A three-dimensional alveolar bone defect repair simulation method based on generative adversarial network

Abstract

The present invention discloses a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network, including S1, collecting alveolar bone CT scan data, oral dynamic pressure map, bite force data and related biomarker information of patients; S2, generating a detailed three-dimensional alveolar bone model with biomechanical properties; S3, designing a generative adversarial network model; S4, using the three-dimensional alveolar bone model generated in step S2 and real alveolar bone injury data and biomechanical data to train the generative adversarial network to generate a three-dimensional alveolar bone injury model consistent with the actual situation; S5, based on the three-dimensional alveolar bone injury model generated in step S4, simulating different repair materials and technologies to generate multiple repair scheme models; S6, selecting the optimal repair scheme. The three-dimensional alveolar bone model in the present invention provides more comprehensive and accurate alveolar bone structure information, which helps doctors understand the damage of the alveolar bone more accurately.

Description

Technical Field

[0001] The present invention relates to the field of alveolar bone technology, and in particular to a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network. Background Art

[0002] Existing alveolar bone defect repair technology mainly relies on the doctor's experience and two-dimensional imaging data for diagnosis and planning. These two-dimensional imaging data, including X-rays and CT slices, can provide certain references, but due to the lack of integrity and accuracy at the three-dimensional spatial level, doctors often have great limitations when formulating repair plans, and it is difficult to fully understand the morphology of alveolar bone damage and the relationship between surrounding tissues, which further leads to inaccurate repair plans, unsatisfactory surgical results, and even increased surgical risks.

[0003] In addition, the repair schemes in the prior art usually adopt a unified standard and lack consideration of the specific conditions of individual patients. The alveolar bone structure and damage conditions of each patient are different. A unified repair method cannot meet individual needs, which may lead to unsatisfactory repair effects, prolonged recovery time, and may even cause secondary damage. With the increase in the demand for personalized medical care, how to formulate accurate repair plans based on the specific conditions of each patient has become an urgent problem to be solved. Therefore, how to provide a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network is a topic that technicians in this field urgently need to solve. Summary of the invention

[0004] The present invention proposes a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network, which provides more comprehensive and accurate alveolar bone structure information, helping doctors to understand the damage of the alveolar bone more accurately.

[0005] A three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network according to an embodiment of the present invention comprises the following steps:

[0006] S1. Collect the patient's alveolar bone CT scan data, oral dynamic pressure map, bite force data and related biomarker information to construct a preliminary three-dimensional alveolar bone model;

[0007] S2. Use a multimodal fusion algorithm based on deep learning to process the preliminary three-dimensional alveolar bone model, integrate different types of biomedical information through feature-level fusion technology, and generate a detailed three-dimensional alveolar bone model with biomechanical properties;

[0008] S3. Design a generative adversarial network model, which includes a generator and a discriminator. The generator is responsible for generating a three-dimensional alveolar bone damage model and a repair model, and the discriminator evaluates the matching degree between the generated damage model and the repair model and the real alveolar bone damage data.

[0009] S4, using the three-dimensional alveolar bone model generated in step S2 and the real alveolar bone injury data and biomechanical data to train a generative adversarial network to generate a three-dimensional alveolar bone injury model consistent with the actual situation;

[0010] S5, based on the three-dimensional alveolar bone injury model generated in step S4, simulating different repair materials and techniques to generate multiple repair solution models;

[0011] S6. Evaluate the practicality and biocompatibility of each repair solution model through the discriminator and select the optimal repair solution.

[0012] Optionally, the S1 specifically includes the following steps:

[0013] S11, obtaining the alveolar bone CT scan data of the patient through a CT device, wherein the acquisition steps include: the patient maintains a stable posture, performs a full oral scan, and generates a three-dimensional data set D including multi-plane tomographic images. CT ;

[0014] S12. For the three-dimensional data set D CT Preprocessing:

[0015] Use a Gaussian filter to remove noise:

[0016]

[0017] Where I(x,y) is the original CT image, σ is the filter standard deviation, and k is the filter kernel radius;

[0018] Use histogram equalization technology to improve image contrast:

[0019]

[0020] Where L is the number of gray levels, N is the total number of pixels, and h(i) is the cumulative histogram of gray level i;

[0021] Apply the Canny edge detection algorithm to extract the boundary contour of the alveolar bone:

[0022]

[0023] Among them, G x and G y are the gradients of the image on the x-axis and y-axis respectively, G is the gradient amplitude, and θ is the gradient direction;

[0024] S13, recording the pressure distribution under different occlusal states by using an oral pressure sensor to obtain the patient's oral dynamic pressure map P dyn :

[0025] Arrange multiple micro pressure sensors in the patient's mouth, and the distribution of the sensor array is represented by S(x, y), where x and y are the position coordinates of the sensors;

[0026] The data of each sensor during the patient's static and dynamic occlusion process are recorded to generate a pressure distribution map P dyn (t), expressed as:

[0027]

[0028] Where n is the number of sensors, f(t) is the pressure function at time t;

[0029] S14. Collect the patient's bite force data F occl , the bite force data is obtained by an electronic bite instrument in multiple bite tests:

[0030] The patient was asked to bite on the electronic bite sensor, and the torque and total force at different tooth positions during each bite were recorded. The torque was expressed as:

[0031] M i =r i ×F i ;

[0032] Among them, r i is the length of the lever arm, F i is the acting force;

[0033] The bite force distribution diagram F is generated by averaging the results of multiple tests. occl (x,y):

[0034]

[0035] Where N is the number of trials;

[0036] S15. Collecting patient biomarker information B bio , including bone density B density , bone metabolism marker B metabolism and inflammatory factor B inflammation :

