Deep-learning-based method and system for extracting finish line of full-crown-restoration preparation

Abstract

A deep-learning-based method and system for extracting a finish line of a full-crown-restoration preparation. The method comprises: acquiring three-dimensional-mesh data of a preparation of a subject (S1); inputting the three-dimensional-mesh data of the preparation of the subject into a three-dimensional-mesh semantic segmentation network model, so as to obtain a three-dimensional-mesh segmentation result for the preparation of the subject (S2); preprocessing the three-dimensional-mesh segmentation result for the preparation, and performing smoothing processing on the preprocessed three-dimensional-mesh segmentation result, so as to obtain an accurate three-dimensional-mesh segmentation result (S3); on the basis of the accurate three-dimensional-mesh segmentation result, calculating segmentation edges, so as to obtain a rough finish line of the preparation (S4); and on the basis of the rough finish line of the preparation, performing smoothing processing and projection processing, so as to obtain a smooth finish line of the preparation on a three-dimensional mesh (S5). The method can assist with marking a finish line of a preparation, thereby shortening the time for the design and preparation of a restoration, and also reducing marking errors caused by subjective factors.

Description

A Deep Learning-Based Method and System for Edge Line Extraction of Crown Repair Preparation Technical Field

[0001] This invention relates to the field of full crown restoration technology, and in particular to a method and system for extracting the edge lines of a full crown restoration preparation body based on deep learning. Background Technology

[0002] Full crown restorations, as a method for restoring the morphology and masticatory function of prepared dental implants, are the preferred treatment option for fixed restorations of tooth defects. With the rapid development of digital technology, optical impressions and computer-aided design / manufacturing (CAD / CAM) have become crucial aspects of restoration design and fabrication. Although digital technologies for restoration design have improved the efficiency of restorative treatment to some extent, in the wave of end-to-end digitalization, especially in the face of challenges such as differences in design capabilities among dentists and dental technicians and the rapid increase in the amount of oral scan data, more efficient and stable solutions are still needed.

[0003] Existing methods for extracting the edge line of a full-crown preparation usually rely on the clinical experience of doctors or technicians. However, this method has problems such as high requirements for the experience of doctors and technicians, different subjective judgments among different doctors and technicians, and the need for a lot of manual interaction, which wastes a lot of time, is complicated to operate, and has low accuracy.

[0004] Therefore, there is an urgent need for a deep learning-based method and system for extracting the edge lines of prepared crown restorations, which can assist in annotating the edge lines of the prepared restorations, save restoration design and fabrication time, and reduce annotation errors caused by subjective factors. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a method and system for extracting the edge lines of prepared crown restorations based on deep learning. This method can assist in annotating the edge lines of the prepared restorations, saving time in restoration design and fabrication, while also reducing annotation errors caused by subjective factors.

[0006] This invention provides a deep learning-based method for extracting the edge lines of a full crown restoration preparation body, comprising the following steps:

[0007] S1. Obtain the three-dimensional mesh data of the test subject's prepared body;

[0008] S2. Input the 3D mesh data of the test subject's prepared body into the 3D mesh semantic segmentation network model to obtain the 3D mesh segmentation result of the test subject's prepared body;

[0009] S3. Preprocess the three-dimensional mesh segmentation result of the prepared body, and smooth the preprocessed three-dimensional mesh segmentation result to obtain an accurate three-dimensional mesh segmentation result.

[0010] S4. Based on the accurate 3D mesh segmentation results, calculate the segmentation edges to obtain the rough edge lines of the prepared body;

[0011] S5. Smooth and project the rough edge lines of the prepared object to obtain smooth edge lines of the prepared object on the 3D mesh.

[0012] Furthermore, in S1, acquiring the 3D mesh data of the test subject's prepared body includes:

[0013] Intraoral scanning 3D mesh data of a single abutment tooth preparation from the test subject was obtained.

[0014] Furthermore, in S2, the training process of the 3D mesh semantic segmentation network model is as follows:

[0015] S2.1 Construct a 3D mesh semantic segmentation sample dataset; wherein, the sample dataset includes 3D mesh data of several non-test subject preparatory bodies and segmentation labels corresponding to the 3D mesh data of several non-test subject preparatory bodies;

[0016] S2.2, Construct a neural network model;

[0017] S2.3. Train the neural network model using the 3D mesh semantic segmentation sample dataset until the loss value of the neural network model is less than the preset loss value and the segmentation accuracy of the neural network model is greater than the preset accuracy, thus obtaining the trained 3D mesh semantic segmentation neural network model.

