A method and system for dental point cloud completion based on a Transformer encoder

Abstract

The present invention discloses a tooth point cloud completion method and system based on a transformer encoder, comprising: S1, processing an incomplete tooth point cloud, and obtaining a feature matrix #imgabs2# according to the obtained cluster center #imgabs0# and local features #imgabs1#; S2, encoding the cluster center #imgabs3# and the local features #imgabs4# around each key point, mining a potential vector #imgabs5# for describing the missing teeth, and predicting a low-resolution point cloud Y of the missing teeth from the potential vector #imgabs6# c ; S3, the low resolution point cloud Y c and the latent vector #imgabs7# to form the reconstructed feature #imgabs8# from #imgabs9# and Y c Recover high-resolution point cloud Y f , combination Y c and Y f Get the final tooth point cloud Y pred ; S4, by minimizing Y pred and the corresponding gold standard Y for real missing teeth gt The Chev distance between the two is used to complete the training of the tooth completion network; S5, the test data is completed to obtain the complete tooth point cloud, and the indicators are used for evaluation. The effectiveness of this method has been proven through extensive experiments and applied to the point cloud completion of incomplete teeth.

Description

Technical Field

[0001] The present invention relates to medical image processing technology, and in particular, to a method and system for dental point cloud completion based on a transformer encoder. Background Art

[0002] Oral health has received increasing attention, and at the same time, people's demand for professional and high-quality dental care has also become higher. Under this demand, high-precision 3D dental models obtained by intraoral scanners (IOS) play an important role in CAD oral diagnosis. In addition, using artificial intelligence technology represented by deep neural networks to process dental models and solve dental problems has gradually become the mainstream method in digital medical diagnosis. In the field of oral medicine, tooth defect filling is an important issue. We need to predict the missing teeth of patients and complete the dental body for subsequent clinical applications. The method described in this article uses the powerful learning ability of deep neural networks to solve the above problems.

[0003] 3D models have various forms of representation, among which point cloud is the simplest one. However, the properties of disorder, correlation, and permutation invariance of point clouds bring difficulties to neural networks. After PointNet and PointNet++ used deep neural networks to achieve point cloud segmentation and point cloud classification, point cloud deep learning has gradually become a popular research field and is also increasingly widely used in the field of digital medicine. The point cloud completion technology has emerged as the times require. It starts from incomplete dental point cloud data and estimates the complete point cloud to achieve the purpose of completion.

[0004] Traditional point cloud completion methods often rely on the structural prior information that can be extracted from the point cloud, such as symmetry information or semantic information. Algorithms are designed based on this prior information to complete the incomplete dental point cloud. However, such methods can only handle incomplete dental point clouds with a very low missing rate and significant structural features. With the in-depth research of researchers, deep learning methods for point cloud analysis and generation have gradually emerged. There have been a series of excellent frameworks such as PCN, GRNet, and PF-Net. These point cloud completion methods take the incomplete dental point cloud as input, learn the structural information in the point cloud based on a deep neural network, and obtain a complete point cloud as output, which can effectively solve the problems of high memory occupancy and point cloud artifacts caused by discrete representation. At the same time, due to the disorder and permutation invariance of the point cloud, traditional convolution operations cannot be directly applied to the point cloud, and the design of deep neural networks still faces great challenges. For example, the existing point cloud completion frameworks focus too much on the overall features of the point cloud and ignore the local geometric information inside the point cloud, resulting in the complete point cloud after completion of these frameworks tending to a certain similar shape and unable to reflect the characteristics of individuals. Directly applying it to the dental point cloud dataset will lose a lot of tooth details and thus does not have clinical significance. Summary of the Invention

[0005] Object of the Invention: One object of the present invention is to provide a dental point cloud completion method based on a transformer encoder, which can ensure the details of the completed teeth to the greatest extent while paying attention to the global features of the point cloud.

[0006] Another object of the present invention is to provide a dental point cloud completion system based on a transformer encoder.

