Oral and maxillofacial three-dimensional model construction method based on oral soft and hard tissue identification

Abstract

The present application relates to the technical field of oral and maxillofacial three-dimensional modeling processing, in particular to an oral and maxillofacial three-dimensional model construction method based on oral soft and hard tissue recognition, comprising: acquiring original three-dimensional voxel data of the oral and maxillofacial region of a patient containing gray value information, constructing a multi-scale feature field and calculating voxel feature gradient intensity, generating an adaptive boundary tracking grid according to the spatial distribution of the feature gradient intensity. In the grid node, the soft and hard tissue boundary energy is calculated by combining the voxel gray value and the feature gradient intensity, the boundary energy field is constructed and the force line propagation is simulated, the high-energy ridge line is extracted and the corresponding grid element is marked as a set of tissue boundary voxels to be classified, and the oral and maxillofacial three-dimensional model is generated by combining a pre-trained tissue attribute classifier. The method can adaptively match the gradient change at the junction of soft and hard tissues, accurately quantify the attribution degree of the voxels at the junction, improve the positioning accuracy and continuity of boundary recognition, and make the three-dimensional model more consistent with the actual anatomical structure of the oral and maxillofacial region.

Description

Technical Field

[0001] This invention relates to the field of oral and maxillofacial three-dimensional modeling technology, and in particular to a method for constructing oral and maxillofacial three-dimensional models based on the identification of oral soft and hard tissues. Background Technology

[0002] The current method of constructing a three-dimensional model of the oral and maxillofacial region is based on three-dimensional voxel data of the oral and maxillofacial region. Conventional techniques use fixed-parameter meshes to complete the voxel space division, extract tissue boundaries through grayscale threshold segmentation or conventional edge detection methods, and distinguish soft and hard tissue attributes based on preset rules or general classification models to complete the construction of the three-dimensional model of the oral and maxillofacial region. This is the current conventional method of oral and maxillofacial tissue segmentation and three-dimensional modeling.

[0003] Fixed-parameter meshes cannot adapt to the spatial distribution variations of feature gradients in voxel data. Boundary tracking accuracy is constrained by preset mesh parameters, and tissue boundary positioning offsets are prone to occur. Conventional boundary extraction relies solely on single voxel grayscale or gradient information, failing to quantify the degree of voxel affiliation at the soft-hard tissue interface. It does not construct an energy field for tissue boundaries, making it impossible to locate continuous tissue boundaries through field force line propagation. Consequently, the continuity and accuracy of soft-hard tissue boundary recognition are insufficient, and the anatomical structure reconstruction effect of the oral and maxillofacial 3D model is limited.

[0004] This invention addresses the problems of existing meshes being unable to adaptively match the spatial distribution of feature gradients and being unable to accurately extract the boundary between soft and hard tissues through boundary energy fields and force line propagation, thereby achieving accurate identification of oral soft and hard tissues and effective construction of three-dimensional models of the oral and maxillofacial region. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology, and the proposed method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues solves this problem.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues, comprising:

[0007] Obtain raw three-dimensional voxel data of the patient's oral and maxillofacial region, wherein the raw three-dimensional voxel data includes grayscale information;

[0008] A multi-scale feature field is constructed on the original three-dimensional voxel data. The feature gradient intensity of each voxel is calculated based on the gray value information in the multi-scale feature field. An adaptive boundary tracking mesh is generated based on the spatial distribution of the feature gradient intensity.

[0009] At the nodes of the adaptive boundary tracking mesh, the soft and hard tissue boundary energy is calculated based on the voxel gray value information and the feature gradient intensity. The soft and hard tissue boundary energy is used to quantify the degree of belonging of the node at the boundary between soft and hard tissue.

[0010] Using the boundary energy of soft and hard tissues as the field strength, a boundary energy field is constructed on an adaptive boundary tracing grid, and the propagation of force lines in the boundary energy field is simulated to trace the high-energy ridges formed by the convergence of force lines.

[0011] Identify and extract the high-energy ridges, and mark the adaptive boundary tracking grid cells traversed by the high-energy ridges as a set of boundary voxels of the tissue to be classified;

[0012] Based on the set of voxels representing the tissue boundaries to be classified, and combined with a pre-trained tissue attribute classifier, a three-dimensional model of the oral and maxillofacial region is generated.

[0013] As a further aspect of the present invention, a multi-scale feature field is constructed on the original three-dimensional voxel data, and the feature gradient intensity of each voxel is calculated based on the grayscale value information in the multi-scale feature field, including:

[0014] The original three-dimensional voxel data is convolved and filtered using three-dimensional Gaussian kernels of different scales to generate multiple smooth voxel data layers corresponding to different spatial resolutions.

[0015] Calculate the gray-level gradient components of each voxel in the three spatial directions in each layer of smoothed voxel data;

[0016] The gray-level gradient components of each voxel at the corresponding position in all smooth voxel data layers are weighted and fused to obtain the multi-scale fused feature gradient vector of the voxel. The magnitude of the multi-scale fused feature gradient vector is defined as the feature gradient intensity of the voxel.

[0017] As a further aspect of the present invention, the step of generating an adaptive boundary tracking mesh based on the spatial distribution of feature gradient intensities includes:

[0018] In three-dimensional space, using voxels as the basic unit, voxels with feature gradient strength higher than the mesh generation threshold are marked as candidate mesh nodes;

[0019] Initial connections are established between candidate grid nodes based on the directional consistency of feature gradient strength, forming a sparse spatial graph structure.

[0020] The spatial graph structure is processed into triangular facets to generate a set of triangular facets that are oriented and connected with candidate mesh nodes as vertices and feature gradient directions as references. The set of triangular facets constitutes the adaptive boundary tracking mesh.

[0021] As a further aspect of the present invention, at the nodes of the adaptive boundary tracking mesh, the boundary energy of soft and hard tissues is calculated based on voxel gray value information and feature gradient intensity, including:

[0022] For each node on the adaptive boundary tracing mesh, obtain the gray value of its corresponding voxel in the original 3D voxel data, as well as the characteristic gradient intensity of the voxel calculated accordingly.