[0037] Bone density was measured by dual-energy X-ray absorptiometry. The formula for measuring bone density is:

[0038]

[0039] Among them, M bone is the mass of bone, V bone is the volume of the bone;

[0040] Blood tests were used to detect the levels of bone metabolism markers and inflammatory factors. The test values ​​were expressed as bone metabolism markers, B metabolismand inflammatory factor B inflammation ;

[0041] S16, CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio Input into the 3D modeling software to generate a preliminary 3D alveolar bone model M init :

[0042] The basic 3D geometric model is constructed using CT data. The model is represented as M geo (x,y,z);

[0043] The dynamic pressure map and bite force data are mapped to the corresponding positions of the three-dimensional model to form a model M with dynamic biomechanical characteristics. bio (x,y,z):

[0044]

[0045] The material properties of the model are adjusted according to the biomarker information to reflect the real biomechanical properties. The material properties of the model are expressed as A bio (x,y,z):

[0046] A bio (x,y,z)=f(B density ,B metabolism ,B inflammation );

[0047] Combine the processed geometric model, dynamic biomechanical properties and material properties to generate a preliminary three-dimensional alveolar bone model:

[0048] M init (x,y,z)=M geo (x,y,z)+α·M bio (x,y,z)+β·A bio (x,y,z);

[0049] Among them, α and β are fusion coefficients, which are used to adjust the weights of geometry, mechanics and material properties.

[0050] Optionally, S2 includes the following steps:

[0051] S21. Use a multimodal fusion algorithm based on deep learning to perform preliminary three-dimensional alveolar bone model M init To optimize, feature-level fusion technology is used to combine feature vectors from different data sources. The feature vector V fusion It is expressed as:

[0052] V fusion =[DCT ,P dyn ,F occl ,B bio ];

[0053] S22. Extract and fuse multimodal features using improved convolutional neural networks:

[0054] Design a multi-input convolutional neural network architecture, where each input corresponds to a type of data, namely CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio ;

[0055] For each type of data, separate convolutional and pooling layers are constructed for feature extraction:

[0056] V CT =CNN CT (D CT );

[0057] V dyn =CNN dyn (P dyn );

[0058] V occl =CNN occl (F occl );

[0059] V bio =CNN bio (B bio );

[0060] Among them, V CT 、V dyn 、V occl 、V bio are the feature vectors of each data type respectively;

[0061] A cross-modal attention mechanism is introduced to interact information between different modal features:

[0062] V cross_attn =CrossAttention(V CT ,V dyn ,V occl ,V bio );

[0063] Among them, CrossAttention represents the cross-modal attention mechanism, which calculates the correlation between each feature vector. The formula is:

[0064]

[0065] Among them, Q, K, and V are query, key, and value vectors respectively, and d k is the dimension of the key vector;

[0066] Improve the accuracy of feature representation through multi-head cross-modal attention mechanism:

[0067] MultiHeadCross(Q,K,V)=[head 1 ,head 2 ,…,head h ]W O ;

[0068] in,

[0069] W O is the output weight matrix, is the weight matrix of query, key, and value vectors;

[0070] The fused features are enhanced, and the auxiliary feature tasks are generated by self-supervised learning methods. The biomechanical properties and anatomical structure of the alveolar bone are combined through joint training and the loss function of the auxiliary feature tasks:

[0071]

[0072] in, The feature representation generated for self-supervised learning, V fusion is the actual feature representation, u i is the node displacement vector of the simulation model, is the node displacement vector in the real data, λ is the regularization coefficient, and ∈ is a small value to prevent the denominator from being zero;

[0073] The fused feature vector is further processed using a fully connected layer to generate a comprehensive feature representation V combined :

[0074] V combined =FC(V cross_attn );

[0075] Among them, FC is the fully connected layer;

[0076] The comprehensive feature representation is input into the subsequent convolutional layer to generate the final 3D alveolar bone model M detailed (x,y,z):

[0077] M detailed (x,y,z)=CNN combined (V combined );

[0078] Among them, CNNcombined (V combined ) is the three-dimensional model after comprehensive feature processing.

[0079] Optionally, S3 includes the following steps:

[0080] S31. Design a generative adversarial network model consisting of a generator G and a discriminator D:

[0081] Design of generator G:

[0082] The generator G receives the random noise vector z and the optimized 3D alveolar bone model M detailed As input, the three-dimensional alveolar bone damage model and repair model G(z,M detailed );

[0083] The input to the generator is represented as:

[0084] z~p z (z),M detailed ;

[0085] Where z is a random noise vector, p z (z) is the noise distribution;

[0086] The generation process of the generator is expressed as:

[0087] G(z,M detailed ) = f 3 (w 3 ·σ(f 2 (W 2 ·σ(f 1 (W 1 ·[z⊕M detailed ]+b 1 ))+b 2 ))+

[0088] b 3 );

[0089] Among them, ⊕ represents the concatenation operation of features, W 1 ,W 2 ,W 3 is the weight matrix, b 1 ,b 2 ,b 3 is the bias vector, f 1 ,f 2 ,f 3 is a multi-layer convolution operation, σ is an activation function;

[0090] Design of discriminator D:

[0091] The discriminator D receives the real alveolar bone injury data x and the repair model G(z,Mdetailed ) as input, and outputs the judgment result D(x) and the damage model D(G(z,M detailed ));

[0092] The input of the discriminator is expressed as:

[0093] x~p data (x),G(z,M detailed );

[0094] Among them, x is the real data, p data (x) is the real data distribution;

[0095] The output of the discriminator is expressed as:

[0096] D(x)=DiscriminatorNetwork(x),D(G(z,M detailed ))=DiscriminatorNetwork(z,M detailed ))];

[0097] Among them, DiscriminatorNetwork is a discriminator network structure, which performs classification judgment through multi-layer convolutional neural networks and fully connected layers. The calculation process is:

[0098]

[0099] in, is a multi-layer convolution operation, W 1 ,W 2 ,W 3 is the weight matrix, b 1 ,b 2 ,b 3 is the bias vector, f 1 ,f 2 ,f 3 is the convolution operation, and σ is the activation function.

