A method and system for segmenting three-dimensional tooth and jaw models

Abstract

The present invention provides a method and system for segmenting a three-dimensional model of a tooth and jaw, wherein the method comprises obtaining manual features of each vertex in the three-dimensional model of a tooth and jaw; splicing the manual features of each vertex with Cartesian coordinates to obtain input features of each vertex; inputting each vertex input feature into a deep learning point cloud segmentation network in turn to obtain a tooth position label of each vertex; mapping the tooth position label of each vertex to the three-dimensional model of the tooth and jaw in turn to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums. The present invention integrates and uses a variety of lightweight manual features. Compared with algorithms that only use normal and curvature information, it can better capture the multi-scale and multi-level local features of the tooth and jaw model, and can play a significant role in improving the accuracy of the tooth and jaw segmentation task; on the basis of manual features, deep learning is used to further extract the features of the three-dimensional tooth and jaw model, which reduces the network learning cost and reduces the dependence of the deep network on the number of training samples and the training cycle.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a method and system for segmenting a three-dimensional tooth and jaw model. Background Art

[0002] Computer-aided design (CAD) is now widely used in the field of dentistry, such as orthodontic diagnosis, tooth restoration, tooth realignment, etc. The segmentation of the three-dimensional model of the jaw is the basis of dental CAD, and accurate tooth segmentation results play a key role in subsequent analysis and design. At present, the segmentation algorithm of the three-dimensional model of the jaw can be divided into traditional segmentation algorithm and deep learning-based segmentation algorithm according to whether pure manual features are used.

[0003] Traditional dental model segmentation algorithms generally rely on manually constructed features, such as normal vectors, curvature, etc. Typical algorithms include curvature-based algorithms, contour-based algorithms, projection-based algorithms, harmonic field-based algorithms, etc. The design of manual features has a great influence on the final segmentation accuracy. In the boundary area where the curvature is not obvious, the traditional segmentation algorithm often shows poor segmentation accuracy due to the similarity of manual features. In addition, the high requirement for the operator's professional experience and the time-consuming calculation of high-dimensional features are also common problems of the current traditional segmentation algorithms. The overall segmentation effect is lower than the later emerging deep learning-based methods.

[0004] At present, the deep learning-based three-dimensional model segmentation algorithm of the dental jaw tends to use the original point cloud coordinates and normal features. Different from the manually constructed features that the traditional segmentation algorithm relies on, the deep learning-based method generally realizes the extraction of local features by establishing multi-layer neighborhoods and feature aggregation. The typical feature extraction algorithm is PointNet++, and many subsequent studies are carried out on this basis. The deep learning-based method is significantly stronger than the manually designed features in feature extraction ability, and can usually achieve higher index values ​​in segmentation accuracy. However, the effectiveness of such methods currently usually depends on the scale of manually annotated data sets, and the acquisition of three-dimensional image data is difficult and the annotation cost is high, which often makes it difficult to meet the data set scale requirements of deep learning. In contrast, the rule-based artificial features used by traditional segmentation algorithms have natural universality, do not rely on large-scale data sets, and have clearer geometric meanings and better interpretability than the features extracted by deep learning. The existing algorithms based on artificial features and deep learning have their own advantages and disadvantages, and have not yet been well combined. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for segmenting a three-dimensional tooth and jaw model.

[0006] In a first aspect, the present invention provides a method for segmenting a three-dimensional tooth and jaw model, comprising:

[0007] Obtain manual features of each vertex in the three-dimensional model of the dental jaw;

[0008] The manual features of each vertex are concatenated with the Cartesian coordinates to obtain the input features of each vertex;

[0009] Input the input features of each vertex into the deep learning point cloud segmentation network in turn to obtain the tooth position label of each vertex;

[0010] The tooth position label of each vertex is mapped to the three-dimensional model of the jaw in turn to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums.

[0011] Furthermore, the step of obtaining the manual features of each vertex in the three-dimensional model of the tooth and jaw includes:

[0012] The Laplace coordinates of each vertex are calculated according to the following formula:

[0013]

[0014] Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood;

[0015] Get the normal vector of each vertex;

[0016] Get the vertex principal curvature, Gaussian curvature and mean curvature;

[0017] Get the RSD feature of each vertex;

[0018] Get the FPFH features for each vertex.

