Tooth intelligent segmentation and parameterized bionic reconstruction method fusing multi-modal data

Abstract

The invention discloses a tooth intelligent segmentation and parameterized bionic reconstruction method fusing multi-modal data, and belongs to the field of artificial intelligence and computer-aided biomedical engineering. The method comprises a CBCT intelligent segmentation module, an intraoral scanning intelligent segmentation module, an internal tissue segmentation module, a point cloud generation registration module and a parameterized bionic reconstruction module. The point cloud generation registration module further comprises a point cloud generation sub-module and a point cloud registration sub-module; the parameterized bionic reconstruction module further comprises a dentin reconstruction sub-module, a cancellous bone reconstruction sub-module and a gingival reconstruction sub-module. The invention aims to realize tooth semantic segmentation, point cloud reconstruction and parameterized bionic modeling of an organization structure under cross-modal data fusion, and provides high-precision bionic internal structure information for teeth in combination with an artificial intelligence model and parameterized modeling.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and computer-aided biomedical engineering, and specifically relates to a method for intelligent tooth segmentation and parametric bionic reconstruction based on multimodal data. Background Art

[0002] In the process of oral clinical diagnosis and treatment, obtaining high-quality three-dimensional tooth models is the basis for achieving personalized restoration design and precise treatment. Traditional three-dimensional modeling methods mostly rely on manual modeling or modeling methods based on anatomical atlases. Not only is the modeling efficiency low, but the modeling results are also greatly affected by the operator's subjective experience, and there are problems such as poor consistency and limited reconstruction accuracy. CBCT has been widely used in stomatology because of its high spatial resolution and low radiation dose, which can clearly present the three-dimensional structure of internal tissues such as enamel, dentin, and cementum. However, due to inherent factors such as metal artifacts, image noise, and similar bone tissue density, traditional image processing algorithms such as threshold segmentation and region growing method still face many challenges in accurately segmenting teeth using CBCT, which can easily lead to problems such as blurred boundaries and loss of details, making it difficult to meet clinical precision requirements.

[0003] With the development of deep learning, the widespread application of convolutional neural networks in medical image segmentation has provided a better approach for automatic tooth segmentation. However, the upsampling and downsampling processes may still lead to the loss of detail information and edge features, resulting in uneven surface of the segmentation results. To this end, some studies have attempted to introduce traditional image processing algorithms for post-processing to smooth the surface, but such methods often simply fill in the potholes and fail to restore the true structural information of the teeth, resulting in significant differences from real teeth. Multimodal data fusion methods make up for the shortcomings of a single modality. Combining the tissue penetration ability of CBCT with the high-resolution crown information provided by intraoral scanning is an important approach to improve the accuracy of tooth modeling. However, due to the large differences between the two data sources in acquisition methods, resolution, and image features, how to achieve efficient and robust data alignment and information fusion remains a difficult problem in current research.

[0004] Existing 3D tooth reconstruction techniques primarily focus on geometric restoration, reconstructing tooth contours while neglecting the complex internal microstructure and biological functions of teeth. As a typical biological structure, the functionality of teeth depends not only on their overall morphology but also on the synergistic effects of different tissues in terms of material properties, arrangement, and microstructure. Traditional geometric modeling methods are unable to express key physiological structures such as the directional arrangement of dentinal tubules, the porous honeycomb structure of cancellous bone, and the connection between gingival fibers, limiting their practical application in areas such as bite force simulation, biomaterial evaluation, and implant design.

[0005] Therefore, there is an urgent need for a method for intelligent segmentation and parametric bionic reconstruction of teeth that integrates multimodal data, can identify the multi-level internal tissue structure characteristics of teeth, and realize parametric bionic modeling based on tissue anatomy and mechanical functions. Intelligent segmentation and parametric bionic reconstruction of teeth is a highly complex systematic project, involving multidisciplinary cross-disciplinary knowledge such as medical image processing, computer vision, and biomechanical modeling. The development of a complete segmentation, registration, and reconstruction process will not only help promote the intelligent and personalized development of oral digital diagnosis and treatment, but also provide strong technical support for multiple scenarios such as clinical implants, orthodontics, restorations, and oral teaching. This direction has become a research hotspot and technical difficulty in the current field of oral medicine. Summary of the Invention