[0100] Optionally, S4 includes the following steps:

[0101] S41. Use the final three-dimensional alveolar bone model M detailed (x, y, z), real alveolar bone injury data x and its biomechanical data training generation adversarial network model GAN;

[0102] S42. Prepare training data set:

[0103] The generated three-dimensional alveolar bone model M detailed Combined with the real alveolar bone injury data x, a training set D is formed train :

[0104]

[0105] in, represents the i-th optimized three-dimensional alveolar bone model, x i Indicates the corresponding real damage data;

[0106] S43, input the optimized three-dimensional alveolar bone model M detailed and random noise vector z, generate the damage model D(G(z,M detailed )), the loss function of the generator is:

[0107]

[0108] Where K(θ) is the parameterized stiffness matrix, is the node displacement vector of the generated model, F i is the force load vector, λ 1 and λ 2 is the weight coefficient, and They are generated and real three-dimensional alveolar bone models respectively;

[0109] S44, input the real alveolar bone injury data x and the generated injury model G(z,M detailed ), output the judgment results D(x) and D(G(z,M detailed ), the loss function of the discriminator is:

[0110]

[0111] Among them, γ is the regularization coefficient, θ D is the parameter of the discriminator, K is the number of layers of the discriminator, is the gradient of the k-th layer parameter;

[0112] S45, by alternately training the generator G and the discriminator D, the generated three-dimensional alveolar bone injury model G(z,M detailed ) is highly matched with the real alveolar bone injury data x;

[0113] The parameter update formula of the generator is:

[0114]

[0115] The parameter update formula of the discriminator is:

[0116]

[0117] in, and are the generator and discriminator parameters of the tth iteration, η is the learning rate, β is the momentum coefficient, m t is the momentum term;

[0118] The final three-dimensional alveolar bone injury model M generated It is expressed as:

[0119] M generated =G(z,M detailed .

[0120] Optionally, S5 includes the following steps:

[0121] S51, based on three-dimensional alveolar bone injury model M generated , simulate different restoration materials and techniques to design multiple restoration scheme models;

[0122] S52. Set of parameters defining restorative materials and techniques:

[0123] For each repair material, define a set of material property parameters θ material , including elastic modulus E, Poisson's ratio and density;

[0124] For each repair technique, define a set of technical parameters θ technique , including surgical approach and repair steps;

[0125] S53, input the generated three-dimensional alveolar bone injury model M generated and parameter set θ of restoration materials and techniques material and θ technique , generate the repair solution model M repair :

[0126]

[0127] Among them, α ijk and β ijk is the weight coefficient, To generate the model M generated The basis function, ψ ijk Basis functions for repair material and technical parameters;

[0128] S54, based on the repair solution model M repair And its material and technical parameters, establish the corresponding finite element analysis model FEA repair , the stress-strain distribution formula of finite element analysis is:

[0129] K repair =∫ Ω B T DBdΩ+∫ Γ N T CNdΓ;

[0130]

[0131] Frepair =∫ Ω N T fdΩ+∫ Γ N T tdΓ;

[0132] Among them, B is the strain-displacement matrix, D is the material stiffness matrix, N is the shape function matrix, C is the contact stiffness matrix, b is the body force density vector, f is the external force vector, and t is the surface force density vector;

[0133] S55. For each repair solution model M repair The biomechanical performance simulation was performed to calculate the stress-strain distribution, repair effect and biocompatibility. The simulation results are expressed as:

[0134]

[0135] Among them, σ i is the node stress value, σ repair is the strain tensor of the repaired model.

[0136] Optionally, S6 includes comparing the repair effects of different materials and techniques, and selecting the best repair solution for subsequent personalized treatment design:

[0137]

[0138]

[0139] Among them, σ target and σ target are the target stress and strain distributions, respectively, and are the stress and strain distributions of the jth repair scheme, respectively, and λ is the regularization coefficient.

[0140] The beneficial effects of the present invention are:

[0141] (1) The present invention collects the patient's CT scan data, oral dynamic pressure map, bite force data and related biomarker information, and uses a multimodal fusion algorithm based on deep learning to generate a detailed three-dimensional alveolar bone model with biomechanical properties. Compared with traditional two-dimensional imaging data, the three-dimensional alveolar bone model provides more comprehensive and accurate alveolar bone structure information, which helps doctors understand the damage of the alveolar bone more accurately.

[0142] (2) The present invention designs and trains a generative adversarial network model, including a generator and a discriminator. The generator is responsible for generating a three-dimensional alveolar bone injury model and a repair model, and the discriminator evaluates the degree of matching between the generated model and the real data. Through the adversarial training process, the accuracy of the generated model and the recognition ability of the discriminant model are improved, and a three-dimensional alveolar bone injury model that is highly matched with the actual situation can be generated.

