature points. The alveolar bone ridge apex is the highest natural ridge of the alveolar bone in the edentulous area, and the key points of nerve pathways are key turning points in the course of nerves; both are fixed landmark points in hard tissue. The apex of the occlusal surface of adjacent teeth is the highest convex point of the occlusal surface of adjacent teeth, and the gingival margin feature points are contour turning points at the junction of the gingiva and teeth; both are stable reference points for the integration of soft tissue and teeth. For example, in the three-dimensional anatomical data of the edentulous area, the most prominent vertex of the alveolar bone can be directly located as the alveolar bone ridge vertex. In the corresponding intraoral surface data, the cusps of the occlusal surfaces of the adjacent teeth around the edentulous area can be located as the occlusal surfaces vertices of the adjacent teeth, ensuring that there is a clear spatial correspondence between the two types of feature points.

[0021] Furthermore, using the preprocessed alveolar bone feature points and oral soft tissue feature points as matching benchmarks, the rigid transformation matrix of the alveolar bone feature points and oral soft tissue feature points is calculated using the least squares method. This allows for preliminary positional calibration of the three-dimensional anatomical data and the intraoral surface data, ensuring that the alveolar bone in the edentulous area initially coincides with the corresponding gingival region. The least squares method is used to calculate the rigid transformation matrix. Its principle is to minimize the sum of Euclidean distances between corresponding feature points to solve for the optimal translation and rotation parameters (rigid transformation only changes the position and angle of the object, not its shape and size), thus forming the rigid transformation matrix. For example, assuming three sets of corresponding feature points are extracted: alveolar bone side A1 (1, 2, 3), A2 (4, 5, 6), A3 (7, 8, 9) and soft tissue side B1 (2, 3, 4), B2 (5, 6, 7), B3 (8, 9, 10), the translation amount (X+1, Y+1, Z+1) and rotation angle (0° around the Z-axis) are calculated using the least squares method, generating the rigid transformation matrix. After transforming the three-dimensional anatomical data according to this matrix, the distance errors between A1 and B1, A2 and B2, and A3 and B3 are minimized, achieving preliminary alignment between the alveolar bone and the gingival region in the edentulous area. Therefore, the rigid transformation matrix calculated using the least squares method helps improve the spatial consistency after the preliminary alignment of the three-dimensional anatomical data and intraoral surface data.

[0022] Next, a non-rigid registration deformation network is constructed using a 3D Swin Transformer encoder and a diffuse deformation field decoder. The initially aligned 3D anatomical data and intraoral surface data are used as inputs to this non-rigid registration deformation network. The encoder extracts multi-scale features from the initially aligned 3D anatomical data and intraoral surface data, and the decoder converts these multi-scale features into a non-rigid deformation field. Optionally, during the forward propagation of the non-rigid registration deformation network, the 3D Swin Transformer module is used as the encoder core in the encoder stage. The input data is first unified in dimension through a voxelization layer, and group normalization is applied to reduce modal differences. The encoder consists of four hierarchical stages. Stage 1 processes high-resolution features (window size 8×8×8), capturing local details such as alveolar bone surface texture and gingival attachment points through a windowed multi-head self-attention mechanism. Stages 2-4 progressively downsample feature maps (window sizes increasing to 16×16×16 and 32×32×32) to capture global structural features such as jawbone morphology and implant angles. Each stage integrates a shift window mechanism and residual connections to enhance long-range dependency modeling and training stability. Consequently, the encoder outputs a multi-scale feature pyramid, including high, medium, and low-resolution feature maps. These features are fused through a cross-modal attention module, calculating correlation weights between 3D anatomical data and intraoral surface data to ensure feature alignment. For example, high-resolution features preserve the contours of adjacent teeth, while low-resolution features encode the overall pattern of bone density distribution. In the decoder stage, a diffuse deformation field decoder generates a non-rigid deformation field conditioned on the encoder features. The decoder is built on a U-Net architecture and further introduces a diffusion process. Specifically, the forward process progressively adds Gaussian noise to the initial deformation field (zero-field initialization) for T steps (e.g., T=1000) to simulate deformation uncertainty. The reverse process iteratively denoises through a conditional denoising network. The network input includes encoder features and the current noisy deformation field, and the output is a corrected displacement field. The denoising network uses deformable convolutional layers to handle local deformations, ensuring a smooth output. By replacing traditional convolution with the attention mechanism of the 3D Swin Transformer, the multi-scale context of 3D data can be processed more efficiently, and the robustness of the deformation field is improved through a diffusion process. After feature extraction, the multi-scale features are directly passed to the decoder, laying the foundation for deformation field generation.

[0023] Finally, the non-rigid deformation field generated by the decoder is used to locally adjust the deformation of the initially aligned 3D anatomical data. This optimizes the fit between the alveolar bone surface and the gingival attachment area in the initially aligned 3D anatomical data, and preserves the bone density distribution and adjacent tooth contours in the initially aligned 3D anatomical data and intraoral surface data as original detail features to construct the initial model. Optionally, the displacement vector map corresponding to the non-rigid deformation field is applied to the CBCT voxel data, and resampling is achieved through trilinear interpolation. The adjustment prioritizes high-gradient regions, such as the alveolar bone-gingival boundary. Differential regions are automatically identified using an attention weight map (derived from the Swing Transformer encoder), and the deformation field applies greater displacement to these regions to ensure a smooth fit between the fit boundaries. Simultaneously, Total Variation Regularization is used to constrain the deformation field to avoid discontinuities or excessive deformation, maintaining the natural continuity of the anatomical structure. Further, optionally, during the deformation process, the multi-scale features of the encoder are directly fused into the output through skip connections in the Feature Pyramid Network (FPN). Specifically, bone density distribution is derived from Hounsfield units in CBCT, preserving local variations through high-resolution feature maps to avoid density distortion caused by deformation. Adjacent tooth contours are extracted from intraoral scan data, maintaining geometric accuracy through medium-resolution feature maps to ensure the contours do not distort after deformation. The adjusted 3D anatomical data is fused with the intraoral scan data to generate a complete 3D digital oral model as the initial model, directly applicable to downstream tasks (such as dental implant design). This preserves original biomechanical features, such as bone density gradients and tooth morphology, while achieving high-precision matching of the alveolar bone-gingival boundary. This step is trained through self-supervised learning, using a loss function that combines similarity loss (such as normalized cross-correlation) and adversarial loss to reduce reliance on labeled data.

[0024] In step S102 above, the alveolar bone and oral soft tissue feature points are first extracted and used as a reference to achieve preliminary alignment of three-dimensional anatomical data and intraoral surface data through rigid registration. Then, a non-rigid registration deformation network is used to extract multi-scale features to generate a deformation field for local adjustment to optimize the fitting boundary between the alveolar bone and the gingival attachment area. At the same time, the original details such as bone density distribution and the outline of adjacent teeth are preserved, and finally the initial model is constructed.

[0025] Optionally, after generating the initial model, in step S102, a three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and the three-dimensional deformation field is superimposed and fused into the initial model to generate a three-dimensional digital oral model. As an optional embodiment, the three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and the scanning blind spots in the initial model are dynamically repaired using the three-dimensional deformation field to generate a three-dimensional digital oral model. This includes: inputting the initial model and blind spot marker map as input into the Swin-UMamba algorithm model; extracting multi-scale features from the initial model using the encoder of the Swin-UMamba algorithm model, and hierarchically encoding the extracted macroscopic structural features and the microscopic features surrounding the blind spots; mapping the hierarchically encoded multi-scale features to the three-dimensional deformation field required for blind spot repair using the decoder and the blind spot marker map; and superimposing and fusion the three-dimensional deformation field into the initial model to dynamically repair the scanning blind spots in the initial model, resulting in an optimized three-dimensional digital oral model.

[0026] The above steps first identify the scanning blind spots in the initial 3D digital oral model. Specifically, the initial model undergoes data integrity checks, and features of areas prone to generating scanning blind spots, such as alveolar bone depressions in edentulous areas, interdental spaces, and the gingival-tooth attachment margins, are extracted. By comparing the point cloud density (the point cloud density in the scanning blind spot is more than 30% lower than that in the normal area), geometric continuity (obvious topological discontinuities exist at the blind spot boundary), and feature matching degree with adjacent areas (such as interruptions in the extension of adjacent tooth texture), the spatial range and boundary coordinates of the scanning blind spots are marked, generating a blind spot marking map. Thus, the location and extent of the scanning blind spots are located using the blind spot marking map, avoiding omissions or errors in subsequent restoration, and providing clear target guidance for targeted restoration using the Swin-UMamba algorithm.