[0006] This paper presents a method for intelligent tooth segmentation and parametric biomimetic reconstruction that fuses CBCT and intraoral scan data. The method aims to achieve semantic segmentation, point cloud reconstruction, and biomimetic modeling of tooth structure using cross-modal data fusion. This method leverages the advantages of CBCT in internal tissue imaging and the high-precision capabilities of intraoral scanning in capturing tooth surface textures. Combining artificial intelligence algorithms with parametric modeling strategies, it provides complete three-dimensional internal structural information for teeth.

[0007] The technology of the present invention is implemented as follows: A method for intelligent tooth segmentation and parametric bionic reconstruction integrating multimodal data includes the following five modules: a CBCT intelligent segmentation module, an intraoral scanning intelligent segmentation module, an internal tissue segmentation module, a point cloud generation and registration module, and a parametric bionic reconstruction module; wherein the point cloud generation and registration module further includes a point cloud generation submodule and a point cloud registration submodule; and the parametric bionic reconstruction module further includes a dentin reconstruction submodule, a cancellous bone reconstruction submodule, and a gingival reconstruction submodule.

[0008] The functions of each module are described as follows: The CBCT intelligent segmentation module, the intraoral scanning intelligent segmentation module, and the internal tissue segmentation module are used to extract semantic segmentation masks for teeth and their internal structures from multimodal images. The CBCT intelligent segmentation module uses an improved 3D convolutional neural network (V-CTSegNet) to segment the external contours of teeth in CBCT volumetric images. The intraoral scanning intelligent segmentation module uses a lightweight neural network (P-IOSSegNet) to finely segment the crown surface model based on point cloud data. The internal tissue segmentation module uses a multi-threshold segmentation strategy combined with Hounsfield unit value distribution to perform 3D fine segmentation of tissues such as enamel, dentin, and cementum.

[0009] Point cloud generation and registration module: This module converts volume data into a mesh using an isosurface extraction algorithm, and then converts it into a point cloud format. It unifies all segmentation results into a three-dimensional point cloud space and constructs a geometric representation of cross-modal data. The registration process starts with coarse registration and then fine registration, minimizing the Euclidean distance error function between different data sources. in and Represent the corresponding point pairs in the source point cloud and the target point cloud, is the rotation matrix, is the translation vector.

[0010] Parametric bionic reconstruction module: import the registered point cloud data into the design software, use rational B-spline to fit the surface, and realize continuous closed structure reconstruction. The module includes three types of bionic structure modeling: Dentin reconstruction submodule: This module simulates the radially arranged microstructural features of dentinal tubules and constructs a microtubule bundle network. This structure has a certain directionality and gradient density distribution, which meets the biomimetic requirements of fluid conduction and sensing functions.

[0011] Cancellous bone reconstruction submodule: Based on its heterogeneity and anisotropy, the stress dispersion behavior of porous bone tissue is simulated by sampling points with large bone density differences in CBCT cancellous bone or Voronoi diagram.

[0012] Gingival reconstruction submodule: Based on its tissue characteristics consisting of a dense collagen fiber network, a fibrous mesh model with directionally arranged structures is constructed to express its role in soft tissue support and biomechanical buffering.

[0013] Specifically, the CBCT intelligent segmentation module uses the V-CTSegNet network to segment the tooth surface model of the CBCT image; the overall structure of the V-CTSegNet includes: an encoder, a decoder, and a fine-grained feature connection module.

[0014] The encoder consists of multiple layers of 3D convolution blocks and downsampling operations. Each layer contains multiple convolution blocks stacked together. A residual structure is used to improve the propagation capability of deep features, and batch normalization is used to make the output of the middle layer of the network more stable. The encoder extracts multi-scale spatial semantic features while reducing the resolution layer by layer, providing a structured high-dimensional expression for tooth segmentation. The decoder gradually restores the spatial resolution of the feature map through multiple layers of 3D transposed convolutions, and after upsampling each layer, it fuses the high-resolution features with the corresponding layer of the encoder to enhance the protection of tooth edge information during the decoding process. The residual structure is introduced inside the convolution module to effectively alleviate the problem of vanishing gradients in deep networks and enhance the stability of model training.