[0143] (3) Based on the generated three-dimensional alveolar bone injury model, the present invention simulates different repair materials and technologies to design multiple repair scheme models, defines the parameter set of repair materials and technologies, and selects the optimal repair scheme through multiple simulations and optimizations. This method fully considers the specific conditions of each patient, provides a personalized repair scheme, and improves the repair effect and patient satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0144] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0145] Figure 1 This is a flow chart of a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network proposed in the present invention. DETAILED DESCRIPTION

[0146] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0147] refer to Figure 1 A three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network comprises the following steps:

[0148] S1. Collect the patient's alveolar bone CT scan data, oral dynamic pressure map, bite force data and related biomarker information to construct a preliminary three-dimensional alveolar bone model;

[0149] S2. Use a multimodal fusion algorithm based on deep learning to process the preliminary three-dimensional alveolar bone model, integrate different types of biomedical information through feature-level fusion technology, and generate a detailed three-dimensional alveolar bone model with biomechanical properties;

[0150] S3. Design a generative adversarial network model consisting of a generator and a discriminator. The generator is responsible for generating a three-dimensional alveolar bone damage model and a repair model, while the discriminator evaluates the matching degree between the generated damage model and the repair model and the real alveolar bone damage data.

[0151] S4, using the three-dimensional alveolar bone model generated in step S2 and the real alveolar bone injury data and biomechanical data to train a generative adversarial network to generate a three-dimensional alveolar bone injury model consistent with the actual situation;

[0152] S5, based on the three-dimensional alveolar bone injury model generated in step S4, generating multiple restoration scheme models, each restoration scheme model simulating different restoration materials and techniques;

[0153] S6. Evaluate the practicality and biocompatibility of each repair solution model through the discriminator and select the optimal repair solution.

[0154] In this implementation, S1 specifically includes the following steps:

[0155] S11, obtaining the alveolar bone CT scan data of the patient through a CT device, wherein the acquisition steps include: the patient maintains a stable posture, performs a full oral scan, and generates a three-dimensional data set D including multi-plane tomographic images. CT ;

[0156] S12. For the three-dimensional data set D CT Preprocessing:

[0157] Use a Gaussian filter to remove noise:

[0158]

[0159] Where I(x,y) is the original CT image, σ is the filter standard deviation, and k is the filter kernel radius;

[0160] Use histogram equalization technology to improve image contrast:

[0161]

[0162] Where L is the number of gray levels, N is the total number of pixels, and h(i) is the cumulative histogram of gray level i;

[0163] Apply the Canny edge detection algorithm to extract the boundary contour of the alveolar bone:

[0164]

[0165] Among them, G x and G y are the gradients of the image on the x-axis and y-axis respectively, G is the gradient amplitude, and θ is the gradient direction;

[0166] S13, recording the pressure distribution under different occlusal states by using an oral pressure sensor to obtain the patient's oral dynamic pressure map P dyn :

[0167] Arrange multiple micro pressure sensors in the patient's mouth, and the distribution of the sensor array is represented by S(x, y), where x and y are the position coordinates of the sensors;

[0168] The data of each sensor during the patient's static and dynamic occlusion process are recorded to generate a pressure distribution map P dyn (t), expressed as:

[0169]

[0170] Where n is the number of sensors, f(t) is the pressure function at time t;

[0171] S14. Collect the patient's bite force data F occl , the bite force data is obtained through multiple bite tests using an electronic bite instrument:

[0172] The patient was asked to bite on the electronic bite sensor, and the torque and total force at different tooth positions during each bite were recorded. The torque was expressed as:

[0173] M i =r i ×F i ;

[0174] Among them, r i is the length of the lever arm, F i is the acting force;

[0175] The bite force distribution diagram F is generated by averaging the results of multiple tests. occl (x,y):

[0176]

[0177] Where N is the number of trials;

[0178] S15. Collecting patient biomarker information B bio , including bone density B density , bone metabolism marker B metabolism and inflammatory factor B inflammation :

[0179] Bone density was measured by dual-energy X-ray absorptiometry. The formula for measuring bone density is:

[0180]

[0181] Among them, M bone is the mass of bone, V bone is the volume of the bone;

[0182] Blood tests were used to detect the levels of bone metabolism markers and inflammatory factors. The test values ​​were expressed as bone metabolism markers, B metabolismand inflammatory factor B inflammation ;

[0183] S16, CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio Input into the 3D modeling software to generate a preliminary 3D alveolar bone model M init :

[0184] The basic 3D geometric model is constructed using CT data. The model is represented as M geo (x,y,z);

[0185] The dynamic pressure map and bite force data are mapped to the corresponding positions of the three-dimensional model to form a model M with dynamic biomechanical characteristics. bio (x,y,z):

[0186]

[0187] The material properties of the model are adjusted according to the biomarker information to reflect the real biomechanical properties. The material properties of the model are expressed as A bio (x,y,z):

[0188] A bio (x,y,z)=f(B density ,B metabolism ,B inflammation );

[0189] Combine the processed geometric model, dynamic biomechanical properties and material properties to generate a preliminary three-dimensional alveolar bone model:

[0190] M init (x,y,z)=M geo (x,y,z)+α·M bio (x,y,z)+β·A bio (x,y,z);

[0191] Among them, α and β are fusion coefficients, which are used to adjust the weights of geometry, mechanics and material properties.

[0192] In this implementation, S2 includes the following steps:

[0193] S21. Use a multimodal fusion algorithm based on deep learning to perform preliminary three-dimensional alveolar bone model M init To optimize, feature-level fusion technology is used to combine feature vectors from different data sources. The feature vector V fusion It is expressed as:

[0194] V fusion =[DCT ,P dyn ,F occl ,B bio ];

[0195] S22. Extracting and fusing multimodal features using improved convolutional neural networks:

[0196] Design a multi-input convolutional neural network architecture, where each input corresponds to a type of data, namely CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio ;

[0197] For each type of data, separate convolutional and pooling layers are constructed for feature extraction:

[0198] V CT =CNN CT (D CT );

[0199] V dyn =CNN dyn (P dyn );

[0200] V occl =CNN occl (F occl );

[0201] V bio =CNN bio (B bio );