[0023] The grayscale values ​​are converted into tissue density feature values ​​using a predefined mapping function;

[0024] Multiplying the tissue density eigenvalue by the eigengradation intensity, and then multiplying by a correction coefficient related to the local curvature of the node, the resulting product is the soft and hard tissue boundary energy of the node.

[0025] As a further aspect of the present invention, the step of constructing a boundary energy field on an adaptive boundary tracing grid using the boundary energy of soft and hard tissues as the field strength, simulating the propagation of force lines in the boundary energy field, and tracing the high-energy ridges formed by the convergence of force lines includes:

[0026] Each node of the adaptive boundary tracing mesh is considered as a field source, and the magnitude of its field strength is defined by the soft and hard organization boundary energy of the node.

[0027] Starting from the node with the highest field strength, the simulated virtual force lines propagate along the triangular facet connection direction of the adaptive boundary tracing mesh towards the direction where the energy of the soft and hard tissue boundary decreases;

[0028] Record the propagation path of all virtual force lines, and extract the nodes and triangle edges through which multiple virtual force lines converge;

[0029] The nodes and triangular facets that converge are connected and smoothed in space to form a continuous spatial curve, which is the high-energy ridge.

[0030] As a further aspect of the present invention, based on the set of boundary voxels of the tissue to be classified, and combined with a pre-trained tissue attribute classifier, a three-dimensional model of the oral and maxillofacial region is generated, including:

[0031] The pre-trained tissue attribute classifier is used to determine the attributes of each voxel in the set of voxels of the tissue boundary to be classified. The tissue attribute classifier outputs the probability distribution of each voxel belonging to bone tissue, tooth tissue, gingival soft tissue or mucosal soft tissue.

[0032] Based on the probability distribution, organizational labels are assigned to all nodes in the adaptive boundary tracking mesh, and each node is given a deterministic organizational category label;

[0033] Based on the organizational category label of the nodes, spatial domain clustering is performed on the set of nodes belonging to the same organizational category to form multiple spatially contiguous initial organizational blocks;

[0034] For each organization's initial block, internal topology consistency optimization is performed to adjust the spatial connection relationship of the nodes within the organization's initial block, ensuring that each organization's initial block constitutes a closed and continuous geometric surface.

[0035] The geometric surfaces of all optimized initial tissue blocks are spatially registered and stitched together, and finally fused to construct a three-dimensional oral and maxillofacial model containing clear soft and hard tissue boundary information.

[0036] The attributes of each voxel in the set of boundary voxels of the tissue to be classified are determined using a pre-trained tissue attribute classifier, specifically including:

[0037] The tissue attribute classifier uses the multi-scale fusion feature gradient vector of voxels, the gray value of the voxels in the original three-dimensional voxel data, and the tissue category label distribution of the voxels in the adaptive boundary tracking grid as input features.

[0038] The organizational attribute classifier performs nonlinear transformation and mapping on the input features through an internal deep neural network;

[0039] The tissue attribute classifier outputs a multidimensional vector, where each dimension of the multidimensional vector represents the probability value of the input voxel belonging to one of four predefined categories: bone tissue, tooth tissue, gingival soft tissue, and mucosal soft tissue.

[0040] As a further aspect of the present invention, according to the probability distribution, label assignment is performed on all nodes in the adaptive boundary tracking mesh, including:

[0041] For each node in the adaptive boundary tracking mesh, obtain the probability distribution corresponding to the voxels associated with the node, based on the output of the organization attribute classifier;

[0042] Select the organization category with the highest probability value from the probability distribution;

[0043] The organization category with the highest probability value is assigned as the deterministic organization category label of the node.

[0044] As a further aspect of the present invention, the step of spatially clustering nodes belonging to the same organizational category based on the organizational category label of the nodes to form multiple spatially contiguous initial organizational blocks includes:

[0045] In three-dimensional space, all nodes with the same organization category label are used as initial seed points;

[0046] Based on the initial seed point, within the topology of the adaptive boundary tracing mesh, all nodes that are directly or indirectly connected to the initial seed point through triangular facet edges and have the same organizational category label are grouped into the same set.

[0047] Each such set of nodes constitutes an internally connected point cloud cluster in three-dimensional space, and the point cloud cluster is defined as an initial block of organization.

[0048] As a further aspect of the present invention, the internal topology consistency optimization for each organization's initial block includes:

[0049] For each point cloud cluster corresponding to the initial block of the organization, calculate its three-dimensional convex hull or use the Poisson surface reconstruction algorithm to generate an initial closed triangular mesh surface;

[0050] Detect non-manifold edges, isolated triangular facets, or internal hole defects present in the initial closed triangular mesh surface;

[0051] By adding or deleting triangular facets, iteratively adjusting vertex positions, and performing local retriangulation, the non-manifold edges, isolated triangular facets, or internal holes are eliminated, resulting in an optimized geometric surface with a topological structure of a simply connected closed manifold.

[0052] As a further aspect of the present invention, the geometric surfaces of all optimized initial tissue blocks are spatially registered and stitched together, and finally fused to construct a three-dimensional oral and maxillofacial model containing clear soft and hard tissue boundary information, including:

[0053] The optimized geometric surfaces of all initial tissue blocks are placed in the same three-dimensional coordinate system;

[0054] For adjacent initial tissue blocks with different tissue category labels, the vertices of their shared boundary regions are iteratively aligned to achieve seamless registration of adjacent geometric surfaces.

[0055] At the registered shared boundary, overlapping triangular patches are deleted, and new triangular patches are generated on the boundary contour to connect the geometric surfaces of the initial blocks of different tissues.

[0056] After connecting all adjacent boundaries, a unified triangular mesh model containing multiple tissue category regions is obtained, which is the final three-dimensional model of the oral and maxillofacial region.

[0057] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0058] Within a multi-scale feature field, the feature gradient intensity is calculated based on the voxel gray value. An adaptive boundary tracking mesh is generated according to the spatial distribution of the feature gradient intensity. The mesh can adaptively adjust its distribution shape and unit density to follow the spatial changes of the feature gradient in the oral and maxillofacial region, closely conforming to the gradient change characteristics of the soft and hard tissue interface. It abandons the parameter constraints of fixed resolution meshes, allowing the boundary tracking mesh to match the spatial position of the actual tissue boundary, reducing the misalignment between mesh units and tissue boundaries.