[0007] Technical Solution: A dental point cloud completion method based on a transformer encoder of the present invention includes the following steps:

[0008] S1. Process the dental point cloud dataset after unifying the scale using the dynamic graph convolutional neural network DGCNN to obtain the clustering centers and extract the local features of the key points in and obtain a feature matrix according to the clustering centers and the local features

[0009] S2. Use the transformer encoder to encode the clustering centers and the local features around each key point to mine the latent vectors for describing the teeth at the missing parts , and use a multi-layer perceptron MLP to predict the low-resolution point cloud Y of the missing teeth from the latent vectors ​c ;

[0010] S3. Combine the low-resolution point cloud Y c and the latent vector to form a reconstruction feature Use a multi-scale point cloud generator to recover the high-resolution point cloud Y from and Y c ; Combine Y f and Y c to obtain the final dental point cloud Y f ; pred ;

[0011] S4. Complete the training of the dental restoration model by minimizing the Chebyshev distance between Y pred and the corresponding true missing tooth gold standard Y gt ;

[0012] S5. Use the trained dental restoration model to complete the test data, obtain the complete dental point cloud, and evaluate it using metrics.

[0013] Furthermore, step S1 includes the following steps:

[0014] S11. Let the dental point cloud dataset be S = {(x1, y1), (x2, y2),..., (x L , y L )}, where x i represents the incomplete dental point cloud, y i represents the corresponding missing tooth gold standard, i = 1,..., L, and L represents the number of samples in the dataset; Use farthest point sampling (FPS) to sample all samples in the dental point cloud dataset to the same number of points and regularize it to obtain a dental point cloud dataset with the same scale;

[0015] S12. Use the dynamic graph convolutional neural network (DGCNN) to cluster the dental point cloud dataset with the unified scale obtained in step S11 to obtain the cluster center , and this cluster center contains the coordinates of several key points, and at the same time extracts the local features around each key point

[0016] S13. Use two different multi-layer perceptrons (MLPs) to process and respectively, and add the processing results together as the feature of the incomplete dental point cloud; The feature matrix of the incomplete dental point cloud is:

[0017]

[0018] where and They are respectively the first multi-layer perceptron and the second multi-layer perceptron for processing and respectively, and the feature matrix representing the incomplete dental point cloud. The feature matrix represents the incomplete dental point cloud.

[0019] Furthermore, step S2 is specifically as follows:

[0020] The transformer encoder queries according to the key point coordinates in the clustering center based on the feature matrix obtained in step S1 and then further learns the local geometric structure through feature aggregation and max pooling of the linear layer; the clustering center provides semantic feature information and encoder learning structure feature information, combines the semantic feature information and the encoder learning structure feature information, and uses the result of the transformer encoder as the latent vector This latent vector describes the geometric structure and numerical range of the missing part; then a multi-layer perceptron is used to predict the low-resolution point cloud Y from the latent vector ; the expressions of the latent vector c and the low-resolution point cloud Y are as follows: c The expressions of the latent vector and the low-resolution point cloud Y are as follows:

[0021]

[0022]

[0023] where is the feature matrix, is the third multi-layer perceptron for predicting the low-resolution point cloud Y c ;

[0024] Furthermore, step S3 is specifically as follows:

[0025] The multi-scale point cloud generator includes a fourth multi-layer perceptron and a fifth multi-layer perceptron. First, the latent vector obtained in step S2 and the low-resolution point cloud are concatenated together from the first dimension as the reconstruction feature The fourth multi-layer perceptron predicts a higher-resolution intermediate point cloud from , and then the intermediate point cloud is further concatenated with from the first dimension as the input of the fifth multi-layer perceptron. The fifth multi-layer perceptron predicts the high-resolution point cloud Y f , and finally Y f and Y c are concatenated together from the first dimension as the final predicted point cloud Y pred ; the reconstruction feature and the final predicted point cloud Y pred is expressed by the formula:

[0026]

[0027]

[0028] wherein, represents the reconstructed feature, and respectively represent the fourth multi-layer perceptron and the fifth multi-layer perceptron, represents the matrix connection operation.