[0202] Among them, V CT 、V dyn 、V occl 、V bio are the feature vectors of each data type respectively;

[0203] A cross-modal attention mechanism is introduced to interact information between different modal features:

[0204] V cross_attn =CrossAttention(V CT ,V dyn ,V occl ,V bio );

[0205] Among them, CrossAttention represents the cross-modal attention mechanism, which calculates the correlation between each feature vector. The formula is:

[0206]

[0207] Among them, Q, K, and V are query, key, and value vectors respectively, and d k is the dimension of the key vector;

[0208] Improve the accuracy of feature representation through multi-head cross-modal attention mechanism:

[0209] MultiHeadCross(Q,K,V)=[head 1 ,head 2 ,…,head h ]W O ;

[0210] in,

[0211] W O is the output weight matrix, is the weight matrix of query, key, and value vectors;

[0212] The fused features are enhanced, and the auxiliary feature tasks are generated by self-supervised learning methods. The biomechanical properties and anatomical structure of the alveolar bone are combined through joint training and the loss function of the auxiliary feature tasks:

[0213]

[0214] in, The feature representation generated for self-supervised learning, V fusion is the actual feature representation, u i is the node displacement vector of the simulation model, is the node displacement vector in the real data, λ is the regularization coefficient, and ∈ is a small value to prevent the denominator from being zero;

[0215] The fused feature vector is further processed using a fully connected layer to generate a comprehensive feature representation V combined :

[0216] V combined =FC(V cross_attn );

[0217] Among them, FC is the fully connected layer;

[0218] The comprehensive feature representation is input into the subsequent convolutional layer to generate the final 3D alveolar bone model M detailed (x,y,z):

[0219] M detailed (x,y,z)=CNN combined (V combined );

[0220] Among them, CNNcombined (V combined ) is the three-dimensional model after comprehensive feature processing.

[0221] In this implementation, S3 includes the following steps:

[0222] S31. Design a generative adversarial network model consisting of a generator G and a discriminator D:

[0223] Design of generator G:

[0224] The generator G receives the random noise vector z and the optimized 3D alveolar bone model M detailed As input, the three-dimensional alveolar bone damage model and repair model G(z,M detailed );

[0225] The input to the generator is represented as:

[0226] z~p z (z),M detailed ;

[0227] Where z is a random noise vector, p z (z) is the noise distribution;

[0228] The generation process of the generator is expressed as:

[0229] G(z,M detailed ) = f 3 (W 3 ·σ(f 2 (W 2 ·σ(f 1 (W 1 ·[z⊕M detailed ]+b 1 ))+b 2 ))+b 3 );

[0230] Among them, ⊕ represents the concatenation operation of features, W 1 ,W 2 ,W 3 is the weight matrix, b 1 ,b 2 ,b 3 is the bias vector, f 1 ,f 2 ,f 3 is a multi-layer convolution operation, σ is an activation function;

[0231] Design of discriminator D:

[0232] The discriminator D receives the real alveolar bone injury data x and the repair model G(z,M detailed) as input, and outputs the judgment result D(x) and the damage model D(G(z,M detailed ));

[0233] The input of the discriminator is expressed as:

[0234] x~p data (x),G(z,M detailed );

[0235] Among them, x is the real data, p data (x) is the real data distribution;

[0236] The output of the discriminator is expressed as:

[0237] D(x)=DiscriminatorNetwork(x),D(G(z,M detailed ))=DiscriminatorNetwork(G(z,M detailed ))];

[0238] Among them, DiscriminatorNetwork is a discriminator network structure, which performs classification judgment through multi-layer convolutional neural networks and fully connected layers. The calculation process is:

[0239]

[0240] in, is a multi-layer convolution operation, W 1 ,W 2 ,W 3 is the weight matrix, b 1 ,b 2 ,b 3 is the bias vector, f 1 ,f 2 ,f 3 is the convolution operation, and σ is the activation function.

[0241] In this implementation, S4 includes the following steps:

[0242] S41. Use the final three-dimensional alveolar bone model M detailed (x, y, z), real alveolar bone injury data x and its biomechanical data training generation adversarial network model GAN;

[0243] S42. Prepare training data set:

[0244] The generated three-dimensional alveolar bone model M detailed Combined with the real alveolar bone injury data x, a training set D is formed train :

[0245]

[0246] in, represents the i-th optimized three-dimensional alveolar bone model, x i Indicates the corresponding real damage data;

[0247] S43, input the optimized three-dimensional alveolar bone model M detailed and random noise vector z, generate the damage model D(G(z,M detailed )), the loss function of the generator is:

[0248]

[0249] Where K(θ) is the parameterized stiffness matrix, is the node displacement vector of the generated model, F i is the force load vector, λ 1 and λ 2 is the weight coefficient, and They are generated and real three-dimensional alveolar bone models respectively;

[0250] S44, input the real alveolar bone injury data x and the generated injury model G(z,M detailed ), output the judgment results D(x) and D(G(z,M detailed ), the loss function of the discriminator is:

[0251]

[0252] Among them, γ is the regularization coefficient, θ D is the parameter of the discriminator, K is the number of layers of the discriminator, is the gradient of the k-th layer parameter;

[0253] S45, by alternately training the generator G and the discriminator D, the generated three-dimensional alveolar bone injury model G(z,M detailed ) is highly matched with the real alveolar bone injury data x;

[0254] The parameter update formula of the generator is:

[0255]

[0256] The parameter update formula of the discriminator is:

[0257]

[0258] in, and are the generator and discriminator parameters of the tth iteration, η is the learning rate, β is the momentum coefficient, m t is the momentum term;

[0259] The final three-dimensional alveolar bone injury model M generated It is expressed as:

[0260] M generated =G(z,M detailed ).