[0027] Furthermore, the Swin-UMamba algorithm is used for model input and parameter configuration. In this application, the Swin-UMamba algorithm is a hybrid deep learning algorithm that integrates the Swin Transformer window attention mechanism and the UMamba linear attention flow mechanism. It is used to solve the balance problem between local fine feature repair and global anatomical logic consistency in 3D oral digital models, while also considering repair accuracy and efficiency, and adapting to the complex anatomical structures of oral data (such as bone density gradient and gingival elastic deformation characteristics). Specifically, the initial model and blind zone marking map are used as input and imported into the Swin-UMamba algorithm model. Based on the bone density distribution (from 3D anatomical data), the outline of adjacent teeth (from intraoral surface data), and the spatial features of the marked blind zones in the initial model, the model parameters of the Swin-UMamba algorithm model are configured. For example, the window attention mechanism of the Swin Transformer module is used, setting the window size to 1 / 4 of the diameter of the adjacent tooth crown in the initial model to accurately capture the geometric texture features of the normal area surrounding the blind zone (such as the pit and fissure morphology of the adjacent tooth occlusal surface and the undulation pattern of the gingival surface). For example, by enabling the linear attention flow mechanism of the UMamba module, the features of the normal area within a 3mm radius around the blind zone are used as a restoration reference. Feature association weights are established based on bone density gradient and gingival elastic deformation trend (derived from gingival morphology features), prioritizing the preservation of feature integrity of key areas related to implant placement (such as alveolar bone implant candidate areas). Thus, through parameter configuration and algorithm mechanism selection adapted to oral anatomy, the Swin-UMamba algorithm ensures efficient association between blind zone and normal area features, while prioritizing the feature integrity of key implant areas.

[0028] Next, dynamic repair and feature preservation of the scanned blind areas are performed. The encoder of the Swin-UMamba algorithm extracts multi-scale features from the initial model, cascading the macroscopic structural features (such as the overall morphology of the alveolar bone) and the microscopic features surrounding the blind areas (such as the enamel texture of adjacent teeth and the curvature of the gingival margin) into layers. The decoder, combined with the blind area marker map, maps the layered multi-scale features into a three-dimensional deformation field required for blind area repair, and dynamically repairs according to the following logic: For alveolar bone depression blind areas, a repair structure conforming to the physiological morphology of the alveolar bone is generated based on the surrounding bone density distribution (the structural strength of high-density areas must match after repair); for interdental gap blind areas, a gap-filling structure continuous with the texture of adjacent teeth is generated, referencing the extension trend of the adjacent tooth outline; for gingival attachment margin blind areas, the angle error between the repaired margin and tooth attachment is controlled within a preset degree, combined with the elastic deformation characteristics of the gingiva, ensuring that the fit conforms to the clinical physiological state of the gingiva. During the restoration process, the original data of bone density distribution and adjacent tooth contours from the initial 3D digital oral model are preserved in real time to avoid interference from the restoration operation on the virtual implant parameters generated by the SegFormer segmentation network. Thus, through these steps, personalized and physiological restoration of different types of scanning blind spots is achieved, ensuring the consistency of the morphology and structure of the restoration area with the surrounding tissues while preserving the original key data, avoiding negative impacts on the subsequent generation of virtual implant parameters, and helping to further improve positioning accuracy.

[0029] Optionally, the 3D digital oral model repaired using the Swin-UMamba algorithm is validated. The geometric continuity between the repaired area and the surrounding normal area is compared, and the connection error between the repair boundary and the normal area is detected using a distance measurement tool (error must be ≤0.05mm). The rationality of the features of the repaired area is verified; for example, the density distribution of the alveolar bone repaired area must be consistent with the surrounding bone density gradient, and the morphology of the gingival repaired area must match the elastic characteristics of the adjacent gingiva. If the validation reveals topological faults or feature deviations in the repaired area, the algorithm is returned to Swin-UMamba to readjust the attention weights and deformation field parameters for secondary repair. After passing the validation, the optimized 3D digital oral model is obtained. Thus, through multi-dimensional validation and secondary repair mechanisms, the error of the repaired model is controlled within the clinically acceptable range, ensuring that the optimized 3D digital oral model meets the accuracy and rationality requirements of subsequent implant design.

[0030] It is understood that, in this embodiment, the model structure of the Swin-UMamba algorithm includes, but is not limited to, 3D feature extraction, attention association, and repair generation modules. Each module is adapted and optimized for the oral cavity scenario. The core function of the 3D feature encoding module, i.e., the encoder, is to transform the voxel data and blind zone marker map of the initial 3D oral digital model into multi-scale feature vectors, while retaining the original information of oral anatomical features such as bone density and adjacent tooth shape. The module structure adopts a hierarchical convolution and a Swin Transformer window attention submodule. First, the input data is downsampled through a 3D convolutional layer to generate feature maps of different scales. For example, large-scale maps are used to capture the overall morphology of the alveolar bone, while small-scale maps focus on details such as the gingival margin and interdental spaces. The window size is set according to the characteristics of oral data, and the window size is adapted to 1 / 4 of the diameter of the adjacent tooth crown. For example, when the diameter of the adjacent tooth crown is 4mm, the window size is set to 1mm, ensuring that each window can completely cover the feature reference units around the blind zone, such as the local area of ​​a single tooth or an elastic deformation unit of the gingiva, avoiding feature fragmentation across windows. Simultaneously, an anatomical feature embedding layer is incorporated, embedding bone density values ​​and gingival elasticity coefficients from three-dimensional anatomical data as additional features into the feature vectors of corresponding voxels, providing physiological attribute references for subsequent repairs. For example, structural strength needs to be preserved after repairing high-density alveolar bone areas.

[0031] The core function of the attention association module, namely the Swin-UMamba hybrid attention layer, is to establish feature associations between the blind area and the surrounding normal area, providing accurate reference for restoration. Its structural details include: the Swin Transformer submodule, responsible for fine-grained local feature association, which calculates the similarity between each blind area voxel and the surrounding normal voxels using window-shifting attention on the feature map of the normal area within a 3mm radius of the blind area. For example, it calculates the feature matching degree between the blind area alveolar bone voxel and the surrounding high-density alveolar bone voxel, generating local attention weights to ensure that restoration details are consistent with surrounding microscopic features such as alveolar bone texture. The UMamba submodule is responsible for global anatomical logic association. Through a linear attention flow based on a selective scanning mechanism, it associates blind area features with global anatomical features such as the density gradient of the entire alveolar bone and the gingival morphology corresponding to the occlusal trajectory of the opposing teeth, generating global constraint weights to avoid restoration results that violate physiological laws, such as the restoration angle of the gingival attachment margin exceeding the normal physiological range. The weight fusion unit fuses the local attention weights and global constraint weights in a 7:3 ratio. Because local details have higher priority in the oral cavity scene, the final output is a restoration reference feature vector for each blind area voxel.

[0032] The core function of the repair generation module, or decoder, is to transform the attention-linked feature vectors into a three-dimensional deformation field, dynamically repairing blind areas while preserving key details of the original model. Its structural details are manifested in the use of deconvolution and residual connections. The deconvolution layer upsamples the feature map with fused attention weights to the original model resolution, while the residual connections directly input the original non-blind area features of the initial model into the decoder, avoiding the loss of original details such as the outline of adjacent teeth and bone density distribution during the repair process. The three-dimensional deformation field generation unit, based on the repair reference feature vector, generates voxel-level deformation parameters for the blind areas, such as the displacement direction and scaling ratio of each voxel in the blind area, forming a three-dimensional deformation field. For alveolar bone depression blind areas, the deformation parameters are adjusted according to the surrounding bone density gradient; for example, the deformation amplitude is smaller in high-density areas. For interdental gap blind areas, filling deformation parameters are generated according to the extension trend of adjacent tooth texture. The medical constraint verification layer embeds oral clinical constraint rules, such as the error of the gingival-tooth attachment angle being less than 5° and the difference between the density of the alveolar bone repair area and the surrounding gradient being less than 100 HU. The generated deformation field is verified in real time. If it exceeds the constraint range, it is fed back to the attention association module to readjust the weights until the clinical requirements are met.

[0033] Compared to traditional 3D model restoration algorithms such as U-Net, simple Swin Transformer, and basic Mamba, the Swin-UMamba used in this application is specifically designed for oral scenarios and employs a hybrid attention mechanism. This resolves the conflict between local precision and global plausibility, while improving restoration efficiency and clinical compliance. Furthermore, relying on 3D adaptability optimization, it avoids voxel information redundancy issues with pure 3D structures, improving 3D restoration accuracy. Through feature association refinement, it incorporates oral anatomical features, enabling the restored model to accurately match the physical properties of oral tissues. By balancing restoration efficiency and supporting iterative optimization processes for local re-restoration, it enhances both the effectiveness and efficiency of oral model restoration.

[0034] Step S103: Load and adapt the SegFormer segmentation network using 3SHAPE software, and input the three-dimensional oral digital model into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters.

[0035] In this embodiment, the initial implantation parameters include the initial implantation position, angle, and depth of the virtual implant.

