Double-path dental crown point cloud completion method

Abstract

The invention relates to a double-path dental crown point cloud complementing method, which comprises the following steps of: obtaining an adjacent tooth point cloud; information in the adjacent tooth point cloud is extracted, and a dental crown structure condition feature vector is obtained; initializing a dental crown noise point cloud; denoising the dental crown noise point cloud by taking the dental crown structure condition feature vector as a guide to obtain an initial dental crown point cloud; inputting the adjacent tooth point cloud and the initial dental crown point cloud into a dual-path multi-scale refining network to obtain a dental crown point cloud; by adopting the technical scheme, the obtained dental crown point cloud fits the adjacent tooth structure, the surface is continuous and smooth, the boundary form is seamlessly matched with the adjacent tooth, and the problems of non-uniform structure distribution and insufficient local details in dental crown point cloud generation are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral cavity, in particular to a double-path crown point cloud completion method. BACKGROUND

[0002] In dental restoration treatment, when teeth are damaged due to factors such as loss, caries or fracture, customized crowns are often used for functional and morphological reconstruction. However, the traditional crown design process highly depends on the experience of doctors, even if computer-aided design systems such as CAD / CAM[1] are introduced, a large amount of manual adjustment is still required, which is low in efficiency and limited in individual adaptability. Existing systems mostly rely on standardized crown databases for template fitting, which is difficult to fully adapt to individual differences in patient dentition, often leading to clinical problems such as malocclusion, improper abutment, and even temporomandibular joint disorders.

[0003] In recent years, with the rapid development of oral scanning equipment, CBCT and three-dimensional reconstruction technology, the acquisition of high-precision digital dentition models has become more efficient and stable, providing a solid data foundation for intelligent denture design. Deep learning methods based on three-dimensional point clouds have gradually become the core means of personalized crown reconstruction. Models based on three-dimensional data use point clouds or voxels as input, which can more directly describe the three-dimensional structure of the crown and have more advantages in capturing complex geometric details. Existing technologies use neighboring tooth geometry information to predict missing areas and generate single molar crowns, verifying the feasibility of artificial intelligence in denture repair. Some of these methods use structure-aware Transformer networks such as PoinTr, which have made some progress in crown shape restoration. However, these methods generally only use Chamfer Distance (CD) as the main reconstruction loss function. CD calculates the distance by finding the nearest neighbor points, and its optimization process tends to make the predicted points gather in the dense areas of the target point cloud, while ignoring the sparse areas, which easily leads to uneven distribution of generated crown point clouds, local aggregation or voids, and difficulty in meeting the strict requirements of restorations for surface smoothness and geometric continuity. SUMMARY

[0004] In view of the above problems, the present application provides a double-path crown point cloud completion method to solve the problems of uneven structure distribution and insufficient local details in existing methods for generating crown point clouds.

[0005] To achieve the above-mentioned purpose, the inventors provide a double-path crown point cloud completion method, which includes the following steps:

[0006] Obtain neighboring tooth point cloud;

[0007] Extract information from the neighboring tooth point cloud to obtain a crown structure condition feature vector;

[0008] Initialize the crown noise point cloud;

[0009] guided by the tooth crown structure condition feature vector, to obtain an initial tooth crown point cloud;

[0010] inputting the adjacent tooth point cloud and the initial tooth crown point cloud into a double-path multi-scale refinement network to obtain the tooth crown point cloud.

[0011] Further, the step of extracting information in the adjacent tooth point cloud to obtain the tooth crown structure condition feature vector adopts a geometric perception condition feature extraction module to extract information in the adjacent tooth point cloud to obtain the tooth crown structure condition feature vector.

[0012] Further, the geometric perception condition feature extraction module comprises an FPS, a dynamic graph convolutional neural network and a Transformer encoder; the FPS performs down-sampling on the adjacent tooth point cloud to obtain a sparse adjacent tooth point cloud; the dynamic graph convolutional neural network dynamically constructs a connection relationship between points based on distance to find local adjacent points according to each point in the sparse adjacent tooth point cloud to obtain an adjacent tooth point cloud vector; and the Transformer encoder introduces a self-attention of a KNN mechanism to perform local and global feature interaction and fusion on the adjacent tooth point cloud vector to obtain the tooth crown structure condition feature vector.

[0013] Further, the step of guided denoising of the tooth crown noise point cloud to obtain the initial tooth crown point cloud guided by the tooth crown structure condition feature vector comprises guiding a condition guided diffusion generation module by the tooth crown structure condition feature vector to perform step-by-step reverse progressive denoising on the tooth crown noise point cloud to obtain the initial tooth crown point cloud.

[0014] Further, the condition guided diffusion generation module comprises six PCNets, each point of the tooth crown noise point cloud is independently input into the six PCNets, and the tooth crown structure condition feature vector and a timestamp T are simultaneously input.