[0019] Furthermore, the manual features of each vertex are concatenated with the Cartesian coordinates to obtain the input features of each vertex, including:

[0020] The input features of each vertex are calculated according to the following formula:

[0021]

[0022] Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M iis the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHi is the FPFH feature of the i-th vertex.

[0023] Furthermore, the input features of each vertex are sequentially input into the deep learning point cloud segmentation network to obtain the tooth position label of each vertex, including:

[0024] The input features of each vertex are sequentially input into the three feature extraction layers of the deep learning point cloud segmentation network. In each feature extraction layer, the farthest point sampling algorithm is used for downsampling. The neighborhood of the downsampling point is determined using the spherical query algorithm with different radius parameters. The PointNet algorithm is used to extract local features from the points in each neighborhood. The features extracted from the points in the neighborhoods of different radii are concatenated as the first output feature of the current feature extraction layer.

[0025] The first output feature of each vertex is sequentially input into the three feature propagation layers of the deep learning point cloud segmentation network. The mth feature propagation layer concatenates the output of the previous layer and the input of the 4th to mth feature extraction layers as the input of the current feature propagation layer. In each feature propagation layer, the feature is diffused from the downsampled point to the surrounding points through trilinear interpolation, and the feature dimension is reduced using the MLP algorithm as the second output feature of the current feature propagation layer.

[0026] The second output feature of each vertex is input into two MLP networks so that the number of output feature channels is equal to the number of semantic categories. The upper and lower gums are regarded as two semantic categories respectively. The reduced-dimensional features are input into the Softmax function to obtain the probability that the current point belongs to different semantic categories. The category corresponding to the maximum probability is taken as the tooth position label.

[0027] In a second aspect, the present invention provides a dental three-dimensional model segmentation system, comprising:

[0028] An acquisition module is used to obtain the manual features of each vertex in the three-dimensional model of the dental jaw;

[0029] The splicing module is used to splice the manual features of each vertex with the Cartesian coordinates to obtain the input features of each vertex;

[0030] The input module is used to input the input features of each vertex into the deep learning point cloud segmentation network in sequence to obtain the tooth position label of each vertex;

[0031] The mapping module is used to map the tooth position label of each vertex to the three-dimensional model of the jaw in sequence to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums.

[0032] Furthermore, the acquisition module includes:

[0033] The first calculation unit is used to calculate the Laplace coordinates of each vertex according to the following formula:

[0034]

[0035] Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood;

[0036] A first acquisition unit, used to acquire a normal vector of each vertex;

[0037] A second acquisition unit is used to acquire vertex principal curvature, Gaussian curvature and mean curvature;

[0038] A third acquisition unit is used to acquire the RSD feature of each vertex;

[0039] The fourth acquisition unit is used to acquire the FPFH feature of each vertex.

[0040] Furthermore, the splicing module includes:

[0041] The second calculation unit is used to calculate the input features of each vertex according to the following formula:

[0042]

[0043] Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M i is the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHiis the FPFH feature of the i-th vertex.

[0044] The present invention provides a method and system for segmenting a three-dimensional model of a tooth and jaw, wherein the method includes obtaining manual features of each vertex in the three-dimensional model of a tooth and jaw; splicing the manual features of each vertex with Cartesian coordinates to obtain input features of each vertex; inputting the input features of each vertex into a deep learning point cloud segmentation network in sequence to obtain a tooth position label of each vertex; mapping the tooth position label of each vertex to the three-dimensional model of the tooth and jaw in sequence to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums. The present invention adopts the above scheme, integrates and uses a variety of lightweight manual features, and can better capture the multi-scale and multi-level local features of the tooth and jaw model compared to the algorithm that only uses normal and curvature information, and can play a significant role in improving the accuracy of the tooth and jaw segmentation task; directly introduces several types of manual features into the deep learning framework, and uses deep learning to further extract the features of the three-dimensional tooth and jaw model based on the manual features, thereby reducing the learning cost of the network and reducing the dependence of the deep network on the number of training samples and the training cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A workflow diagram of a method for segmenting a three-dimensional tooth and jaw model provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the structure of a dental three-dimensional model segmentation system provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the structure of a method for segmenting a three-dimensional tooth and jaw model provided by an embodiment of the present invention;

[0049] Figure 4 A segmentation network framework diagram based on deep learning in a method for segmenting a three-dimensional tooth and jaw model provided in an embodiment of the present invention;

[0050] Figure 5 An upper and lower jaw model diagram in a method for segmenting a three-dimensional tooth and jaw model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] As described in the background technology, the existing algorithms based on artificial features and deep learning have their own advantages and disadvantages, and have not been well combined. Therefore, in order to solve the above problems, the embodiment of the present invention provides a method for segmenting a three-dimensional model of a tooth and jaw, such as Figure 3 As shown, Figure 3 Schematic diagram of the structure of the dental three-dimensional model segmentation method.