[0036] Specifically, SegFormer is a semantic segmentation network that combines a Transformer architecture with a lightweight MLP decoder. It employs a hierarchical Transformer encoder to output multi-scale features, eliminating the need for positional encoding and avoiding performance degradation caused by differences in test resolution compared to training resolution. These characteristics make it well-suited for medical image analysis, including automatic segmentation of medical images such as CT and MRI. Based on this, in step S103, since SegFormer is based on the PyTorch framework, and 3SHAPE software typically has its own internal data processing flow, the core idea for loading and applying SegFormer in 3SHAPE is to build an intermediate node that allows 3SHAPE to call the externally trained SegFormer model for processing and then import the results back into 3SHAPE.

[0037] For example, in practical implementation, the first step is to configure and prepare the environment. It's essential to confirm whether the 3SHAPE software provides an API interface, supports plugin development, or allows calling external scripts; this is a prerequisite for integration. Simultaneously, a Python environment needs to be set up, installing Python 3.8 or later, PyTorch, the Transformers library, and other necessary SegFormer dependencies, ensuring compatibility between the environment and 3SHAPE's external calls. A SegFormer model also needs to be obtained. Pre-trained models can be acquired from open-source platforms such as HuggingFaceModelHub, and then domain-adapted training or fine-tuning can be performed based on the characteristics of dental images. This requires collecting labeled dental image data to train the model, adapting it to segmentation tasks involving teeth, gums, implants, and other structures. Furthermore, data preprocessing and integration are crucial steps. The image data to be segmented needs to be exported from the 3SHAPE software, such as CBCT DICOM files, intraoral scan STL or PLY files, and then converted to PNG or JPEG formats supported by SegFormer. For 3D data, layer-by-layer slicing is typically required to convert it into a 2D image sequence. Next, the SegFormer Processor is used to preprocess the image, including normalization and resizing, to meet the model's input requirements. Then, a script is written in the Python environment. First, the trained SegFormer model and processor are loaded. Then, the image exported from 3SHAPE is read and preprocessed. Prediction is performed using the model, and the output logits need post-processing, including upsampling to the original image size and using argmax to obtain the predicted class for each pixel. The results can then be saved as an image or an array. Finally, the predicted segmentation mask is converted to a 3SHAPE-recognizable STL, DICOM, or software-specific markup format, which may require writing additional conversion scripts. The processed results are then loaded back into the software via 3SHAPE's API or manual import function, and the accuracy of the segmentation results is checked within the software, with minor manual corrections performed if necessary.

[0038] Then, in step S103, the 3D digital oral model is input into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters. This includes: First, converting the optimized 3D digital oral model into a voxelized data compatible with the SegFormer segmentation network by converting the model's 3D point cloud data into voxelized data. Then, combining the original detailed features, the voxelized data is enhanced. For example, bone density labels are added to the voxelized data, such as marking high-density areas as preferred implant areas and low-density areas as implant risk areas. A spatial mask of nerve distribution is overlaid (e.g., marking a 2mm radius around a nerve channel as a prohibited implant area), preserving the edge features of adjacent teeth's contours for subsequent implant-to-adjacent tooth gap verification, resulting in enhanced 3D oral model data.

[0039] Furthermore, in step S103, after feature enhancement of the voxelized data, the enhanced voxelized data is input into the SegFormer segmentation network. The hierarchical feature extraction architecture of the SegFormer network identifies candidate implant regions. Combined with the optimized and matched boundary between the alveolar bone surface and the gingival attachment area, feasible implant regions in the edentulous area are segmented from the candidate implant regions and converted into an implant region segmentation mask. For example, a Hierarchical Transformer encoder is used to extract model features at different scales (large scale capturing the overall morphology of the alveolar bone, small scale focusing on the details of the edentulous area), with a focus on strengthening the voxel feature weights of the alveolar bone in the edentulous area to ensure priority identification of candidate implant regions during segmentation. During feature extraction, a dynamic weight allocation mechanism is used to strengthen the feature weights of voxels corresponding to the alveolar bone in the edentulous area (e.g., weight values ​​are 30%-50% higher than in non-edentulous areas), ensuring that the network prioritizes implant-related regions during feature learning and avoids interference from non-critical region features in the identification of candidate implant regions. Then, the Feature Pyramid Fusion decoder is activated to fuse the multi-scale features output by the encoder. Combined with the optimized matching alveolar bone and gingival attachment boundaries, the feasible implantation area of ​​the edentulous region is accurately segmented (excluding nerve-prohibited areas and low-density risk areas), generating an implantation area segmentation mask map. For example, the feature maps of different scales output by the Feature Pyramid Fusion decoder are fused. First, large-scale global features and small-scale local features are integrated by spatial alignment and feature dimension fusion. Then, combined with the optimized matching boundary between the alveolar bone surface and the gingival attachment area (such as the coordinates of the transition area between the alveolar bone and the gingiva), "prohibited implantation areas" (within 2mm of the nerve periphery) and "implantation risk areas" (low-density areas) are removed from the initially identified implantation candidate areas, and feasible implantation areas in the edentulous region that combine safety (away from nerves) and stability (high-density bone) are selected. Finally, the spatial range of the feasible implantation area is converted into a binary implantation area segmentation mask map (the voxel value of feasible areas is 1, and that of infeasible areas is 0), visually marking the specific spatial location suitable for implant placement.

[0040] Next, in step S103, a virtual implant and initial implantation parameters are generated based on the implant region segmentation mask and clinical implant selection rules. For example, based on the implant region segmentation mask and clinical implant selection rules (such as matching the implant diameter to alveolar bone width and determining the implant length based on bone density), a parameter generation module is constructed in the output layer of the SegFormer segmentation network. Specifically, the initial implant placement location is determined by selecting the coordinate point within the feasible implantation area that has the highest bone density, is furthest from the nerve pathway, and has a distance from adjacent teeth that meets clinical standards (≥1.5mm) as the center point. The initial implant placement angle is determined by referencing the occlusal relationship in the intraoral surface data to control the deviation of the implant axis from the perpendicular direction of the opposing tooth's occlusal surface within a preset degree, while avoiding the roots of adjacent teeth. The initial implant placement depth is determined by matching the alveolar bone height (from alveolar bone feature point measurement data) with the implant length, ensuring that the distance between the implant apex and the alveolar bone floor is greater than 2mm, thus generating the initial implantation parameters for the virtual implant.

[0041] Finally, in step S103, the generated virtual implant and initial implantation parameters are validated for compatibility. If the validation is successful, the final virtual implant and its corresponding initial implantation parameters are output. For example, by virtually assembling the virtual implant model with a 3D digital oral model, the gap between the implant and adjacent teeth (≥1.5mm), the distance to the nerve pathway (≥2mm), and the fit with the gingival attachment area are validated. Simultaneously, bone density data is used to validate the bone-bearing capacity of the implantation area (ensuring the implant body is located in a high-density bone region). If the validation reveals parameter deviations (such as insufficient gap with adjacent teeth or implantation angle exceeding the occlusal fit range), feedback is sent to the parameter generation module for readjustment. After successful validation, the final 3D virtual implant model and its corresponding initial implantation parameters are output, providing basic data for subsequent osseointegration prediction and parameter optimization.

[0042] Step S104: Based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, a long-sequence prediction network is used to predict the positional displacement and degree of bone resorption of the virtual implant after osseointegration through a segmented attention mechanism matched with the oral monitoring cycle and a network memory caching mechanism associated with the context. The long-sequence prediction network is a Transformer-XL network that integrates a medical constraint model.

[0043] As an optional embodiment, in step S104, firstly, historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters are standardized and integrated to form a multi-source temporal feature matrix. Then, combining clinical treatment guidelines and physiological characteristics of oral implantology, a medical constraint model is constructed, including position offset constraint rules and bone resorption degree constraint rules, to define the physiological rationality boundary of the prediction results. The medical constraint model is then used as a regularization term in the Transformer-XL network loss function. For example, the position offset adjustment rule could be that if the offset in a certain axis exceeds the clinical safety threshold (e.g., an X-axis offset of 0.4 mm exceeds the 0.3 mm threshold in high-density areas), the initial implantation position is adjusted in the opposite direction of the offset (if the X-axis offset is 0.4 mm, the initial position is offset 0.35 mm in the opposite direction of the X-axis, reserving 0.05 mm for bone integration offset redundancy). The adjustment range needs to refer to bone density; the adjustment range in low-density areas is 20% smaller than that in high-density areas to avoid affecting implant stability. For example, the adjustment rules for bone resorption are as follows: if vertical bone resorption is >0.5mm / 6 months, or horizontal bone resorption is >0.3mm / 6 months, then adjust the initial implantation depth (increase the initial depth by 0.15mm for every 0.1mm increase in vertical bone resorption to ensure an effective anchorage length of ≥3mm after osseointegration). Simultaneously, adjust the initial implantation angle according to the rate of bone resorption (e.g., rate >0.1mm / month) (rotate 2°-3° towards the direction of slower bone resorption to reduce stress on the resorbed area). A multi-parameter synergistic rule is as follows: when positional deviation and bone resorption coexist, prioritize adjusting the initial implantation angle (angle adjustment improves both), then fine-tune the initial position and depth. This avoids new fitting problems caused by adjusting a single parameter; for example, adjusting only the position might increase the gap deviation with adjacent teeth.