[0015] The intraoral scanning intelligent segmentation module uses the P-IOSSegNet network to segment the crown part of the intraoral scanning data; the P-IOSSegNet structure includes: a multi-scale sampling module, a feature extraction module, and a point feature decoding module.

[0016] The intraoral scan point cloud is input into the network in the form of an unordered N * 3 feature matrix, where N is the number of points and 3 is the 3D coordinate of each point. First, the multi-scale sampling module downsamples the input original point cloud data layer by layer through hierarchical set abstraction operations, and constructs the adjacency relationship of local areas, thereby effectively capturing the local geometric features and overall morphological information of the crown surface. Subsequently, the feature propagation module performs feature aggregation and context encoding on the point cloud subset in each local area, introduces a spherical neighborhood search mechanism and a multi-scale receptive field design, to enhance the model's robustness and generalization ability to complex tooth surface structure changes. The point feature decoding module gradually upsamples the high-dimensional features in the low-resolution layer through a layer-by-layer feature propagation mechanism and maps them back to the original point cloud coordinate system, realizing point-level classification prediction of the crown area.

[0017] The internal tissue segmentation module uses the grayscale differences of CBCT images to perform multi-threshold segmentation. After the initial segmentation, a smoothing algorithm is used to optimize the surface geometric continuity. The smoothing operation is based on the following algorithm: in is the vertex position, is its neighborhood point set, is the smoothing coefficient.

[0018] The segmentation results are meshed using an isosurface generation algorithm, and then uniformly processed to create a standard point cloud model. After registration, the multimodal point cloud models are fused to obtain a spatially consistent representation of the tooth structure.

[0019] The point cloud data is fitted through NURBS surface to form a closed surface, which is then input into the parametric modeling module for internal structure modeling. Gum and dentin are used as surface inputs respectively, and random points are generated in the gum battery to simulate the density of dentinal tubules. The random points of the two are further connected based on the nearest distance and lines are established. The bionic morphology of dentinal tubules is simulated by adjusting the tubular radius.

[0020] Using cancellous bone as the input surface, a Voronoi cell network was constructed based on sampling points with significant HU value differences in CBCT. This simulated the irregular distribution of trabeculae and voids within cancellous bone, reflecting its natural honeycomb structure. Finally, a biomimetic cancellous bone model with porous structural characteristics was generated from the original bone tissue surface using Boolean subtraction.

[0021] For gingival reconstruction, based on its connective tissue properties of dense collagen fibers, the gingival surface is first input and random points are generated to simulate the distribution of collagen fiber bundles. A fiber network structure is then established by connecting these points, constructing a flexible path. This path is then solidified by assigning a radius. Finally, Boolean operations are used to integrate the fiber structure into the gingival volume model, achieving biomimetic gingival modeling.

[0022] Compared with the prior art, the technical solution of the present invention has the following advantages: CBCT is used to obtain high-precision surface information of the internal tissue structure of teeth and intraoral scanning, realizing multimodal data complementarity and significantly improving reconstruction accuracy; deep learning networks for different modalities are introduced to intelligently segment multi-source data with higher segmentation accuracy; through the parametric bionic modeling process, the microstructural characteristics of dentin, cancellous bone, gums, etc. can be finely simulated, realizing continuous modeling from macromorphology to microstructure, and improving the application value of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic flow chart of an embodiment of a method for intelligent tooth segmentation and parametric bionic reconstruction integrating multimodal data provided by the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. It should be understood that the embodiments described here are only exemplary embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] Step S101 uses CBCT image data as input. CBCT is three-dimensional image data acquired by dental equipment and is widely used in dental diagnosis and treatment planning. CBCT images have high resolution and can clearly display the external morphology of teeth as well as the density of their internal tissues, such as enamel, dentin, and cementum. CBCT image data is typically input in the form of DICOM format or similar three-dimensional volume data, containing a large number of voxels. Each voxel contains its spatial coordinates in three-dimensional space and a Hounsfield unit (HU) value, which is proportional to the density of the tissue.

