Two-stage automatic appliance generation method and device based on tooth edge enhancement and readable storage medium thereof

Abstract

The invention provides a two-stage automatic appliance generation method and device based on tooth edge enhancement and a readable storage medium thereof, and belongs to the field of oral medicine digitization. According to the method, jaw pad generation is converted into a point cloud completion task, upper and lower jaw point clouds are preprocessed and then input into a point cloud completion network, appliance point clouds are generated through a coarse-to-fine stage, an interference area is removed through surface reconstruction, lower jaw line enhancement and Boolean difference operation, and a personalized appliance model is generated. Through combination of deep learning and geometric processing, manual operation is reduced, the precision, efficiency and standardization degree of jaw pad generation are improved, and the problem that an existing digital technology depends on manual boundary sketching is solved.

Description

Technical Field

[0001] The present invention relates to the field of oral medicine digital technology, and in particular to a method and device for generating a two-stage automatic orthodontic appliance (jaw pad) based on tooth edge enhancement, and a readable storage medium thereof. Background Art

[0002] Jaw pads are removable orthodontic appliances used to adjust occlusion. Traditional manual production suffers from low precision and inefficiency. While digital jaw pads improve production accuracy through oral scans, existing technology still relies on extensive manual work, such as manually delineating jaw pad boundaries, adjusting contact zone parameters, and manually processing occlusal interference areas.

[0003] In recent years, artificial intelligence and deep learning have provided new ideas for three-dimensional geometry generation. However, how to efficiently apply them to the automated generation of jaw pads to solve the problems of cumbersome manual operations and insufficient process standardization has become a technical difficulty that urgently needs to be overcome in this field.

[0004] Therefore, there is an urgent need for a new two-stage automatic orthodontic appliance generation method, device and readable storage medium based on tooth edge enhancement to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a two-stage automatic orthodontic appliance generation method, device and readable storage medium based on tooth edge enhancement, which addresses the problem that current technology requires manual participation in jaw pad boundary delineation, contact area setting and parameter adjustment, resulting in heavy workload, low efficiency and insufficient standardization.

[0006] The core technology of this invention is to convert the generation of jaw pads (orthodontic appliances) into a point cloud completion task. Through a two-stage algorithm (point cloud generation and surface reconstruction), combined with preprocessing, point cloud completion network, adaptive denoising, mandibular line enhancement and Boolean operations, the fully automatic generation of jaw pads (orthodontic appliances) is achieved.

[0007] In a first aspect, the present invention provides a two-stage automatic appliance generation method based on tooth edge enhancement, the method comprising the following steps: Perform geometric preprocessing on the input maxillary and mandibular point clouds to obtain standardized point clouds; The standardized point cloud is input into the point cloud completion network, and the appliance point cloud is generated through a coarse-to-fine two-stage structure; Perform surface reconstruction and geometry optimization on the appliance point cloud, and fuse the original contour to enhance the mandibular line; The interference areas between the appliance point cloud and the upper and lower jaws were removed through Boolean difference operation to generate a personalized appliance model.

[0008] Furthermore, geometric preprocessing includes: Calculate the centroid of the input maxillary and mandibular point clouds and translate them to the origin; Scale the point cloud to fit within the unit sphere; Align the main directions of the point cloud through principal component analysis; Downsample the point cloud.

[0009] Furthermore, the coarse-to-fine two-stage structure of the point cloud completion network includes: The first stage: extract point cloud block-level features through farthest sampling and dynamic graph convolution, combine position encoding and input Transformer encoder to generate global features, and generate initial coarse point cloud through MLP; The second stage: The initial coarse point cloud is divided into high-confidence point clouds through the point filtering module, and then merged with the original point cloud to generate a fine point cloud through FoldingNet.

[0010] Furthermore, the point filtering module divides the initial coarse point cloud into a high-confidence set and a low-confidence set by calculating the average self-attention score of the point cloud, and takes the high-confidence set as the filtered point cloud.

[0011] Furthermore, after generating the appliance point cloud, it also includes: Principal component analysis was performed on the appliance point cloud, and the point cloud was segmented along the principal axis direction with the minimum variance. The upper half of the point cloud was offset along this direction to generate an enhanced sample of the occlusal contact area.

[0012] Furthermore, surface reconstruction and geometry optimization include: Adaptive denoising of appliance point clouds is performed based on a joint scoring function of local density, curvature, and normal consistency; Estimating the normal vector of the denoised point cloud; Poisson surface reconstruction based on normal vector consistency weights.

[0013] Further, the jawline is enhanced by integrating the original contours including: Apply the AlphaWrap algorithm to the initial cropped model to generate the expanded model; Perform Boolean operation on the expanded model and the appliance model after surface reconstruction, and obtain the optimized appliance model after resampling; Boolean difference operations include: First, the optimized appliance model was subjected to a Boolean subtraction operation with the undercut mandibular model, and then with the maxillary model to remove the interference area.