[0015] The six PCNets output noise residual independent update point cloud of each point:

[0016] Starting from the maximum noise timestamp T, the following steps are repeated until T = 0.

[0017] The noisy point cloud X t The condition guided diffusion generation module is input, and the six PCNets are used for point-by-point denoising to obtain X t-1 .

[0018] Further, the PCNet comprises hierarchical residual connection and a gating mechanism; the feature channel dimension dynamically changes according to a progressive rule of [3, 128, 256, 512, 256, 128, 3] to form a symmetric structure of "contraction-expansion"; the reset gate in the gating mechanism dynamically controls the reservation ratio of historical features of each layer, and the calculation method is:

[0019] H l = R l (W lh H l-1 + b lh ) + W lb y, R l = σ(W lr y + b lr );

[0020] where y is the concatenation of the conditional feature and the time embedding, σ(·) is the Sigmoid activation function, and the weights W . and the bias b . are learnable parameters.

[0021] Further, the step of denoising the dental crown noisy point cloud guided by the dental crown structure conditional feature vector to obtain an initial dental crown point cloud comprises the following steps:

[0022] The denoising process continues to denoise the dental crown noisy point cloud X T for T time steps to obtain an initial dental crown point cloud X0, which satisfies the following conditions:

[0023] Joint distribution of the denoising process:

[0024]

[0025] where θ represents the learnable parameters of the conditional guided diffusion generation module; p θ (·) represents the conditional probability distribution of the reverse denoising; X 0:T represents the sequence from the initial dental crown point cloud X0to the dental crown noisy point cloud X T ; p θ (X 0:T |c) represents the joint probability distribution of the sequence from the initial dental crown point cloud X0to the dental crown noisy point cloud X T given the dental crown structure conditional feature vector c, which is parameterized by the parameters θ; p(X T ) represents the initial noise distribution; c represents the dental crown structure conditional feature vector; X t represents the noisy point cloud at the t-th step; X t-1 represents the noisy point cloud at the (t-1)-th step; represents the conditional probability distribution of the previous time step point cloud X t given the current noisy point cloud X t-1 and the dental crown structure conditional feature vector c during the process from time step t = 1 to t = T;

[0026] Conditional distribution of the denoising process:

[0027] p θ (X t-1 |Xt c)=N(X t-1 |μ θ (X t ,t,c),β t I);

[0028] in, μ represents a Gaussian distribution. θ μ represents the denoised mean predicted by the conditionally guided diffusion generation module. θ (X t ,t,c) represents the mean function of a Gaussian distribution, predicted by the conditionally guided diffusion generation module; β t Represents the noise scheduling parameters, where I is the identity matrix; β t I represents the covariance matrix;

[0029] The training objective is to maximize the lower bound of the log-likelihood.

[0030] E q [log p θ (X θ |X T ,c)];

[0031] By minimizing μ of the standard normal distribution θ The mean square error (MSE) loss between μ and μ is used to calculate and obtain the initial crown point cloud.

[0032] Furthermore, the dual-path multi-scale refinement network includes a feature extraction subnetwork, a feature transfer module, and a decoding module. The feature extraction subnetwork includes adjacent tooth paths and initial crown paths, which independently perform hierarchical downsampling and feature extraction on the input adjacent tooth point clouds and initial crown point clouds, respectively. The feature transfer module uses an attention mechanism to focus the points on the initial crown path and fuse relevant geometric features from the corresponding regions of the adjacent tooth paths. The decoding module performs hierarchical upsampling, fuses features from the corresponding levels of the encoder through skip connections, refines the features using MLP, and gradually restores the point cloud density to obtain the crown point cloud.

[0033] Furthermore, the number of sampling points in the feature extraction subnet is increased, the receptive field radius of the corresponding layer is reduced, the number of backbone channels is expanded, and the number of channels in the decoding module is expanded synchronously with the backbone network.

[0034] Furthermore, the attention-weighted neighborhood construction parameter K of the feature transfer module is expanded based on the default value.

[0035] Differing from the prior art, the technical scheme extracts multi-scale structural features from the adjacent tooth point cloud, generates a tooth crown structure conditional feature vector fusing the local occlusion morphology and the global dental arch pose, and then denoises the tooth crown noise point cloud guided by the tooth crown structure conditional feature vector, so as to make the obtained initial tooth crown point cloud fit the adjacent tooth structure and avoid'meaningless noise'; then the double-path multi-scale refinement network is adopted to fuse the adjacent tooth point cloud and the initial tooth crown point cloud, the geometric information of the adjacent tooth point cloud is used to improve the details and structural continuity of the initial tooth crown point cloud, and the tooth crown point cloud is obtained, the surface continuity of which is smooth and the boundary morphology thereof is seamlessly matched with the adjacent tooth, thereby solving the problems of uneven structure distribution and insufficient local details in the tooth crown point cloud generation.