[0026] When using CBCT equipment to obtain patient imaging data, problems such as blurred tooth area boundaries, noise, and metal artifacts may occur due to the influence of scanning slice thickness, metal artifacts, and the absorption and transmittance of different tissue areas. Before segmentation, the image data needs to be subjected to related operations such as boundary enhancement, noise removal, and internal hole filling.

[0027] In an embodiment of the present invention, CBCT is input in DICOM format and smoothed by Gaussian filtering to reduce the influence of noise, and the tooth contour is enhanced by edge enhancement technology to make the subsequent intelligent segmentation more accurate.

[0028] In step S102, the CBCT intelligent segmentation module intelligently segments the input CBCT image data, extracting the 3D structure of the tooth region and its internal tissues. To achieve this, this module uses the deep learning-based V-CTSegNet model for automated semantic segmentation of CBCT images. V-CTSegNet is an improved 3D convolutional neural network based on the U-Net architecture, specifically designed for segmenting teeth and their internal tissues in CBCT volumetric images.

[0029] The V-CTSegNet encoder consists of multiple layers of 3D convolutional blocks and downsampling operations. Each encoder layer consists of multiple stacked convolutional blocks and uses a residual structure to improve the propagation of deep features. The encoder extracts multi-scale spatial semantic features of the input image by gradually reducing the resolution. The encoder's role is to extract high-dimensional feature representations from CBCT images and compress the information by gradually reducing the resolution.

[0030] The decoder gradually restores the spatial resolution of the feature map through multiple layers of 3D transposed convolutions, and performs upsampling on each layer, gradually restoring the low-resolution features to high-resolution feature maps. Each upsampling operation is then fused with the high-resolution features of the corresponding encoder layer, thereby enhancing the protection of tooth edges and details during the decoding process.

[0031] In this embodiment of the present invention, the training dataset consists of 165 CBCT images from patients of different ages and genders. Each data sample is stored in DICOM format with an image resolution of 512×512×256. Each image in the dataset is manually annotated. During training, the V-CTSegNet model uses batch normalization to improve training stability and reduce internal covariate shift. The cross-entropy loss function and the Dice coefficient are used as evaluation criteria. The Dice coefficient measures the degree of overlap between the model segmentation results and the true label. Its formula is as follows: in is the point set of the crown in the prediction result, is the point set of the tooth crown in the ground truth.

[0032] After 100 rounds of training, the model achieved a Dice coefficient of 0.92 and an average loss of 0.12. The Adam optimizer was used for optimization with a learning rate of 0.001. Resampling and random rotation were used during network training to improve the generalization ability of the model.

[0033] Step S103 aims to accurately segment the different tooth tissues in three dimensions using CBCT image data, including internal structures such as enamel, dentin, and cementum. This module utilizes the Hounsfield unit density distribution characteristic in CBCT images, which is proportional to tissue density, combined with a multi-threshold segmentation strategy to finely delineate the different tissue layers of the tooth. After segmentation, a Laplace smoothing algorithm is used to optimize the geometric continuity of the segmentation results and generate a mask of the internal tooth tissue.

[0034] In the embodiment of the present invention, the dentin, pulp and cancellous bone regions are segmented by multi-thresholding, and the continuity of the geometric surface is enhanced by the Laplace smoothing algorithm.

[0035] The CBCT mask image of step S104 is obtained by outputting the CBCT mask image of step S102 and step S103. The point cloud generation and registration module of step S105 converts the CBCT mask image processed by the CBCT intelligent segmentation module into three-dimensional point cloud data and fuses and aligns the tooth tissue information with the tooth surface point cloud.

[0036] Step S106: Convert the CBCT image into point cloud data. In this embodiment, the different tissues contained in the CBCT mask image are analyzed using the Marching Cubes algorithm. This algorithm calculates the isosurface of each voxel based on the grayscale value of the voxel and generates a mesh by connecting adjacent points on the isosurface. The mesh is then converted into point cloud data to represent the geometric shape of the tooth.