[0014] In a second aspect, the present invention provides a two-stage automatic jaw pad generation device based on tooth edge enhancement, comprising: The preprocessing module performs geometric preprocessing on the input upper and lower jaw point clouds to obtain standardized point clouds; The point cloud completion module inputs the standardized point cloud into the point cloud completion network and generates the appliance point cloud through a coarse-to-fine two-stage structure; The reconstruction module performs surface reconstruction and geometric optimization on the appliance point cloud, integrating the original contour to enhance the mandibular line; The output module removes the interference area between the appliance point cloud and the upper and lower jaws through Boolean difference operation to generate a personalized appliance model.

[0015] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the above-mentioned two-stage automatic orthodontic appliance generation method based on tooth edge enhancement.

[0016] In a fourth aspect, the present invention provides a readable storage medium having a computer program stored therein, the computer program including a program code for controlling a process to execute a process, the process including the two-stage automatic appliance generation method based on tooth edge enhancement described above.

[0017] The main contributions and innovations of the present invention are as follows: 1. Improved automation and reduced manual intervention: A two-stage algorithm transforms jaw pad generation into a point cloud completion task, eliminating the need for manual boundary drawing and contact zone parameter setting. This significantly reduces the doctor's workload and improves the efficiency and standardization of the digital process.

[0018] 2. Optimizing model accuracy and geometric quality: The point cloud completion network combines dynamic graph convolution with the Transformer encoder to improve point cloud density and uniformity through a coarse-to-fine generation mechanism. An adaptive denoising algorithm and Poisson reconstruction, combined with local density, curvature, and other features, enhance the model's surface smoothness and geometric continuity.

[0019] 3. Enhanced matching of jawline contour and occlusion: The AlphaWrap expansion model and Boolean fusion operations resolve the blurred jawline problem in traditional methods and ensure a clear contour. The Boolean difference operation accurately removes the interference area between the upper and lower jaws, generating a personalized model that conforms to the actual occlusion relationship.

[0020] 4. Data enhancement and robustness improvement: A PCA-based data enhancement method for the occlusal contact area improves the point cloud's ability to express occlusal features; adaptive weighted Poisson reconstruction is more adaptable to sparse / inhomogeneous point clouds, enhancing the robustness of the algorithm.

[0021] The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below so that other features, objects, and advantages of the invention are more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 A processing flow chart of a point cloud completion network according to an embodiment of the present invention; Figure 2 is a structural diagram of a dynamic graph convolution module according to an embodiment of the present invention; Figure 3 is a structural diagram of a point filtering module according to an embodiment of the present invention; Figure 4 is a point cloud image of an appliance with more uniform point cloud distribution according to an embodiment of the present invention; Figure 5 is a point cloud rendering of an appliance after data enhancement according to an embodiment of the present invention; Figure 6 is a schematic diagram of an appliance model ModelA after surface reconstruction according to an embodiment of the present invention; Figure 7 is a schematic diagram of applying the AlphaWrap algorithm to the initial cropped model to generate an expanded model Model B according to an embodiment of the present invention; Figure 8 is a schematic diagram of an optimized appliance model with a clear jaw line according to an embodiment of the present invention; Figure 9 is a schematic diagram of an appliance model conforming to an occlusal relationship according to an embodiment of the present invention; Figure 10 FIG. 4 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The implementations described in the following exemplary embodiments are not intended to represent all implementations consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of this specification, as detailed in the appended claims.

[0024] It should be noted that in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the method may include more or fewer steps than those described in this specification. In addition, a single step described in this specification may be broken down into multiple steps for description in other embodiments, and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0025] The generation of existing digital orthodontic appliances (jaw pads) requires manual boundary drawing and contact area setting, which is inefficient and lacks standardization.

[0026] Based on this, the present invention solves the problems existing in the prior art based on a two-stage process of "point cloud generation → surface reconstruction and geometry optimization".

[0027] Example 1 The present invention aims to propose a two-stage automatic orthodontic appliance (jaw pad) generation method based on tooth edge enhancement. Specifically, the method comprises the following steps: Step 1: Perform geometric preprocessing on the input upper and lower jaw point clouds to obtain standardized point clouds; In this embodiment, the raw data of the mandibular point cloud includes scanned models of the upper and lower teeth or manually cropped models. Because manually cropped models may have uncertainties in position and orientation, which may affect the point cloud completion effect, they need to be geometrically preprocessed. The specific steps are as follows: 1) Centroid translation (moving the center of the point cloud to the origin) Assume that the original point cloud is P={p1,p2,...,p N}, p i ∈R 3 , calculate the center of mass C of the original point cloud:

[0028] Where N is the total number of points in the point cloud; p i = (x i , y i , z i ) represents the 3D coordinates of the i-th point in the point cloud (i = 1, 2, ..., N)); Then translate the point cloud by subtracting the centroid from the points in the cloud:

[0029] in, Represents the coordinates of the translated points, that is, the constituent points of the point cloud after translation.