[0036] The above content is only a summary of the technical scheme of the present application, in order to enable those skilled in the art to more clearly understand the technical scheme of the present application, and then can be implemented according to the content recorded in the specification and the drawings, and in order to make the above-mentioned purpose and other purposes, characteristics and advantages of the present application more easily understood, the following will be described in combination with the specific embodiments of the present application and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0037] The drawings only serve to show the principles, implementation modes, applications, characteristics and effects of the specific embodiments and other related contents of the present application, and cannot be considered as a limitation of the present application.

[0038] In the drawings of the specification:

[0039] Figure 1 The flowchart of the double-path tooth crown point cloud completion method described in the specific embodiments;

[0040] Figure 2 The structure diagram of the geometric perception conditional feature extraction module;

[0041] Figure 3 The structure diagram of the PCNet;

[0042] Figure 4 The structure diagram of the tooth crown point cloud generated by combining the conditional diffusion and the double-path refinement network. DETAILED DESCRIPTION

[0043] In order to specifically describe the possible application scenarios, technical principles, implementable specific schemes, and the purposes and effects that can be achieved of the present application, the following will be described in detail in combination with the specific embodiments listed and the drawings. The embodiments recorded in the present text are only used to more clearly illustrate the technical scheme of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application.

[0044] The term "embodiment" is mentioned herein means that the specific features, structures, or characteristics described in connection with the embodiment can be included in at least one embodiment of the present application. The term "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit the independence or association between other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, each technical feature mentioned in each embodiment can be combined in any way to form a corresponding implementable technical solution.

[0045] Unless otherwise defined, the meanings of the technical terms used herein are the same as those commonly understood by those skilled in the art to which the present application belongs; the use of related terms herein is only for the purpose of describing specific embodiments, and is not intended to limit the present application.

[0046] In the description of the present application, the phrase "and / or" is a description of the logical relationship between the objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases: A exists, B exists, and A and B exist at the same time. In addition, the character " / " herein generally represents that the associated objects before and after are a "or" logical relationship.

[0047] In the present application, terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, primary and secondary, or order relationship between the entities or operations.

[0048] In the present application, without more limitation, the "includes", "contains", "has" or other similar open expressions used in the sentence are intended to cover non-exclusive inclusion, and these expressions do not exclude the presence of other elements in the process, method or product including the described elements, so that the process, method or product including a series of elements can not only include those limited elements, but also include other elements not explicitly listed, or also include the elements inherent to such process, method or product.

[0049] As the same as the understanding in the "Guidelines for Examination", in the present application, the expressions "greater than", "less than", "exceed" and the like are understood as not including the number; the expressions "above", "below", "within" and the like are understood as including the number. In addition, in the description of the embodiments of the present application, the meaning of "multiple" is more than two (including two), and similar expressions related to "multiple" are also understood in this way, for example, "multiple groups", "multiple times" and the like, unless otherwise explicitly limited.

[0050] In the description of the embodiments of the present application, the spatial relative expressions used, such as "central", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", and the like, indicate the orientation or positional relationship shown in the specific embodiments or the drawings, and are only used for the convenience of describing the specific embodiments of the present application or for the convenience of the reader to understand, and do not indicate or imply that the indicated device or component must have a particular position, a particular orientation, or be constructed or operated in a particular orientation, and therefore cannot be understood as a limitation on the embodiments of the present application.

[0051] The processor described in the embodiments of the present application can be implemented by hardware, firmware, software or a combination thereof, and can use at least one of circuit, single or multiple application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), central processing units (CPU), controllers, microcontrollers, microprocessors, and other physical, biological or chemical structures that can realize the same or equivalent functions as the above-mentioned processors, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute part or all of the steps or any combination of the steps mentioned in the computer programs or methods of the embodiments of the present application.

[0052] The computer program involved in the embodiments can be stored in a computer device readable storage medium, including but not limited to magnetic disk, magnetic tape, magnetic card, floppy disk, flash memory, optical disc, optical card, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM) and electrically erasable programmable ROM (EEPROM) and the like, and also includes other biological, physical or chemical structures that can realize the same or equivalent functions as the above-mentioned storage medium, such as DNA, RNA, protein and the like units with information storage capability. In specific embodiments, the storage medium can be one of the above-mentioned medium types, or a combination of the above-mentioned medium types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributedly stored in multiple media. The storage medium containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built-in in the device, or connected with the device as an external device or part of the external device. In some embodiments, the storage medium with the computer device readable storage medium is deployed locally; in other embodiments, the storage medium can also be deployed remotely from the processor, such as network-attached storage accessed via RF circuit or external port and communication network, wherein the communication network can be the Internet, one or more intranets, local area network (LAN), wide area network (WAN), storage area network (SAN) and the like, or appropriate combination thereof, as long as the access of the computer device to the storage medium can be realized. In addition, the computer program involved in the embodiments can be stored in plaintext / encrypted form, or can be designed as training data, and integrated and reorganized in the parameter state of the deep neural network or other machine learning model by model training.