[0037] In step S107, the different point cloud data are precisely aligned to obtain a three-dimensional tooth model with spatially consistent internal tissue information. In this embodiment, the point cloud registration uses manual coarse registration and the Iterative Closest Point algorithm, which is the most commonly used point cloud registration method. It aims to precisely align two point clouds by minimizing the Euclidean distance error between the point clouds.

[0038] In the initial stage of point cloud registration, a coarse registration method is used to roughly align the CBCT point cloud and the intraoral scan point cloud. In this embodiment, the coarse registration tool in Geomagic Studio software is used to achieve this process. After the coarse registration, the ICP algorithm will be applied to perform fine registration. The ICP algorithm minimizes the spatial error between the two point clouds, thereby obtaining a high-precision registration result.

[0039] After the fine registration is completed, the obtained CBCT point cloud and the intraoral scan point cloud will be fused together to generate the CBCT tooth and interior point cloud of step S108.

[0040] Step S109 uses the intraoral scan data as input. This data is three-dimensional point cloud data acquired by an intraoral scanner, providing high-precision geometric information about the tooth surface and surrounding soft tissue. In this embodiment, the input intraoral scan data is preprocessed using Poisson resampling to avoid clustering of sampling points and ensure that the distribution of points is as uniform as possible.

[0041] Step S110 is responsible for intelligently segmenting the input intraoral scan point cloud data and extracting the crown region. This step uses the deep learning-based P-IOSSegNet model to automatically perform semantic segmentation on the intraoral scan data. P-IOSSegNet is based on the PointNet architecture and is specifically designed to handle crown region segmentation in intraoral scan point cloud data.

[0042] The core of the P-IOSSegNet model includes the following three modules: a multi-scale sampling module, a feature extraction module, and a point feature decoding module; the multi-scale sampling module uses a hierarchical set abstraction operation to downsample the input point cloud data and construct local area relationships through neighborhood information to ensure that the crown edge is highly preserved; in the feature extraction module, features are aggregated for each local point cloud subset through a hierarchical network structure. The point set of each local area will be encoded according to its spatial relationship to obtain a high-dimensional feature representation of the area; the point feature decoding module performs binary classification on the high-dimensional features, determines whether each point belongs to the crown area, and outputs the category label corresponding to each point.

[0043] In this example, the training dataset is the open-source Teeth3DS dataset, which consists of 1,800 intraoral scans of the upper and lower jaws of 900 patients. The dataset has an accuracy of 10-90 microns and a point resolution of 30-80 points per square micron. Teeth segmentation and annotation were performed by clinical assessors with more than 10 years of experience in orthodontics, dental surgery, and endodontic treatment. The Spherical Cross-Entropy loss function is used, which takes into account the spatial distribution of points, making the classification loss more stable. It is defined as follows: in, is the number of points, For the The labels of the points, For the The predicted probability of a point.

[0044] The AdamW optimizer was used to optimize the parameters of the P-IOSSegNet model. An improved version of the Adam optimizer, the AdamW optimizer offers enhanced regularization and is suitable for high-dimensional data types such as point clouds. In this example, the AdamW optimizer's learning rate was set to 0.001, and optimization was performed using β thresholds (0.9, 0.999). After 80 rounds of training, a Dice coefficient of 0.88 was achieved.

[0045] The intraoral scanned crown point cloud of step S110 is predicted and outputted in step S111. The multi-source point cloud data of step S108 and step S111 are passed through step S112. Step S112 is the same as step S105, including two sub-steps of step S113 and step S114. The complete point cloud of the tooth structure of step S115 is outputted. In this embodiment, it includes the precise point cloud of the tooth surface, dentin, cancellous bone, pulp, and gum point cloud.

[0046] The complete point cloud of the tooth structure is sent as input to the parametric bionic modeling and reconstruction module of step S116, which includes three submodules from step S117 to step S119, which are used for bionic reconstruction of dentin, cancellous bone, and gum respectively.