[0030] 2) Unit sphere normalization (scaling the point cloud to fit within the unit sphere) For all points , calculate its distance from the origin:

[0031] Among them, d i express Euclidean distance to the origin; Then find the maximum distance: Finally, all points are divided by the maximum distance and scaled to the unit sphere:

[0032] in, Represents the coordinates of the scaled point, that is, the points that make up the point cloud after normalization of the unit sphere.

[0033] After scaling to the unit sphere, all models are processed at the same geometric scale (coordinate range [-1, 1]), which can avoid scale interference during network learning.

[0034] 3) Principal Component Analysis (PCA) rotation (alignment of principal directions) Assume that the normalized point cloud matrix is ​​X (usually Matrix (N is the number of points, each row corresponds to a three-dimensional point coordinate (x i , y i , z i )), perform covariance analysis on X: Among them, C is the covariance matrix of the point cloud, and the dimension is , characterizes the distribution characteristics of point clouds in three-dimensional space; is the transposed matrix of X, and the dimension becomes ; Calculate the eigenvalues ​​and eigenvectors C = UΛUT: Where U is the eigenvector matrix ( , orthogonal matrix, U T =U -1 ), whose column vector is the eigenvector of C, corresponding to the principal component direction (PCA axis) of the point cloud; is the eigenvalue diagonal matrix ( ), the diagonal elements λ1 ≥ λ2 ≥ λ3 are the eigenvalues ​​of C, sorted from largest to smallest. PCA transforms the point cloud from the original coordinate system to a new coordinate system spanned by the principal components through eigendecomposition of the covariance matrix, achieving "decorrelation" and "main direction alignment."

[0035] Finally, project the point cloud to the new coordinate system:

[0036] in, is the point cloud matrix after projection to the new coordinate system ( ), which is the aligned point cloud.

[0037] Because the axes of the new coordinate system are composed of the principal component directions (PCA axes) of the point cloud, the principal directions of the projected point cloud (e.g., the direction of the tooth row) coincide with the principal axes of the new coordinate system (e.g., the z-axis), eliminating rotational deviations from the original point cloud. For example, if the principal direction of the original point cloud was oblique (e.g., the tilt angle of a patient's dental model), this direction will be parallel to the z-axis of the coordinate system after projection, ensuring consistent orientation across different patients' dental models.

[0038] Therefore, it can ensure that the upper and lower mandibular point clouds of different patients are input into the network in the same direction, so that the generated orthodontic appliance model is more in line with actual clinical needs in terms of occlusal relationship, for example, the direction of the mandibular line is consistent with the curvature of the dental arch.

[0039] 4) Point cloud downsampling (reduce the number of points and improve processing efficiency) Set the number of points to N, and randomly sample M points from it (M < <N):

[0040] In this way, the downsampled point cloud serves as the network input. The completion network learns the distribution patterns of the original point cloud and recovers the complete appliance point cloud (including detailed features) from M points, achieving a "complete with less" effect. Downsampling is chosen for its simplicity and low computational complexity (O(N) time complexity, requiring only random number generation and index extraction). Most importantly, it requires minimal hardware and is suitable for real-time preprocessing (i.e., the immediate generation of jaw pads after clinical oral scans, as targeted by this invention).

[0041] Of course, random sampling may lose local key features (such as tooth tips and edge points). Therefore, the present invention enhances the local feature extraction capability through "farthest sampling + dynamic graph convolution" in the subsequent network to make up for the shortcomings of random sampling.

[0042] Step 2: Input the standardized point cloud into the point cloud completion network and generate the appliance point cloud through a coarse-to-fine two-stage structure; In this embodiment, the original point cloud obtained through the above-mentioned preprocessing steps is sent to the point cloud completion network, and the original point cloud can be used to obtain the completed detail point cloud through the point cloud completion network.

[0043] Specifically, if Figure 1 As shown in the figure, the input base point cloud is the point cloud of the lower teeth obtained through preprocessing, which serves as the geometric basis for the generation of the jaw pad. The point cloud completion network first performs the farthest sampling on the input original point cloud (base point cloud) (randomly selects an initial point, and each time selects the point farthest from the selected point set and adds it to the set until K points are selected), obtains K seed points, and uses each seed point as the center to define a spherical neighborhood with a fixed radius r, dividing the point cloud into K local blocks. Each block contains the seed point and the points in its neighborhood, achieving the "local-global" decomposition of the point cloud, and each block enters the dynamic graph convolution module to extract block-level features. The seed point of each block is converted into a position code through a sine function, and the features and position codes corresponding to each block are concatenated as the input of the Transformer encoder, and the unified features of the original point cloud are output after maximum pooling. :

[0044] in, is the i-th block; is the i-th seed point; is the local geometric features of the i-th block (such as edge and curvature features); is the position code (spatial coordinate mapping) of the i-th seed point; concat(·) is the concatenation of local features and position codes to form a composite feature containing "geometry + position"; T P (·) is the Transformer encoder that processes composite features through the self-attention mechanism to generate global correlation features; MaxPool(·) extracts key information from global features through maximum pooling and outputs a unified feature F P .