[0018] Furthermore, in S2.1, the process of constructing segmentation labels corresponding to the 3D mesh data of several non-test subjects includes:

[0019] S2.11. Connect the ordered points on the edge line of any non-test subject's preparation body to form a closed curve.

[0020] S2.12. Based on the closed curve, the three-dimensional mesh of the non-test subject's preparation body is divided into a preparation region and a non-preparation region. Each triangular mesh in the preparation region and the non-preparation region is labeled to obtain the segmentation label corresponding to the three-dimensional mesh of the non-test subject's preparation body. The preparation region refers to the region that contains only the tooth body, and the non-preparation region refers to the region that contains other tissues besides the tooth body.

[0021] Furthermore, in S3, the three-dimensional mesh segmentation results of the prepared body are preprocessed, including outlier removal and hole filling.

[0022] The 3D mesh segmentation results of the prepared body are processed by clustering algorithm to remove outliers and fill holes.

[0023] Furthermore, in S3, the preprocessed 3D mesh segmentation results are smoothed to obtain accurate 3D mesh segmentation results, including:

[0024] Based on the preprocessed 3D mesh segmentation results, a smoothing process is performed using a fully connected conditional random field algorithm to obtain accurate 3D mesh segmentation results.

[0025] Furthermore, in S4, based on the accurate 3D mesh segmentation results, the segmentation edges are calculated, and the rough edge lines of the prepared body are obtained, including:

[0026] S41. Based on the accurate 3D mesh segmentation results, obtain all segmentation edge triangle meshes to form a triangle set;

[0027] S42. Based on the vertices of the segmented edge triangle meshes in the triangle set, calculate the centroid of each segmented edge triangle mesh, use the centroids of the segmented edge triangle meshes as rough edge points to form a rough edge line scatter set, and use the edge lines formed by the rough edge points as the rough edge lines of the preparatory body.

[0028] Furthermore, in S5, the rough edge lines of the prepared body are smoothed and projected to obtain the smooth edge lines of the prepared body on the 3D mesh, including:

[0029] S51. Downsample the set of scattered points on the rough edge line into an unordered set of points.

[0030] S52. Sort the unordered set of points to obtain an ordered set of points that form a closed curve by connecting the beginning and end of the unordered set of points.

[0031] S53. Based on the ordered set of points, a smooth point cloud of the pre-body edge line is obtained by smoothing the point cloud through a multinomial regression algorithm.

[0032] S54. Project the smooth edge line point cloud of the preparatory body onto the 3D mesh of the preparatory body to obtain the smooth edge line of the preparatory body on the 3D mesh.

[0033] This invention also provides a deep learning-based edge line extraction system for full crown restoration preparations, used to execute the deep learning-based edge line extraction method for full crown restoration preparations described in any of the above claims. The system includes the following modules:

[0034] The data acquisition module is used to acquire the three-dimensional mesh data of the test subject's prepared body;

[0035] The 3D mesh data segmentation module, connected to the data acquisition module, is used to input the 3D mesh data of the test subject's pre-body into the 3D mesh semantic segmentation network model to obtain the 3D mesh segmentation result of the test subject's pre-body.

[0036] The segmentation result refinement module is connected to the 3D mesh data segmentation module. It is used to preprocess the 3D mesh segmentation results of the prepared body and smooth the preprocessed 3D mesh segmentation results to obtain accurate 3D mesh segmentation results.

[0037] The rough edge line calculation module is connected to the segmentation result refinement module. It is used to calculate the segmentation edge based on the accurate 3D mesh segmentation result and obtain the rough edge line of the prepared body.

[0038] The smooth edge line extraction module, connected to the rough edge line calculation module, is used to perform smoothing and projection processing on the rough edge lines of the prepared body to obtain the smooth edge lines of the prepared body on the 3D mesh.

[0039] The embodiments of the present invention have the following technical effects:

[0040] This invention targets intraoral scanning 3D mesh data of prostheses. It accurately segments the 3D mesh of the prosthesis using a 3D mesh semantic segmentation network model, and performs preprocessing, smoothing, and projection processing on the segmentation results. This achieves accurate, rapid, and fully automated extraction of smooth edge lines from the prosthesis. The extracted smooth edge lines are then used for subsequent crown design and fabrication, ensuring a perfect fit between the fabricated crown and the prosthesis edge. This invention assists in annotating prosthesis edge lines, saving design and fabrication time and reducing annotation errors caused by subjective factors. Attached Figure Description

[0041] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1 is a flowchart of the deep learning-based method for extracting the edge line of a full crown repair preparation provided in an embodiment of the present invention;

[0043] Figure 2 is a schematic diagram of the structure of the three-dimensional mesh semantic segmentation network model provided in an embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of the rough edge line of the preparatory body provided in an embodiment of the present invention;