[0047] Step S117 is for dentin reconstruction. In this embodiment, after the dental pulp and cancellous bone are input into Rhino, they are converted into NURBS surfaces to facilitate parametric modeling in Grasshopper. The gingiva and dentin are input as polysurfaces via two Brep cells. Randomly generated points are used to simulate the density of the dental tubules. The corresponding points are connected via Line and sent to the Pipe cell to generate the biomimetic dentinal tubule morphology.

[0048] Step S118 is used for cancellous bone reconstruction. In this embodiment, sampling points with lower bone density in the cancellous bone are extracted based on the HU value difference in the CBCT and used as the input of the Point battery. In order to simulate the honeycomb structure of cancellous bone, the distribution structure of trabeculae and void areas is constructed using the Voronoi diagram method. The Voronoi 3D component is used to generate a Voronoi cell network. Each Voronoi unit represents a trabecular area or a void area. By adjusting the distribution parameters of the Voronoi diagram, the distribution density and morphology of the trabeculae can be controlled, so that the generated model has the natural honeycomb pore structure of cancellous bone.

[0049] Step S119 is used for gingival reconstruction. In this embodiment, the gingival data scanned from the oral cavity is fed into Grasshopper via the Brep cell, and random points are generated using the Populate 3D cell to simulate the distribution of collagen fiber bundles. Flexible paths are then created using the Curve and Flex cells to simulate the natural curvature and arrangement of collagen fiber bundles in the gingiva. The point sets are then connected to establish a fiber network structure, which is then materialized using the Pipe cell and fused with the original gingiva through Boolean operations to generate a biomimetic fibrous structure gingival model.

[0050] Through the above steps, step S120 outputs a bionic model of the internal tissue of the tooth that is similar to the real tooth situation in this example. The model includes the bionic structure of the tooth surface, dentin, cancellous bone, pulp and gums for subsequent tasks such as clinical and teaching use.

[0051] Although the present invention has been described in detail through specific embodiments, the present invention is not limited to these embodiments. Any person skilled in the art may make various modifications and variations to the present invention without departing from the technical concept and scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope defined in the claims.

Claims

1. A method for intelligent tooth segmentation and parametric bionic reconstruction based on multimodal data, characterized in that: include: CBCT intelligent segmentation module, intraoral scanning intelligent segmentation module, internal tissue segmentation module, point cloud generation and registration module, parametric bionic reconstruction module; wherein the point cloud generation and registration module includes a point cloud generation submodule and a point cloud registration submodule; the parametric bionic reconstruction module includes a dentin reconstruction submodule, a cancellous bone reconstruction submodule, and a gingival reconstruction submodule; The CBCT intelligent segmentation module inputs the CBCT image into the V-CTSegNet deep learning network model to perform automatic semantic segmentation on the crown and root areas, and outputs a CBCT predicted mask image containing tooth boundary information; The intraoral scanning intelligent segmentation module inputs the intraoral scanning point cloud into the P-IOSSegNet deep learning network model for unordered point sets to achieve segmentation and extraction of the crown surface model and obtain the intraoral scanning prediction point cloud; The internal tissue segmentation module uses a multi-threshold segmentation strategy to segment different tooth tissues based on the Hounsfield unit density distribution characteristics of the CBCT image, and further processes its morphology through a 3D image processing algorithm to obtain a predicted point cloud of the tooth internal tissue; The point cloud generation and registration module converts the CBCT prediction mask image obtained by the CBCT intelligent segmentation module into a three-dimensional surface mesh and samples it to generate a CBCT prediction point cloud, and spatially registers and fuses the CBCT prediction point cloud with the intraoral scan prediction point cloud to obtain a high-precision tooth prediction point cloud; The parametric bionic reconstruction module divides the structural regions and performs functional bionic design on the fused high-precision tooth prediction point cloud and internal tissue prediction point cloud using Grasshopper, a parametric modeling plug-in for Rhino. This module constructs, based on the biological functional characteristics of different tissues, a tubular structure to simulate the micro-fluid conduction function of the dentinal tubule area, an irregular honeycomb structure to simulate the support and elastic function of cancellous bone, and a fibrous cell network structure to simulate the flexible connection characteristics of gingival tissue.