[0029] Further, step S4 is specifically as follows:

[0030] The Chebyshev distance is used to describe the distance between two dental point clouds, and the Chebyshev distance calculation formula is as follows:

[0031]

[0032] wherein, S1 and S2 represent two dental point clouds, CD(S1, S2) represents the Chebyshev distance between dental point clouds S1 and S2, x and y respectively represent the point coordinates belonging to point clouds S1 and S2, and the calculation complexity of the Chebyshev distance is

[0033] Use the loss function to represent the objective function of the dental point cloud completion model, and complete the training of the dental completion model;

[0034] The loss function is calculated through the Chebyshev distance as:

[0035]

[0036] wherein, α is a weight parameter used to balance the two terms, CD(Y c , Y gt ) represents the distance between the low-resolution point cloud and the gold standard, αCD(Y pred , Y gt ) represents the distance between the finally predicted dental point cloud and the gold standard; the low-resolution point cloud characterizes the global feature information, and the high-resolution point cloud ensures the details of the completed tooth.

[0037] Further, step S5 is specifically as follows:

[0038] Using the tooth restoration model trained in step S4, predict the test samples to restore the missing teeth. To evaluate the performance of the tooth restoration results, the Chebyshev distance, Earth Mover's Distance (EMD), and F-score are used as evaluation metrics to evaluate the tooth restoration results. Among them, the Chebyshev distance has two forms: CD-l1 and CD-l2. CD-l1 uses the L1 norm to calculate the Chebyshev distance between two points, and CD-l2 uses the L2 norm to calculate the Chebyshev distance between two points. The calculation formula of the Earth Mover's Distance is as follows:

[0039]

[0040] where S1 and S2 represent two dental point clouds, and EMD(S1, S2) represents the Earth Mover's Distance between the dental point clouds S1 and S2. φ represents the bijective mapping between S1 and S2, which is used to minimize the distance between corresponding points.

[0041] The F-score is defined as the harmonic mean between the precision and recall of the prediction results.

[0042] A tooth point cloud restoration system based on a Transformer encoder according to the present invention includes a data processing module, an encoding module, a multi-scale point cloud generator, a tooth restoration model training module, and a test evaluation module. Among them, the data processing module samples all samples in the tooth point cloud dataset to the same number of points, regularizes them to unify tooth point clouds of different scales, and clusters the incomplete tooth point clouds with unified scales to obtain cluster centers , and at the same time extracts local features around the key points in the cluster centers , and then and are processed and added respectively to obtain the feature matrix of the incomplete tooth point cloud ; the encoding module queries according to the obtained feature matrix according to the key point coordinates in , and then further learns the local geometric structure through the feature aggregation of the linear layer and max pooling , provides semantic feature information and encoder learning structure feature information, combines the two kinds of information, and takes the result of the encoder as the latent vector , and then predicts the low-resolution point cloud Y from c ; the multi-scale point cloud generator predicts the high-resolution point cloud Y through the latent vector and the low-resolution point cloud Y c , and combines Y f and Y f together as the final predicted tooth point cloud Y c pred ​​; The tooth restoration model training module completes the training of the tooth restoration network by minimizing the Chebyshev distance between Y pred and the corresponding true missing tooth gold standard Y gt ; The test and evaluation module completes the restoration of the missing teeth in the missing part according to the prediction results of the test samples and evaluates the tooth restoration results.

[0043] Preferably, the encoding module includes a Transformer encoder and a multi-layer perceptron. The multi-layer perceptron is used to predict the output result of the Transformer encoder to obtain a low-resolution point cloud Y c .

[0044] A tooth point cloud restoration device based on a Transformer encoder of the present invention includes a memory and a processor, wherein:

[0045] The memory is used to store a computer program that can run on the processor;

[0046] The processor is used to execute the steps of the above-mentioned tooth point cloud restoration method based on the Transformer encoder when running the computer program.

[0047] A storage medium of the present invention stores a computer program, and when the computer program is executed by at least one processor, it implements the steps of the above-mentioned tooth point cloud restoration method based on the Transformer encoder.