[0059] The boundary energy of soft and hard tissues is calculated by fusing voxel grayscale information and feature gradient intensity at the nodes of the adaptive boundary tracking mesh. The boundary energy is used as the field strength to build a boundary energy field. The propagation of force lines in the boundary energy field is simulated and the high-energy ridges formed by the convergence of force lines are extracted. The corresponding mesh cells are marked as the set of voxels of the boundary of the tissue to be classified. Then, a three-dimensional model is generated by combining the pre-trained tissue attribute classifier. The boundary energy of soft and hard tissues can quantify the degree of belonging of mesh nodes at the boundary between soft and hard tissues. The boundary energy field can clearly present the energy distribution state of the boundary area. The propagation of force lines can automatically form continuous tissue boundary ridges, weaken the interference of noise on boundary recognition, improve the continuity and accuracy of tissue boundary extraction, and make the soft and hard tissue boundaries in the constructed oral and maxillofacial three-dimensional model fit the actual anatomical structure. Attached Figure Description

[0060] Figure 1 This is a flowchart of the method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues as described in this invention;

[0061] Figure 2 A flowchart for constructing a multi-scale feature field and calculating the feature gradient intensity;

[0062] Figure 3 A flowchart for constructing the boundary energy field and tracing high-energy ridgelines. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0065] See Figure 1 This invention provides a method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues. The implementation process is as follows:

[0066] The system acquires raw 3D voxel data containing grayscale information of the patient's oral and maxillofacial region. Based on this, a multi-scale feature field is constructed. In this field, the feature gradient intensity is calculated based on the voxel grayscale value, and an adaptive boundary tracking mesh is generated based on the spatial distribution of the feature gradient intensity. At this mesh node, the soft and hard tissue boundary energy is calculated by combining the voxel grayscale value and the feature gradient intensity to quantify the degree of belonging of the node at the tissue boundary. Then, a boundary energy field is constructed on the mesh using this energy as the field strength. The propagation of force lines in the field is simulated, and the high-energy ridges formed by their convergence are tracked. The mesh cells through which the high-energy ridges pass are extracted and marked as the set of voxel boundaries of the tissue to be classified. Finally, based on this set, a pre-trained tissue attribute classifier is used to complete the discrimination of tissue attributes and the final generation of the model.

[0067] In one embodiment of the present invention, a multi-scale feature field is constructed on the original three-dimensional voxel data, and the feature gradient intensity of each voxel is calculated in the field. (See reference...) Figure 2This process is achieved by convolutionally filtering the original data using multiple 3D Gaussian kernels of different scales, generating a series of smoothed voxel data layers corresponding to different spatial resolutions. For each smoothed data layer, the gray-level gradient components of each voxel in three spatial directions are calculated. The gradient components of each voxel at corresponding positions in all smoothed layers are weighted and fused to obtain the multi-scale fused feature gradient vector of that voxel. The magnitude of this vector is defined as the feature gradient intensity of that voxel. An adaptive boundary tracking mesh is generated based on the spatial distribution of the feature gradient intensity. The operation involves marking voxels with feature gradient intensities exceeding a preset mesh generation threshold as candidate mesh nodes in 3D space. Initial connections are established between these candidate nodes based on the directional consistency of their feature gradient intensities, forming a sparse spatial graph structure. This spatial graph structure is then triangulated to generate a set of triangular patches with candidate nodes as vertices and oriented connections based on the feature gradient directions. This set of triangular patches constitutes the adaptive boundary tracking mesh.

[0068] In the specific implementation, the construction of a multi-scale feature field and the generation of an adaptive boundary tracking mesh are performed on the original 3D voxel data. The original 3D voxel data comes from medical image scans of the patient's oral and maxillofacial region. Each voxel unit in the 3D voxel data contains its spatial coordinates and grayscale values ​​representing tissue density. In the specific implementation, the first step in constructing the multi-scale feature field is to perform convolution filtering on the original 3D voxel data using a set of 3D Gaussian kernels with different scale parameters. Each 3D Gaussian kernel corresponds to a specific spatial smoothing scale. After convolution, a series of smoothed voxel data layers with the same spatial dimensions as the original 3D voxel data but different degrees of detail smoothing are generated. It can be understood that 3D Gaussian kernels with larger scale parameters have a stronger smoothing effect on the original 3D voxel data, which can suppress noise and reflect the grayscale trend of the tissue over a large range, while 3D Gaussian kernels with smaller scale parameters have a weaker smoothing effect on the original 3D voxel data, which can retain more local details and edge information.

[0069] In some embodiments, after obtaining multiple smooth voxel data layers, it is necessary to calculate the gray-level gradient of each voxel in each smooth voxel data layer. For any smooth voxel data layer, for each voxel within it, the gray-level change of that voxel in the three positive directions of the three-dimensional Cartesian coordinate system X, Y, and Z is calculated. These three changes together constitute the gray-level gradient component vector of that voxel in the current smooth voxel data layer. In specific implementations, for a given voxel, multiple gray-level gradient component vectors will be calculated in different smooth voxel data layers. Fusing this information from different scales is the key to constructing robust features. The calculation of the multi-scale fused feature gradient vector involves weighting and summing the gray-level gradient component vectors of the voxel at the corresponding positions in all smooth voxel data layers according to the scale of its layer. Optionally, the set of weights used for weighting can be inversely proportional to the scale parameter of the Gaussian kernel, giving higher weights to gradient components generated by smaller-scale Gaussian kernels that contain more high-frequency edge information. The magnitude of the multi-scale fused feature gradient vector, i.e. its Euclidean norm, is defined as the final feature gradient intensity of the voxel. The feature gradient intensity is a scalar, and its value reflects the degree of gray-scale change in the image at the location of the voxel. Near the boundary between soft and hard tissues, the feature gradient intensity usually exhibits a higher value.