[0045] Figure 4 is a schematic diagram of the downsampled disordered point set provided in an embodiment of the present invention;

[0046] Figure 5 is a schematic diagram of the smooth edge line of a preparatory body provided in an embodiment of the present invention;

[0047] Figure 6 is a schematic diagram of the smooth edge line of a pre-body on a three-dimensional mesh according to an embodiment of the present invention;

[0048] Figure 7 is a schematic diagram of the edge line extraction system for full crown repair preparation based on deep learning provided in an embodiment of the present invention. Embodiments of the present invention

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0050] This invention proposes a deep learning-based method for extracting the edge lines of a full-crown restoration preparation. Figure 1 is a flowchart of the deep learning-based method for extracting the edge lines of a full-crown restoration preparation provided by an embodiment of this invention. Referring to Figure 1, the method specifically includes:

[0051] S1. Obtain the three-dimensional mesh data of the test subject's prepared body.

[0052] Specifically, the three-dimensional mesh data of the test subject's prepared body includes intraoral scan three-dimensional mesh data of a single abutment tooth prepared body; wherein, a single abutment tooth prepared body includes the portion of a single tooth from the gingiva to the tooth body; the three-dimensional mesh data is a three-dimensional model file, including information such as the vertex number and coordinates of all triangle meshes.

[0053] S2. Input the 3D mesh data of the test subject's prepared body into the 3D mesh semantic segmentation network model to obtain the 3D mesh segmentation result of the test subject's prepared body.

[0054] Specifically, Figure 2 is a schematic diagram of the structure of the three-dimensional mesh semantic segmentation network model provided in an embodiment of the present invention. Referring to Figure 2, the training process of the three-dimensional mesh semantic segmentation network model is as follows:

[0055] S2.1 Construct a 3D mesh semantic segmentation sample dataset; wherein, the sample dataset includes 3D mesh data of several non-testers' pre-built bodies and segmentation labels corresponding to the 3D mesh data of several non-testers' pre-built bodies.

[0056] Furthermore, intraoral scanning 3D mesh data of single abutment tooth preparations from several non-test subjects were obtained, along with corresponding manually annotated preparation edge lines.

[0057] S2.11 Connect the ordered points on the edge line of any non-test subject's preparation body to form a closed curve.

[0058] S2.12. Based on the closed curve, the 3D mesh of the non-test subject's preparation body is divided into a preparation region and a non-preparation region. Each triangular mesh in the preparation region and the non-preparation region is labeled to obtain the segmentation label corresponding to the 3D mesh of the non-test subject's preparation body.

[0059] The preparation area refers to the region containing only the tooth structure, while the non-preparation area refers to the region containing other tissues (such as the gingiva) in addition to the tooth structure. For example, the triangular mesh label in the preparation area is set to 1, and the triangular mesh label in the non-preparation area is set to 0.

[0060] Furthermore, data augmentation is performed on the constructed 3D mesh semantic segmentation sample dataset; the data augmentation includes random flipping, random scaling, random rotation, random jittering, and random discarding.

[0061] S2.2, Construct a neural network model.

[0062] Specifically, Picasso++ is used as the neural network model in this embodiment. The neural network includes an encoder and a decoder. The output channels of the encoder are 32, 64, 96, 128, 192, and 256, respectively, and the output channels of the decoder are 128, 128, 96, 96, and 96, respectively.

[0063] The encoder extracts features mainly through an initial convolutional module, a grid encoder module, and a dual encoder module. The decoder module mainly consists of upsampling, feature concatenation, and 1×1 convolution. Max pooling is used at the encoder for downsampling five times with pooling strides of 4, 3, 3, 2, 2. Upsampling is performed five times at the decoder to restore the size. At each stage, skip connections are used to concatenate the corresponding coding layer features with the upsampled decoding layer features.

[0064] Furthermore, the specific operations involved in the Picasso++ neural network model include:

[0065] ① Use spherical harmonic functions to represent the convolution filter as a continuous function of azimuth and elevation angles.

[0066] The expression for a continuous function is as follows:

[0067] ;

[0068] ;

[0069] Where F represents the convolution filter function, ! represents the factorial, and θ∈[0,2π) is the azimuth angle. ∈[0,π) is the pitch angle, a spherical harmonic function. Let l be the basis function, and let the integers l and m represent the degree and order of the spherical harmonic function, respectively. These are linear coefficients, where, The vector space representing complex numbers, where a is a lm The range of values ​​for is a set of complex numbers; This represents a Legendre polynomial without a Condon–Shortley phase.