[0053] A double-path dental crown point cloud completion method aims to reconstruct the three-dimensional structure of the missing middle dental crown based on the adjacent dental point cloud. Mainly extract multi-scale structural features from the adjacent dental point cloud, generate dental crown structure conditional feature vectors that fuse local occlusion morphology and global dental arch pose, and then guide the noise point cloud of the dental crown to denoise with the dental crown structure conditional feature vector, so as to obtain an initial dental crown point cloud that fits the adjacent dental structure and avoids "meaningless noise"; then a double-path multi-scale refinement network is used to fuse the adjacent dental point cloud and the initial dental crown point cloud, and the geometric information of the adjacent dental point cloud is used to improve the details and structural continuity of the completed initial dental crown point cloud, to obtain a dental crown point cloud with smooth surface continuity and seamless matching of the adjacent dental boundary morphology, solving the problems of uneven structure distribution and insufficient local details in the generation of the dental crown point cloud.

[0054] The following embodiments of a double-path dental crown point cloud completion method are provided in combination with Figure 1 , which includes the following steps:

[0055] S10, obtain the adjacent tooth point cloud;

[0056] S20, extract information in the adjacent tooth point cloud to obtain a crown structure condition feature vector;

[0057] S30, initialize the crown noise point cloud;

[0058] S40, denoise the crown noise point cloud guided by the crown structure condition feature vector to obtain an initial crown point cloud;

[0059] S50, input the adjacent tooth point cloud and the initial crown point cloud into a double-path multi-scale refinement network to obtain a crown point cloud.

[0060] In step S10, the adjacent tooth point cloud refers to a high-precision spatial coordinate set of the surface of the healthy teeth on both sides of the missing crown, which fully represents the geometric shape and spatial pose of the adjacent teeth in the form of discrete points. The adjacent tooth point cloud can be obtained by three-dimensional digital scanning technology, and can be obtained by directly collecting by an intraoral scanner, scanning and reconstructing a plaster model, etc.

[0061] In step S10, the crown structure condition feature vector not only contains the geometric shape constraint of the adjacent teeth (to ensure that the missing crown and the adjacent teeth are complementary in shape), but also contains the spatial position relationship (to ensure that the reconstruction result naturally connects with the adjacent teeth in the three-dimensional space), thereby providing accurate local constraint basis for subsequent refinement reconstruction of the missing crown, and effectively avoiding physiological structure contradictions such as interference or excessive gap between the adjacent teeth and the missing crown.

[0062] The above extraction of information in the adjacent tooth point cloud to obtain the crown structure condition feature vector is the core of accurately capturing key constraint information related to the reconstruction of the missing crown from the adjacent tooth three-dimensional point cloud. It can be realized through the cooperative process of multi-scale feature coding and geometric feature analysis. Specifically, first, the adjacent tooth point cloud is preprocessed, including denoising, downsampling, and key point extraction (such as focusing on the adjacent surface contact area, the crown edge, the occlusal surface ridge, and other areas closely related to the missing crown); then, multi-scale feature extraction is performed through a multi-layer perceptron and a convolutional neural network, capturing local curvature, normal vector, and other microscopic geometric information (such as the arc-shaped depression of the adjacent surface contact area and the sharpness of the cusp) of the adjacent tooth surface at a low scale, and aggregating the overall spatial pose (such as the relative position and inclination angle of the missing crown) and the topological structure (such as the crown height and the neck edge shape) of the adjacent tooth at a medium-high scale; finally, the multi-scale features are compressed and integrated into a fixed-dimensional crown structure condition feature vector.

[0063] Referring to Figure 2 In some embodiments, a geometric perception condition feature extraction module is used to extract information in the adjacent tooth point cloud to obtain a crown structure condition feature vector. Specifically, the geometric perception condition feature extraction module can include an FPS, a dynamic graph convolutional neural network, and a Transformer encoder.

[0064] The FPS (Farthest Point Sampling) down-samples the neighboring tooth point cloud to obtain a sparse neighboring tooth point cloud; that is, an initial point is randomly selected from the input original neighboring tooth point cloud (high-density, unordered point set) as a "seed", and the point farthest from the selected point set is iteratively selected until the target number of center points is reached, so as to reduce the number of neighboring tooth point clouds, reduce the subsequent calculation amount, retain the sparse neighboring tooth point cloud with "uniform distribution and full coverage", and avoid local point redundancy.

[0065] The dynamic graph convolutional neural network (DGCNN) dynamically constructs the connection relationship between points based on the distance of each point in the sparse neighboring tooth point cloud to find the local neighboring points (such as KNN neighborhood), aggregates the neighborhood point features to make the features of each point no longer isolated, but fused with the local geometric structure (such as neighborhood shape, distance relationship), and forms a neighboring tooth point cloud vector. The above neighboring tooth point cloud vector not only contains local geometric information, but also has position coding capability, which is helpful for spatial structure understanding.