[0045] Among them, Figure 2 As shown in the figure, the dynamic graph convolution module first performs a k-nearest neighbor search on the input point to construct the adjacency graph for the current layer. It then uses a multi-layer perceptron (MLP) to extract edge features and feeds them into the attention module to enhance expressiveness. In the attention module, one path undergoes global average pooling, followed by an MLP+Sigmoid function to generate weights. Another path uses max pooling and a ReLU activation function to extract salient features. These features are then multiplied by the weights and concatenated with the residual of the original features to produce enhanced point features, the desired block-level features.

[0046] After obtaining the features of the point cloud (unified features ), a multi-layer perceptron (MLP) is used to transform the point cloud features into the initial coarse point cloud coordinates P c (Mapping high-dimensional features into three-dimensional coordinates), only retaining the basic shape outline of the jaw pad, providing a structural basis for subsequent fine generation.

[0047] In order to obtain a more reliable coarse point cloud, a point filtering module is proposed. By calculating the average self-attention score of each point, the point cloud is divided into a high confidence set and a low confidence set (high confidence set 75%, low confidence set 25%), and the high confidence set is used as the filtered point cloud P f .like Figure 3 As shown, in the point filtering module, the point coordinates first pass through the MLP layer to obtain the corresponding features, and then compare them with Splicing . Then calculate the average self-attention score between all connected features:

[0048] Based on the calculated confidence score, P c Divide into high confidence set and low confidence set, and use the high confidence set as the filtered intermediate point cloud P f。At the point cloud decoder, a denser and finer point cloud P is generated r 。

[0049] Merge P f with the original point cloud P. The merged point cloud is sampled at the farthest end to obtain a new point cloud P de , and a final point cloud P is generated through a FoldingNet de 。

[0050] For example, the farthest end sampling pseudocode: Given a point cloud P = {p1, p2, …, p N}}, the goal is to sample M << N points: 1. Randomly select an initial point p i1 , and add it to the sampling set S.

[0051] 2. For each unselected point , calculate its minimum distance d(p, S) from all points in the set S.

[0052] 3. Select the point p* = argmax d(p, S) with the maximum distance and add it to S.

[0053] 4. Repeat steps 2 and 3 until M points are selected.

[0054] The FoldingNet maps low-dimensional features to three-dimensional space through a "folding" operation. The core steps are: mapping the point coordinates of P de to a low-dimensional latent space (such as 128 dimensions); remapping the latent space features to three-dimensional coordinates through multiple folding operations (Folding Layer) to generate a dense point cloud P r . The formula is expressed as: P r = FoldingNet(P de ); After obtaining the final point cloud P de , the preprocessed point cloud needs to be divided into 20 parts and input into the point cloud completion network respectively to obtain 20 corresponding jaw pad generation point clouds. Subsequently, these 20 generated point clouds are merged, and the merged result is downsampled to obtain a more evenly distributed orthodontic appliance point cloud. Finally, an inverse process is performed on this point cloud to restore it to the same scale and coordinate system as the original cropped model, obtaining a figure as shown in Figure 4 , Figure 4 showing a more evenly distributed orthodontic appliance point cloud.

[0055] The next step is data augmentation. Principal component analysis (PCA) is performed on the appliance point cloud. Eigenvalue decomposition is used to project the point cloud onto orthogonal principal axes. The principal axis direction with the smallest variance is selected as the principal direction for segmentation (maximum variance direction: the direction in which the point cloud is most dispersed (such as the length or width of the jaw pad); minimum variance direction: the direction in which the point cloud is most compact (i.e., the thickness direction), representing the approximate perpendicular direction of the upper and lower surfaces of the jaw pad (the main direction of occlusal contact)). The point cloud is then divided into two parts along this principal direction, with the upper half used to construct data augmentation samples for the occlusal contact points.

[0056] The reason why the principal axis direction with the smallest variance is selected as the main segmentation direction is that the occlusal contact points are usually distributed on the upper surface of the jaw pad (the contact area with the upper teeth), and the coordinate difference of this area in the thickness direction (the direction of minimum variance) is the smallest. Segmentation along this direction can accurately extract the contact surface.

[0057] In this embodiment, the logical segmentation basis for the upper and lower half segmentation is: calculating the projection values ​​of all points along the minimum variance direction d , take the median As the segmentation threshold; the upper part is selected: only the points with projection values ​​greater than the median are retained ( > ), these points correspond to the upper surface of the jaw pad (the occlusal contact area close to the upper teeth). Therefore, the key area to focus on is to eliminate interference from the jaw pad edge and non-contact area, and only enhance the occlusal contact point samples to obtain the precise occlusal surface between the jaw pad and the upper teeth.