[0044] The above steps eliminate the dimensional differences of different parameters through the standardization and integration of multi-source data, providing a unified feature foundation for network input. The construction of the medical constraint model and the fusion of regularization terms limit the prediction boundary from a clinical standard perspective, avoiding network outputs that violate physiological laws. The clarification of specific adjustment rules provides an executable clinical basis for the dynamic optimization of subsequent initial implantation parameters, ensuring that the direction of parameter adjustment is consistent with the requirements for implantation safety and stability.

[0045] Optionally, when the Transformer-XL network prediction exceeds the constraint threshold, the regularization term increases the loss value, guiding network parameter adjustment through backpropagation to bring the Transformer-XL network prediction convergence towards a clinically reasonable range. The above steps establish a dynamic feedback correction mechanism for the prediction results. When the network prediction shows a trend of deviating from the clinically safe range, the regularization term, through loss value adjustment, forcibly guides parameter optimization, effectively reducing the clinical irrationality of the prediction results, improving the reliability and accuracy of network prediction, and providing precise data support for subsequent parameter adjustments based on the prediction results.

[0046] Next, the multi-source temporal feature matrix is ​​input into the Transformer-XL network, which integrates a medical constraint model. A segmented attention mechanism is used to divide the multi-source temporal feature matrix into temporal segments, with the segment length matched to the oral monitoring cycle. A network memory caching mechanism is employed to capture the long-term dependencies between the previous and current temporal segments on feature parameters. These long-term dependencies include the correlation between long-term occlusal force and alveolar bone resorption, as well as the dynamic correlation between the initial implantation angle and positional offset. Based on these long-term dependencies, the predicted positional offset and degree of bone resorption of the virtual implant during the current osseointegration cycle are obtained. The positional offset includes the offset value in a three-dimensional spatial coordinate system, and the degree of bone resorption includes vertical bone resorption, horizontal bone resorption, and their corresponding rates of change.

[0047] It can be understood that, in the above steps, when inputting the multi-source temporal feature matrix into the Transformer-XL network that integrates the medical constraint model, it is first necessary to ensure that the input data is compatible with the format of the network's input layer. This is typically achieved by adjusting the dimensions of the feature matrix using the network's built-in data preprocessing module to meet the Transformer-XL network's requirements for receiving temporal data. When using a segmented attention mechanism to divide temporal segments, the segment length is set in conjunction with the clinical oral monitoring cycle of virtual implant osseointegration. For example, if clinical osseointegration monitoring collects data monthly for a total of 6 months, the temporal segment length is set to the feature data span corresponding to a single month. This ensures that each segment can fully reflect the parameter changes within a certain stage of osseointegration. During the segmentation process, the network's temporal segmentation unit automatically identifies the time dimension markers in the feature matrix, ensuring the continuity and integrity of the segment division and avoiding feature breaks or repetitions.

[0048] When employing a network memory caching mechanism, a dedicated memory caching unit is set up after each attention module in the Transformer-XL network. This unit stores the feature information of the previous time segment after attention calculation in real time. When processing the current time segment, the network calls the previous feature data stored in the memory caching unit and performs joint attention calculation with the feature data of the current segment. In this way, the long-term dependency relationship between the two feature parameters is captured. For example, when processing the time segment of the third month, the memory caching unit will retrieve the feature information of the occlusal force parameters in the second month segment and perform correlation analysis with the alveolar bone resorption features in the third month segment to identify the long-term effect of occlusal force on alveolar bone resorption. At the same time, the feature data corresponding to the initial implantation angle is stored in the memory caching unit and continuously correlated with the positional offset features in the processing of subsequent time segments to explore the dynamic correlation pattern between the two as the osseointegration process changes.

[0049] When the Transformer-XL network makes predictions based on long-term dependencies, it first fuses the feature data of each time segment after attention calculation and memory association to form a global feature representation covering the entire osseointegration cycle. This global feature representation is then input into the network's prediction output layer. The output layer, combined with the fused medical constraint model, analyzes the global feature representation and, based on the captured long-term dependencies, generates predictions of the virtual implant's positional offset and bone resorption degree within the current osseointegration cycle. The positional offset prediction is based on feature mapping in a three-dimensional spatial coordinate system, clarifying the implant's offset trend and specific state in the X, Y, and Z axes. The bone resorption degree prediction distinguishes between vertical and horizontal resorption and analyzes the rate of change for each resorption type by combining the feature change patterns of each time segment. This ultimately forms a prediction result that comprehensively reflects the implant's state after osseointegration, providing a basis for adjusting subsequent initial implantation parameters.

[0050] The segmented attention mechanism and the adaptation to the oral monitoring cycle in the above steps enable the network to focus on the characteristic changes at different stages of osseointegration, thereby improving the accuracy of staged predictions. The memory caching mechanism captures long-term dependencies, effectively uncovering the potential correlation between parameters such as occlusal force and implantation angle and osseointegration results, avoiding prediction bias caused by ignoring long-term effects. The prediction results cover multi-dimensional data of three-dimensional offset and bone resorption, providing detailed and comprehensive decision-making basis for the subsequent comprehensive adjustment of initial implantation parameters, ensuring that the adjusted parameters can adapt to the dynamic occlusal relationship after osseointegration.

[0051] In one optional example, the adjusted implantation parameters (candidate optimal implantation parameters) are validated in multiple dimensions to ensure they meet clinical implantation requirements. Biomechanical validation: Combining occlusal force parameters from intraoral surface data, the force distribution of the implant at the adjusted position, angle, and depth is simulated in a 3D oral model. The peak stress at the implant-bone interface is calculated (≤30MPa, meeting the bone load-bearing threshold). If the peak stress exceeds the threshold, the parameter adjustment is returned for re-optimization. Osteointegration validation: The adjusted parameters are fed into the Transformer-XL long-sequence prediction network with medical constraints to re-predict the post-osseointegration positional shift and bone resorption. If the prediction results are both within clinically safe thresholds (positional shift ≤0.3mm, bone resorption ≤0.5mm / 6 months), the parameter is deemed satisfactory in terms of osseointegration. Denture fit verification: A virtual denture model is generated based on the adjusted parameters and assembled with the 3D oral model to verify the shape coordination between the denture and adjacent teeth (smooth edge transition, gap ≤0.1mm) and the fit with the gums (fit gap ≤0.05mm). If the fit meets the standard, the parameter is determined as the final optimal implantation parameter.

[0052] Understandably, besides occlusal force and alveolar bone resorption itself, the positional displacement of virtual implants and the degree of bone resorption are also affected by various factors such as local oral conditions, implant characteristics, patient's overall condition, and prosthetic design. If the overall alveolar bone density is low (e.g., bone tissue in patients with osteoporosis) or there is insufficient local bone volume (e.g., alveolar bone atrophy in the edentulous area), the contact area and anchoring force between the implant and bone tissue will decrease. Even if the initial placement is accurate, positional displacement is likely to occur during osseointegration due to insufficient stability. At the same time, the bone tissue's resistance to resorption weakens, significantly aggravating the degree of bone resorption. Bone homogeneity also has an impact. If there are local weak areas with uneven alveolar bone density in the edentulous area, bone resorption is likely to occur in the weak areas after the implant is subjected to force, leading to implant displacement towards the weak areas. Furthermore, the memory cache stores feature information of initial signs of bone resorption in weak areas from earlier segments. In subsequent segment processing, this information is correlated with the positional offset features of the current segment to uncover the long-term dependency that "accumulated bone resorption in weak areas leads to implant displacement towards weak areas," avoiding the omission of the gradual influence of bone factors due to neglecting temporal continuity. In addition, the memory cache can also correlate the interaction between bone features and other factors (such as occlusal force). For example, it stores early features of "increased bone resorption in low-density areas under moderate occlusal force." If the same combination of occlusal force and low-density features is detected in subsequent segments, historical correlation patterns are quickly invoked, improving prediction efficiency and accuracy. When integrating medical constraint models, clinical rules such as "lower bone resorption threshold in low-density areas" and "higher risk of positional displacement in weak bone areas" are transformed into regularization terms. When prediction results show situations that violate the laws of bone influence, such as "excessively low bone resorption in low-density areas" or "implant not shifting towards weak areas," the regularization terms increase the loss value. Backpropagation is used to adjust network parameters, ensuring that the prediction results align with the actual clinical impact of bone factors on implant status. Ultimately, the location offset and bone resorption prediction results output by the network will fully reflect the influence of alveolar bone density, bone volume, and uniformity. For example, in cases with localized weak areas, the predicted degree of bone resorption will be significantly higher in the weak areas than in other areas, and the location offset will also point towards the weak areas. Furthermore, the overall predicted offset and resorption amount of low-density, insufficient bone volume cases will be higher than those of normal bone cases, which is consistent with actual clinical patterns.