[0070] In some embodiments, an adaptive boundary tracking mesh can be generated based on the spatial distribution of the calculated feature gradient intensities. The generation process first requires setting a mesh generation threshold in 3D space, marking all voxels in the original 3D voxel data with feature gradient intensities higher than the threshold as candidate mesh nodes. Initial connections are established between candidate mesh nodes based on the consistency of feature gradient directions. Specifically, the cosine of the angle between the multi-scale fused feature gradient vectors corresponding to any two spatially adjacent candidate mesh nodes is calculated. When the cosine value is greater than a preset direction consistency threshold, a connection edge is established between these two candidate mesh nodes. In this way, a sparse spatial graph structure is formed in 3D space, with candidate mesh nodes as vertices and connection edges as topological relationships. In a specific implementation, this sparse spatial graph structure is subjected to triangulation. Triangulation takes the spatial coordinates of all candidate mesh nodes as input and generates a set of triangular patches, where the three vertices of each triangular patch are candidate mesh nodes. It can be understood that when generating triangular patches, the algorithm references the direction information of the feature gradient vector, preferentially forming triangular patch connections between nodes with similar feature gradient directions. This allows the structure of the generated triangular patch set to adaptively conform to the trend of grayscale changes in the image. This triangular patch set constitutes the adaptive boundary tracking mesh used for subsequent boundary energy field calculation and propagation. In specific implementations, multi-scale fusion of feature gradient vectors... The weighted fusion calculation can be expressed by the following formula:

[0071]

[0072] Where: symbol The gradient vector representing the multi-scale fused features of a voxel. (Symbol) This indicates all preset scales. The summation operation. (Symbol) The representative corresponds to the scale. The fusion weight coefficient. (Symbol) , , Representing the scale In the smoothed voxel data layer below, the voxel is in , , The gray-level gradient components are located in three spatial directions. This formula describes how to integrate gradient information at different scales into a unified feature description vector.

[0073] In one embodiment of the present invention, see [reference] Figure 3 At the nodes of the adaptive boundary tracking mesh, the soft and hard tissue boundary energy is calculated based on the voxel gray value information and the feature gradient intensity. For each node on the mesh, the gray value of its corresponding voxel in the original data and the calculated feature gradient intensity are obtained. The gray value is converted into a tissue density feature value through a predefined mapping function. The tissue density feature value is multiplied by the feature gradient intensity and then multiplied by a correction coefficient related to the local curvature of the node. The resulting product is defined as the soft and hard tissue boundary energy of the node. Using the boundary energy of soft and hard tissues as the field strength, a boundary energy field is constructed on an adaptive boundary tracing mesh, and force line propagation is simulated to track high-energy ridges. In this process, each node of the mesh is regarded as a field source, and its field strength is defined by the soft and hard tissue boundary energy of that node. The simulation starts from the node with the highest field strength, and the virtual force lines propagate along the connection direction of the mesh triangular facets in the direction of decreasing soft and hard tissue boundary energy. The propagation path of all virtual force lines is recorded, and the nodes and triangular facet edges that are jointly converged by multiple virtual force lines are extracted. These nodes and edges that are jointly converged are connected and smoothed in space to form a continuous spatial curve, which is the high-energy ridge being tracked.

[0074] In practice, for each node on the generated adaptive boundary tracing mesh, its corresponding soft and hard tissue boundary energy needs to be calculated. The calculation process requires obtaining the grayscale value of the original voxel associated with the node in the original 3D voxel data and reading the pre-calculated feature gradient intensity of that voxel. The grayscale value is converted into a tissue density feature value through a predefined mapping function. This mapping function can be a linear piecewise function used to map the grayscale range of the medical image to a numerical range that better reflects the physical tissue density. Optionally, the mapping function can also be a nonlinear lookup table based on prior knowledge. In practice, the converted tissue density feature value is directly multiplied by the feature gradient intensity to obtain a preliminary boundary energy estimate. This preliminary boundary energy estimate is then multiplied by a correction coefficient related to the local curvature of the node. The local curvature can be estimated by analyzing the normal changes of the triangular facets surrounding the node in the adaptive boundary tracing mesh. The correction coefficient enhances the energy weight of nodes located in high-curvature boundary regions. The final product is defined as the soft and hard tissue boundary energy value of that node.

[0075] In some embodiments, using the calculated soft and hard tissue boundary energy as the field strength, a boundary energy field can be constructed on the adaptive boundary tracing mesh. In this boundary energy field, each node of the adaptive boundary tracing mesh is considered a point field source, and the magnitude of the field strength it excites is directly defined by the soft and hard tissue boundary energy value of that node. Nodes with higher values ​​represent stronger field sources. In a specific implementation, the propagation of force lines in the simulated boundary energy field begins with identifying the node with the highest field strength. This node is taken as the starting point for the propagation of virtual force lines, which propagate along the triangular facet connection direction of the adaptive boundary tracing mesh. It can be understood that the direction of propagation is determined by the soft and hard tissue boundary energy distribution of the current node. Virtual force lines always propagate from the current node towards its directly connected neighboring nodes with lower soft and hard tissue boundary energy values. This process iterates until the virtual force lines propagate to a local energy minimum point or boundary. (Soft and hard tissue boundary energy) The calculation formula is expressed as follows:

[0076]

[0077] Where: symbol This represents the final calculated boundary energy between the soft and hard tissues. (Symbol) This represents the tissue density feature value obtained by transforming the original grayscale values ​​of voxels using a mapping function. (Symbol) Multi-scale fused feature gradient vector representing voxels The magnitude, i.e., the characteristic gradient strength. (Symbol) This represents a correction coefficient related to the local geometric curvature of the node. This formula quantifies the significance of the node at the tissue boundary, where tissue density difference, gray-level gradient intensity, and local geometry jointly determine the energy value.

[0078] In some embodiments, it is necessary to record the complete propagation paths of all virtual force lines originating from different high-field-strength nodes. Each virtual force line's path is defined by the sequence of adaptive boundary tracing grid nodes and triangular facet edges it passes through sequentially. In specific implementations, tracing high-energy ridges involves analyzing the spatial distribution of these propagation paths and extracting nodes and triangular facet edges that are traversed by more than one virtual force line. Optionally, a convergence threshold can be set; when a node or edge is traversed by more than this threshold number of virtual force line paths, it is marked as a candidate component of a high-energy ridge. These discrete nodes and triangular facet edges that are traversed by the convergence are connected in three-dimensional space based on their inherent topological adjacency in the adaptive boundary tracing grid. The resulting spatial polygons are then smoothed using a method such as B-spline curve fitting, ultimately forming one or more continuous, smooth spatial curves. These spatial curves represent the high-energy ridges that characterize potential boundaries between soft and hard tissues.