[0058] Then the upper half of the point cloud is offset by 2 units along the main direction to simulate the deformation caused by the compression of the jaw pad during occlusion and artificially increase the spatial diversity of the contact points: Where P is the original point cloud, PCA minimum variance direction; Δ=2 is the offset distance; P aug This is the enhanced point cloud, the effect is as follows Figure 5 shown.

[0059] Step 3: Surface reconstruction and geometry optimization of the appliance point cloud are performed, and the original contour is integrated to enhance the mandibular line; In this embodiment, surface reconstruction is performed on the appliance point cloud after data enhancement, and post-processing operations are performed, including smoothing, denoising, and mesh refinement, to improve the surface quality and geometric continuity of the model. The specific steps are as follows: 1. Point cloud preprocessing Adaptive point cloud denoising algorithm based on local geometric features: combines the local density, local curvature, and normal consistency of the point cloud to adaptively determine whether each point is a noise point.

[0060] First, for each point p in the point cloud i , define a sphere with it as the center and radius r. All points p within the range of this sphere j This constitutes point p i The “neighborhood point set” , this neighborhood is used to calculate local geometric features:

[0061] 1) Local density estimate:

[0062] It's point p i The local density estimate of the point. Calculate the distance from each point in its neighborhood to itself and find the average of these distances. The larger the average distance, the sparser the surrounding points. The smaller the distance, the denser the local area. This value is used to determine whether the point is in a sparse area and whether it may be a noise point. The prior knowledge that "low density = possible noise" can be converted into a computable numerical indicator, combined with curvature and normal consistency to form a multi-feature joint judgment, which can not only eliminate scanning noise, but also retain sparse but key geometric structures such as tooth edges and cusps, laying the foundation for high-quality point clouds for subsequent surface reconstruction.

[0063] 2) Local curvature estimate:

[0064] Principal component analysis (PCA) on point p i The covariance analysis of the neighborhood points is performed to obtain three eigenvalues ​​λ1, λ2, λ3, λ min It can be used to estimate curvature: if the curvature is large, it means that the point is located at a boundary, sharp corner, or mutation area. These areas may be real structures even if the density is low and cannot be easily deleted as noise points.

[0065] 3) Normal consistency estimate: Used to characterize the directional consistency of the normal vector within its neighborhood; n i Represents point p i Normal vector (three-dimensional unit vector, perpendicular to the local surface of the point cloud); n j Represents the neighborhood point p j Normal vector of the neighborhood point; by calculating the normal vector of each neighborhood point and p i The cosine value of the angle between the normal vectors and the average is obtained . The closer it is to 1, the higher the normal consistency is, and the point is on a smooth surface; if the consistency is low, it may be an edge, corner or abnormal point, such as the point on the incisal edge of a tooth, where the normal vectors in the neighborhood point in different directions. Low; the neighborhood of the outlier may contain real points and noise points, and the normal vector direction is chaotic, Also low.

[0066] 4) Joint scoring function: The above three features (density , curvature , normal consistency ) into a comprehensive scoring function , used to measure point p i Is it a noise point? Each item is normalized (to make different features in the same numerical range) and weighted summed by learnable weights α, β, γ: in Characterizes normal inconsistency and density , curvature Together they make up the noise score: like Low and High (sparse), it may be a noise point; if Low but High (large curvature), possibly an edge or a sharp point (real structure, preserved).

[0067] Finally, a threshold θ is set to determine whether it is a noise point. If the threshold is exceeded, it is considered a noise point and should be removed from the point cloud: is a noise point; This allows for adaptive denoising (e.g., outliers and scanning error points) while retaining true geometric points. After this, homogenization is required on the denoised point cloud, which is now free of significant noise. The purpose of downsampling is to optimize point distribution rather than denoising. Because point clouds generated by point cloud completion networks or after data augmentation may be locally dense (e.g., occlusal contact areas) or sparse (e.g., jawline edges), direct reconstruction can lead to: The mesh in dense areas is overly refined, increasing the computational effort; The reconstructed surface in sparse areas is discontinuous (such as holes).

[0068] In this embodiment, point cloud downsampling is: voxel grid downsampling, which divides the space into cubic voxels of a fixed size, and retains only one point in each voxel, which is the centroid of the corresponding voxel.

[0069] Assume that the length of the voxel side is s, and divide the three-dimensional space into a voxel grid. For each voxel V, if it contains the point set , then replace the original points with the centroids of these points:

[0070] Among them, if there is no point or only one point in the voxel, the empty voxel or the original point is retained.

[0071] The non-uniform point cloud generated by deep learning is converted into high-quality input suitable for traditional geometric algorithms (Poisson reconstruction), ensuring the surface smoothness and geometric accuracy of the orthodontic appliance model. This operation, in conjunction with adaptive denoising, creates a point cloud optimization chain from "eliminating artifacts and retaining true features to uniform representation," ultimately improving the clinical applicability of the digital jaw cushion.