[0070] ② Following the expression of spherical harmonic functions, we obtain the features of vertices after passing through the surface-to-point convolution module.

[0071] The formula for calculating the eigenvalues ​​of a vertex is:

[0072] ;

[0073] Among them, g v The characteristics of vertex v, h f Indicate the relevant features of adjacent faces, Let v represent the set of adjacent faces (i.e., triangular faces) of vertex v, and the related characteristics of these adjacent faces are: Where f represents any adjacent face, Indicates the azimuth angle of any adjacent face. It represents the pitch angle of any adjacent plane.

[0074] ③ Following the expression method of spherical harmonic functions, the features of the triangular surface after passing through the point-to-surface convolution module are obtained.

[0075] The formula for calculating the eigenvalues ​​of a triangular face is:

[0076] ;

[0077] Among them, g f Let {h1,h2,h3} represent the characteristics of the triangular face f, and let {h1,h2,h3} represent the characteristics of the three vertices {v1,v2,v3} of the triangular face f.

[0078] ④ Following the expression method of spherical harmonic functions, the features of the triangular face after passing through the face-to-face convolution module are obtained.

[0079] The formula for calculating the eigenvalues ​​of a triangular face is:

[0080] ;

[0081] Where k represents any point within the triangle face, and K represents the texture resolution. , A min and A max α and β represent the minimum and maximum areas of the triangular mesh face, respectively, where A represents the area of ​​the current triangular mesh face, and α and β are hyperparameters. , θ represents the set of non-negative integers. k This represents the azimuth angle of any point within the plane of the triangle. Represents the pitch angle of any point within the plane of the triangle. This represents the set of color vectors for all points on the face f of triangle. This indicates the range of RGB color values.

[0082] ⑤ Calculate the geometric features of the surface.

[0083] For any triangular face in a 3D mesh, its geometric features are represented as [l, θ, n], where l = (l1, l2, l3) represents the side lengths of the three sides of the triangular face, n represents the normal to the triangular face, and θ = (θ1, θ2, θ3) represents the interior angles of the triangular face. The calculation formula is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] Where x1, x2, and x3 represent the three-dimensional coordinate vectors of the three vertices of the triangle face, and <·> represents the inner product operation.

[0088] S2.3. Train the neural network model using the 3D mesh semantic segmentation sample dataset until the loss value of the neural network model is less than the preset loss value and the segmentation accuracy of the neural network model is greater than the preset accuracy, thus obtaining the trained 3D mesh semantic segmentation neural network model.

[0089] Specifically, the training strategies for Picasso++ neural network models include:

[0090] ①The loss function of the Picasso++ neural network is the cross-entropy loss function.

[0091] The loss function formula is:

[0092] ;

[0093] Among them, L CELet p represent the cross-entropy loss function, q represent the predicted value, C represent the number of detected categories, and j represent the j-th detected category. In this embodiment, C=2 detected categories.

[0094] ② During training, the optimizer is Adam, the initial learning rate is set to 0.001, and the weight decay is set to 1×e. -5 The learning rate decreases continuously as the number of training rounds increases, and is dynamically adjusted according to the following formula:

[0095] ;

[0096] Where lr represents the learning rate adjusted in each round of training, lr init This represents the initial learning rate, and t represents the current training epoch. It is the attenuation coefficient.

[0097] Preferably, lr init =0.001, =0.98.

[0098] Furthermore, the training process of the Picasso++ neural network model includes:

[0099] A. Obtain a series of data from the 3D mesh semantic segmentation sample dataset as input to the neural network model; the data includes the coordinates of the 3D mesh vertices, the vertex index of each triangle in the 3D mesh, the number of vertices, and the number of triangles.

[0100] B. Use the series of data obtained in step A to perform feature calculation and spherical harmonic function coefficient calculation; among which, feature calculation mainly includes extracting the geometric features of the surface and calculating the normal vector of the surface.

[0101] C. The initial convolutional layer performs preliminary processing on the input geometric features, extracts low-level features, and maps them to a high-dimensional feature space.

[0102] D. The output features of the initial convolutional layer are input into the encoder. The encoder extracts features layer by layer through a series of convolutional layers and increases the receptive field by max pooling downsampling.

[0103] E. Input the encoder's low-level output features into the decoder, decode them through the decoder module, restore the image size through upsampling, and fuse the output features of the corresponding layer of the encoder using skip connections to fuse shallow and deep information.

[0104] F. Input the decoder output features into the prediction layer to obtain the label of each triangle face of the predicted 3D mesh data, which is used as the final prediction result.