[0079] In one embodiment of the present invention, a three-dimensional model of the oral and maxillofacial region is generated based on a set of voxels representing the boundary of the tissue to be classified, combined with a pre-trained tissue attribute classifier. The pre-trained tissue attribute classifier performs attribute discrimination on each voxel in the set. This classifier uses the voxel's multi-scale fusion feature gradient vector, its grayscale value in the original data, and the distribution of tissue category labels of its neighboring nodes in the adaptive boundary tracking grid as input features. Through its internal deep neural network, it performs nonlinear transformation and mapping on these input features, ultimately outputting a multi-dimensional vector. Each dimension of this vector represents the probability value that the input voxel belongs to one of four predefined categories: bone tissue, dental tissue, gingival soft tissue, and mucosal soft tissue. Based on this probability distribution, tissue labels are assigned to all nodes in the adaptive boundary tracking grid, giving each node a deterministic tissue category label. Based on the node's tissue category label, spatial domain clustering is performed on the set of nodes belonging to the same tissue category, forming multiple spatially continuous initial tissue blocks. Internal topological consistency optimization is performed on each initial tissue block, adjusting the spatial connectivity of nodes within the block. The geometric surfaces of all optimized initial tissue blocks are spatially registered and stitched together, and finally fused to construct a three-dimensional oral and maxillofacial model containing clear soft and hard tissue boundary information.

[0080] In practical implementation, the pre-trained tissue attribute classifier is a deep neural network model, such as a 3D convolutional neural network or a graph neural network. Before being used to construct a 3D model of the oral and maxillofacial region, this classifier has undergone supervised training using a large amount of labeled oral and maxillofacial medical image voxel data. When the tissue attribute classifier performs attribute discrimination on each voxel in the "set of voxels representing the boundary of the tissue to be classified," the input feature it relies on is a combined feature vector. This combined feature vector specifically includes: the "multi-scale fusion feature gradient vector" calculated for the voxel in the above embodiment, the original grayscale value of the voxel in the "original 3D voxel data," and the tissue category label distribution of the voxel among all its neighboring nodes in the "adaptive boundary tracking grid." It can be understood that the tissue category label distribution of neighboring nodes can be null or uniformly distributed in the initial iteration, while subsequent iterations use the label information generated in the previous classification round, thus introducing spatial context constraints.

[0081] In some embodiments, the deep neural network inside the organizational attribute classifier consists of multiple fully connected layers, nonlinear activation function layers, and a final normalized exponential function layer. After the combined feature vector is input into the deep neural network, it undergoes a series of nonlinear transformations and feature mappings, and the network finally outputs a four-dimensional probability vector. This probability vector Each dimension represents the probability that the input voxel belongs to one of four predefined tissue categories, i.e., the probability of bone tissue. Probability of tooth structure The probability of soft tissue in the gums and the probability of mucosal soft tissue The sum of the four probability values ​​is 1. See Table 1, which shows the possible output probability distribution of a voxel after being classified by the tissue attribute classifier.

[0082] Table 1: Output Probability Distribution of Organization Attribute Classifier

[0083]

[0084] In practice, based on the probability distribution output by the tissue attribute classifier, tissue labels are assigned to all nodes in the adaptive boundary tracking mesh. For each node in the mesh, the probability distribution vector corresponding to its associated voxel is queried, and the tissue category with the highest probability value is selected. This category is then assigned as the deterministic tissue category label for that node. For example, Voxel_001 in Table 1 is assigned the label "dental tissue," and Voxel_002 is assigned the label "bone tissue." Optionally, to avoid misclassification when probabilities are similar, a probability difference threshold can be set. When the difference between the highest and second-highest probabilities is less than this threshold, the label of that node can be temporarily marked as "pending" for later processing. Based on the deterministic tissue category labels of all nodes, spatial domain clustering is performed on the set of nodes belonging to the same tissue category. In three-dimensional space, all nodes with the label "bone tissue" form an initial seed set, all nodes with the label "dental tissue" form another initial seed set, and so on. Taking the initial seed point set of "bone tissue" as an example, within the topological connectivity of the adaptive boundary tracing grid, the adjacent nodes directly connected to any "bone tissue" node through the edge of the triangular facet are checked. If the label of the adjacent node is also "bone tissue", it is added to the current set, and the search continues from this newly added node until no new connected nodes with the same label can be found. Through this region growing algorithm, multiple spatially continuous and internally connected initial tissue blocks are finally formed.

[0085] It is understandable that optimizing the internal topology consistency of each initial organizational block is a necessary step, because the node set obtained by clustering may only be continuous in the sense of a point set, and its geometric surface may have defects. In specific implementation, for the point cloud cluster corresponding to an initial organizational block, a Poisson surface reconstruction algorithm is used to generate an initial, closed triangular mesh surface, and non-manifold edges, isolated triangular faces, or internal holes are detected in this initial surface. By deleting isolated triangular faces, filling internal holes, performing Laplacian smoothing iterative adjustment on vertex positions, and local retriangulation, these topological defects are eliminated, so that the geometric surface of the initial organizational block meets the topological requirement of a simply connected closed manifold. The optimized geometric surfaces of all initial organizational blocks are placed in a unified global 3D coordinate system. For initial organizational blocks that are spatially adjacent and have different organizational category labels, the vertices of their shared boundary region are iteratively aligned to achieve seamless registration. At the registered shared boundary, the overlapping triangular faces generated from the perspectives of different blocks are deleted, and new triangular faces are generated between the exposed boundary contours of the two blocks to connect them. Optionally, the vertex density at shared boundaries can be resampled before final suturing to ensure the quality of the mesh at the connection points. After completing the connection operations of all adjacent boundaries, a unified triangular mesh model containing clearly defined regions of bone tissue, tooth tissue, gingival soft tissue, and mucosal soft tissue is constructed. This triangular mesh model is the final three-dimensional model of the oral and maxillofacial region.