[0053] Specifically, Figure 1 As shown, the segmentation method includes:

[0054] Step S101, obtaining the manual features of each vertex in the three-dimensional model of the tooth and jaw.

[0055] In this step, the three-dimensional model of the dental jaw usually contains tens to hundreds of thousands of vertices and triangular faces. According to the vertex coordinates contained in the model and the adjacency information provided by the triangular faces, various manual features can be extracted to characterize the local or global information of the model.

[0056] The Laplace coordinates of each vertex are calculated according to the following formula:

[0057]

[0058] Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, which is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood.

[0059] Get the normal vector of each vertex, which can be calculated by cross-producting any two sides of the triangle and normalizing them. Then traverse all vertices in turn, find all triangles adjacent to the target vertex, and use the average normal vector of all triangles adjacent to the target vertex as the normal vector of the target vertex.

[0060] Get the vertex principal curvature, Gaussian curvature and mean curvature. Assume N i In a model composed of triangles as the smallest unit, vertex v iNeighborhood, A(v i ) represents the vertex v i The area of ​​the local average region, f(v i ) is a function defined by a model composed of triangles as the smallest unit at vertex v i The value at . Then the cotangent discretization form of the Laplace-Beltrami operator is:

[0061]

[0062] Among them, v j ∈N i (v i ) represents the vertex v i All vertices of the domain; α ij and β ij They are the neighboring triangles with v j -v i The two triangles with the same edge as v j -v i Two opposite corners.

[0063] The calculation formula for the average curvature of the i-th vertex is:

[0064] Assume θ j For vertex v i The neighboring triangle patch is composed of vertex v i is the angle of the vertex, then the Gaussian curvature of the i-th vertex is: The two principal curvatures can be calculated from the mean curvature and the Gaussian curvature:

[0065]

[0066] Among them, k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex.

[0067] Get the RSD feature of each vertex. Traverse all points in the neighborhood, calculate the distance d between all points in the neighborhood and the difference α between the normal vectors between points, and estimate the radius r of the circle where the two points are located based on the Taylor expansion of the cosine theorem. The simplified calculation formula is: Vertex v i The RSD characteristic of vertex v i The maximum and minimum values ​​of the radius r among all pairs of points in the neighborhood, that is, F RSDi =[r max ,r min ].

[0068] Get the FPFH feature of each vertex. First calculate the normal vectors of all points, and then define point p t and point p s The local coordinate system (u, v, w) of the point pair is as follows:

[0069]

[0070] Among them, × represents the outer product; n s For point p s The normal vector of the point; u, v, w are the basis vectors of the local coordinate system of the point.

[0071] Based on the local coordinate system of the point pair, construct the eigenvectors α, φ, θ:

[0072]

[0073] Among them, · is the inner product; n t For point p t The normal vector of point p t and point p s The Euclidean distance.

[0074] The SPFH (Simplified Point Feature Histogram) feature is defined as: placing the <α, φ, θ> feature set of the point pairs in the neighborhood into three corresponding histograms, counting the number of votes, and the number of histogram intervals is n. The SPFH feature vector has 3n dimensions. If n is 11 here, the corresponding SPFH feature vector has 33 dimensions.

[0075] Define vertex v i The neighborhood of consists of two parts: one part is the vertex v i The point pair consisting of the surrounding k points, and the other part is the vertex v i Each neighboring point v k The statistics of the second part are weighted averaged to get the final F FPFHi Calculation method:

[0076]

[0077] Among them, the weight ω k Represents the vertex and its neighboring points v k distance.

[0078] Step S102, concatenate the manual features of each vertex with the Cartesian coordinates to obtain the input features of each vertex.

[0079] In this step, the input features of each vertex are calculated according to the following formula:

[0080]

[0081] Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M i is the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHi is the FPFH feature of the i-th vertex.

[0082] The Cartesian coordinates v i There are 3-dimensional features, Laplace coordinates δ i There are 3-dimensional features, the normal vector M i There are 3-dimensional features, the principal curvature k 1i and k 2i , Gaussian curvature K i and the mean curvature H i Both are 1-dimensional features, RSD features F RSDi There are 2 dimensions, FPFH features F FPFHi has 33 dimensions, so F in A total of 48 features.