[0048] Beneficial effects: Compared with the prior art, the present invention applies the Transformer encoder to tooth point cloud restoration, uses the self-attention mechanism in the Transformer to balance the inductive bias brought by the three-dimensional geometric relationship in the point cloud, models the local relationship by mining the potential feature vectors of the missing teeth, and uses a rough-to-fine method to generate the final point cloud. In addition, the loss function designed by the present invention emphasizes the global features represented by the key points in the low-resolution point cloud, and can ensure the details of the restored teeth to the greatest extent while focusing on the global features of the point cloud. Extensive experiments have proved the effectiveness of the present method and applied it to the point cloud restoration of incomplete teeth. And through experiments on a real tooth loss data set, it can be concluded that the method proposed by the present invention has good performance. Description of the Drawings

[0049] Figure 1 is the flowchart of the method of the present invention;

[0050] Figure 2 is the framework diagram of the point cloud restoration network of the present invention.

[0051] Figure 3 is the effect comparison diagram of the point cloud restoration method of the present invention. Specific implementation manner

[0052] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] As Figure 1 and Figure 2 shown, a tooth point cloud completion method based on a Transformer encoder includes the following steps:

[0054] S1. Use the dynamic graph convolutional neural network DGCNN to process the incomplete tooth point cloud to obtain the clustering centers and extract the local features of the key points in ; and obtain the feature matrix according to the clustering centers and the local features

[0055] The specific process is as follows:

[0056] Let the tooth point cloud data set be S = {(x1, y1), (x2, y2),..., (x L , y L )}, where x i represents the incomplete tooth point cloud, y i represents the corresponding missing tooth gold standard, i = 1,..., L, and L represents the number of samples in the data set. First, use the farthest point sampling FPS to sample all samples in the tooth point cloud data set to the same number of points, and regularize it to unify point clouds of different scales to obtain a tooth point cloud data set with the same scale. Then use the dynamic graph convolutional neural network DGCNN to cluster the incomplete tooth point cloud to obtain the clustering centers The clustering centers contain the coordinates of several key points, and at the same time extract the local features around each key point Finally, use two different multi-layer perceptrons MLP to and process respectively, and add the processing results together as the feature of the incomplete tooth point cloud. The feature matrix of the incomplete tooth point cloud is:

[0057]

[0058] wherein, wherein and are the first multi-layer perceptron and the second multi-layer perceptron for processing and respectively, represents the feature matrix obtained by adding their processing results, that is, the feature matrix of the incomplete tooth point cloud.

[0059] S2. Use a geometric structure-sensitive Transformer encoder to process the clustering centers and the local features around each key point for encoding, and extract the latent vectors that can describe the missing teeth . Use a multi-layer perceptron (MLP) to predict the low-resolution point cloud Y of the missing teeth from the latent vector V c ;

[0060] The specific process is as follows:

[0061] The self-attention in the Transformer is used to solve the inductive bias problem existing in training incomplete dental point cloud data. At the same time, the present invention designs a geometrically sensitive module, which can better model the geometric structure in the point cloud. The geometric structure-sensitive Transformer encoder is composed of six consecutive modules. The encoder queries according to the key point coordinates in the clustering center in the feature matrix obtained in step S1, and then further learns the local geometric structure through feature aggregation of the linear layer and max pooling. The clustering center provides semantic feature information and the encoder learns structural feature information. The present invention combines these two types of information and uses the result of the encoder as the latent vector . This latent vector describes the geometric structure and numerical range of the missing part. Then use a multi-layer perceptron to predict the low-resolution point cloud Y from the latent vector . The expressions of the latent vector c and the low-resolution point cloud Y are as follows: c where

[0062]

[0063]

[0064] is the feature matrix obtained in step S1, and is the third multi-layer perceptron for predicting the low-resolution point cloud Y . c

[0065] S3. Combine the low-resolution point cloud Y c and the latent vector to form the reconstruction feature . Use a multi-scale point cloud generator to recover the high-resolution point cloud Y from c and Y f . Combine Y c and Y f to obtain the final dental point cloud Y pred ​;

[0066] The specific process is as follows:

[0067] The multi-scale point cloud generator consists of two similar fourth multi-layer perceptrons and fifth multi-layer perceptrons, and this generator restores the final high-resolution point cloud in a coarse-to-fine manner. First, the latent vector obtained in step S2 and the low-resolution point cloud Y c are concatenated together from the first dimension as the reconstruction feature , the fourth multi-layer perceptron predicts a higher-resolution intermediate point cloud from , then the intermediate point cloud is concatenated with further from the first dimension as the input of the fifth multi-layer perceptron, and the fifth multi-layer perceptron predicts the high-resolution point cloud Y f . Finally, Y f and Y c are concatenated together from the first dimension as the final predicted dental point cloud Y pred . The reconstruction feature and the final predicted point cloud Y pred are expressed by the formula:

[0068]

[0069]

[0070] where, represents the reconstruction feature, and represent the fourth multi-layer perceptron and the fifth multi-layer perceptron in the multi-scale point cloud generator respectively, represents the matrix concatenation operation.

[0071] S4. By minimizing the Chamfer distance between Y pred and the corresponding true missing tooth gold standard Y gt , the training of the dental restoration model is completed;

[0072] The specific process is as follows:

[0073] This method uses the Chamfer Distance to describe the distance between two dental point clouds. The Chamfer distance calculation formula is as follows:

[0074]

[0075] where, S1 and S2 represent two dental point clouds, CD(S1, S2) represents the Chamfer distance between the dental point clouds S1 and S2, and x and y represent the point coordinates belonging to the point clouds S1 and S2 respectively. The calculation complexity of the Chamfer distance is , it can greatly shorten the training time of this method. As described in step S3, the final point cloud is obtained in a way from rough to fine. The loss function of the point cloud completion network represents the objective function of the dental point cloud completion model, and the training of the dental completion model is completed;

[0076] The loss function of the point cloud completion network of the present invention includes the following two parts:

[0077]

[0078] Among them, α is a weight parameter used to balance the two terms. The first term calculates the distance between the low-resolution point cloud and the gold standard, that is, CD(Y c , Y gt ) represents the distance between the low-resolution point cloud and the gold standard; the second term calculates the distance between the final predicted point cloud and the gold standard, that is, αCD(Y pred , Y gt ) represents the distance between the final predicted dental point cloud and the gold standard; the low-resolution point cloud characterizes the global feature information, and the high-resolution point cloud ensures the details of the completed teeth. Designing the loss function in this way increases the proportion of feature points, so that the final model can ensure the details of the completed teeth to the greatest extent while focusing on the global features of the point cloud.

[0079] In the training of the dental completion model in this embodiment, we use 5469 examples as the training set and 640 examples as the test set. All the codes are implemented in PyTorch. All modules in the network use the ADAM optimizer. The initial learning rate is set to 0.0001, and the learning rate decays by 70% every 40 epochs. The batch size is set to 64, and the total number of training epochs is 500. The model is trained on two NVIDIA RTX3090 by minimizing the objective function.

[0080] S5. Use the trained dental completion model to complete the test data to obtain the complete dental point cloud, and evaluate it using metrics.

[0081] The specific process is as follows:

[0082] Use the dental completion model trained in step S4 to predict the test samples and complete the teeth in the missing parts. In order to evaluate the performance of the tooth completion results, this method selects the Chebyshev distance, the Earth Mover Distance, and the F-score as evaluation metrics, where two forms of the Chebyshev distance are used: CD-l1 and CD-l2. CD-l1 uses the L1 norm to calculate the distance between two points, and CD-l2 uses the L2 norm to calculate the distance between two points. The calculation formula of the Earth Mover Distance is as follows:

[0083]

[0084] Among them, S1 and S2 represent two dental point clouds, EMD(S1, S2) represents the Earth Mover's Distance between the dental point clouds S1 and S2, and φ represents the bijective mapping between S1 and S2, which is used to minimize the distance between corresponding points.

[0085] The F-score is defined as the harmonic mean between the precision and recall of the prediction results; the precision refers to the number of correctly predicted points divided by all points in the gold standard compared with the gold standard, and the recall refers to the number of correctly predicted points divided by all points in the input compared with the input.

