A method and system for reconstructing dental crown restorations based on diffusion model
Through the combination of diffusion model and neural network structure, the problem of existing crown restoration technology relying on artificial proficiency and generation instability is solved, and efficient and fine crown restoration reconstruction is achieved, improving the quality and efficiency of restoration.
Patent Information
- Application Number
- CN202310135822.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-20
- Publication Date
- 2025-08-12
- Estimated Expiration
- 2043-02-20
AI Technical Summary
The existing crown restoration technology relies on the proficiency of the dentist. The generative adversarial network has problems such as instability in training and poor generation effect, resulting in poor quality of the maxillofacial morphology of crown restoration.
Using the diffusion model, the depth map is generated by projecting the mesh model of the preparatory teeth and occlusion teeth, and using the noise predictor and denoising image generator, combined with neural network structures such as U-net, autoencoder, and transformer, the target tooth depth map is gradually denoised, and finally the crown restoration is obtained through point cloud reconstruction.
A fully automatic crown restoration process is realized, the quality and efficiency of the restoration is improved, the cost of training labor is eliminated, and a more refined modeling of the crown is formed.
Smart Images
Figure CN116167219B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of dental crown restoration, and in particular to a method and system for reconstructing a dental crown restoration based on a diffusion model. Background Art
[0002] Dental caries, a common chronic oral disease, can cause a range of complications, including endodontic disease and oral and maxillofacial inflammation. Crown restorations can restore normal chewing function to broken teeth. Currently, most crown restorations utilize computer-aided geometric design systems, such as 3Shape, Duret, and OrthoCAD, using standard tooth template libraries as a key component of oral restoration software. However, the quality of the maxillofacial morphology of these crown restorations depends on the dentist's proficiency. Some studies have attempted to generate depth information for crowns by designing generative adversarial networks. However, these networks suffer from training instability and poor generation performance. Summary of the Invention
[0003] The purpose of the present invention is to provide a method and system for reconstructing a dental crown restoration based on a diffusion model, in order to address the deficiencies of the prior art.
[0004] The technical solutions adopted in the present invention are as follows:
[0005] A method for reconstructing a dental crown restoration based on a diffusion model, comprising:
[0006] The mesh models of the prepared teeth and the occlusal teeth are projected according to a fixed projection rule to generate a depth map, thereby obtaining a depth map of the prepared teeth and a depth map of the occlusal teeth;
[0007] Random noise map, preparation tooth depth map and occlusal tooth depth map Figure 1 The noise map is input into the trained diffusion model, and the diffusion model is used to continuously diffuse and denoise the noise map to obtain the target tooth depth map. Then, according to the projection rule of the generated depth map, the target tooth depth map is back-projected to obtain the target tooth 3D point cloud. Finally, the point cloud is used to reconstruct the target tooth mesh, i.e., the reconstructed crown restoration.
[0008] The diffusion model includes a noise predictor and a denoised image generator; the noise predictor is used to perform noise prediction based on the input noise map, the prepared tooth depth map, and the occlusal tooth depth map, and output the predicted noise; the denoised image generator is used to generate a denoised image based on the predicted noise and serve as the noise map predicted by the noise predictor in the next step:
[0009]
[0010] Where, X t is the noise map of step t, X t-1 is the noise map at step t-1, β tis the noise variance coefficient of the current diffusion step t, γ t is the diffusion coefficient of the current diffusion step t, ε t is the prediction noise of the current diffusion step t, and Z is a normal distribution.
[0011] Furthermore, the mesh models of the prepared tooth and the occlusal tooth are projected according to a fixed projection rule to generate a depth map, and the method for obtaining the prepared tooth depth map and the occlusal tooth depth map is:
[0012] The maximum and minimum coordinates of the mesh model in the x and y planes are used as the projection range. The truncation plane and projection plane are then determined based on the crown height on the z axis. The projection plane is located above the positive z-axis of the truncation plane. The projection plane is divided into pixel grids according to the projection range. The depth information value (i, j) on the pixel grid is given by the following formula:
[0013]
[0014] Where i, j represent the positions of the pixel grid, d is the projection distance, which is the shortest perpendicular distance from the center point of each pixel grid to the tooth grid;
[0015] The pixel value p on the pixel grid records the z-axis information z_value, which is expressed as follows:
[0016]
[0017] z_0 is the z-axis coordinate of the projection plane.