[0066] The Transformer encoder introduces a KNN mechanism for self-attention to interact and fuse the local and global features of the neighboring tooth point cloud vector to obtain a dental crown structure condition feature vector. Specifically, the neighboring tooth point cloud is first converted into a high-dimensional vector representation, and each vector contains the spatial coordinates, normal vector, curvature and other geometric properties of the point; then, in the calculation of the attention weight, the self-attention mechanism does not perform global correlation on all point cloud vectors, but selects the K nearest points (such as 32 neighboring points) around each point through the KNN mechanism, and only calculates the attention interaction within the local neighborhood, which enables the model to accurately capture the local details of the neighboring tooth surface (such as the arc-shaped depression of the adjacent surface contact area and the curvature mutation of the dental crown edge); at the same time, through the stacking of multiple encoders, the local features are transmitted and aggregated layer by layer, gradually integrating the global spatial pose information of the neighboring tooth (such as the relative position of the neighboring tooth and the missing dental crown, the overall inclination angle, and the occlusion relationship), and finally compressing these multi-scale information into a fixed-dimensional dental crown structure condition feature vector. This vector not only contains the local geometric constraints of the neighboring tooth and the missing dental crown directly related to each other (to ensure morphological complementarity), but also contains global spatial layout information (to ensure the coordination of the three-dimensional position), providing accurate structural condition guidance for subsequent dental crown reconstruction.

[0067] In step S30, the crown noise point cloud refers to generating a set of three-dimensional point sets containing random disturbances in the initial stage of three-dimensional reconstruction of the crown. The distribution, coordinates and density of the crown noise point cloud have randomness (i.e., "noise") and do not have the fine geometric structure of the real crown (such as the features of the cusp, the pit and groove, and the proximal contact area). For example, the crown noise point cloud can be random Gaussian noise without specific structure, providing an initial structure "skeleton" for subsequent refinement. The crown noise point cloud helps to better learn the topological structure and geometric constraints of the crown, avoids falling into a local optimal solution, and is especially suitable for the crown completion scene with large missing areas or insufficient initial information.

[0068] In step S40, the step of denoising the crown noise point cloud guided by the crown structure condition feature vector to obtain the initial crown point cloud learns the neighboring tooth constraints and global morphological rules through the crown structure condition feature vector, iteratively optimizes the point coordinates in the crown noise point cloud, eliminates invalid noise, and strengthens effective structure features, and finally evolves into an initial crown point cloud that conforms to the physiological morphology. In some embodiments, the conditional guided diffusion generation module can be used to perform denoising on the crown noise point cloud. That is, the step-by-step reverse progressive denoising of the crown noise point cloud is performed by using the crown structure condition feature vector to guide the conditional guided diffusion generation module, and the initial crown point cloud is obtained. The conditional guided diffusion generation module realizes the denoising reconstruction of the crown point cloud based on the conditional diffusion generation framework (CDPM), and the core logic module takes the crown noise point cloud as the starting point, corrects the point cloud at each step by using the crown structure condition feature vector (such as the local shape and spatial layout of the neighboring teeth), continuously reduces the noise, and after multiple rounds of denoising iteration, the crown noise point cloud gradually converges, and the initial crown point cloud with the basic morphology of the crown is output. Specifically, an architecture based on a point-by-point conditional network is used, which includes six independent PCNet modules to realize controllable generation at the point level. That is, the conditional guided diffusion generation module includes six PCNets, each point of the crown noise point cloud is independently input into the six PCNets, and the crown structure condition feature vector and the timestamp T are simultaneously input;

[0069] The six PCNets output the noise residual of each point to update the point cloud independently:

[0070] Starting from the maximum noise timestamp T, repeat the following steps until T = 0;

[0071] The noisy point cloud X t The input conditional guided diffusion generation module is denoised point by point by the six PCNets to obtain X t-1 .

[0072] In the six PCNets, each point updates the features independently without relying on the "average / aggregation" of neighboring points (unlike U-Net and other convolution-dependent architectures). Independent updating can more accurately preserve / recover details and avoid "cross-point interference" (such as the features of dense points being incorrectly transmitted to sparse points) caused by uneven point cloud density (such as dense points on the convex and concave surfaces of the crown and sparse points in the sparse area). Combined with the condition feature vector of the crown structure, the denoising of each point is consistent with the complex surface of the crown (such as the smooth surface and gully details of the crown). See Figure 3 As shown, the core structure of PCNet includes hierarchical residual connection and gating mechanism to achieve efficient feature extraction and conditional constraints. Among them, the feature channel dimension dynamically changes according to the progressive rule of [3, 128, 256, 512, 256, 128, 3], forming a "shrinkage-expansion" symmetric structure: the input point cloud coordinates (X, Y, Z) are first upgraded by 128-dimensional and 256-dimensional layers, and the depth aggregation of global and local geometric features (such as the spatial distribution of adjacent contact areas and the density gradient of the initial point cloud) is completed in the 512-dimensional channel layer. Then, through the 256-dimensional and 128-dimensional layers, the dimension is reduced, and finally the corrected point cloud coordinates are output. In this process, the hierarchical residual connection directly transmits the shallow features (such as the position information of the original point cloud) to the deep layer through the jump connection, avoiding the loss of information during multi-scale feature extraction, and ensuring that the geometric structure at different scales (from the global outline to the local details) can be effectively captured. The reset gate in the gating mechanism dynamically controls the retention ratio of historical features in each layer, and its calculation method is:

[0073] H l =R l (W lh H l-1 +b lh )+W lb y,R l =σ(W lr y+b lr );

[0074] Where y is the concatenation result of the condition feature and the time embedding, σ(·) is the Sigmoid activation function, and the weights W . and the bias b . are learnable parameters. By directly injecting condition information into the feature calculation process, the model can more finely adjust the point cloud restoration behavior at different time steps, focusing on the global outline in the high-noise stage and gradually refining the local geometric details in the low-noise stage, achieving more accurate point cloud restoration.

[0075] The condition-guided diffusion generation module can use Markov chains to make the noisy point clouds X t and X t-1 at adjacent time stamps satisfy the "point-by-point conditional probability" (i.e., X t-1The distribution of only depends on X t ). This point-wise mapping guarantees the "progressiveness" of the denoising process, making the initial crown point cloud generation more stable.

[0076] The step of denoising the crown noise point cloud guided by the crown structure condition feature vector to obtain the initial crown point cloud comprises the following steps:

[0077] The denoising process continues for T time steps to the crown noise point cloud X T The denoising process continues for T time steps to the crown noise point cloud X

[0078] The joint distribution of the denoising process:

[0079]

[0080] Where θ represents the learnable parameters of the condition-guided diffusion generation module; p θ (·) represents the conditional probability distribution of the reverse denoising; X 0:T represents the sequence from the initial crown point cloud X0to the crown noise point cloud X T ; p θ (X 0:T |c) represents the joint probability distribution of the sequence from the initial crown point cloud X0to the crown noise point cloud X T given the crown structure condition feature vector c, parameterized by the parameters θ; p(X T ) represents the initial noise distribution; c, z represents the crown structure condition feature vector; X t represents the noisy point cloud at the t-th step; X t-1 represents the noisy point cloud at the (t-1)-th step; represents the conditional probability distribution of the previous time step point cloud X t given the current noisy point cloud X t-1 and the crown structure condition feature vector c, during the process from time step t=1 to t=T;

[0081] The conditional distribution of the denoising process:

[0082] p θ (X t-1 |X t ,c)=N(X t-1 |μ θ (X t ,t,c),β t I);

[0083] Where, represents a Gaussian distribution; μ θ represents the denoising mean predicted by the condition-guided diffusion generation module; μ θ (X tμt,c) denotes the mean function of the Gaussian distribution, predicted by the conditional guided diffusion generation module; β t denotes the noise scheduling parameter, I is the identity matrix; β t I denotes the covariance matrix;

[0084] The training objective aims to maximize the lower bound of the log-likelihood:

[0085] E q [logp θ (X θ |X T ,c)];

[0086] By minimizing the mean square error (MSE) loss between μ θ and μ of the standard normal distribution, the initial crown point cloud is obtained.

[0087] In step S50, the adjacent tooth point cloud and the initial crown point cloud are input into the double-path multi-scale refinement network to obtain the crown point cloud. This is achieved by two parallel and interactive feature processing paths that fuse adjacent tooth constraints and initial morphological information at multiple scales to realize the structural reconstruction from the initial crown point cloud (coarse) to the crown point cloud (fine). One path focuses on analyzing the geometric features of the adjacent tooth point cloud, extracting spatial position, surface curvature, adjacent surface contact area morphology and other constraint information of the adjacent tooth through multi-scale convolution or graph neural network (such as capturing the micron-level gap between the adjacent tooth and the target crown contact edge at a small scale, and obtaining the three-dimensional pose of the overall arrangement of the adjacent tooth at a large scale); the other path extracts features from the initial crown point cloud, and analyzes its existing profile at different scales (such as correcting the overall volume deviation of the initial point cloud at a large scale, and optimizing the density distribution of the local point cloud at a small scale). The two paths realize feature interaction through cross-layer attention mechanism. The contact area constraint extracted by the adjacent tooth path will guide the initial crown path to adjust the point cloud distribution in the corresponding area (to avoid interference with the adjacent tooth), and the global profile features of the initial crown path will react on the weight distribution of the adjacent tooth constraint (highlight the constraint priority of key areas). After multiple rounds of multi-scale feature fusion and iterative optimization, the network gradually corrects the noise points in the initial point cloud, supplements the missing structure (such as generating the corresponding cusp and sulcus according to the occlusion relationship of the adjacent tooth), and finally outputs the crown point cloud that not only conforms to the physiological morphological consistency (such as crown height and neck margin line position), but also maintains a natural adjacent relationship with the adjacent tooth (such as seamless fitting of the contact area and uniform gap).