[0086] This method uses the above four metrics to evaluate the completion results, and selects four representative algorithms in the field of point cloud completion: Point Completion Network (PCN), Point Fractal Network (PF-Net), Variational Relational Point Completion Network (VRCNet), Geometry-Aware Transformers (PoinTr). The metric results are shown in Table 1. The method of the present invention has achieved the best results in the four metrics of CD-l1, CD-l2, EMD, and F-score, indicating that the completion results of the present invention are closer to the true results. Figure 3 The visual comparison results with the four algorithms are shown. It can be seen that the completion results of the present invention can more finely restore the dental shape, avoid a certain degree of noise, and are closer to the gold standard.

[0087] Table 1 Dental Point Cloud Completion Index Results

[0088]

[0089] A dental point cloud completion system based on a transformer encoder according to the present invention includes a data processing module, an encoding module, a multi-scale point cloud generator, a dental completion model training module, and a test evaluation module. Among them, the data processing module samples all samples in the dental point cloud dataset to the same number of points, regularizes them to unify point clouds of different scales, and clusters the incomplete dental point clouds with unified scales using the dynamic graph convolutional neural network DGCNN to obtain the clustering centers , and at the same time extracts the local features around the key points in the clustering centers , and then and are processed and added respectively to obtain the features of the incomplete dental point cloud ; The encoding module includes a Transformer encoder and a multi-layer perceptron. The Transformer encoder queries according to the key point coordinates obtained from , and then further learns the local geometric structure through feature aggregation of the linear layer and max pooling. It provides semantic feature information and encoder learning structure feature information, combines the two kinds of information, and takes the result of the Transformer encoder as the latent vector . The multi-layer perceptron predicts the low-resolution point cloud Y from . According to the key point coordinates in , and then further learns the local geometric structure through feature aggregation of the linear layer and max pooling. It provides semantic feature information and encoder learning structure feature information, combines the two kinds of information, and takes the result of the Transformer encoder as the latent vector , and the multi-layer perceptron predicts the low-resolution point cloud Y from . c ; The multi-scale point cloud generator predicts the high-resolution point cloud Y c through the latent vector and the low-resolution point cloud Y c . and the low-resolution point cloud Y c to predict the high-resolution point cloud Y f . f Combines Y f f and Y c c together as the final predicted point cloud Y pred . pred ; The tooth restoration model training module completes the training of the tooth restoration network by minimizing the Chebyshev distance between Y pred pred and the corresponding true missing tooth gold standard Y gt . gt The test and evaluation module completes the tooth restoration of the missing part according to the prediction result of the test sample and evaluates the restoration result.

[0090] A tooth point cloud restoration device based on a Transformer encoder, including a memory and a processor, wherein:

[0091] The memory is used to store a computer program that can run on the processor;

[0092] The processor is used to execute the steps of the above-mentioned tooth point cloud restoration method based on a Transformer encoder when running the computer program.

[0093] A storage medium stores a computer program, and when the computer program is executed by at least one processor, it implements the steps of the above-mentioned tooth point cloud restoration method based on a Transformer encoder.