[0105] G. Calculate the loss value between the predicted result and the true label, backpropagate to update the network weights, and repeat the training for 500 rounds to obtain the trained 3D mesh semantic segmentation neural network model.

[0106] S3. Preprocess the three-dimensional mesh segmentation result of the prepared body, and smooth the preprocessed three-dimensional mesh segmentation result to obtain an accurate three-dimensional mesh segmentation result.

[0107] The preprocessing includes outlier removal and hole filling.

[0108] The 3D mesh segmentation results of the prepared body are processed by clustering algorithm to remove outliers and fill holes.

[0109] Specifically, based on the three-dimensional mesh segmentation results of the prepared body, a clustering algorithm based on triangular mesh topology is executed to remove outliers and fill holes. The clustering algorithm based on triangular mesh topology is an algorithm suitable for processing triangular mesh data to find connected clusters. This algorithm is used to cluster connected triangles in the triangular mesh, where connected triangles refer to triangles connected by edges.

[0110] Further steps for outlier removal and hole filling include:

[0111] S31. Based on the three-dimensional mesh segmentation results of the preparatory body, map back to the original triangular mesh, filter out the triangular meshes with the predicted segmentation label of 1, and form a new triangular mesh data.

[0112] S32. For the new triangular mesh data in S31, perform a clustering algorithm to find connected triangle clusters. Select the largest triangle cluster, which is the 3D mesh after outlier removal. Map the index of this 3D mesh back to the original 3D mesh data. Set the triangle label value corresponding to these indices to 1 and the rest of the label values ​​to 0 to obtain the predicted segmentation label after outlier removal.

[0113] S33. Based on the predicted segmentation labels after outlier removal, map them back to the original triangular mesh, filter out the triangular meshes with predicted labels of 0, and form a new triangular mesh data.

[0114] S34. For the new triangular mesh data in S33, perform a clustering algorithm to find connected triangle clusters. Select the largest triangle cluster, which is the 3D mesh after hole filling. Map the index of this 3D mesh back to the original 3D mesh data. Set the triangle label values ​​corresponding to these indices to 0 and the remaining label values ​​to 1 to obtain the predicted segmentation labels after hole filling.

[0115] S35. Based on the preprocessed 3D mesh segmentation results, smoothing is performed using a fully connected conditional random field algorithm to obtain accurate 3D mesh segmentation results.

[0116] Specifically, the core idea of ​​the fully connected conditional random field (CRF) algorithm is to optimize label allocation by maximizing the conditional probability P(X|I), where X is the label set and I represents the input grid data. The fully connected CRF model uses an energy function E(X) to describe the cost of label configuration, and the goal is to find the label configuration that minimizes the energy. The energy function E(X) typically includes a univariate potential energy function and a bivariate potential energy function.

[0117] ;

[0118] in, Let represent the univariate potential energy function, which indicates the value of the observed energy y for the index i of a triangular face. i At that time, the triangle face belongs to category label x i The probability of, where The index i of the triangle face is assigned to the label x. i The probability is given by I, which represents the input grid data, and i and j, which represent the i-th and j-th indices of the triangle face, respectively. The values ​​of i and j range from 1 to N, where N is the total number of triangle faces in the input grid data (i.e., the total number of samples). Let x represent the binary potential energy function. i , x j The specific formula for the binary potential function, representing two distinct triangular faces, is as follows:

[0119] ;

[0120] in, This indicates a label compatibility term, which constrains the conditions for propagation between adjacent triangular faces; f i and f j Let K represent the features of points i and j, respectively. G Represents the Gaussian kernel function. It is a characteristic function. denoted by , where m represents different feature channels or feature types, and each m represents an independent Gaussian kernel function.

[0121] Furthermore, the post-processing steps of the fully connected conditional random field algorithm are as follows:

[0122] The input is the segmentation label after outlier removal and hole filling, i.e., the input is a one-dimensional vector. A fully connected conditional random field model is constructed, where the label confidence of the univariate energy term is set to 0.95 and the compatibility parameter of the binary energy term is set to 1. After 50 iterations, the optimized label is output as a one-dimensional vector.

[0123] S4. Based on the accurate 3D mesh segmentation results, calculate the segmentation edges to obtain the rough edge lines of the prepared body.

[0124] S41. Based on the accurate 3D mesh segmentation results, obtain all segmented edge triangle meshes to form a triangle set.