[0053] Step S105: Based on the positional offset of the virtual implant after osseointegration and the degree of bone resorption, dynamically adjust the initial implantation parameters of the virtual implant to obtain the optimal implantation parameters.

[0054] As an optional embodiment, in step S105, parameter adjustment rules are constructed based on the positional offset, degree of bone resorption, initial implantation parameters, and bone density data. These rules include positional offset adjustment rules, bone resorption degree adjustment rules, and multi-parameter collaborative rules. Based on these rules, the initial coordinates of the virtual implant are located in the three-dimensional digital oral model. According to the positional offset adjustment rules, the implant coordinates are moved in the opposite direction of the offset, and the gap between the adjusted implant and adjacent teeth, and the distance to the nerve pathway, are detected simultaneously. If the gap or distance is not up to standard, the current position of the virtual implant is iteratively fine-tuned. Based on the bone resorption degree adjustment rules, the implant axis angle is rotated in the three-dimensional digital oral model. Using occlusal relationship data, the occlusal contact state between the adjusted virtual implant and the opposing teeth is simulated, ensuring a uniform distribution of occlusal contact points and that each occlusal surface has more than three contact points. If the contact points are abnormal, the angle is corrected. Based on the initial depth calculated from bone resorption, the length of the virtual implant placed into the alveolar bone was adjusted using bone density distribution data in a 3D digital oral model. This ensured that the implant apex was located in a high-density bone region and at least 2 mm from the alveolar bone floor. Multidimensional compatibility verification was then performed on the adjusted initial implantation parameters to obtain the optimal implantation parameters that passed the verification.

[0055] Specifically, in step S105, when constructing parameter adjustment rules, it is necessary to deeply integrate the correlation logic between positional offset, bone resorption degree, initial implantation parameters, and bone density data. This ensures that positional offset adjustment rules can specifically counteract offset trends, bone resorption adjustment rules can proactively address resorption risks, and multi-parameter collaborative rules can coordinate the mutual influence between parameters, avoiding new problems caused by a single adjustment. When adjusting the position in the 3D digital oral model, the initial coordinates of the virtual implant are first locked using the model positioning function. Then, the coordinates are moved in the opposite direction of the predicted offset according to the positional offset adjustment rules. Simultaneously, the gaps and distances between adjacent teeth and nerve pathways in the model are detected in real time. If the requirements are not met, small-amplitude iterative fine-tuning is used to gradually optimize the position, ensuring that the adjusted position avoids offset risks while meeting safe distance requirements. In the angle adjustment stage, the implant axis angle is changed in the model using the rotation function. The contact state between the prosthesis and opposing teeth is simulated by combining the occlusal relationship in the intraoral surface data. The distribution of contact points is observed using visualization technology. If abnormalities such as concentration or absence are found, the angle is continuously corrected until the contact points are uniform and the number meets the requirements. Depth adjustment requires calculating the necessary increase in implant length based on bone resorption. By incorporating a bone density distribution layer into the model, the implant depth is precisely controlled to ensure the root apex rests in a high-density bone region while maintaining a safe distance from the alveolar bone floor, balancing anchoring force with the space reserved for bone resorption. Finally, multidimensional fit verification integrates spatial relationships, occlusal data, bone parameters, and other multi-source information from the model. This comprehensive evaluation of the adjusted parameters from the perspectives of safety, functionality, and stability ensures that all indicators meet clinical standards.

[0056] Thus, the systematic construction of parameter adjustment rules provides clear clinical logic support for initial parameter adjustments, avoiding the subjectivity and randomness of empirical adjustments. Dynamic adjustment and real-time monitoring within the 3D model enable visualization and precision in parameter optimization, ensuring that adjustments to implant position, angle, and depth not only offset the effects of post-osseointegration displacement and resorption but also strictly adhere to safety constraints such as interdental spaces and nerve distances. Angle adjustment optimizes the occlusal contact point, improving the functional fit of the implant restoration and reducing the risk of secondary damage caused by occlusal abnormalities. Depth adjustment, combined with precise control of bone density, enhances the initial stability and long-term anchoring effect of the implant, reducing the risk of loosening caused by bone resorption. Multidimensional fit verification comprehensively ensures the clinical feasibility of the adjusted parameters. By integrating multiple evaluation results, the final optimal implantation parameters achieve a balance between safety, functionality, and stability, providing reliable digital guidance for actual implant surgery and improving implant success rates and patient outcomes.

[0057] As an optional embodiment, after step S105, based on the intraoral surface data, a graph neural network combined with reinforcement learning algorithms can be used to simulate the dynamic contact process in the oral cavity to obtain predicted data of the occlusal surface of the prosthesis; a reward function is set based on occlusal comfort and force distribution, and combined with the predicted data of the occlusal surface of the prosthesis, the dynamic occlusal relationship of the virtual implant in the oral cavity is optimized; candidate modification schemes for optimal implantation parameters are generated according to the dynamic occlusal relationship, and the candidate modification schemes are dynamically simulated and displayed through a three-dimensional digital oral model, allowing the user to select the final target modification scheme. This process captures the topological association and dynamic contact features of the oral structure through a graph neural network, and continuously optimizes the occlusal relationship by combining the reward mechanism of reinforcement learning. This not only improves the fitting accuracy of the virtual implant with the dynamic function of the oral cavity, but also gives the user decision-making participation through visual interaction, making the final scheme more in line with individual oral functional needs and clinical operation reality.

[0058] Further optionally, in the above optional embodiments, the prosthesis occlusal surface prediction data is obtained by simulating the dynamic contact process in the oral cavity using a graph neural network combined with reinforcement learning algorithms based on the intraoral surface data. This includes: extracting features from the intraoral surface data, extracting the three-dimensional contour features of adjacent teeth, the contact point distribution features of the occlusal relationship, and the elastic deformation features of the gingival morphology; and standardizing the extracted multidimensional features to obtain standardized multidimensional intraoral surface features. Extracting the three-dimensional contour features of adjacent teeth, the contact point distribution features of the occlusal relationship, and the elastic deformation features of the gingival morphology essentially extracts the core elements directly related to dynamic contact from complex intraoral surface data. Standardization, by unifying the units and data range, eliminates numerical differences between different features, ensuring that the multidimensional intraoral surface features can be fairly analyzed by the model, providing a consistent and effective input basis for subsequent modeling.

[0059] Furthermore, based on the aforementioned multidimensional intraoral surface features, using individual teeth, gingival functional areas, and the denture area corresponding to the virtual implant as nodes, the core structures involved in occlusal contact within the oral cavity can be accurately located. Using the spatial positional relationships and contact associations between nodes as edges, an initial oral structure association graph is constructed. This allows for the quantification of functional connections between various structures. This graph structure transforms dispersed oral features into an interconnected whole, enabling the model to capture the coordinated changes of various parts during occlusion from a global perspective, providing a structured computational framework for dynamic contact simulation.

[0060] Furthermore, the initial occlusal surface morphology of the prosthesis is generated based on the optimal implantation parameters. This drives a dynamic graph neural network model optimized by a reinforcement learning algorithm to perform multiple rounds of dynamic contact simulation according to a preset oral motion trajectory. In each round of simulation, the reinforcement learning agent outputs occlusal surface adjustment actions, and the dynamic graph neural network updates the contact state prediction values ​​in real time. The initial morphology is generated based on the optimal implantation parameters, ensuring that the simulation starting point meets clinical requirements. The preset oral motion trajectory (such as typical action paths like chewing and swallowing) provides a realistic motion scenario for the simulation. The occlusal surface adjustment actions output by the reinforcement learning agent in each round are essentially strategy optimizations based on the current contact state, while the dynamic graph neural network dynamically reflects the impact of the adjustment actions on the overall occlusal relationship by updating the contact state prediction values ​​between nodes in real time (such as contact point position and force). The collaboration between the two ensures that the simulation process not only follows the physiological laws of oral movement but also approximates the ideal contact state through intelligent optimization.