[0261] In this implementation, S5 includes the following steps:

[0262] S51, based on three-dimensional alveolar bone injury model M generated , design multiple restoration scheme models, each of which simulates different restoration materials and techniques;

[0263] S52. Set of parameters defining restorative materials and techniques:

[0264] For each repair material, define a set of material property parameters θ material , including elastic modulus E, Poisson's ratio and density;

[0265] For each repair technique, define a set of technical parameters θ technique , including surgical approach and repair steps;

[0266] S53, input the generated three-dimensional alveolar bone injury model M generated and parameter set θ of restoration materials and techniques material and θ technique , generate the repair solution model M repair :

[0267]

[0268] Among them, α ijk and β ijk is the weight coefficient, To generate the model M generated The basis function, ψ ijk Basis functions for repair material and technical parameters;

[0269] S54, based on the repair solution model M repair And its material and technical parameters, establish the corresponding finite element analysis model FEA repair , the stress-strain distribution formula of finite element analysis is:

[0270] K repair =∫ Ω B T DBdΩ+∫ Γ N T CNdΓ;

[0271]

[0272] F repair =∫ Ω N T fdΩ+∫ Γ N T tdΓ;

[0273] Among them, B is the strain-displacement matrix, D is the material stiffness matrix, N is the shape function matrix, C is the contact stiffness matrix, b is the body force density vector, f is the external force vector, and t is the surface force density vector;

[0274] S55. For each repair solution model M repair The biomechanical performance simulation was performed to calculate the stress-strain distribution, repair effect and biocompatibility. The simulation results are expressed as:

[0275]

[0276] Among them, σ i is the node stress value, ε repair is the strain tensor of the repaired model.

[0277] In this embodiment, S6 includes comparing the repair effects of different materials and technologies, and selecting the best repair solution for subsequent personalized treatment design:

[0278]

[0279] Among them, σ target and ε target are the target stress and strain distributions, respectively, and are the stress and strain distributions of the jth repair scheme, respectively, and λ is the regularization coefficient.

[0280] In order to overcome the shortcomings of the above-mentioned prior art, this embodiment proposes a three-dimensional alveolar bone defect repair simulation method based on a generative adversarial network. A high-quality three-dimensional alveolar bone model is generated by a generative adversarial network to accurately simulate different defects and repair schemes, helping doctors better understand the alveolar bone damage and develop personalized repair plans:

[0281] Collect the patient's alveolar bone CT scan data, oral dynamic pressure map, bite force data and related biomarker information to construct a preliminary three-dimensional alveolar bone model;

[0282] The preliminary 3D alveolar bone model was processed using a multimodal fusion algorithm based on deep learning, and different types of biomedical information were integrated through feature-level fusion technology to generate a detailed 3D alveolar bone model with biomechanical properties.

[0283] Design a generative adversarial network model consisting of a generator and a discriminator. The generator is responsible for generating a 3D alveolar bone damage model and a repair model, while the discriminator evaluates the matching degree between the generated damage model and the repair model and the real alveolar bone damage data.

[0284] Use a detailed 3D alveolar bone model and real alveolar bone injury data and biomechanical data to train a generative adversarial network to generate a 3D alveolar bone injury model that is consistent with the actual situation;

[0285] Based on the generated three-dimensional alveolar bone injury model, multiple restoration scheme models are generated, and each restoration scheme model simulates different restoration materials and techniques;

[0286] The discriminator is used to evaluate the practicality and biocompatibility of each repair solution model and select the optimal repair solution.

[0287] Through the above method, the limitations of two-dimensional imaging data in the existing technology are overcome, and high-precision three-dimensional visualization and simulation of alveolar bone injury and repair process are achieved, which improves the accuracy of the repair plan and meets the needs of personalized medical care. A more suitable treatment plan can be tailored for each patient, significantly improving the effect of alveolar bone defect repair.

[0288] The present invention collects the patient's CT scan data, oral dynamic pressure map, bite force data and related biomarker information, and uses a multimodal fusion algorithm based on deep learning to generate a detailed three-dimensional alveolar bone model with biomechanical properties. Compared with traditional two-dimensional imaging data, the three-dimensional alveolar bone model provides more comprehensive and accurate alveolar bone structure information, which helps doctors understand the damage of the alveolar bone more accurately.

[0289] The present invention designs and trains a generative adversarial network model including a generator and a discriminator. The generator is responsible for generating a three-dimensional alveolar bone damage model and a repair model, and the discriminator evaluates the matching degree between the generated model and the real data. Through the adversarial training process, the accuracy of the generated model and the recognition ability of the discriminant model are improved, and a three-dimensional alveolar bone damage model that is highly matched with the actual situation can be generated.

[0290] Based on the generated three-dimensional alveolar bone injury model, the present invention simulates different repair materials and technologies to design multiple repair scheme models, defines the parameter set of repair materials and technologies, and selects the optimal repair scheme through multiple simulations and optimizations. This method fully considers the specific conditions of each patient, provides a personalized repair scheme, and improves the repair effect and patient satisfaction.