[0086] In one embodiment of the present invention, organizational labels are assigned to all nodes in the adaptive boundary tracing mesh according to the probability distribution output by the organizational attribute classifier. For each node in the mesh, the probability distribution corresponding to the voxels associated with that node is obtained from the classifier output. The organizational category with the highest probability value is selected from the probability distribution and assigned as the deterministic organizational category label of that node. Based on the organizational category label of the nodes, spatial domain clustering is performed on the set of nodes belonging to the same organizational category to form multiple spatially continuous initial organizational blocks. In three-dimensional space, all nodes with the same organizational category label are used as initial seed points. Based on these initial seed points, within the topology of the adaptive boundary tracing mesh, all nodes that are directly or indirectly connected to an initial seed point through triangular facet edges and have the same organizational category label are grouped into the same set. Each such set of nodes constitutes an internally connected point cloud cluster in three-dimensional space, and this point cloud cluster is defined as an initial organizational block.

[0087] In practice, for each node in the adaptive boundary tracking grid, its associated voxel has been assigned a four-dimensional probability distribution vector by a pre-trained tissue attribute classifier:

[0088]

[0089] in: Represents the probability of bone tissue formation. Represents the probability of tooth structure. Represents the probability of soft tissue in the gums. This represents the probability of mucosal soft tissue. The core operation of tissue labeling is to select the tissue category with the highest probability value from this probability distribution. Specifically, this is done by traversing the probability distribution vector. The four components are used to find the component with the largest value. The organization category corresponding to this component is selected as the deterministic organization category label for the current node. For example, for a node, if its probability vector is... The component with the highest probability value is The corresponding tooth tissue category will be assigned to that node. See Table 2 for the probability distribution and label assignment results of the three different nodes.

[0090] Table 2: Node Probability Distribution and Label Allocation Table

[0091]

[0092] It's understandable that in some cases, the highest probability value and the second highest probability value in the probability distribution may be very close. In such cases, directly selecting the highest probability category might lead to noise in subsequent clustering of labels. Optionally, a probability difference threshold can be introduced. To handle this type of situation, compare the highest probability and the second highest probability The difference between the values ​​determines the label allocation strategy. Label allocation decision function. It can be defined as:

[0093]

[0094] Where: symbol Represents the probability vector The label assignment decision results, symbols Indicates taking The category index corresponding to the maximum value in the middle, symbol Represents a probability vector The maximum value in, sign Represents a probability vector The second largest value in, sign This represents a preset probability difference threshold. For example, in Table 2, Node_6593 has a bone tissue probability... Probability with tooth structure The difference is only ,like Set as If the node is not specified, it will be marked as "pending" and will not participate in the current initial clustering, but will be left for subsequent processing.

[0095] In some embodiments, the process of spatial domain clustering of nodes belonging to the same tissue category based on their tissue category labels is performed within the topology of a three-dimensional space and an adaptive boundary tracing mesh. In a specific implementation, the first step of spatial domain clustering is to collect all nodes with the same deterministic tissue category label as the initial seed set for that tissue category. For example, all nodes labeled "bone tissue" constitute the initial seed set for bone tissue, and all nodes labeled "gingival soft tissue" constitute the initial seed set for gingival soft tissue. Taking the "bone tissue" initial seed set as an example, the clustering algorithm traverses every node in this set, searching for all adjacent nodes directly connected to the current node via triangular facet edges within the topological connectivity of the adaptive boundary tracing mesh. It can be understood that if an adjacent node's tissue category label is also "bone tissue" and it has not yet been assigned to any bone tissue point cloud cluster, then that adjacent node is assigned to the same point cloud cluster as the current node. This process is recursive or iterative. Starting from any initial bone tissue seed point, it searches and merges all nodes directly or indirectly connected to it via triangular facet edges and labeled "bone tissue," until no new nodes satisfying the conditions are found. At this point, an initial bone tissue block is formed, which is represented in three-dimensional space as an internally connected cluster of nodes. Optionally, nodes marked "pending" are not processed in the initial clustering stage. They can be reassigned after the main initial tissue blocks are formed, based on their spatial location and the labels of adjacent blocks, using rules such as majority voting. The above region growth process is repeated for all categories of initial seed point sets, ultimately forming multiple spatially continuous and internally connected initial tissue blocks in the oral and maxillofacial region, such as independent tooth tissue blocks, maxillary bone tissue blocks, gingival soft tissue blocks, etc. Each initial tissue block is the basic unit for subsequent surface reconstruction and suturing operations.

[0096] In one embodiment of the present invention, internal topological consistency optimization is performed on each initial tissue block. For the point cloud cluster corresponding to each initial tissue block, its three-dimensional convex hull is calculated or a Poisson surface reconstruction algorithm is used to generate an initial closed triangular mesh surface. Non-manifold edges, isolated triangular facets, or internal hole defects that may exist in the initial closed triangular mesh surface are detected. Through the addition and deletion of triangular facets, iterative adjustment of vertex positions, and local retriangulation, the discovered non-manifold edges, isolated triangular facets, or internal hole defects are eliminated, thereby obtaining an optimized geometric surface with a topological structure of a simply connected closed manifold. The optimized geometric surfaces of all initial tissue blocks are spatially registered and stitched together to ultimately fuse and construct a three-dimensional model of the oral and maxillofacial region. The optimized geometric surfaces of all initial tissue blocks are placed in the same three-dimensional coordinate system. For initial tissue blocks that are adjacent in space and have different tissue category labels, the vertices of their shared boundary regions are iteratively aligned to achieve seamless registration of adjacent geometric surfaces. At the registered shared boundary, overlapping triangular patches are deleted, and new triangular patches are generated on the clear boundary contours to connect the geometric surfaces of different initial tissue blocks. After completing the connection operation of all adjacent boundary regions, a unified triangular mesh model containing multiple tissue category regions is obtained. This triangular mesh model is the final three-dimensional model of the oral and maxillofacial region.

[0097] In practical implementation, internal topological consistency optimization is performed on each initial organizational block formed in the spatial domain clustering. Each initial organizational block corresponds to a cluster of internally connected node points in three-dimensional space. First, a geometric surface needs to be generated for this point cloud cluster. This can be achieved using algorithms that calculate the three-dimensional convex hull or Poisson surface reconstruction algorithms to generate an initial, closed triangular mesh surface. The Poisson surface reconstruction algorithm can reconstruct a smooth and watertight surface from the point cloud. It is understandable that regardless of the algorithm used to generate the initial closed triangular mesh surface, various topological defects may exist. These defects include non-manifold edges, isolated triangular facets, or internal holes. Non-manifold edges refer to edges shared by more than two triangular facets; isolated triangular facets refer to free facets that are not connected to the main surface mesh; and internal holes refer to unwanted gaps appearing on a continuous surface mesh.