[0018] Furthermore, the noise variance coefficient β t It is proportional to the diffusion step number t and its value range is [0.000001,0.01].
[0019] Furthermore, the diffusion coefficient of the current diffusion step s represents the symbol in the current diffusion step multiplication operation.
[0020] Furthermore, the diffusion model is trained based on the collected training data set to minimize the loss function, and each sample of the training data set includes the corresponding tooth preparation depth Figure X p , occlusal depth Figure X op and target tooth depth Figure X 0.
[0021] Furthermore, the loss function is expressed as follows:
[0022]
[0023] Where ε represents random noise, which obeys the normal distribution of N(0,1), || || represents the L1 norm; γ is the diffusion coefficient, f θ () represents the noise predictor.
[0024] Furthermore, the network structure of the noise predictor is U-net, autoencoder, and transformer.
[0025] A dental crown restoration reconstruction system based on a diffusion model, for implementing the dental crown restoration reconstruction method based on a diffusion model, comprising:
[0026] A projection unit is used to project the mesh models of the prepared teeth and the occlusal teeth according to a fixed projection rule to generate a depth map, thereby obtaining a depth map of the prepared teeth and a depth map of the occlusal teeth;
[0027] Target tooth depth map reconstruction unit is used to reconstruct random noise map, prepared tooth depth map and occlusal tooth depth map. Figure 1 The noise image is input into the trained diffusion model, and the noise image is continuously denoised by the diffusion model to obtain the target tooth depth map;
[0028] The dental crown restoration reconstruction unit is used to back-project the target tooth depth map according to the projection rule for generating the depth map to obtain the three-dimensional point cloud of the target tooth, and finally use the point cloud to reconstruct the target tooth mesh, that is, the reconstructed dental crown restoration.
[0029] The beneficial effects of the present invention are as follows: the present invention utilizes the diffusion model to predict the Markov state of the data to achieve the characteristics of capturing the semantic structure and fine-grained features of the image, thereby providing a fully automatic crown restoration method, eliminating the cost of training labor and greatly improving efficiency; the present invention reconstructs the crown restoration based on the diffusion model, forming a more refined modeling of the crown, thereby greatly improving the quality of the crown restoration. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 This is a mesh model diagram of the prepared tooth;
[0031] Figure 2 This is a mesh model diagram of the occlusal teeth;
[0032] Figure 3 is the target tooth mesh model diagram;
[0033] Figure 4 This is a flow chart of a method for reconstructing a dental crown restoration based on a diffusion model provided by the present invention;
[0034] Figure 5 This is a schematic diagram of a projection rule provided by the present invention;
[0035] Figure 6This is a denoising process diagram of a diffusion model provided by the present invention;
[0036] Figure 7 This is a structural diagram of a noise predictor provided by the present invention;
[0037] Figure 8 is the target crown three-dimensional mesh;
[0038] Figure 9 is the corresponding actual target crown depth map;
[0039] Figure 10 Target crown depth map (a) and target crown restoration (b) reconstructed using the method of the present invention. DETAILED DESCRIPTION
[0040] The present invention provides a method for reconstructing a crown restoration based on a diffusion model. The crown restoration is obtained by reconstructing a target tooth mesh model based on a mesh model of the prepared tooth and the occlusal tooth using a diffusion model. Figure 1-3 As shown in the figure, the coordinate axis is the tooth axis, wherein the x-axis is from the labial side to the lingual side, the y-axis is the mesial direction, and the z-axis is from the mandible to the maxilla and is perpendicular to the xy plane.