[0072] 2. Normal vector estimation For each point p i , search for its k nearest neighbors (find the k nearest points through kd tree or ball query , construct a local neighborhood point set), perform principal component analysis (PCA) on its neighborhood point set, and obtain eigenvalues ​​λ1, λ2, λ3 and corresponding eigenvectors v1, v2, v3. For example, the eigenvector v3 corresponding to the minimum eigenvalue λ3 is the normal vector n i , because it represents the direction with the least local change in the point cloud (perpendicular to the surface), a minimum spanning tree traversal is used to ensure consistency.

[0073] Preferably, if k is too small, it is easily affected by noise (for example, when k=5, the neighborhood contains noise points), and if k is too large, local features will be blurred (such as the sharpness of the tooth edge). In the present invention, k is usually set to 15~25.

[0074] Among them, the minimum spanning tree (MST) traversal is: 1) Graph construction: Build a fully connected graph with all points as nodes and edge weights as the distance between points; 2) MST generation: Use Prim or Kruskal algorithm to generate a minimum spanning tree, ensuring that any two points in the tree are connected and the total edge weight is minimal; 3) Traversal adjustment: Select the reference point p0 and manually set its normal vector direction (e.g. outward); traverse the adjacent points along the MST. If the adjacent point p i If the angle between the normal vector and p0 is greater than 90°, reverse n i direction to ensure consistency.

[0075] 4) Mathematical expression: If adjacent points p i With p j The cosine of the angle between the normal vector and , then n i =-n i, so that cosθ>0.

[0076] In this embodiment, the minimum spanning tree from graph theory is used to optimize the normal vector orientation, ensuring that the normal vectors of the appliance point cloud accurately reflect the local surface orientation while maintaining global uniformity (e.g., all facing outward). This lays the foundation for subsequent Poisson reconstruction to generate a high-quality mesh. This operation is crucial for clinical jaw braces because it ensures that the model surface conforms to actual physical laws (e.g., an outward-facing normal vector represents the outer surface of the jaw brace), thereby forming a correct occlusal contact relationship with the patient's teeth when worn.

[0077] 3. Adaptive Weight Poisson Reconstruction According to factors such as the local density and confidence of the point cloud, different regions are assigned different weights in the Poisson equation to improve the adaptability of reconstruction to sparse / inhomogeneous point clouds.

[0078] Introducing weight terms into the Poisson equation , reflecting the consistency of the local normals of the point cloud. The more consistent the normals (indicating higher credibility), the greater the reconstruction weight assigned; conversely, the lower the weight is, the more it is used to suppress erroneous normals that interfere with the Poisson solution.

[0079] in, For point p i The normal consistency value (obtained by averaging the cosine of the neighborhood normal vector angles in the previous step, in the range [0,1]); γ is the weight amplification factor (γ ≥ 1, typical value γ = 2), which is used to enhance the weight advantage of high consistency points; the denominator is the normalization term to ensure that the sum of all weights is 1. When N i ≈1 (normal height is consistent): w(p i ) is close to 1, the point has a strong constraint on the reconstructed surface; when N i ≈0 (normal chaos): w(p i ) is close to 0, and the influence of this point on the surface is suppressed.

[0080] By introducing weights into the Poisson reconstruction equation, the goal becomes to solve the weighted Poisson field:

[0081] Where V(p) is the approximation of the “gradient field” constructed by the point and the normal vector, and the weight The intensity of the gradient field V(p) is modulated so that the "indicator field" in high-confidence areas contributes more to the Poisson field, thereby guiding the surface to fit closer to the point cloud in these areas. Specifically, the three-dimensional space is divided into voxel grids, and the finite difference method is used to discretize the equation: For each voxel center v, calculate the weighted gradient sum of its neighborhood points; construct a linear equation system A w= b, where A is the coefficient matrix, w is the voxel value vector, and b is the constant term on the right. The implicit function value is obtained by using the multigrid method. The zero level set is the reconstructed surface. The final surface reconstructed appliance model is as follows: Figure 6 shown.

[0082] Step 4: Use Boolean difference operation to remove the interference area between the appliance point cloud and the upper and lower jaws to generate a personalized appliance model.

[0083] In this example, the model output by the point cloud completion network is too smooth in the jawline area, making it difficult to show a clear contour. Therefore, it is necessary to further optimize the jawline morphology by combining traditional geometric processing methods to achieve jawline data enhancement. The specific steps are as follows: 1) Apply the AlphaWrap algorithm to the initial cropped model to generate an expanded model, denoted as ModelB, as follows Figure 7 ; 2) The appliance model after surface reconstruction is recorded as ModelA, such as Figure 6 ; 3) Perform Boolean operations on ModelA and ModelB to obtain a fusion model, resample the fusion model and reconstruct the surface to obtain an optimized appliance model with a clear mandibular line, such as Figure 8 shown.