Claims

1. A method for dental point cloud completion based on a transformer encoder, characterized in that, Including the following steps: S1. Process the unified-scale dental point cloud dataset using the dynamic graph convolutional neural network DGCNN to obtain the clustering centers and extract the local features of the key points and obtain the feature matrix based on the clustering centers and the local features S2. Use a Transformer encoder to process the clustering centers and the local features around each key point to encode and extract potential vectors for describing the missing teeth Use a multi-layer perceptron (MLP) to predict the low-resolution point cloud Y of the missing teeth from the potential vectors ; c ; S3. Combine the low-resolution point cloud Y c and the latent vector to form a reconstruction feature Use a multi-scale point cloud generator to recover the high-resolution point cloud Y from and Y c to obtain the high-resolution point cloud Y f . Combine Y c and Y f to obtain the final dental point cloud Y pred ; specifically: The multi-scale point cloud generator includes a fourth multi-layer perceptron and a fifth multi-layer perceptron. First, the latent vector obtained in step S2 and the low-resolution point cloud are concatenated together in the first dimension as the reconstruction feature The fourth multi-layer perceptron predicts a higher-resolution intermediate point cloud from Then, the intermediate point cloud is further concatenated together in the first dimension with as the input of the fifth multi-layer perceptron, and the fifth multi-layer perceptron predicts the high-resolution point cloud Y f , finally, Y f and Y c are concatenated together in the first dimension as the final dental point cloud Y pred ; The formulas for the reconstruction feature and the final dental point cloud Y pred are expressed as: Among them, represents a reconstruction feature, and respectively represent a fourth multi-layer perceptron and a fifth multi-layer perceptron, represents a matrix connection operation; S4. Complete the training of the dental restoration model by minimizing the pred Chebyshev distance between Y and the corresponding gold standard Y of the truly missing teeth; specifically: gt ​ Use the Chebyshev distance to describe the distance between two dental point clouds. The Chebyshev distance calculation formula is as follows: Among them, S1 and S2 represent two dental point clouds, CD(S1, S2) represents the Chebyshev distance between the dental point clouds S1 and S2, x and y respectively represent the point coordinates belonging to the point clouds S1 and S2, and the computational complexity of the Chebyshev distance is Use the loss function to represent the objective function of the dental restoration model, and complete the training of the dental restoration model; The loss function is calculated by the Chebyshev distance as: where α is a weight parameter for balancing the two terms, and CD(Y c , Y gt ) represents the distance between the low-resolution point cloud and the gold standard, and αCD(Y pred , Y gt ) represents the distance between the final predicted dental point cloud and the gold standard; the low-resolution point cloud represents the global feature information, and the high-resolution point cloud ensures the details of the restored tooth; S5. Use the trained dental restoration model to complete the restoration of the test data, obtain the complete dental point cloud, and use indicators for evaluation.

2. The method for dental point cloud completion based on a Transformer encoder according to claim 1, wherein Step S1 includes the following steps: S11. Let the dental point cloud data set be \(S=\{(x_1,y_1),(x_2,y_2),...,(x L ,y L )\}, where \(x i \) represents the incomplete dental point cloud, and \(y i \) represents the corresponding gold standard of missing teeth. \(i = 1,\ldots,L\), and \(L\) represents the number of samples in the data set. Use the farthest point sampling (FPS) to sample all samples in the dental point cloud data set to the same number of points, and regularize it to obtain a dental point cloud data set with the same scale. S12. Use the dynamic graph convolutional neural network DGCNN to cluster the unified-scale tooth point cloud dataset finally obtained in step S11 to obtain the clustering centers The clustering centers contain the coordinates of several key points, and at the same time extract the local features around each key point S13. Use two different multi-layer perceptrons (MLPs) to process and respectively, and add the processing results together as the features of the incomplete dental body point cloud. The feature matrix of the incomplete dental body point cloud is: Among them, among them and are respectively the first multi-layer perceptron and the second multi-layer perceptron for processing and , and represents the feature matrix of the incomplete tooth point cloud.

3. A method for dental point cloud completion based on a Transformer encoder according to claim 1, characterized in that Step S2 is specifically: The transformer encoder is based on the feature matrix obtained in step S1 Query according to the key point coordinates in the clustering center , and then further learn the local geometric structure through feature aggregation and max pooling of the linear layer; the clustering center provides semantic feature information and encoder learning structure feature information, combines the semantic feature information and the encoder learning structure feature information, and uses the result of the transformer encoder as the latent vector This latent vector describes the geometric structure and numerical range of the missing part; then use a multi-layer perceptron to predict the low-resolution point cloud Y from the latent vector c ; the expressions of the latent vector and the low-resolution point cloud Y c are as follows: Among them, is the feature matrix, is the third multi-layer perceptron for predicting the low-resolution point cloud Y c of.

4. A method for dental point cloud completion based on a Transformer encoder according to claim 1, characterized in that, Step S5 is specifically: Use the dental restoration model trained in step S4 to predict the test samples and restore the missing teeth. To evaluate the performance of the tooth restoration results, use the selected Chebyshev distance, Earth Mover's Distance, and F-score as evaluation indicators to evaluate the tooth restoration results; among them, the Chebyshev distance includes two forms: CD-l1 and CD-l2. CD-l1 uses the L1 norm to calculate the Chebyshev distance between two points, and CD-l2 uses the L2 norm to calculate the Chebyshev distance between two points. The calculation formula of the Earth Mover's Distance is as follows: Where S1 and S2 represent two dental point clouds, EMD(S1, S2) represents the Earth Mover's Distance between the dental point clouds S1 and S2, φ represents the bijective mapping between S1 and S2, which is used to minimize the distance between corresponding points, and x represents the point coordinates belonging to the point cloud S1; F-score is defined as the harmonic mean between the precision and recall of the prediction results.