[0098] In some embodiments, detecting defects in the surface of an initial closed triangular mesh requires a specialized mesh analysis algorithm. This algorithm traverses all edges in the mesh and counts the number of times each edge is referenced by a facet to identify non-manifold edges. It then examines face adjacency to identify isolated triangular facet clusters and locates internal holes by searching for boundary edges. In practice, eliminating identified non-manifold edges, isolated triangular facets, or internal holes requires a series of mesh editing operations. For non-manifold edges, redundant facets sharing the edge are typically deleted, or local regions are retriangulated to eliminate redundant connections. For isolated triangular facets, if their area is less than a threshold or far from the main surface, they are directly deleted from the mesh. For internal holes, the sequence of vertices constituting the hole boundary is identified, and triangulation is performed within the planar polygon defined by this set of vertices to generate new triangular facets to fill the hole. Through the addition and deletion of triangular facets, iterative adjustment of vertex positions, and retriangulation of local regions, the geometric surface of each initial block of the organization ultimately satisfies the topological requirements of a simply connected closed manifold—that is, a geometric shell without boundaries, isolated parts, and uniform thickness. Optionally, after internal topology consistency optimization, the geometric surfaces of each tissue initial block can be smoothed, for example, by applying a Laplacian smoothing algorithm, to improve mesh quality and prepare for subsequent registration. In practice, the optimized geometric surfaces of all tissue initial blocks are spatially registered and stitched. First, the optimized geometric surfaces of all tissue initial blocks need to be imported and placed in the same global 3D coordinate system. Since these surfaces originate from the same adaptive boundary tracking mesh, they are roughly aligned in space. However, for adjacent tissue initial blocks with different tissue category labels, their shared boundary regions may have slight gaps or overlaps due to independent surface reconstruction processes. The goal of iterative alignment of the vertex positions of the shared boundary region is to minimize the distance between two adjacent surfaces at the boundary, and its energy function... It can be defined as the sum of the squares of the distances between the pairs of vertices that need to be aligned:

[0099]

[0100] Where: symbol The sign represents the alignment error energy that registration needs to minimize. Indicates the shared boundary The summation of predefined corresponding vertex pairs, sign This represents the initial block of the organization. The first on the boundary Vertex positions, symbols This represents the initial block belonging to an adjacent organization. The corresponding first boundary Vertex positions, symbols This represents the calculation of the Euclidean distance between two points. It involves iterative adjustments. and Position to minimize It can achieve seamless registration of adjacent geometric surfaces.

[0101] It is understandable that at the shared boundary after registration, the geometric surfaces of the initial blocks from two different tissues often have overlapping faces. In some embodiments, these overlapping triangular faces need to be removed. This is done by traversing all faces in the boundary region; if all three vertices of a triangular face are extremely close to the surface of the other block, the face is considered overlapping and removed. In a specific implementation, after removing overlapping faces, the geometric surfaces of the two initial tissue blocks will have an open, contour-matched loop at the boundary. New triangular faces need to be generated on this loop to connect the two surfaces. High-quality connected meshes can be generated by performing Delaunay triangulation between the opposite boundary vertices or by using a striping algorithm. Optionally, after generating new faces, vertices near the seam can be locally relaxed to smooth the geometry at the seam. After connecting all adjacent boundary regions, the geometric surfaces of the original independent initial tissue blocks are merged into a unified, topologically continuous overall triangular mesh model. This unified triangular mesh model accurately includes different categories of regions such as bone tissue, dental tissue, gingival soft tissue, and mucosal soft tissue, which is the final three-dimensional model of the oral and maxillofacial region containing clear soft and hard tissue boundary information.

[0102] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues, characterized in that, include: Obtain raw three-dimensional voxel data of the patient's oral and maxillofacial region, wherein the raw three-dimensional voxel data includes grayscale information; A multi-scale feature field is constructed on the original three-dimensional voxel data. The feature gradient intensity of each voxel is calculated based on the gray value information in the multi-scale feature field. An adaptive boundary tracking mesh is generated based on the spatial distribution of the feature gradient intensity. On the nodes of the adaptive boundary tracking mesh, the soft and hard tissue boundary energy is calculated based on the voxel gray value information and the feature gradient intensity. The soft and hard tissue boundary energy is used to quantify the degree of belonging of the node at the boundary between soft and hard tissue. Using the boundary energy of soft and hard tissues as the field strength, a boundary energy field is constructed on an adaptive boundary tracing grid, and the propagation of force lines in the boundary energy field is simulated to trace the high-energy ridges formed by the convergence of force lines. Identify and extract the high-energy ridges, and mark the adaptive boundary tracking grid cells traversed by the high-energy ridges as a set of boundary voxels of the tissue to be classified; Based on the set of voxels representing the tissue boundaries to be classified, and combined with a pre-trained tissue attribute classifier, a three-dimensional model of the oral and maxillofacial region is generated.

2. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 1, characterized in that, A multi-scale feature field is constructed on the original three-dimensional voxel data. Within this multi-scale feature field, the feature gradient intensity of each voxel is calculated based on its grayscale value information, including: The original three-dimensional voxel data is convolved and filtered using three-dimensional Gaussian kernels of different scales to generate multiple smooth voxel data layers corresponding to different spatial resolutions. Calculate the gray-level gradient components of each voxel in the three spatial directions in each layer of smoothed voxel data; The gray-level gradient components of each voxel at the corresponding position in all smooth voxel data layers are weighted and fused to obtain the multi-scale fused feature gradient vector of the voxel. The magnitude of the multi-scale fused feature gradient vector is defined as the feature gradient intensity of the voxel.

3. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 2, characterized in that, The process of generating an adaptive boundary tracking mesh based on the spatial distribution of feature gradient intensities includes: In three-dimensional space, using voxels as the basic unit, voxels with feature gradient strength higher than the mesh generation threshold are marked as candidate mesh nodes; Initial connections are established between candidate grid nodes based on the directional consistency of feature gradient strength, forming a sparse spatial graph structure. The spatial graph structure is processed into triangular facets to generate a set of triangular facets that are oriented and connected with candidate mesh nodes as vertices and feature gradient directions as references. The set of triangular facets constitutes the adaptive boundary tracking mesh.

4. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 3, characterized in that, At the nodes of the adaptive boundary tracking mesh, the boundary energy of soft and hard tissues is calculated based on voxel gray value information and feature gradient intensity, including: For each node on the adaptive boundary tracing mesh, obtain the gray value of its corresponding voxel in the original 3D voxel data, as well as the characteristic gradient intensity of the voxel calculated accordingly. The grayscale values ​​are converted into tissue density feature values ​​using a predefined mapping function; Multiplying the tissue density eigenvalue by the eigengradation intensity and then by a correction coefficient related to the local curvature of the node, the resulting product is the soft and hard tissue boundary energy of the node.

5. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 4, characterized in that, The process involves using the boundary energy of soft and hard tissues as the field strength, constructing a boundary energy field on an adaptive boundary tracing grid, simulating the propagation of force lines within the boundary energy field, and tracing the high-energy ridges formed by the convergence of these force lines. Each node of the adaptive boundary tracing mesh is considered as a field source, and the magnitude of its field strength is defined by the soft and hard organization boundary energy of the node. Starting from the node with the highest field strength, the simulated virtual force lines propagate along the triangular facet connection direction of the adaptive boundary tracing mesh towards the direction where the energy of the soft and hard tissue boundary decreases; Record the propagation path of all virtual force lines, and extract the nodes and triangle edges through which multiple virtual force lines converge; The nodes and triangular facets that converge are connected and smoothed in space to form a continuous spatial curve, which is the high-energy ridge.

6. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 5, characterized in that, Based on the set of boundary voxels of the tissue to be classified, and combined with a pre-trained tissue attribute classifier, a three-dimensional model of the oral and maxillofacial region is generated, including: The pre-trained tissue attribute classifier is used to determine the attributes of each voxel in the set of voxels of the tissue boundary to be classified. The tissue attribute classifier outputs the probability distribution of each voxel belonging to bone tissue, tooth tissue, gingival soft tissue or mucosal soft tissue. Based on the probability distribution, organizational labels are assigned to all nodes in the adaptive boundary tracking mesh, and each node is given a deterministic organizational category label; Based on the organizational category label of the nodes, spatial domain clustering is performed on the set of nodes belonging to the same organizational category to form multiple spatially contiguous initial organizational blocks; For each organization's initial block, internal topology consistency optimization is performed to adjust the spatial connection relationship of the nodes within the organization's initial block, ensuring that each organization's initial block constitutes a closed and continuous geometric surface. The geometric surfaces of all optimized initial tissue blocks are spatially registered and stitched together, and finally fused to construct a three-dimensional oral and maxillofacial model containing clear soft and hard tissue boundary information. The attributes of each voxel in the set of boundary voxels of the tissue to be classified are determined using a pre-trained tissue attribute classifier, specifically including: The tissue attribute classifier uses the multi-scale fusion feature gradient vector of voxels, the gray value of the voxels in the original three-dimensional voxel data, and the tissue category label distribution of the voxels in the adaptive boundary tracking grid as input features. The organizational attribute classifier performs nonlinear transformation and mapping on the input features through an internal deep neural network; The tissue attribute classifier outputs a multidimensional vector, where each dimension of the multidimensional vector represents the probability value of the input voxel belonging to one of four predefined categories: bone tissue, tooth tissue, gingival soft tissue, and mucosal soft tissue.

7. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 6, characterized in that, Based on the probability distribution, label assignment is performed on all nodes in the adaptive boundary tracking mesh, including: For each node in the adaptive boundary tracking mesh, obtain the probability distribution corresponding to the voxels associated with the node, based on the output of the organization attribute classifier; Select the organization category with the highest probability value from the probability distribution; The organization category with the highest probability value is assigned as the deterministic organization category label of the node.

8. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 7, characterized in that, Based on the organizational category labels of the nodes, spatial domain clustering is performed on the set of nodes belonging to the same organizational category to form multiple spatially contiguous initial organizational blocks, including: In three-dimensional space, all nodes with the same organization category label are used as initial seed points; Based on the initial seed point, within the topology of the adaptive boundary tracing mesh, all nodes that are directly or indirectly connected to the initial seed point through triangular facet edges and have the same organizational category label are grouped into the same set. Each such set of nodes constitutes an internally connected point cloud cluster in three-dimensional space, and the point cloud cluster is defined as an initial block of organization.

9. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 8, characterized in that, The process of optimizing the internal topology consistency of each organization's initial block includes: For each point cloud cluster corresponding to the initial block of the organization, calculate its three-dimensional convex hull or use the Poisson surface reconstruction algorithm to generate an initial closed triangular mesh surface; Detect non-manifold edges, isolated triangular facets, or internal hole defects present in the initial closed triangular mesh surface; By adding or deleting triangular facets, iteratively adjusting vertex positions, and performing local retriangulation, the non-manifold edges, isolated triangular facets, or internal holes are eliminated, resulting in an optimized geometric surface with a topological structure of a simply connected closed manifold.

10. The method for constructing a three-dimensional model of the oral and maxillofacial region based on the identification of oral soft and hard tissues according to claim 9, characterized in that, The process involves spatially registering and stitching the geometric surfaces of all optimized initial tissue blocks, ultimately fusing them to construct a three-dimensional oral and maxillofacial model containing clearly defined soft and hard tissue boundary information, including: The optimized geometric surfaces of all initial tissue blocks are placed in the same three-dimensional coordinate system; For adjacent initial tissue blocks with different tissue category labels, the vertices of their shared boundary regions are iteratively aligned to achieve seamless registration of adjacent geometric surfaces. At the registered shared boundary, overlapping triangular patches are deleted, and new triangular patches are generated on the boundary contour to connect the geometric surfaces of the initial blocks of different tissues. After connecting all adjacent boundaries, a unified triangular mesh model containing multiple tissue category regions is obtained, which is the final three-dimensional model of the oral and maxillofacial region.

Images