[0083] Step S103, input the input features of each vertex into the deep learning point cloud segmentation network in sequence to obtain the tooth position label of each vertex.

[0084] In this step, the input features of each vertex are sequentially input into the three feature extraction layers of the deep learning point cloud segmentation network. In each feature extraction layer, the farthest point sampling algorithm is used for downsampling. The neighborhood of the downsampling point is determined using the spherical query algorithm with different radius parameters. The PointNet algorithm is used to extract local features from the points in each neighborhood. The features extracted from the points in the neighborhoods of different radii are concatenated as the first output feature of the current feature extraction layer.

[0085] The first output feature of each vertex is sequentially input into the three feature propagation layers of the deep learning point cloud segmentation network. The mth feature propagation layer concatenates the output of the previous layer and the input of the 4th to mth feature extraction layers as the input of the current feature propagation layer. In each feature propagation layer, the feature is diffused from the downsampled point to the surrounding points through trilinear interpolation, and the feature dimension is reduced using the MLP algorithm as the second output feature of the current feature propagation layer.

[0086] The second output feature of each vertex is input into two MLP networks so that the number of output feature channels is equal to the number of semantic categories. The upper and lower gums are regarded as two semantic categories respectively. The reduced-dimensional features are input into the Softmax function to obtain the probability that the current point belongs to different semantic categories. The category corresponding to the maximum probability is taken as the tooth position label.

[0087] like Figure 4 As shown in 12, since the model contains N nodes, each node has 48-dimensional features, the network input can be composed of a matrix of N×48. For all input points, the calculated node features F in Input into three feature extraction layers SA1-SA3 in sequence, namely Figure 4 6-8 of them.

[0088] Specifically, each feature extraction layer SA includes the following steps:

[0089] Use the farthest point sampling algorithm to downsample the input points to npoint points, such as Figure 4 As shown in 20, SA1 is downsampled to 512 points. The neighborhood of the downsampled points is determined using a spherical query algorithm with different radius parameters, such as Figure 4 As shown in 14-15, R=0.1 means that the radius of the spherical neighborhood is 0.1, and S=32 means searching for 32 points in the spherical neighborhood.

[0090] Furthermore, the PointNet algorithm is used to extract local features of the points in each neighborhood, including multiple cascaded MLP-BatchNorm-ReLU layers. Figure 4 17 represents the MLP layer, which is used to transform the dimension of the input features, such as Figure 4 As shown in Figure 13, this layer changes the input feature dimension to 32, and the same applies to other MLP layers. Figure 4 18 in the middle represents BatchNorm, Figure 4 19 represents the activation function ReLU.

[0091] Furthermore, the features extracted from points in neighborhoods of different radii are concatenated as the first output feature of the current feature extraction layer, and the dimension of the first output feature is determined by the last MLP layer of the feature extraction branch in the current neighborhood.

[0092] Specifically, each feature propagation layer FP includes the following steps:

[0093] The mth feature propagation layer concatenates the output of the previous layer and the input of the 4th to mth feature extraction layers as the input of the current feature propagation layer, such as Figure 4 As shown in 16.

[0094] Furthermore, in each feature propagation layer, the features are diffused from the downsampled points to the surrounding points through trilinear interpolation, that is, upsampled to npoint points.

[0095] Furthermore, a multi-layer cascaded MLP-BatchNorm-ReLU layer is used to reduce the feature dimension as the second output feature of the current feature propagation layer.

[0096] The second output feature of each vertex is sequentially input into the two-layer MLP network, so that the number of output feature channels is equal to the number of semantic categories. The third molars of the upper and lower jaws are not considered here, and the upper and lower gums are regarded as two semantic categories respectively, so the total number of semantic categories is 30, including the maxillary gum area, that is, Figure 5 The 22 in the figure corresponds to the semantic number 0; the teeth 11-17 in the upper right area, i.e. Figure 5 The 23-29 in the figure correspond to the semantic numbers 1-7; the teeth 21-27 in the upper left area, i.e. Figure 5 30-36 in the figure correspond to semantic numbers 8-14; the mandibular gingival area, i.e. Figure 5 The 37 in the figure corresponds to the semantic number 15; the teeth 31-37 in the lower left area, i.e. Figure 5 The 38-44 in the figure correspond to the semantic numbers 16-22; the teeth 41-47 in the lower right area, i.e. Figure 5 45-51 in the table correspond to semantic numbers 23-29.