5. A dental point cloud completion system based on a transformer encoder, characterized in that, It includes a data processing module, an encoding module, a multi-scale point cloud generator, a dental restoration model training module, and a test evaluation module. Among them, the data processing module samples all samples in the dental point cloud dataset to the same number of points, regularizes them to unify dental point clouds of different scales, and clusters the incomplete dental point clouds with unified scales to obtain cluster centers. At the same time, local features around the key points in the cluster centers are extracted. Then and are processed separately and added together to obtain the feature matrix of the incomplete dental point cloud. The encoding module queries according to the obtained feature matrix in accordance with the key point coordinates, and then further learns the local geometric structure through feature aggregation of the linear layer and max pooling. It provides semantic feature information and encoder learning structure feature information, combines the two kinds of information, and takes the result of the encoder as the latent vector. Then a low-resolution point cloud Y is predicted from c ; the multi-scale point cloud generator predicts a high-resolution point cloud Y through the latent vector c and the low-resolution point cloud Y f , combines Y f and Y c together as the final predicted dental point cloud Y pred ; the dental restoration model training module completes the training of the dental restoration network by minimizing the Chebyshev distance between Y pred and the corresponding true missing tooth gold standard Y gt ; the test evaluation module completes the teeth in the missing part according to the prediction result of the test sample and evaluates the teeth restoration result. The multi-scale point cloud generator includes a fourth multi-layer perceptron and a fifth multi-layer perceptron. First, the latent vector obtained in step S2 and the low-resolution point cloud are concatenated together in the first dimension as the reconstruction feature The fourth multi-layer perceptron predicts a medium-resolution point cloud from Then, the medium-resolution point cloud is further concatenated together with in the first dimension as the input of the fifth multi-layer perceptron, and the fifth multi-layer perceptron predicts a high-resolution point cloud Y f , Finally, Y f and Y c are concatenated together in the first dimension as the final dental point cloud Y pred ; The formulas for the reconstruction feature and the final dental point cloud Y pred are expressed as: Among them, represents the reconstruction feature, and respectively represent the fourth multi-layer perceptron and the fifth multi-layer perceptron, represents the matrix connection operation; Use the Chebyshev distance to describe the distance between two dental point clouds. The Chebyshev distance calculation formula is as follows: Among them, S1 and S2 represent two dental point clouds, CD(S1, S2) represents the Chebyshev distance between the dental point clouds S1 and S2, x and y respectively represent the point coordinates belonging to the point clouds S1 and S2, and the computational complexity of the Chebyshev distance is Use the loss function to represent the objective function of the dental point cloud restoration network, and complete the training of the dental restoration model; The loss function is calculated by the Chebyshev distance as: where α is a weight parameter for balancing the two terms, and CD(Y c , Y gt ) represents the distance between the low-resolution point cloud and the gold standard, and αCD(Y pred , Y gt ) represents the distance between the finally predicted dental point cloud and the gold standard; the low-resolution point cloud characterizes the global feature information, and the high-resolution point cloud ensures the details of the completed tooth.

6. The dental point cloud completion system based on the transformer encoder according to claim 5, characterized in that The encoding module includes a Transformer encoder and a multi-layer perceptron, and the multi-layer perceptron is used to predict the result output by the Transformer encoder to obtain the low-resolution point cloud Y c .

7. A dental point cloud completion device based on a Transformer encoder, characterized in that, Including a memory and a processor, where: The memory is used to store a computer program that can run on the processor; The processor is used to execute the steps of a dental point cloud restoration method based on a transformer encoder as described in any one of claims 1-4 when running the computer program.

8. A storage medium, characterized in that, The computer program is stored on the storage medium, and when the computer program is executed by at least one processor, it implements the steps of a dental point cloud restoration method based on a transformer encoder as described in any one of claims 1-4.

Images