[0084] In this embodiment, the core logic of the Boolean operation is: Operation type: Typically a union of "ModelA∪ModelB" or a difference of "ModelA trims ModelB", depending on the optimization objective: If ModelA's jawline is too smooth, ModelB's clear jawline contour is "carved" onto ModelA by union fusion with ModelB; If Model A interferes with the mandibular teeth, the conflicting areas are eliminated through the subtraction operation to ensure the spatial adaptability of the jaw pad when worn. For example, the optimized orthodontic appliance model is subjected to a Boolean subtraction operation with the mandibular model that has been processed for undercuts to remove the overlapping areas; then, the resulting model is subjected to a Boolean subtraction operation with the maxillary model to further eliminate the interference areas with the maxillary teeth, and finally the orthodontic appliance model that meets the occlusal relationship is obtained, such as Figure 9 .

[0085] In summary, using the anatomical model (Model B) as a benchmark, algorithm-driven geometric fusion and reconstruction organically combine the "functional jaw pad" with the "personalized mandibular contour," ensuring both effective occlusal function and improved wearer comfort. The functional and anatomical models are fused through Boolean operations, combined with resampling and resurfacing, to achieve precise optimization of the mandibular line.

[0086] For ease of understanding, the professional terms of the present invention are explained as follows: 1. Minimum Spanning Tree Traversal (MST Traversal): Constructs a minimum spanning tree (a tree structure connecting all points with minimal edge weights) within a point cloud neighborhood. Traversing the tree ensures consistent propagation of normals or features. During PCA post-processing, if the normals of the neighborhood point cloud are ambiguous (e.g., due to noise causing normal flipping), MST traversal propagates the correct normals based on distance weights, avoiding sudden changes in local features.

[0087] 2. Gradient Field Approximation: Construct a virtual field from the point cloud coordinates p and the normal vector n , simulating the gradient distribution of the physical field as the source term of the Poisson equation. This field characterizes the local "directional trend" of the point cloud. Regions with high normal consistency have smooth field distributions, making it easier to solve the Poisson equation for continuous surfaces.

[0088] 3. Resampling: Resampling the fused point cloud to redistribute the sampling density to even out the point distribution and eliminate sparse or clustered areas that may appear after Boolean operations.

[0089] 4. Poisson Reconstruction: By solving the Poisson equation Generate an implicit surface from a point cloud, where χ is the source term of the signed distance function and the surface gradient is constrained by the point cloud normal. After weighting, the equation becomes , so that the normal consistent area contributes more to the surface shape and suppresses the influence of noise points.

[0090] 5. Voxelization: Divide the three-dimensional space into a cubic grid (voxel) with a side length of s, and use the centroid of the internal point cloud for each voxel. Representation, realize point cloud dimensionality reduction and structured processing.

[0091] 6. AlphaWrap algorithm: An algorithm used to process three-dimensional geometric data. It is mainly used to convert the input three-dimensional geometric data (such as triangular meshes, triangle chains or point sets) into a valid triangular surface mesh that strictly contains the input and has the characteristics of watertightness, no intersection and two-dimensional manifold.

[0092] 7. Occlusal Pressure Direction: The direction of the force lines perpendicular to the occlusal plane generated when the upper and lower teeth occlude. In the appliance point cloud, it corresponds to the minimum variance direction (thickness direction) obtained by PCA analysis.

[0093] 8. Anatomical definition of the mandibular line: The outline of the lower edge of the mandible is the key reference for the design of the jaw pad edge. It must remain clear after reconstruction to ensure the edge fit during wear.

[0094] 9. Boolean Operation: A Boolean operation is a method of generating new models by performing logical operations on two or more geometric models. It is mainly divided into three types: union, intersection, and difference.

[0095] Example 2 Based on the same concept, the present invention also proposes a two-stage automatic jaw pad generation device based on tooth edge enhancement, comprising: The preprocessing module performs geometric preprocessing on the input upper and lower jaw point clouds to obtain standardized point clouds; The point cloud completion module inputs the standardized point cloud into the point cloud completion network and generates the appliance point cloud through a coarse-to-fine two-stage structure; The reconstruction module performs surface reconstruction and geometric optimization on the appliance point cloud, integrating the original contour to enhance the mandibular line; The output module removes the interference area between the appliance point cloud and the upper and lower jaws through Boolean difference operation to generate a personalized appliance model.

[0096] Example 3 This embodiment also provides an electronic device, referring to Figure 10 , includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program, and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0097] Specifically, the processor 402 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiment of the present invention.

[0098] Memory 404 may include a large-capacity memory 404 for data or instructions. By way of example, and not limitation, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to the data processing device. In certain embodiments, memory 404 is non-volatile memory. In certain embodiments, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. In appropriate circumstances, the RAM may be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM may be a fast page mode dynamic random access memory 404 (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0099] The memory 404 may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor 402 .

[0100] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any one of the two-stage automatic appliance generation methods based on tooth edge enhancement in the above embodiments.