[0097] Furthermore, the reduced-dimensional features are fed into the Softmax function to obtain the probability that the current point belongs to different semantic categories, that is, Figure 4 The category corresponding to the maximum probability is taken as the semantic category to which the current point belongs, that is, the tooth position label.

[0098] Step S104, mapping the tooth position label of each vertex to the three-dimensional model of the jaw in sequence, and obtaining the boundary line between two adjacent teeth and the boundary line between the teeth and the gums. Figure 3 As shown in Figure 5.

[0099] The present invention integrates and uses a variety of lightweight manual features. Compared with the algorithm that only uses normal and curvature information, it can better capture the multi-scale and multi-level local features of the dental model, and can play a significant role in improving the accuracy of the dental segmentation task; several types of manual features are directly introduced into the deep learning framework, and deep learning is used to further extract the three-dimensional dental model features based on the manual features, which reduces the learning cost of the network and reduces the dependence of the deep network on the number of training samples and training cycles.

[0100] like Figure 2 As shown, the embodiment of the present invention also provides a tooth and jaw three-dimensional model segmentation system, including:

[0101] An acquisition module is used to obtain the manual features of each vertex in the three-dimensional model of the dental jaw;

[0102] The splicing module is used to splice the manual features of each vertex with the Cartesian coordinates to obtain the input features of each vertex;

[0103] The input module is used to input the input features of each vertex into the deep learning point cloud segmentation network in sequence to obtain the tooth position label of each vertex;

[0104] The mapping module is used to map the tooth position label of each vertex to the three-dimensional model of the jaw in sequence to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums.

[0105] Furthermore, the acquisition module includes:

[0106] The first calculation unit is used to calculate the Laplace coordinates of each vertex according to the following formula:

[0107]

[0108] Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood;

[0109] A first acquisition unit, used to acquire a normal vector of each vertex;

[0110] A second acquisition unit is used to acquire vertex principal curvature, Gaussian curvature and mean curvature;

[0111] A third acquisition unit is used to acquire the RSD feature of each vertex;

[0112] The fourth acquisition unit is used to acquire the FPFH feature of each vertex.

[0113] Optionally, the splicing module includes:

[0114] The second calculation unit is used to calculate the input features of each vertex according to the following formula:

[0115]

[0116] Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ iis the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M i is the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHi is the FPFH feature of the i-th vertex.

[0117] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.

[0118] The present invention has been described in detail above in conjunction with specific implementations and exemplary examples, but these descriptions cannot be understood as limiting the present invention. Those skilled in the art understand that, without departing from the spirit and scope of the present invention, a variety of equivalent substitutions, modifications or improvements may be made to the technical solution of the present invention and its implementation methods, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be subject to the attached claims.

Claims

1. A method for segmenting a three-dimensional tooth and jaw model, characterized in that: include: Obtain manual features of each vertex in the three-dimensional model of the dental jaw; The manual features of each vertex are concatenated with the Cartesian coordinates to obtain the input features of each vertex; Input the input features of each vertex into the deep learning point cloud segmentation network in turn to obtain the tooth position label of each vertex; The tooth position label of each vertex is mapped to the three-dimensional model of the jaw in turn to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums; The input features of each vertex are sequentially input into the deep learning point cloud segmentation network to obtain the tooth position label of each vertex, including: The input features of each vertex are sequentially input into the three feature extraction layers of the deep learning point cloud segmentation network. In each feature extraction layer, the farthest point sampling algorithm is used for downsampling. The neighborhood of the downsampling point is determined using the spherical query algorithm with different radius parameters. The PointNet algorithm is used to extract local features from the points in each neighborhood. The features extracted from the points in the neighborhoods of different radii are concatenated as the first output feature of the current feature extraction layer. The first output feature of each vertex is sequentially input into the three feature propagation layers of the deep learning point cloud segmentation network. The mth feature propagation layer concatenates the output of the previous layer and the input of the 4th to mth feature extraction layers as the input of the current feature propagation layer. In each feature propagation layer, the feature is diffused from the downsampled point to the surrounding points through trilinear interpolation, and the feature dimension is reduced using the MLP algorithm as the second output feature of the current feature propagation layer. The second output feature of each vertex is input into two MLP networks so that the number of output feature channels is equal to the number of semantic categories. The upper and lower gums are regarded as two semantic categories respectively. The reduced-dimensional features are input into the Softmax function to obtain the probability that the current point belongs to different semantic categories. The category corresponding to the maximum probability is taken as the tooth position label.