[0125] Specifically, the labels corresponding to the vertices of the mesh are obtained by segmenting the triangle mesh. Let the original preparatory body mesh be O. For any triangle T in it, its label is u, where u can be 0 or 1. The label u of triangle T is mapped to the three vertices v1, v2, and v3 of the triangle, that is, the vertex label is set to be the same as the triangle label. At this time, the label of each vertex of the original preparatory body 3D mesh data is obtained. Traverse each triangle in the preparatory body 3D mesh data. If the three vertex labels of a triangle are different, it is considered to be a segmentation edge triangle. All segmentation edge triangle meshes are obtained, forming a triangle set C. edge .

[0126] S42. Based on the vertices of the segmented edge triangle meshes in the triangle set, calculate the centroid of each segmented edge triangle mesh. Based on the centroids of the segmented edge triangle meshes, construct a set of scattered points for the rough edge lines, which serves as the rough edge lines of the preparatory body.

[0127] Specifically, Figure 3 is a schematic diagram of the rough edge line of the preparatory body provided in the embodiment of the present invention. Referring to Figure 3, for the triangle set C edge For any cut-edge triangular mesh in the array, assuming its three vertices are v1, v2, and v3, according to the formula... Calculate the centroid of the segmented edge triangular mesh, and use the centroid of the segmented edge triangular mesh as the rough edge points to form a set C of scattered points for the rough edge line. centor The edge line formed by the rough edge points is used as the rough edge line of the preparatory body.

[0128] S5. Smooth and project the rough edge lines of the prepared object to obtain smooth edge lines of the prepared object on the 3D mesh.

[0129] S51. Downsample the set of scattered points on the rough edge line into an unordered set of points.

[0130] Specifically, Figure 4 is a schematic diagram of the downsampled disordered point set provided in an embodiment of the present invention. Referring to Figure 4, the set of scattered points C of the rough edge line is... centor The data is placed into a large voxel grid, where each voxel is a cube, effectively dividing the entire space of the scattered points into several voxels. For any point v... centor (x, y, z) Calculates the voxel index based on the given parameter `voxel_size`. Within each voxel, calculates the center of all points contained within that voxel, which is the sampled point `v`. sample All v sample The set of points after downsampling is called the unordered set of points C. sample .

[0131] S52. Sort the unordered set of points to obtain an ordered set of points that form a closed curve by connecting the beginning and end of the unordered set of points.

[0132] Specifically, the unordered set of points C sample The scattered points in the set are unordered, which is not conducive to operations such as smoothing. The overall shape of the scattered points forms the outline of the preparatory volume's edge line, which is a closed curve. For any point v in the set... 01 As the initial point, use the k-dimensional tree algorithm to search for its nearest neighbor v. 02 Similarly, the k-dimensional tree algorithm is used to search for v. 02 Except v 01 The nearest neighbor v outside 03 As v 03 At the same time, to prevent v 03 and v 01 On the same side, calculate vector (v) 01 -v 02 ) and vector (v 03 -v 02 The angle between ) is greater than 90°, then v 03 Add to the point set; otherwise, continue searching except for v. 01 v 03 Find the nearest neighbor outside of the given area, until the included angle is greater than 90°, and repeat the above steps to find v. 03 And so on, until the nth point v is found. 0n The nearest neighbor and v 01 The distance is used to obtain an ordered set of points that can be connected end to end to form a closed curve.

[0133] S53. Based on the ordered set of points, a smooth point cloud of the pre-body edge line is obtained by smoothing the point cloud through a multinomial regression algorithm.

[0134] Specifically, the sorted set of three-dimensional points includes v 01 v 02 ... v 0nGiven n points, calculate the Euclidean distance between each pair of points and sum them up to obtain a set of distances: {d0=0, d1=|v 02 -v 01 |,d2=d1+|v 03 -v 02 |,...,d n-1 =d n-2 +|v 0n -v 0n-1 The set of points {x1, x2, ..., xn} contains n elements, and the sorted set of points also contains n elements. Each element has three coordinates: x, y, and z. Let {x1, x2, ..., xn} be a set of points with n elements. n}、{y1,y2,...,y n} and {z1,z2,...,z n} respectively with {d0,d1,...,d n-1 Perform a 12th-order polynomial regression to obtain the regression curve expression. Interpolate the regression curve to obtain a smooth preliminary body edge point cloud P.

[0135] S54. Project the smooth edge line point cloud of the preparatory body onto the 3D mesh of the preparatory body to obtain the smooth edge line of the preparatory body on the 3D mesh.