[0291] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A three-dimensional alveolar bone defect repair simulation method based on generative adversarial network, characterized in that: The following steps are involved: S1. Collect the patient's alveolar bone CT scan data, oral dynamic pressure map, bite force data and related biomarker information to construct a preliminary three-dimensional alveolar bone model; S2. Use a multimodal fusion algorithm based on deep learning to process the preliminary three-dimensional alveolar bone model, integrate different types of biomedical information through feature-level fusion technology, and generate a detailed three-dimensional alveolar bone model with biomechanical properties; S3. Design a generative adversarial network model consisting of a generator and a discriminator. The generator is responsible for generating a three-dimensional alveolar bone damage model and a repair model, while the discriminator is responsible for evaluating the matching degree between the generated damage model and the repair model and the real alveolar bone damage data. S4, using the three-dimensional alveolar bone model generated in step S2 and the real alveolar bone injury data and biomechanical data to train a generative adversarial network to generate a three-dimensional alveolar bone injury model consistent with the actual situation; S5, based on the three-dimensional alveolar bone injury model generated in step S4, simulating different repair materials and techniques to generate multiple repair solution models; S6, evaluating the practicality and biocompatibility of each repair solution model through the discriminator, and selecting the optimal repair solution; The S1 specifically includes the following steps: S11, obtaining the alveolar bone CT scan data of the patient through a CT device, wherein the acquisition steps include: the patient maintains a stable posture, performs a full oral scan, and generates a three-dimensional data set D including multi-plane tomographic images. CT ; S12. For the three-dimensional data set D CT Preprocessing: Use a Gaussian filter to remove noise: Where I(x,y) is the original CT image, σ is the filter standard deviation, and k is the filter kernel radius; Use histogram equalization technology to improve image contrast: Where L is the number of gray levels, N is the total number of pixels, and h(i) is the cumulative histogram of gray level i; Apply the Canny edge detection algorithm to extract the boundary contour of the alveolar bone: Among them, G x and G y are the gradients of the image on the x-axis and y-axis respectively, G is the gradient amplitude, and θ is the gradient direction; S13, recording the pressure distribution under different occlusal states by using an oral pressure sensor to obtain the patient's oral dynamic pressure map P dyn : Arrange multiple micro pressure sensors in the patient's mouth, and the distribution of the sensor array is represented by S(x, y), where x and y are the position coordinates of the sensors; The data of each sensor during the patient's static and dynamic occlusion process are recorded to generate a pressure distribution map P dyn (t), expressed as: Where n is the number of sensors, f(t) is the pressure function at time t; S14. Obtaining the patient's bite force data F through multiple bite tests using an electronic bite instrument occl : The patient was asked to bite on the electronic bite sensor, and the torque and total force at different tooth positions during each bite were recorded. The torque was expressed as: M i =r i ×F i ; Among them, r i is the length of the lever arm, F i is the acting force; The bite force distribution diagram F is generated by averaging the results of multiple tests. occl (x,y): Where N is the number of trials; S15. Collecting biomarker information of patients B bio , including bone density B density , bone metabolism marker B metabolism and inflammatory factor B inflammation : Bone density was measured by dual-energy X-ray absorptiometry. The formula for measuring bone density is: Among them, M bone is the mass of bone, V bone is the volume of the bone; Blood tests were used to detect the levels of bone metabolism markers and inflammatory factors. The test values ​​were expressed as bone metabolism markers, B metabolism and inflammatory factor B inflammation ; S16, CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio Input into the 3D modeling software to generate a preliminary 3D alveolar bone model M init : The basic 3D geometric model is constructed using CT data. The model is represented as M geo (x,y,z); The dynamic pressure map and bite force data are mapped to the corresponding positions of the three-dimensional model to form a model M with dynamic biomechanical characteristics. bio (x,y,z): The material properties of the model are adjusted according to the biomarker information to reflect the real biomechanical properties. The material properties of the model are expressed as A bio (x,y,z): A bio (x,y,z)=f(B density ,B metabolism ,B inflammation ); Combine the processed geometric model, dynamic biomechanical properties and material properties to generate a preliminary three-dimensional alveolar bone model: M init (x,y,z)=M geo (x,y,z)+α·M bio (x,y,z)+β·A bio (x,y,z)? Among them, α and β are fusion coefficients, which are used to adjust the weights of geometry, mechanics, and material properties; The S2 comprises the following steps: S21. Use a multimodal fusion algorithm based on deep learning to perform preliminary three-dimensional alveolar bone model M init Optimize and use feature-level fusion technology to fuse feature vectors from different data sources. The feature vector V fusion It is expressed as: V fusion =[D CT ,P dyn ,F occl ,B bio ]; S22. Extract and fuse multimodal features using improved convolutional neural networks: Design a multi-input convolutional neural network architecture, where each input corresponds to a type of data, namely CT scan data D CT , Dynamic pressure diagram P dyn , bite force data F occl and biomarker information B bio ; For each type of data, separate convolutional and pooling layers are constructed for feature extraction: V CT =CNN CT (D CT ); V dyn =CNN dyn (P dyn ); V occl =CNN occl (F occl ); V bio =CNN bio (B bio ); Among them, V CT 、V dyn 、V occl 、V bio are the feature vectors of each data type respectively; A cross-modal attention mechanism is introduced to interact information between different modal features: V cross_attn =CrossAttention(V CT ,V dyn ,V occl ,V bio ); Among them, CrossAttention represents the cross-modal attention mechanism, which calculates the correlation between each feature vector. The formula is: Among them, Q, K, and V are query, key, and value vectors respectively, and d k is the dimension of the key vector; Improve the accuracy of feature representation through multi-head cross-modal attention mechanism: MultiHeadCross(Q,K,V)=[head1,head2,…,head h ]W O ; in, W O is the output weight matrix, is the weight matrix of query, key, and value vectors; The fused features are enhanced, and the auxiliary feature tasks are generated by self-supervised learning methods. The biomechanical properties and anatomical structure of the alveolar bone are combined through joint training and the loss function of the auxiliary feature tasks: in, The feature representation generated for self-supervised learning, V fusion is the actual feature representation, u i is the node displacement vector of the simulation model, is the node displacement vector in the real data, λ is the regularization coefficient, and ∈ is a small value to prevent the denominator from being zero; The fused feature vector is further processed using a fully connected layer to generate a comprehensive feature representation V combined : In combined =FC(V cross_attn ); Among them, FC is the fully connected layer; The comprehensive feature representation is input into the subsequent convolutional layer to generate the final 3D alveolar bone model M detailed (x,y,z): M detailed (x,y,z)=CNN combined (V combined ); Among them, CNN combined (V combined ) is the three-dimensional model after comprehensive feature processing.