[0061] Finally, the model converges through multiple rounds of iterative training. When the reward function value reaches a preset threshold, the predicted occlusal surface data of the prosthesis during the complete oral motion cycle is output. This predicted occlusal surface data includes at least: the three-dimensional morphological coordinates of the occlusal surface, the sequence of contact points with adjacent / opposing teeth at different times, the contact force change curve, and the deformation trend of the gingival contact area. Specifically, the reward function uses occlusal comfort and force distribution as core indicators, awarding higher rewards for better contact conditions, which in turn guides the agent to adjust its action strategy, driving the model to gradually converge. When the reward function value reaches the preset threshold, it means that the simulated dynamic contact process has met clinical functional requirements. At this point, the output complete cycle prediction data (covering the three-dimensional coordinates of the occlusal surface, the contact point sequence, the force curve, and the gingival deformation trend) comprehensively depicts the dynamic characteristics of the prosthesis in various oral movements, providing fine-grained, temporal decision-making basis for subsequent optimization of the dynamic occlusal relationship of the virtual implant. This ensures the physiological rationality of the predicted data and improves the accuracy of dynamic occlusal optimization.

[0062] Further optionally, in the above optional embodiments, before initializing the occlusal surface morphology of the prosthesis to an initial morphology generated based on the optimal implantation parameters, and driving the dynamic graph neural network model optimized by the reinforcement learning algorithm to perform multiple rounds of dynamic contact simulation according to the preset oral motion trajectory, a dynamic graph neural network model can also be constructed. The initial oral structure association graph is input into the initial model, and the static spatial dependencies between nodes are learned through graph convolutional layers. The dynamic interaction features of nodes within the oral motion cycle are captured through temporal recurrent layers, and the predicted values ​​of occlusal contact states at different times are output. The predicted values ​​of occlusal contact states include contact position and contact force. For example, the spatial distance weight between the prosthesis node and the adjacent tooth node, and the contact probability weight between the gingival node and the prosthesis node are calculated to capture the inherent spatial association of the oral structure in a static state, laying a static structural foundation for dynamic simulation. Then, a temporal recurrent layer (such as an LSTM or GRU layer) is connected. This layer combines the static features output by the graph convolutional layer with the motion parameters (such as tooth displacement and gingival deformation coefficient) at each step within the oral motion cycle, and captures the dynamic interaction features of the nodes through recurrent memory units. For example, it records the temporal changes in the contact position between the denture nodes and the opposing tooth nodes during chewing movements, as well as the deformation trend of the gingival nodes with occlusal force, and finally outputs the predicted values ​​of occlusal contact state at different times. The contact position accurately locates the actual contact coordinates between each node, and the contact force quantifies the pressure value at contact, providing a basis for state feedback for subsequent reinforcement learning.

[0063] Furthermore, a decision-making framework is constructed using reinforcement learning algorithms. The predicted values ​​of occlusal contact states at different times are used as state inputs. The morphological adjustment parameters of the occlusal surface of the denture are used as the action space. The reward function is designed with indicators such as the uniformity of occlusal contact points, control of peak force, and gingival compression. The morphological adjustment parameters include the fine-tuning of cusp height and pit and fissure depth.

[0064] Specifically, the predicted values ​​of occlusal contact state (contact position, contact force) at different times output by the dynamic graph neural network are used as state inputs to ensure that the reinforcement learning agent can perceive the current occlusal state in real time. The morphological adjustment parameters of the denture occlusal surface are defined as the action space, specifically including the fine adjustment of cusp height (e.g., ±0.1-0.3mm) and the fine adjustment of pit and fissure depth (e.g., ±0.05-0.2mm). These parameters not only conform to the actual operational dimensions of clinical denture morphological correction, but also allow for precise optimization of the contact state through fine adjustments. At the same time, a reward function is designed with occlusal contact point uniformity (e.g., the standard deviation of the distribution of contact points on the occlusal surface), peak force control (e.g., whether the maximum peak contact force is lower than the tolerance threshold of bone and gingiva), and gingival compression (e.g., whether the contact force at the gingival node is within the safe range) as core indicators. When the contact points are more uniform, the peak force is lower, and the gingival compression is smaller, the reward function outputs a higher score, and vice versa, thereby clarifying the optimization direction of reinforcement learning.

[0065] In summary, the dynamic graph neural network, through its two-layer architecture of static spatial learning and dynamic temporal capture, ensures both accurate understanding of the inherent structural relationships of the oral cavity and real-time tracking of interactive changes during movement. This allows the predicted occlusal contact state to possess both spatial accuracy and temporal continuity, providing reliable state feedback for subsequent simulations. The reinforcement learning decision framework, by using action spaces aligned with clinical reality and multi-dimensional reward indicators, transforms abstract occlusal optimization goals into quantifiable and executable algorithmic logic, preventing optimization directions from deviating from clinical needs. The algorithmic foundation jointly constructed by these two approaches provides stable model support and a clear optimization path for subsequent multi-round dynamic contact simulations, ensuring that the simulation process continuously converges towards the preset goal. This lays the technical foundation for ultimately outputting high-quality prosthesis occlusal surface prediction data and implant parameter correction schemes.

[0066] Step S106: Push a visual dynamic model containing candidate positioning strategies to the user based on the optimal implantation parameters.

[0067] Among them, the candidate positioning strategy is a variety of feasible implant positioning schemes derived from the optimal implantation parameters (position, angle, and depth) determined in step S105. In addition to core parameters such as implant three-dimensional coordinates, axis angle, and implantation depth, it also covers implantation path planning (such as a safe implantation trajectory that avoids nerve channels and adjacent tooth roots) and spatial adaptation schemes between implants and surrounding tissues (such as distance from the gingival margin and control of gaps with adjacent teeth). Each candidate strategy meets the clinical safety threshold and adapts to different clinical operation scenarios (such as differences in surgical approach and patient oral space limitations), providing users with a multi-dimensional selection space.

[0068] The visualized dynamic model is based on an optimized 3D digital oral model and integrates an interactive dynamic display model constructed with candidate positioning strategies. Through 3D modeling technology, it visualizes structures such as implants, alveolar bone, adjacent teeth, and gingiva. It supports dynamic simulation of the implantation process (such as step-by-step display of the implantation path and tissue contact relationship) and oral motion scenarios (such as the dynamic occlusal contact between the implant and the opposing tooth during chewing). It can also intuitively present the spatial relationship between the implant and nerves, adjacent teeth, and high-density bone areas through interactive operations such as zooming, rotation, and cross-section viewing, transforming abstract parameter data into concrete visual information.

[0069] Specifically, in step S106, firstly, based on the optimal implantation parameters and the flexibility requirements of clinical operation, 2-3 candidate positioning strategies are generated. For example, based on the optimal implantation angle, a fine-tuning scheme with preset dimensions is derived, or an offset scheme of 0.1-0.2mm is designed around the optimal position, ensuring that each strategy meets the clinical safety standards of a distance ≥2mm from the nerve, a gap ≥1.5mm from adjacent teeth, and the root apex located in a high-density bone area. Secondly, a visual dynamic model is constructed, integrating the candidate positioning strategies into the optimized three-dimensional digital oral model. Using professional medical modeling software such as Meshmixer and 3-Matic, the three-dimensional implant models corresponding to each candidate strategy, including dimensions and thread morphology, are loaded into the oral model. A dynamic display module is set up to simulate the complete implantation path of the implant from outside the oral cavity into the alveolar bone, and to mark the minimum distance between the implant and the nerve and adjacent teeth in real time during the implantation process. At the same time, chewing movements are simulated, driving the upper and lower jaw models to move according to the preset occlusal trajectory, and dynamically displaying the distribution of contact points and changes in force areas between the implant-corresponding prosthesis and the opposing teeth. Finally, the visualized dynamic model is pushed to the user through a dedicated medical interaction platform, such as local software or a cloud system. Users can zoom in on local areas to view the fitting details between the gum and the implant neck, switch between different candidate strategies to compare the dynamic occlusal effects of each strategy, and export key parameter annotation diagrams to present annotation documents of implant angle and depth, helping users to intuitively evaluate the feasibility of each candidate strategy.

[0070] In step S107, after the user confirms the candidate positioning strategy in the visualized dynamic model, the oral model printer is driven to print the implant guide model with resin liquid and the oral environment model with plaster to assist the user in verifying the installation effect of the candidate positioning strategy.

[0071] Among them, the dental model printer is a high-precision 3D printing device specifically designed for the field of dental implantology. It supports printing with a variety of biocompatible materials and employs differentiated printing technologies for different model needs. When printing implant guides, light-curing molding technologies (such as SLA and DLP) are often used, using liquid resin as the material and ultraviolet light to cure the resin layer by layer, achieving an accuracy of 0.02-0.05mm to ensure precise matching between the guide positioning holes and the implant path. When printing oral environment models, powder molding technologies (such as 3DP) are often used, using plaster as the material. Plaster powder is cured by spraying a binder layer by layer to restore the three-dimensional morphology of the alveolar bone, the gingival contour, and the positional relationship of adjacent teeth, simulating the real oral anatomy.