[0041] Figure 4 Shown is a flow chart of the method of the present invention, as shown in Figure 4 As shown, the specific steps include:
[0042] (1) The mesh models of the prepared teeth and the occlusal teeth are projected according to a fixed projection rule to generate a depth map, thereby obtaining the depth map of the prepared teeth and the depth map of the occlusal teeth.
[0043] In this step, it is necessary to determine the projection rules for projecting the three-dimensional grid onto the two-dimensional plane. As an optional implementation scheme, first use the maximum and minimum coordinates of the obtained tooth grid (the grid model of the prepared teeth and the occlusal teeth) in the x and y planes as the projection range, and then determine the truncation plane (Far Plane) and the projection plane according to the crown height of the z axis (the difference between the maximum and minimum coordinates of the z axis). The truncation plane is located in the grid model. For example, the 1 / 3 position of the crown height can be selected, where the projection plane is located above the positive direction of the z axis of the truncation plane. The position where the truncation plane is translated 8mm along the positive direction of the z axis can be selected. The projection plane is then divided into pixel grids according to the projection range. Generally, the size of the pixel grid is 256*256. Finally, the shortest vertical distance from the center point of each pixel grid to the tooth grid is used as the projection distance d, as shown in the following example. Figure 5 As shown, the depth information on the pixel grid is as follows:
[0044]
[0045] Where i and j represent the positions of the pixel grid.
[0046] The pixel value p on the pixel grid records the z-axis information z_value, which is expressed as follows:
[0047]
[0048] z_0 is the z-axis coordinate of the projection plane.
[0049] The tooth mesh is projected into a depth map according to the above formula.
[0050] (2) Random noise map, prepared tooth depth map and occlusal tooth depth map Figure 1 The noise image is input into the trained diffusion model, and the noise image is continuously denoised by the diffusion model to obtain the target tooth depth map.
[0051] like Figure 6 As shown, the diffusion model gradually denoises the noise image through a T-step diffusion denoising process, and finally obtains a predicted image. In the present invention, the idea of the conditional diffusion model is adopted, that is, the preparation tooth depth map and the occlusal tooth depth map are used as conditional information of the noise predictor to assist the noise prediction process of the noise predictor. Specifically, the diffusion model includes a noise predictor and a denoised image generator; the noise predictor is used to perform noise prediction based on the input noise map, preparation tooth depth map and occlusal tooth depth map and output predicted noise; the network structure of the noise predictor can adopt a general neural network structure, such as U-net, autoencoder, transformer, etc. Figure 7 The figure shows a network structure of a noise predictor, which includes three convolutional downsampling layers, an intermediate feature layer, and a convolutional upsampling layer. A self-attention layer is superimposed after each sampling layer, and a symmetrical residual connection structure is added. Its input is a 3*256*256 image (the concatenation of the prepared tooth depth map, the occlusal tooth depth map, and the random noise map), and the output is the predicted 1*256*256 noise image ε. At the same time, the diffusion process needs to determine the number of diffusion steps T and the noise variance coefficient β corresponding to the diffusion step. t , where t is the number of the current diffusion step, and the diffusion coefficient γ is calculated t As an optional implementation scheme, the parameters are set as follows:
[0052] The current diffusion step number is t, t∈{1,2,3,…,T}.
[0053] The noise variance coefficient β of the current diffusion step t , β t ∈{β1,β2,β3,…,β T}, the setting method is as follows: β in the interval [0.000001,0.01] tIt is proportional to the diffusion step number t.
[0054] Calculate the diffusion coefficient for the current diffusion step s represents the symbol in the current diffusion step multiplication operation.
[0055] The denoised image generator is used to generate a denoised image based on the predicted noise and serves as the noise map predicted by the noise predictor in the next step:
[0056]
[0057] Where, X t is the noise map of step t, X t-1 is the noise map at step t-1, β t is the noise variance coefficient of the current diffusion step t, γ t is the diffusion coefficient of the current diffusion step t, ε t is the prediction noise of the current diffusion step t, and Z is a normal distribution.