[0136] Specifically, Figure 5 is a schematic diagram of a smooth edge line of a pre-body provided in an embodiment of the present invention, and Figure 6 is a schematic diagram of a smooth edge line of a pre-body on a three-dimensional mesh provided in an embodiment of the present invention. Referring to Figures 5 and 6 (wherein, the three-dimensional mesh of the pre-body in Figure 6 is not rendered), based on the point cloud P of the smooth edge line of the pre-body, traversing the points in the point cloud, for any point v p Find the triangle T in the original pre-body 3D mesh that has the shortest distance to it. near and for point v p To T near Draw perpendicular lines, the intersection of which is the projection point v. Then iterate through all points v in the point cloud P. p The set of projection points v is the smooth edge line of the final prepared body on the 3D mesh.

[0137] This invention targets intraoral scanning 3D mesh data of prostheses. It accurately segments the 3D mesh of the prosthesis using a 3D mesh semantic segmentation network model, and performs preprocessing, smoothing, and projection processing on the segmentation results. This achieves accurate, rapid, and fully automated extraction of smooth edge lines from the prosthesis. The extracted smooth edge lines are then used for subsequent crown design and fabrication, ensuring a perfect fit between the fabricated crown and the prosthesis edge. This invention assists in annotating prosthesis edge lines, saving design and fabrication time and reducing annotation errors caused by subjective factors.

[0138] Figure 7 is a schematic diagram of the deep learning-based edge line extraction system for full crown restoration preparation provided in an embodiment of the present invention. This system is used to execute the deep learning-based edge line extraction method for full crown restoration preparation described in the above embodiment. As shown in Figure 7, the system includes the following modules:

[0139] The data acquisition module is used to acquire the three-dimensional mesh data of the test subject's prepared body;

[0140] The 3D mesh data segmentation module, connected to the data acquisition module, is used to input the 3D mesh data of the test subject's pre-body into the 3D mesh semantic segmentation network model to obtain the 3D mesh segmentation result of the test subject's pre-body.

[0141] The segmentation result refinement module is connected to the 3D mesh data segmentation module. It is used to preprocess the 3D mesh segmentation results of the prepared body and smooth the preprocessed 3D mesh segmentation results to obtain accurate 3D mesh segmentation results.

[0142] The rough edge line calculation module is connected to the segmentation result refinement module. It is used to calculate the segmentation edge based on the accurate 3D mesh segmentation result and obtain the rough edge line of the prepared body.

[0143] The smooth edge line extraction module, connected to the rough edge line calculation module, is used to perform smoothing and projection processing on the rough edge lines of the prepared body to obtain the smooth edge lines of the prepared body on the 3D mesh.

[0144] It should be noted that the terminology used in this invention is for describing specific embodiments only and is not intended to limit the scope of this application. As shown in this specification, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may include the plural. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0145] It should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Unless otherwise expressly specified and limited, the terms "installed," "connected," "linked," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components. For those skilled in the art, the specific meaning of the above terms in the present invention can be understood according to the specific circumstances.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting the edge line of a full crown repair preparation body based on deep learning, characterized in that, Includes the following steps: S1. Obtain the three-dimensional mesh data of the test subject's prepared body; S2. Input the three-dimensional mesh data of the test subject's prepared body into the three-dimensional mesh semantic segmentation network model to obtain the three-dimensional mesh segmentation result of the test subject's prepared body; The three-dimensional mesh segmentation result of the prepared body includes the prepared region and the unprepared region; the prepared region refers to the region containing only the tooth body, and the unprepared region refers to the region containing other tissues besides the tooth body. S3. The three-dimensional mesh segmentation result of the prepared body is preprocessed, and the preprocessed three-dimensional mesh segmentation result is smoothed to obtain an accurate three-dimensional mesh segmentation result. The preprocessing includes outlier removal and hole filling. The three-dimensional mesh segmentation results of the prepared body are processed by clustering algorithm to remove outliers and fill holes, specifically including: S31. Based on the three-dimensional mesh segmentation results of the preparatory body, map back to the original triangular mesh, filter out the triangular mesh of the preparatory area, and form a new triangular mesh data. S32. For the new triangular mesh data in S31, perform a clustering algorithm to find connected triangle clusters, select the largest triangle cluster as the 3D mesh after outlier removal, map the index of the 3D mesh back to the original 3D mesh data, set the triangle label value corresponding to these indices to 1, and set the other label values ​​to 0, to obtain the predicted segmentation label after outlier removal. S33. Based on the predicted segmentation labels after outlier removal, map them back to the original triangular mesh, filter out the triangular meshes with predicted labels of 0, and form a new triangular mesh data. S34. For the new triangular mesh data in S33, perform a clustering algorithm to find connected triangle clusters, select the largest triangle cluster as the three-dimensional mesh after hole filling, map the index of the three-dimensional mesh back to the original three-dimensional mesh data, set the triangle label value corresponding to these indices to 0, and set the other label values ​​to 1 to obtain the predicted segmentation label after hole filling. S35. Based on the preprocessed 3D mesh segmentation results, smoothing is performed using a fully connected conditional random field algorithm to obtain accurate 3D mesh segmentation results. S4. Based on the precise three-dimensional mesh segmentation results, calculate the segmentation edges to obtain the rough edge lines of the prepared body; S41. Based on the precise three-dimensional mesh segmentation results, obtain all segmentation edge triangle meshes to form a triangle set; S42. Based on the vertices of the segmented edge triangle meshes in the triangle set, calculate the centroid of each segmented edge triangle mesh, use the centroids of the segmented edge triangle meshes as rough edge points to form a rough edge line scatter set, and use the edge lines formed by the rough edge points as the rough edge lines of the preparatory body. S5. Smoothing and projection processing are performed on the rough edge lines of the prepared body to obtain smooth edge lines of the prepared body on the three-dimensional mesh.