[0072] In the specific implementation of step S107 above, optionally, firstly, after the user confirms the optimal candidate positioning strategy through the visual dynamic model, the system exports two types of model data corresponding to the strategy: a 3D model of the implant guide and a 3D model of the oral environment. The 3D model of the implant guide designs positioning holes and guiding channels according to the candidate strategy to ensure that the implant is inserted along the preset path. The 3D model of the oral environment restores the alveolar bone morphology, adjacent tooth positions, and gingival contours of the patient's edentulous area. Both types of model data are converted into STL format compatible with oral model printers. Secondly, materials and printing technologies are selected according to the model functions. For example, the implant guide uses a biocompatible light-curing resin liquid, such as medical-grade photosensitive resin, which is cured layer by layer by a DLP light-curing printer. After printing, the support structure is removed and the surface is polished to ensure that the edges of the guide are smooth and burr-free, ensuring the fit between the guide and the oral environment model. The oral environment model uses medical-grade plaster powder, which is cured layer by layer by spraying adhesive through a 3DP printer. After molding, it is dried to avoid deformation during subsequent use. At the same time, the density distribution difference of the alveolar bone is restored by adjusting the bulk density of the plaster powder to simulate high-density and low-density areas. Finally, the installation effect is verified by assisting the user. The implant guide is placed in the corresponding position on the oral environment model to check whether the guide perfectly conforms to the gingival contour and whether the positioning holes are aligned with the implant site in the edentulous area. Then, a resin simulated implant, identical in size to the actual implant, is inserted into the oral environment model through the positioning holes on the guide. This verifies whether the implant's insertion angle and depth are consistent with the candidate strategy. Simultaneously, the gap between the implant and adjacent teeth, and the distance from the simulated nerve pathway, are checked to ensure they meet safety requirements, guaranteeing the feasibility of the candidate positioning strategy in the physical model and providing a physical reference for the actual surgery.

[0073] In this embodiment, the collaborative operation of oral 3D scanning technology, SegFormer segmentation network, long sequence prediction network and digital model printing technology realizes the construction process of individualized denture installation and positioning model for patients' oral cavity, assists users in verifying the installation effect of candidate positioning strategies, greatly improves denture installation and positioning efficiency, and enhances the intelligence level, positioning accuracy and safety redundancy of denture installation.

[0074] After introducing the methods of exemplary embodiments of this application, the following references are made. Figure 2 This application describes an exemplary embodiment of a digital denture placement and positioning device based on three-dimensional scanning. The device includes: an oral scanner for acquiring three-dimensional anatomical data and intraoral surface data of a patient's oral cavity; the three-dimensional anatomical data includes bone density and nerve distribution, and the intraoral surface data includes adjacent tooth morphology, occlusal relationship, and gingival morphology; a user terminal for loading a SegFormer segmentation network using 3SHAPE software, elastically aligning the three-dimensional anatomical data and the intraoral surface data to generate an initial model; constructing a three-dimensional deformation field using the Swin-UMamba algorithm, and superimposing and fusing the three-dimensional deformation field into the initial model to generate a three-dimensional digital oral model; loading and adapting the SegFormer segmentation network using 3SHAPE software, and inputting the three-dimensional digital oral model into the SegFormer segmentation network to generate a virtual implant for the patient and corresponding initial implantation parameters; the SegFormer segmentation network is a semantic segmentation network combining a Transformer architecture and a lightweight MLP decoder. The system comprises a network; initial implantation parameters including the initial implantation position, angle, and depth of the virtual implant; a user terminal, which, based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, uses a long-sequence prediction network with a segmented attention mechanism matched to the oral monitoring cycle and a network memory caching mechanism associated with the context to predict the positional offset and degree of bone resorption of the virtual implant after osseointegration; and dynamically adjusts the initial implantation parameters of the virtual implant to obtain optimal implantation parameters based on the positional offset and degree of bone resorption of the virtual implant after osseointegration; wherein the long-sequence prediction network is a Transformer-XL network incorporating a medical constraint model; the user terminal, which also pushes a visual dynamic model containing candidate positioning strategies to the user based on the optimal implantation parameters; and an oral model printer, which, after the user confirms the candidate positioning strategies in the visual dynamic model, drives the oral model printer to print an implant guide model using resin and an oral environment model using plaster, assisting the user in verifying the installation effect of the candidate positioning strategies. The above device can implement the steps described in the above method implementation, and the specific implementation methods of each step will not be repeated here.

[0075] After introducing the methods and apparatus of the exemplary embodiments of this application, a terminal device of the exemplary embodiments of this application will be described next. The terminal device can implement the steps described in the above method embodiments, and the specific implementation of each step will not be repeated here.

[0076] After introducing the methods, apparatus, and terminal devices of exemplary embodiments of this application, the following references are made. Figure 3 The computer-readable storage medium of exemplary embodiments of this application will be described, please refer to... Figure 3 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above method embodiments. The specific implementation methods of each step will not be repeated here.

[0077] It should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Therefore, the protection scope of this application should be determined by the scope of the claims.

Claims

1. A digital denture installation and positioning method based on three-dimensional scanning, characterized in that, The method includes: Three-dimensional anatomical data of the patient's oral cavity and intraoral surface data are obtained through an oral scanner; The three-dimensional anatomical data and the intraoral surface data are elastically aligned to generate an initial model; a three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and the three-dimensional deformation field is superimposed and fused into the initial model to generate a three-dimensional digital oral cavity model. The SegFormer segmentation network is loaded and adapted using 3SHAPE software, and the 3D digital oral model is input into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters. The SegFormer segmentation network is a semantic segmentation network that combines the Transformer architecture and a lightweight MLP decoder. Based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, a long sequence prediction network is used to predict the positional displacement and degree of bone resorption of the virtual implant after osseointegration through a segmented attention mechanism matched with the oral monitoring cycle and a network memory caching mechanism associated with the context. The long sequence prediction network is a Transformer-XL network that integrates a medical constraint model. Based on the positional offset of the virtual implant after osseointegration and the degree of bone resorption, the initial implantation parameters of the virtual implant are dynamically adjusted to obtain the optimal implantation parameters. Based on the optimal implantation parameters, push a visual dynamic model containing candidate positioning strategies to the user; After the user confirms the candidate positioning strategy in the visualized dynamic model, the dental model printer is driven to print the implant guide model using resin liquid and the oral environment model using plaster.

2. The digital denture installation and positioning method based on three-dimensional scanning according to claim 1, characterized in that, After dynamically adjusting the initial implantation parameters of the virtual implant to obtain optimal implantation parameters based on the positional offset of the virtual implant after osseointegration and the degree of bone resorption, the method further includes: Based on the intraoral surface data, a graph neural network combined with reinforcement learning algorithm is used to simulate the dynamic contact process in the oral cavity to obtain the occlusal surface prediction data of the denture; A reward function is set based on occlusal comfort and force distribution, and combined with the predicted data of the occlusal surface of the prosthesis, the dynamic occlusal relationship of the virtual implant in the oral cavity is optimized. Based on the dynamic occlusal relationship, candidate modification schemes for optimal implantation parameters are generated, and the candidate modification schemes are dynamically simulated and displayed through a three-dimensional digital oral model, allowing the user to select the final target modification scheme.

3. The digital denture installation and positioning method based on three-dimensional scanning according to claim 2, characterized in that, The method of obtaining prosthesis occlusal surface prediction data by simulating the dynamic contact process in the oral cavity using a graph neural network combined with reinforcement learning algorithms based on the intraoral surface data includes: Feature extraction is performed on the intraoral surface data to extract the three-dimensional contour features of adjacent teeth, the contact point distribution features of occlusal relationships, and the elastic deformation features of gingival morphology. The extracted multidimensional features are then standardized to obtain standardized multidimensional intraoral surface features. Based on the aforementioned multidimensional intraoral surface features, an initial oral structure association diagram is constructed using individual teeth, gingival functional areas, and the denture area corresponding to the virtual implant as nodes, and the spatial positional relationship and contact association between nodes as edges. The initial occlusal surface morphology of the denture is generated based on the optimal implantation parameters. The dynamic graph neural network model optimized by the reinforcement learning algorithm performs multiple rounds of dynamic contact simulation according to the preset oral motion trajectory. In each round of simulation, the occlusal surface adjustment action is output by the reinforcement learning agent, and the dynamic graph neural network updates the contact state prediction value in real time. The model is trained through multiple rounds of iterations until it converges. When the reward function value reaches the preset threshold, the predicted data of the denture occlusal surface during the complete oral movement cycle is output. Among them, the predicted data of the occlusal surface of the denture during the complete oral movement cycle includes at least: the three-dimensional morphological coordinates of the occlusal surface, the sequence of contact points with adjacent / opposing teeth at different times, the curve of contact force change, and the deformation trend of the gingival contact area.