2. The method for segmenting a three-dimensional tooth-jaw model according to claim 1, characterized in that: The method of obtaining the manual features of each vertex in the three-dimensional model of the tooth and jaw includes: The Laplace coordinates of each vertex are calculated according to the following formula: Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood; Get the normal vector of each vertex; Get the vertex principal curvature, Gaussian curvature and mean curvature; Get the RSD feature of each vertex; Get the FPFH features for each vertex.

3. The method for segmenting a three-dimensional tooth-jaw model according to claim 2, characterized in that: The manual features of each vertex are concatenated with the Cartesian coordinates to obtain the input features of each vertex, including: The input features of each vertex are calculated according to the following formula: Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M i is the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHi is the FPFH feature of the i-th vertex.

4. A dental three-dimensional model segmentation system, characterized in that: include: An acquisition module is used to obtain the manual features of each vertex in the three-dimensional model of the dental jaw; The splicing module is used to splice the manual features of each vertex with the Cartesian coordinates to obtain the input features of each vertex; The input module is used to input the input features of each vertex into the deep learning point cloud segmentation network in sequence to obtain the tooth position label of each vertex; A mapping module is used to map the tooth position label of each vertex to the three-dimensional model of the jaw in sequence to obtain the boundary lines of two adjacent teeth and the boundary lines of the teeth and gums; The input features of each vertex are sequentially input into the deep learning point cloud segmentation network to obtain the tooth position label of each vertex, including: The input features of each vertex are sequentially input into the three feature extraction layers of the deep learning point cloud segmentation network. In each feature extraction layer, the farthest point sampling algorithm is used for downsampling. The neighborhood of the downsampling point is determined using the spherical query algorithm with different radius parameters. The PointNet algorithm is used to extract local features from the points in each neighborhood. The features extracted from the points in the neighborhoods of different radii are concatenated as the first output feature of the current feature extraction layer. The first output feature of each vertex is sequentially input into the three feature propagation layers of the deep learning point cloud segmentation network. The mth feature propagation layer concatenates the output of the previous layer and the input of the 4th to mth feature extraction layers as the input of the current feature propagation layer. In each feature propagation layer, the feature is diffused from the downsampled point to the surrounding points through trilinear interpolation, and the feature dimension is reduced using the MLP algorithm as the second output feature of the current feature propagation layer. The second output feature of each vertex is input into two MLP networks so that the number of output feature channels is equal to the number of semantic categories. The upper and lower gums are regarded as two semantic categories respectively. The reduced-dimensional features are input into the Softmax function to obtain the probability that the current point belongs to different semantic categories. The category corresponding to the maximum probability is taken as the tooth position label.

5. The dental three-dimensional model segmentation system according to claim 4, characterized in that: The acquisition module comprises: The first calculation unit is used to calculate the Laplace coordinates of each vertex according to the following formula: Among them, δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; v i is the i-th vertex in the three-dimensional model of the jaw; d i is the number of vertices directly connected to the i-th vertex; v j is the jth vertex in the three-dimensional model of the jaw; N i ={j(i,j)∈K}, is the set of vertices in the neighborhood of vertex i, and K is the set of all vertices in the neighborhood; A first acquisition unit, used to acquire a normal vector of each vertex; A second acquisition unit is used to acquire vertex principal curvature, Gaussian curvature and mean curvature; A third acquisition unit is used to acquire the RSD feature of each vertex; The fourth acquisition unit is used to acquire the FPFH feature of each vertex.

6. The dental three-dimensional model segmentation system according to claim 5, characterized in that: The splicing module comprises: The second calculation unit is used to calculate the input features of each vertex according to the following formula: Among them, F in is the input feature of the i-th vertex; v i The i-th vertex in the three-dimensional model of the jaw; δ i is the Laplace coordinate of the i-th vertex in the three-dimensional model of the jaw; M i is the normal vector of the i-th vertex; k 1i , k 2i are the principal curvatures of the ith vertex; K i is the Gaussian curvature of the ith vertex; H i is the mean curvature of the ith vertex; F RSDi is the RSD feature of the i-th vertex; F FPFHi is the FPFH feature of the i-th vertex.

Images