[0088] Referring to Figure 4As shown, the above double-path multi-scale refinement network can adopt a double-path ReFinement Network (RFNet) structure, and realize local refinement of the missing dental crown by cooperating with the coding of the coarse point cloud and the adjacent tooth geometry. Specifically, it can include a feature extraction subnetwork, a feature transfer module, and a decoding module; the feature extraction subnetwork includes an adjacent tooth path and an initial dental crown path, which independently performs hierarchical down-sampling and feature extraction on the input adjacent tooth point cloud and initial dental crown point cloud, respectively; the feature transfer module uses an attention mechanism to make the points on the initial dental crown path focus and fuse relevant geometric features from the corresponding region of the adjacent tooth path; and the decoding module performs hierarchical up-sampling, fuses the features of the corresponding level of the encoder through a skip connection, refines the features using an MLP, and gradually restores the point cloud density to obtain the dental crown point cloud.

[0089] However, the dental crown point cloud has higher surface continuity requirements and more complex boundary morphology than the general object completion task. In order to better adapt to the geometric characteristics of the dental crown scene, the key structural parameters of the RFNet are systematically reconstructed and optimized. Specifically, in the feature extraction subnetwork, the sampling number of the original conditional point cloud is first increased to enable the network to more fully perceive the adjacent tooth structure information; at the same time, the receptive field radius of the corresponding layer is reduced to enhance the ability to depict local details. In terms of feature dimension, the number of channels of all main stems and decoding modules is expanded to improve the expression ability and nonlinear modeling depth of the network. That is, the sampling point number of the feature extraction subnetwork is increased, the receptive field radius of the corresponding layer is reduced, the number of channels of the main stem is expanded, and the number of channels of the decoding module is expanded synchronously with the main network.

[0090] In addition, the attention-weighted neighborhood construction parameter K is expanded on the default basis (the default is 8, which can be expanded to 32), that is, the attention-weighted neighborhood construction parameter K of the feature transfer module is 32. This enables the feature transfer module (FT) to obtain more rich context information in the local structure. Not only does this improve the refinement accuracy of the dental crown point cloud, but it also improves the modeling ability of the model in terms of the continuity of the dental crown surface and the sharpness of the boundary.

[0091] The dental crown point cloud generated by the present application is closer to the real dental crown point cloud in terms of geometric accuracy, has better shape restoration ability and structural consistency in the dental crown point cloud completion task, and exhibits better geometric accuracy and matching degree. More importantly, from the point cloud distribution quality and the surface reconstruction effect based on the ball pivoting algorithm, the dental crown point cloud generated by the present application has a uniform and reasonable spatial distribution, avoids local over-densification and abnormal structure arrangement, has a moderate point spacing, and the reconstructed surface is continuous and smooth. This feature significantly improves the surface reconstruction effect based on ball pivoting, generates a complete and natural dental crown mesh, and better restores the geometric structure of the real dental crown.

[0092] Finally, it should be noted that the above-described embodiments are described in the specification and drawings of the application, but this does not limit the patent protection scope of the application. Any equivalent structure or equivalent process replacement or modification based on the essential concept of the application, using the content described in the specification and drawings of the application, and directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, etc., are all included in the patent protection scope of the application.

Claims

1. A dual-path crown point cloud completion method, characterized in that, Includes the following steps: Obtain the point cloud of adjacent teeth; Extract information from the point cloud of adjacent teeth to obtain the conditional feature vector of the crown structure; Initialize the crown noise point cloud; Denoising of the crown point cloud was performed using the crown structure condition feature vector as a guide to obtain the initial crown point cloud; The point clouds of adjacent teeth and the initial crown point cloud are input into a dual-path multi-scale refinement network to obtain the crown point cloud.

2. The dual-path crown point cloud completion method according to claim 1, characterized in that, The step of extracting information from the point cloud of adjacent teeth to obtain the conditional feature vector of the crown structure involves using a geometric perception conditional feature extraction module to extract information from the point cloud of adjacent teeth to obtain the conditional feature vector of the crown structure.

3. The dual-path crown point cloud completion method according to claim 2, characterized in that, The geometric perception conditional feature extraction module includes an FPS, a dynamic graph convolutional neural network, and a Transformer encoder. The FPS downsamples the adjacent tooth point cloud to obtain a sparse adjacent tooth point cloud. The dynamic graph convolutional neural network finds local neighbors based on distance for each point in the sparse adjacent tooth point cloud, dynamically constructs the connection relationship between points, and obtains the adjacent tooth point cloud vector. The Transformer encoder introduces the self-attention mechanism of KNN to perform local and global feature interaction and fusion on the adjacent tooth point cloud vector to obtain the crown structure conditional feature vector.