4. The digital denture installation and positioning method based on three-dimensional scanning according to claim 3, characterized in that, The initial occlusal surface morphology of the prosthesis is an initial morphology generated based on optimal implantation parameters. Before the dynamic graph neural network model optimized by the reinforcement learning algorithm performs multiple rounds of dynamic contact simulation according to a preset oral motion trajectory, the following steps are also included: A dynamic graph neural network model is constructed by inputting the initial oral structure association graph into the initial model, learning the static spatial dependencies between nodes through graph convolutional layers, and capturing the dynamic interaction features of nodes within the oral movement cycle through temporal recurrent layers, and outputting the predicted values ​​of occlusal contact state at different times; the predicted values ​​of occlusal contact state include contact position and contact force. A decision-making framework is constructed using reinforcement learning algorithms. The predicted values ​​of occlusal contact states at different times are used as state inputs. The morphological adjustment parameters of the occlusal surface of the denture are used as the action space. The reward function is designed with indicators such as the uniformity of occlusal contact points, control of peak force, and gingival compression. The morphological adjustment parameters include the fine-tuning of cusp height and pit and fissure depth.

5. The digital denture installation and positioning method based on three-dimensional scanning according to claim 1, characterized in that, The method employs the Swin-UMamba algorithm to construct a three-dimensional deformation field, and combines this field to dynamically repair scanning blind spots in the initial model, generating a three-dimensional digital oral model, including: The initial model and blind zone marking map are used as inputs to import the Swin-UMamba algorithm model; The encoder of the Swin-UMamba algorithm model extracts multi-scale features from the initial model, and encodes the extracted macroscopic structural features and the microscopic features around the blind zone in a hierarchical manner. By combining the decoder with the blind spot marker map, the multi-scale features after hierarchical encoding are mapped into the three-dimensional deformation field required for blind spot repair. The three-dimensional deformation field is superimposed and fused into the initial model to dynamically repair the scanning blind spots in the initial model, resulting in an optimized three-dimensional digital oral model.

6. The digital denture installation and positioning method based on three-dimensional scanning according to claim 1, characterized in that, The step of elastically aligning the three-dimensional anatomical data and the intraoral surface data to generate an initial model includes: The three-dimensional anatomical data and the intraoral surface data are preprocessed to extract alveolar bone feature points from the three-dimensional anatomical data and oral soft tissue feature points from the intraoral surface data. Using the pre-processed alveolar bone feature points and oral soft tissue feature points as matching benchmarks, the rigid transformation matrix of the alveolar bone feature points and oral soft tissue feature points is calculated by the least squares method. The three-dimensional anatomical data and the intraoral surface data are initially calibrated to make the alveolar bone in the edentulous area initially coincide with the corresponding gingival area. A non-rigid registration deformation network was constructed using a 3D Swin Transformer encoder and a diffuse deformation field decoder. The pre-aligned three-dimensional anatomical data and intraoral surface data were used as inputs to the non-rigid registration deformation network. The encoder extracted multi-scale features from the pre-aligned three-dimensional anatomical data and intraoral surface data, and the decoder converted the multi-scale features into a non-rigid deformation field. The non-rigid deformation field is used to locally deform and adjust the pre-aligned three-dimensional anatomical data, so that the fitting boundary between the alveolar bone surface and the gingival attachment area in the pre-aligned three-dimensional anatomical data is optimized and matched. The bone density distribution and the outline of adjacent teeth in the pre-aligned three-dimensional anatomical data and intraoral surface data are retained as original detailed features to construct the initial model.

7. The digital denture installation and positioning method based on three-dimensional scanning according to claim 6, characterized in that, The process of inputting a three-dimensional digital oral model into the SegFormer segmentation network to generate a virtual implant for the patient and corresponding initial implantation parameters includes: The optimized 3D digital oral model undergoes data format conversion, transforming the model's 3D point cloud data into voxelized data compatible with the SegFormer segmentation network; combined with the original detailed features, the voxelized data is then enhanced in terms of features. The enhanced voxelized data is input into the SegFormer segmentation network. The hierarchical feature extraction architecture of the SegFormer segmentation network is used to identify the implant candidate region. Combined with the optimized and matched boundary between the alveolar bone surface and the gingival attachment region, the implantable region of the edentulous area is segmented from the implant candidate region and converted into an implant region segmentation mask map. Based on the implantation region segmentation mask map, virtual implants and initial implantation parameters are generated in combination with clinical implant selection rules; The generated virtual implant and initial implantation parameters are validated for compatibility. If the validation is successful, the final virtual implant and corresponding initial implantation parameters are output.

8. The digital denture installation and positioning method based on three-dimensional scanning according to claim 1, characterized in that, Based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters, a long-sequence prediction network is used. This network employs a segmented attention mechanism matched to the oral monitoring cycle and a network memory caching mechanism associated with the context to predict the virtual implant's positional displacement and degree of bone resorption after osseointegration. Historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters are standardized and integrated to form a multi-source temporal feature matrix; Combining clinical treatment guidelines and physiological characteristics of dental implants, a medical constraint model was constructed, which includes position offset constraint rules and bone resorption degree constraint rules, to define the physiological rationality boundary of the prediction results; and the medical constraint model was used as a regularization term in the loss function of the Transformer-XL network. The multi-source temporal feature matrix is ​​input into the Transformer-XL network that integrates medical constraint models. The multi-source temporal feature matrix is ​​divided into temporal segments using a segmented attention mechanism, and the segment length is set to match the oral monitoring cycle. A network memory caching mechanism is employed to capture the long-term dependencies between the previous and current time segments on feature parameters. These long-term dependencies include the correlation between the long-term effect of occlusal force and alveolar bone resorption, as well as the dynamic correlation between the initial implantation angle and positional offset. The prediction results of the virtual implant positional offset and degree of bone resorption in the current patient during the osseointegration period are obtained through the prediction of the long-term dependencies.

9. The digital denture installation and positioning method based on three-dimensional scanning according to claim 1, characterized in that, The process of dynamically adjusting the initial implantation parameters of the virtual implant to obtain optimal implantation parameters based on the positional offset of the virtual implant after osseointegration and the degree of bone resorption includes: Based on the positional offset, degree of bone resorption, initial implantation parameters, and bone mineral density data, parameter adjustment rules are constructed. These rules include positional offset adjustment rules, bone resorption degree adjustment rules, and multi-parameter synergy rules. Based on parameter adjustment rules, the initial coordinates of the virtual implant are located in the three-dimensional oral digital model. According to the position offset adjustment rules, the implant coordinates are moved in the opposite direction of the offset. The gap between the implant and the adjacent teeth and the distance to the nerve channel are detected simultaneously. If the gap or distance does not meet the standard, the current position of the virtual implant is iteratively fine-tuned. Based on the bone resorption adjustment rules, the implant axis angle is rotated in the three-dimensional oral digital model. Through occlusal relationship data, the occlusal contact state between the corresponding prosthesis and the opposing tooth after adjustment is simulated to make the occlusal contact points evenly distributed and the number of contact points on each occlusal surface greater than 3. If the contact points are abnormal, the angle is corrected. Based on the initial depth calculated and adjusted according to the amount of bone resorption, the length of the virtual implant inserted into the alveolar bone is adjusted in the three-dimensional oral digital model in combination with the bone density distribution data, so that the apex of the implant is located in the high-density bone area and is more than 2mm away from the bottom of the alveolar bone. The adjusted initial implantation parameters are subjected to multidimensional compatibility verification to obtain the optimal implantation parameters that pass the verification.

10. A digital denture mounting and positioning device based on three-dimensional scanning, characterized in that, include: An oral scanner is used to acquire three-dimensional anatomical data of a patient's oral cavity and data on the intraoral surface. The user terminal is used to load the SegFormer segmentation network using 3SHAPE software, and elastically align the three-dimensional anatomical data and the intraoral surface data to generate an initial model; a three-dimensional deformation field is constructed using the Swin-UMamba algorithm, and the three-dimensional deformation field is superimposed and fused into the initial model to generate a three-dimensional digital oral model; the SegFormer segmentation network is loaded and adapted using 3SHAPE software, and the three-dimensional digital oral model is input into the SegFormer segmentation network to generate the patient's virtual implant and corresponding initial implantation parameters; the SegFormer segmentation network is a semantic segmentation network that combines a Transformer architecture and a lightweight MLP decoder; The user terminal is also used to predict the positional displacement and degree of bone resorption of the virtual implant after osseointegration based on historical surgical data, initial implantation parameters, bone density, preoperative patient attributes, and occlusal force parameters. This is achieved using a long-sequence prediction network with a segmented attention mechanism matched to the oral monitoring cycle and a network memory caching mechanism associated with the context. Based on the positional displacement and degree of bone resorption of the virtual implant after osseointegration, the initial implantation parameters of the virtual implant are dynamically adjusted to obtain the optimal implantation parameters. The long-sequence prediction network is a Transformer-XL network that integrates a medical constraint model. Based on the optimal implantation parameters, a visual dynamic model containing candidate positioning strategies is pushed to the user. The dental model printer is used to print implant guide models using resin liquid and dental environment models using plaster after the user confirms the candidate positioning strategy in the visualized dynamic model.

Images