2. A three-dimensional alveolar bone defect repair simulation method based on generative adversarial network according to claim 1, characterized in that: The S3 comprises the following steps: S31. Design a generative adversarial network model consisting of a generator G and a discriminator D: Design of generator G: The generator G receives the random noise vector z and the optimized 3D alveolar bone model M detailed As input, the three-dimensional alveolar bone damage model and repair model G(z,M detailed ); The input to the generator is represented as: z~p z (z),M detailed ; Where z is a random noise vector, p z (z) is the noise distribution; The generation process of the generator is expressed as: in, represents the concatenation operation of features, W1, W2, W3 are weight matrices, b1, b2, b3 are bias vectors, f1, f2, f3 are multi-layer convolution operations, and σ is the activation function; Design of discriminator D: The discriminator D receives the real alveolar bone injury data x and the repair model G(z,M detailed ) as input, and outputs the judgment result D(x) and the damage model D(G(z,M detailed )); The input of the discriminator is expressed as: x~p data (x),G(z,M detailed ); Among them, x is the real data, p data (x) is the real data distribution; The output of the discriminator is expressed as: D(x)=DiscriminatorNetwork(x),D(G(z,M detailed ))=DiscriminatorNetwork(G(z,M detailed ))]; Among them, DiscriminatorNetwork is a discriminator network structure, which performs classification judgment through multi-layer convolutional neural networks and fully connected layers. The calculation process is: in, is a multi-layer convolution operation, W1, W2, W3 are weight matrices, b1, b2, b3 are bias vectors, f1, f2, f3 are convolution operations, and σ is the activation function.

3. A three-dimensional alveolar bone defect repair simulation method based on generative adversarial network according to claim 2, characterized in that: The S4 comprises the following steps: S41. Use the final three-dimensional alveolar bone model M detailed (x, y, z), real alveolar bone injury data x and its biomechanical data training generation adversarial network model GAN; S42. Prepare training data set: The generated three-dimensional alveolar bone model M detailed Combined with the real alveolar bone injury data x, a training set D is formed train : in, represents the i-th optimized three-dimensional alveolar bone model, x i Indicates the corresponding real damage data; S43, input optimized three-dimensional alveolar bone model M detailed and random noise vector z, generate the damage model D(G(z,M detailed )), the loss function of the generator is: Where K(θ) is the parameterized stiffness matrix, is the node displacement vector of the generated model, F i is the force load vector, λ1 and λ2 are weight coefficients, and They are generated and real three-dimensional alveolar bone models respectively; S44, input the real alveolar bone injury data x and the generated injury model G(z,M detailed ), output the judgment results D(x) and D(G(z,M detailed ), the loss function of the discriminator is: Among them, γ is the regularization coefficient, θ D is the parameter of the discriminator, K is the number of layers of the discriminator, is the gradient of the k-th layer parameter; S45, by alternately training the generator G and the discriminator D, the generated three-dimensional alveolar bone injury model G(z,M detailed ) is highly matched with the real alveolar bone injury data x; The parameter update formula of the generator is: The parameter update formula of the discriminator is: in, and are the generator and discriminator parameters of the tth iteration, η is the learning rate, β is the momentum coefficient, m t is the momentum term; The final three-dimensional alveolar bone injury model M generated It is expressed as: M generated =G(z,M detailed )。 4. The three-dimensional alveolar bone defect repair simulation method based on generative adversarial network according to claim 3, characterized in that: The S5 comprises the following steps: S51, based on three-dimensional alveolar bone injury model M generated , simulate different restoration materials and techniques to design multiple restoration scheme models; S52. Set of parameters defining restorative materials and techniques: For each repair material, define a set of material property parameters θ material , including elastic modulus E, Poisson's ratio and density; For each repair technique, define a set of technical parameters θ technique , including surgical approach and repair steps; S53, input the generated three-dimensional alveolar bone injury model M generated and parameter set θ of restoration materials and techniques material and θ technique , generate the repair solution model M repair : Among them, α ijk and β ijk is the weight coefficient, To generate the model M generated The basis function, ψ ijk Basis functions for repair material and technical parameters; S54, based on the repair solution model M repair And its material and technical parameters, establish the corresponding finite element analysis model FEA repair , the stress-strain distribution formula of finite element analysis is: K repair =∫ Ω B T DBdΩ+∫ Γ N T CNdΓ; F repair =∫ Ω N T fdΩ+∫ Γ N T tdΓ: Among them, B is the strain-displacement matrix, D is the material stiffness matrix, N is the shape function matrix, C is the contact stiffness matrix, b is the body force density vector, f is the external force vector, and t is the surface force density vector; S55. For each repair solution model M repair The biomechanical performance simulation was performed to calculate the stress-strain distribution, repair effect and biocompatibility. The simulation results are expressed as: Among them, σ i is the node stress value, ε repair is the strain tensor of the repaired model.

5. The three-dimensional alveolar bone defect repair simulation method based on generative adversarial network according to claim 4, characterized in that: S6 includes comparing the restorative effects of different materials and techniques and selecting the best restorative solution for subsequent personalized treatment design: Among them, σ target and ε target are the target stress and strain distributions, respectively, and are the stress and strain distributions of the jth repair scheme, respectively, and λ is the regularization coefficient.

Images