[0058] Among them, the noise image input to the noise predictor for the first time is a random image X T After T steps of cyclic denoising process, X0 is obtained, which is the final predicted target tooth depth map.
[0059] The diffusion model is trained based on a collected training dataset to minimize the loss function, where each sample in the training dataset includes a corresponding tooth preparation depth. Figure X p , occlusal depth Figure X op and target tooth depth Figure X 0. Specifically, the neural network parameters f of the noise predictor are randomly initialized θ , given the corresponding preparation depth Figure X p , occlusal depth Figure X op and target tooth depth Figure X 0, perform gradient descent operation according to the following loss function to optimize and update the noise predictor f θ Neural network parameters:
[0060]
[0061] Where N(0,1) represents a normal distribution with a mean of 0 and a variance of 1, ε represents random noise that follows the normal distribution of N(0,1), and || || represents the L1 norm. The gradient descent operation is repeated, and the neural network parameters of the noise predictor are updated in each cycle. The updated neural network parameters are used as the noise predictor f in the next cycle. θ The neural network parameters are set until the loss function converges, and the noise predictor f θTraining completed.
[0062] (3) The depth map of the target tooth is then back-projected according to the projection rule for generating the depth map to obtain a three-dimensional point cloud of the target tooth, and finally the target tooth mesh, i.e., the reconstructed crown restoration, is obtained by using the Poisson point cloud reconstruction;
[0063] After cyclic denoising, the final predicted target tooth depth is obtained Figure X 0, according to the projection rules of the depth map, the pixel information is projected back to the three-dimensional space to obtain the point cloud. For each pixel grid, its corresponding x and y axis coordinates have been determined according to the projection rules, and the z axis information is as shown in the formula. For each pixel value p,
[0064]
[0065] Where z_0 is the z-axis coordinate of the projection plane. After obtaining the 3D point cloud, Poisson reconstruction is used to obtain a 3D mesh. The initial bounding box of the original point cloud is calculated. All surfaces outside the bounding box of the 3D mesh are filtered based on the bounding box to obtain the final 3D mesh of the target tooth, i.e., the crown restoration.
[0066] Figure 8 is the three-dimensional mesh of the target crown, where the circled tooth mesh corresponds to the target crown that needs to be reconstructed. Figure 9 is the corresponding actual target crown depth map constructed manually, Figure 10 Figure 1 shows the target crown depth map (a) and target crown restoration (b) reconstructed using the method of the present invention. Restoring proper masticatory function of a broken tooth is fundamental to crown restoration, and the occlusal groove and crown height are crucial information. As shown in the figure, the crown restoration reconstructed using the method of the present invention has relatively accurate occlusal groove information and a crown height that meets occlusal requirements. This demonstrates that the present invention enables more detailed crown modeling, significantly improving the quality of the crown restoration.
[0067] Corresponding to the aforementioned embodiment of a method for reconstructing a dental crown restoration based on a diffusion model, the present invention also provides an embodiment of a system for reconstructing a dental crown restoration based on a diffusion model.
[0068] A dental crown restoration reconstruction system based on a diffusion model of the present invention is used to implement the above-mentioned dental crown restoration reconstruction method based on a diffusion model, comprising:
[0069] A projection unit is used to project the mesh models of the prepared teeth and the occlusal teeth according to a fixed projection rule to generate a depth map, thereby obtaining a depth map of the prepared teeth and a depth map of the occlusal teeth;
[0070] Target tooth depth map reconstruction unit is used to reconstruct random noise map, prepared tooth depth map and occlusal tooth depth map. Figure 1The noise image is input into the trained diffusion model, and the noise image is continuously denoised by the diffusion model to obtain the target tooth depth map;
[0071] The dental crown restoration reconstruction unit is used to back-project the target tooth depth map according to the projection rule for generating the depth map to obtain the three-dimensional point cloud of the target tooth, and finally use the Poisson point cloud reconstruction to obtain the target tooth mesh, that is, the reconstructed dental crown restoration.