2. The method for extracting the edge line of a full crown repair preparation body based on deep learning according to claim 1, characterized in that, In step S1, obtaining the three-dimensional mesh data of the test subject's prepared body includes: Intraoral scanning 3D mesh data of a single abutment tooth preparation from the test subject was obtained.

3. The method for extracting the edge line of a full crown repair preparation body based on deep learning according to claim 2, characterized in that, In step S2, the training process of the three-dimensional mesh semantic segmentation network model is as follows: S2.1 Construct a 3D mesh semantic segmentation sample dataset; wherein, the sample dataset includes 3D mesh data of several non-test subjects' preparatory bodies and segmentation labels corresponding to the 3D mesh data of several non-test subjects' preparatory bodies; S2.2, Construct a neural network model; S2.

3. Train the neural network model using the three-dimensional mesh semantic segmentation sample dataset until the loss value of the neural network model is less than the preset loss value and the segmentation accuracy of the neural network model is greater than the preset accuracy, thereby obtaining a trained three-dimensional mesh semantic segmentation neural network model.

4. The method for extracting the edge line of a full crown repair preparation body based on deep learning according to claim 3, characterized in that, In S2.1, the process of constructing segmentation labels corresponding to the 3D mesh data of several non-test subjects includes: S2.

11. Connect the ordered points in the edge line of any of the non-test subject's prepared body to form a closed curve; S2.

12. Based on the closed curve, the three-dimensional mesh of the non-test subject's preparation body is divided into a preparation region and a non-preparation region. Each triangular mesh in the preparation region and the non-preparation region is labeled to obtain the segmentation label corresponding to the three-dimensional mesh of the non-test subject's preparation body. The preparation region refers to the region containing only the tooth body, and the non-preparation region refers to the region containing other tissues besides the tooth body.

5. The method for extracting the edge line of a full crown restoration preparation body based on deep learning according to claim 1, characterized in that, In step S5, smoothing and projection processing are performed on the rough edge lines of the prepared body to obtain smooth edge lines of the prepared body on the three-dimensional mesh, including: S51. Downsample the set of scattered points on the rough edge line into an unordered set of points; S52. Sort the unordered set of points to obtain an ordered set of points that are connected end to end to form a closed curve. S53. Based on the ordered set of points, a smooth point cloud of the pre-body edge line is obtained by smoothing the points using a multinomial regression algorithm. S54. Project the smooth edge line point cloud of the prepared body onto the three-dimensional mesh of the prepared body to obtain the smooth edge line of the prepared body on the three-dimensional mesh.

6. A deep learning-based edge line extraction system for full crown restoration preparations, used to execute the deep learning-based edge line extraction method for full crown restoration preparations as described in any one of claims 1-5, characterized in that, The system includes the following modules: The data acquisition module is used to acquire the three-dimensional mesh data of the test subject's prepared body; A 3D mesh data segmentation module, connected to the data acquisition module, is used to input the 3D mesh data of the test subject's pre-body into the 3D mesh semantic segmentation network model to obtain the 3D mesh segmentation result of the test subject's pre-body. The segmentation result refinement module is connected to the three-dimensional mesh data segmentation module. It is used to preprocess the three-dimensional mesh segmentation result of the prepared body and smooth the preprocessed three-dimensional mesh segmentation result to obtain an accurate three-dimensional mesh segmentation result. The rough edge line calculation module is connected to the segmentation result refinement module and is used to calculate the segmentation edge based on the refined three-dimensional mesh segmentation result to obtain the rough edge line of the prepared body. The smooth edge line extraction module is connected to the rough edge line calculation module and is used to perform smoothing and projection processing on the rough edge line of the pre-body to obtain the smooth edge line of the pre-body on the three-dimensional mesh.

Images