4. The dual-path crown point cloud completion method according to claim 1, characterized in that, The step of denoising the crown noise point cloud with the crown structure condition feature vector as a guide to obtain the initial crown point cloud includes using the crown structure condition feature vector to guide the condition-guided diffusion generation module to perform progressive reverse denoising on the crown noise point cloud to obtain the initial crown point cloud.

5. The dual-path crown point cloud completion method according to claim 4, characterized in that, The condition-guided diffusion generation module includes six PCNets. Each point of the crown noise point cloud is independently input into the six PCNets, and the crown structure conditional feature vector and timestamp T are input simultaneously. The point cloud is updated independently using the noise residual of each of the six PCNet outputs: Starting from the timestamp of maximum noise, T, repeat the following steps until T = 0; Noisy point cloud X t The input condition-guided diffusion generation module, after point-by-point denoising using six PCNets, yields X. t-1 .

6. The dual-path crown point cloud completion method according to claim 5, characterized in that, The PCNet includes hierarchical residual connections and gating mechanisms; Its feature channel dimensions dynamically change according to a progressive pattern of [3, 128, 256, 512, 256, 128, 3], forming a symmetrical "contraction-expansion" structure. The reset gate in the gating mechanism dynamically controls the retention ratio of historical features in each layer, and its calculation method is as follows: H l =R l (W lh H l-1 +b lh )+W lb y,R l =σ(W lr y+b lr ); Where y is the concatenation result of conditional features and temporal embedding, σ(·) is the Sigmoid activation function, and the weights W and biases b are learnable parameters.

7. The dual-path crown point cloud completion method according to claim 4, characterized in that, The step of denoising the crown noise point cloud guided by the crown structure feature vector to obtain the initial crown point cloud includes the following steps: The denoising process continuously moves towards the crown noise point cloud X within time step T. T Denoising was performed to obtain the initial crown point cloud X0, which satisfies the following conditions: Joint distribution of the denoising process: Where θ represents the learnable parameters of the conditionally guided diffusion generation module; p θ (·) represents the conditional probability distribution for inverse denoising; X 0:T This represents the transition from the initial crown point cloud X0 to the crown noise point cloud X. T The sequence; p θ (X 0:T |c) represents the given crown structural feature vector c, from the initial crown point cloud X0 to the crown noise point cloud X0. T The joint probability distribution of the sequences is parameterized by parameter θ; p(X) T ) represents the initial noise distribution; c represents the crown structure conditional eigenvector; X t X represents the noisy point cloud at step t; t-1 This represents the noisy point cloud at step t-1; This represents the process from time step t=1 to t=T, given the current noisy point cloud X. t And the crown structure condition feature vector c, to obtain the point cloud X of the previous time step. t-1 The conditional probability distribution; Conditional distribution of the denoising process: p θ (X t-1 |X t ,c)=N(X t-1 |μ θ (X t ,t,c),β t I); in, μ represents a Gaussian distribution. θ μ represents the denoised mean predicted by the conditionally guided diffusion generation module. θ (X t ,t,c) represents the mean function of a Gaussian distribution, predicted by the conditionally guided diffusion generation module; β t Represents the noise scheduling parameters, where I is the identity matrix; β t I represents the covariance matrix; The training objective is to maximize the lower bound of the log-likelihood. E q [log p θ (X θ |X T ,c)]; By minimizing μ of the standard normal distribution θ The mean square error (MSE) loss between μ and μ is used to calculate and obtain the initial crown point cloud.

8. The dual-path crown point cloud completion method according to claim 1, characterized in that, The dual-path multi-scale refinement network includes a feature extraction subnetwork, a feature transfer module, and a decoding module; the feature extraction subnetwork includes adjacent tooth paths and initial crown paths, which independently perform hierarchical downsampling and feature extraction on the input adjacent tooth point clouds and initial crown point clouds respectively; The feature transfer module uses an attention mechanism to focus points on the initial crown path and fuse relevant geometric features from the corresponding regions of adjacent tooth paths; the decoding module performs hierarchical upsampling, fuses features from the corresponding level of the encoder through skip connections, refines features using MLP, and gradually restores point cloud density to obtain crown point cloud.

9. The dual-path crown point cloud completion method according to claim 1, characterized in that, The number of sampling points in the feature extraction subnet is increased, the receptive field radius of the corresponding layer is reduced, the number of backbone channels is expanded, and the number of channels in the decoding module is expanded synchronously with the backbone network.

10. The dual-path crown point cloud completion method according to claim 1, characterized in that, The attention-weighted neighborhood construction parameter K of the feature transfer module is expanded based on the default value.

Images