[0072] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The system embodiment described above is only illustrative, in which the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without expending creative work.
[0073] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications derived therefrom remain within the scope of protection of the present invention.
Claims
1. A method for reconstructing a crown restoration based on a diffusion model, characterized in that: include: The mesh models of the prepared teeth and the occlusal teeth are projected according to a fixed projection rule to generate a depth map, and the depth map of the prepared teeth and the depth map of the occlusal teeth are obtained as follows: The maximum and minimum coordinates of the mesh model in the x and y planes are used as the projection range. The truncation plane and projection plane are then determined based on the crown height on the z axis. The projection plane is located above the positive z-axis of the truncation plane. The projection plane is divided into pixel grids according to the projection range. The depth information value (i, j) on the pixel grid is given by the following formula: Where i, j represent the positions of the pixel grid, d is the projection distance, which is the shortest perpendicular distance from the center point of each pixel grid to the tooth grid; The pixel value p on the pixel grid records the z-axis information z_value, which is expressed as follows: z_0 is the z-axis coordinate of the projection plane; The random noise map, the prepared tooth depth map, and the occluded tooth depth map are input into the trained diffusion model. The noise map is continuously diffused and denoised by the diffusion model to obtain the target tooth depth map. The target tooth depth map is then back-projected according to the projection rule for generating the depth map to obtain the target tooth's three-dimensional point cloud. Finally, the point cloud is used to reconstruct the target tooth mesh, i.e., the reconstructed crown restoration. The diffusion model includes a noise predictor and a denoised image generator; the noise predictor is used to perform noise prediction based on the input noise map, the prepared tooth depth map, and the occlusal tooth depth map, and output the predicted noise; the denoised image generator is used to generate a denoised image based on the predicted noise and serve as the noise map predicted by the noise predictor in the next step: Where, X t is the noise map of step t, X t-1 is the noise map at step t-1, β t is the noise variance coefficient of the current diffusion step t, γ t is the diffusion coefficient of the current diffusion step t, ε t is the prediction noise of the current diffusion step t, and Z is a normal distribution.
2. The method according to claim 1, characterized in that The noise variance coefficient β t It is proportional to the diffusion step number t and its value range is [0.000001,0.01].
3. The method according to claim 1, characterized in that Diffusion coefficient of the current diffusion step s represents the symbol in the current diffusion step multiplication operation.
4. The method according to claim 1, wherein The diffusion model is trained based on the collected training data set to minimize the loss function, and each sample of the training data set includes the corresponding prepared tooth depth map X p 、Occlusal depth diagram X op and target tooth depth map X0.
5. The method according to claim 4, characterized in that The loss function is expressed as follows: Where ε represents random noise, which obeys the normal distribution of N(0,1), || || represents the L1 norm; γ is the diffusion coefficient, f θ () represents the noise predictor.
6. The method according to claim 1, wherein The network structure of the noise predictor is U-net, autoencoder, and transformer.
7. A dental crown restoration reconstruction system based on a diffusion model, characterized in that: A method for reconstructing a dental crown restoration based on a diffusion model according to any one of claims 1 to 6, comprising: A projection unit is used to project the mesh models of the prepared teeth and the occlusal teeth according to a fixed projection rule to generate a depth map, thereby obtaining a depth map of the prepared teeth and a depth map of the occlusal teeth; The target tooth depth map reconstruction unit is used to input the random noise map, the prepared tooth depth map and the occlusal tooth depth map into the trained diffusion model, and continuously denoise the noise map through the diffusion model to obtain the target tooth depth map; The dental crown restoration reconstruction unit is used to back-project the target tooth depth map according to the projection rule for generating the depth map to obtain the three-dimensional point cloud of the target tooth, and finally use the point cloud to reconstruct the target tooth mesh, that is, the reconstructed dental crown restoration.