[0101] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408 , wherein the transmission device 406 is connected to the processor 402 , and the input / output device 408 is connected to the processor 402 .

[0102] Transmission device 406 can be used to receive or transmit data via a network. Specific examples of such networks may include wired or wireless networks provided by the electronic device's communications provider. In one embodiment, the transmission device includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 406 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0103] The input / output device 408 is used to input or output information.

[0104] Example 4 This embodiment also provides a readable storage medium, in which a computer program is stored. The computer program includes a program code for controlling a process to execute a process, and the process includes a two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to the first embodiment.

[0105] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be repeated here.

[0106] In general, various embodiments may be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention may be implemented in hardware, while other aspects may be implemented in firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that, as non-limiting examples, the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0107] The embodiments of the present invention may be implemented by computer software that is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets and / or macros may be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer executable components that are configured to perform an embodiment when the program is run. One or more computer executable components may be at least one software code or a portion thereof. In addition, it should be noted at this point that, for example, Figure 1 Any block of the logic flow in the program may represent program steps, or interconnected logic circuits, blocks, and functions, or a combination of program steps and logic circuits, blocks, and functions. The software may be stored on physical media such as memory chips or memory blocks implemented within the processor, magnetic media such as hard disk or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. Physical media are non-transitory media.

[0108] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] The above embodiments merely illustrate several embodiments of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of the present invention. Therefore, the scope of the present invention shall be determined by the appended claims.

Claims

1. A two-stage automatic orthodontic appliance generation method based on tooth edge enhancement, characterized in that: The following steps are involved: Perform geometric preprocessing on the input maxillary and mandibular point clouds to obtain standardized point clouds; Inputting the standardized point cloud into a point cloud completion network, generating an appliance point cloud through a coarse-to-fine two-stage structure; Performing surface reconstruction and geometric optimization on the appliance point cloud, fusing the original contour to enhance the mandibular line; The interference areas between the appliance point cloud and the upper and lower jaws are removed by Boolean difference operation to generate a personalized appliance model.

2. The two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to claim 1, characterized in that: The geometric preprocessing includes: Calculate the centroid of the input maxillary and mandibular point clouds and translate them to the origin; Scale the point cloud to fit within the unit sphere; Align the main directions of the point cloud through principal component analysis; Downsample the point cloud.

3. The two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to claim 1, characterized in that: The coarse-to-fine two-stage structure of the point cloud completion network includes: The first stage: extract point cloud block-level features through farthest sampling and dynamic graph convolution, combine position encoding and input Transformer encoder to generate global features, and generate initial coarse point cloud through MLP; The second stage: the initial coarse point cloud is divided into high-confidence point clouds through the point filtering module, and then merged with the original point cloud to generate a fine point cloud through FoldingNet.

4. The two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to claim 3, characterized in that: The point filtering module divides the initial coarse point cloud into a high confidence set and a low confidence set by calculating the average self-attention score of the point cloud, and takes the high confidence set as the filtered point cloud.

5. The two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to claim 1, characterized in that: After generating the appliance point cloud, the method further includes: A principal component analysis is performed on the appliance point cloud, the point cloud is segmented along the principal axis direction with the minimum variance, and the upper half of the point cloud is offset along the direction to generate an enhanced sample of the occlusal contact area.

6. The two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to claim 1, characterized in that: The surface reconstruction and geometry optimization include: Adaptively denoising the appliance point cloud based on a joint scoring function of local density, curvature, and normal consistency; Estimating the normal vector of the denoised point cloud; Poisson surface reconstruction is performed based on the normal vector consistency weights.

7. A two-stage automatic orthodontic appliance generation method based on tooth edge enhancement according to any one of claims 1 to 6, characterized in that: The fusion of the original contour to enhance the jawline includes: Apply the AlphaWrap algorithm to the initial cropped model to generate the expanded model; Performing Boolean operations on the expanded model and the surface-reconstructed appliance model, and obtaining an optimized appliance model after resampling; The Boolean difference operation includes: First, the optimized appliance model was subjected to a Boolean subtraction operation with the undercut mandibular model, and then with the maxillary model to remove the interference area.

8. A two-stage automatic jaw pad generation device based on tooth edge enhancement, characterized in that: include: The preprocessing module performs geometric preprocessing on the input upper and lower jaw point clouds to obtain standardized point clouds; The point cloud completion module inputs the standardized point cloud into the point cloud completion network and generates the appliance point cloud through a coarse-to-fine two-stage structure; The reconstruction module performs surface reconstruction and geometric optimization on the appliance point cloud, integrating the original contour to enhance the mandibular line; The output module removes the interference area between the appliance point cloud and the upper and lower jaws through Boolean difference operation to generate a personalized appliance model.

9. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the two-stage automatic appliance generation method based on tooth edge enhancement according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes a program code for controlling a process to execute a process, wherein the process includes the two-stage automatic appliance generation method based on tooth edge enhancement according to any one of claims 1 to 7.

Images