2. The method for intelligent segmentation and parametric bionic reconstruction of teeth based on multimodal data fusion according to claim 1, characterized in that: The V-CTSegNet model in the CBCT intelligent segmentation module includes an encoder, a decoder, and a fine-grained feature connection module. The workflow is as follows: The CBCT image is input to the encoder, which performs feature extraction. The encoder includes multi-stage convolution, voxel-level downsampling, and residual connection to extract features from the CBCT and gradually reduce the resolution. The extracted feature vector passes through the decoder, which includes 3D transposed convolution, feature map fusion, and residual path mechanism to achieve layer-by-layer upsampling of the feature map and spatial resolution restoration. The fine-grained feature connection module connects the encoder and decoder of the corresponding layer, and directly passes the feature map in the encoder to the corresponding layer of the decoder.

3. The method for intelligent segmentation and parametric bionic reconstruction of teeth based on fusion of multimodal data according to claim 1, characterized in that: The P-IOSSegNet model in the intraoral scanning intelligent segmentation module includes: a multi-scale sampling module, a feature extraction module, and a point feature decoding module. The workflow is as follows: The intraoral scan point cloud constructs spatial structure information through the multi-scale sampling module; then the feature extraction module aggregates the features of each local subset; finally, the point feature decoding module obtains the point classification result and outputs the intraoral scan point cloud prediction result.

4. The method for intelligent segmentation and parametric bionic reconstruction of teeth based on fusion of multimodal data according to claim 1, characterized in that: The internal tissue segmentation module is characterized by using a multi-threshold segmentation method to distinguish enamel, dentin, and cementum based on the density differences of various types of internal tooth tissues in Hounsfield units in CBCT images, and extracting their 3D surface models. To improve the surface continuity and geometric smoothness of the model, the Laplace smoothing algorithm is further used to remove noise and artifacts, generating an internal tissue mask image with clear structural boundaries and reasonable topology.

5. The method for intelligent segmentation and parametric bionic reconstruction of teeth based on fusion of multimodal data according to claim 1, characterized in that: The point cloud generation and registration module includes a point cloud generation submodule and a point cloud registration submodule; The point cloud generation submodule performs three-dimensional surface reconstruction on the predicted mask image obtained by the CBCT intelligent segmentation module and the smooth internal tissue mask image output by the internal tissue segmentation module. The isosurface mesh information of each structure is extracted using the Marching Cubes isosurface extraction algorithm and converted into dense 3D point cloud coordinate data. The point cloud data is processed using the Open3D graphics library to generate a unified format of CBCT predicted point clouds, which contains complete information such as crowns, roots, and internal tissue structures. The point cloud registration submodule is used to achieve spatial fusion of three-dimensional data from different sources. The CBCT predicted point cloud and the intraoral scan predicted point cloud are coarsely registered through Geomagic Studio's coarse registration combined with the iterative closest point algorithm. The two are then finely registered with high-precision point clouds to correct errors in the spatial dimension and obtain a high-precision tooth prediction point cloud.

6. The method for intelligent segmentation and parametric bionic reconstruction of teeth based on fusion of multimodal data according to claim 1, characterized in that: The parametric bionic reconstruction module performs parametric bionic reconstruction design through Grasshopper, a parametric modeling plug-in of Rhino, and includes: a dentin reconstruction submodule, a cancellous bone reconstruction submodule, and a gingival reconstruction submodule; The dentin reconstruction submodule combines the anatomical microstructural characteristics of dentinal tubules, such as radial tubular arrangement, small diameter, and gradient density distribution, to reconstruct dentinal tubules with microtubule bundle structure in the dentin area, simulating their microscopic fluid conduction and perception functions. The cancellous bone reconstruction submodule reconstructs an irregular honeycomb network structure model based on its porous, heterogeneous, and anisotropic properties, simulating its good mechanical response and stress dispersion capabilities; The gingival reconstruction submodule, based on its biological characteristics of being mainly composed of fibrous connective tissue, reconstructs a fiber cell grid structure with directionality, simulating the physiological function of gingival tissue in soft tissue support and cushioning.

Images