Three-dimensional tooth model segmentation method of double-branch geometric attention network based on centroid guidance

Abstract

A three-dimensional tooth model segmentation method of a double-branch geometric attention network based on centroid guidance is oriented to oral cavity three-dimensional scanning point cloud data and comprises the steps that firstly, normalization and normal vector estimation are conducted on the oral cavity three-dimensional point cloud data, a double-branch encoder for coordinate and normal decoupling is constructed, and a global topological structure and local geometric features are extracted respectively; secondly, a separable attention mechanism guided by the mass center is introduced into a coordinate branch so as to improve the distinguishing ability of adjacent teeth, and a graph convolution attention mechanism is introduced into a normal branch so as to strengthen the boundary expression of the teeth and gingiva; further, in the fusion stage, a mass center thermodynamic diagram and multi-scale feature aggregation are combined, and joint modeling of global and local structures is achieved; and finally, collaborative optimization of instance segmentation and centroid localization tasks is carried out through a dynamically weighted joint loss function. According to the method, the segmentation precision and robustness under complex cases are remarkably improved while the light weight is kept, and the automation and clinical practicability of oral diagnosis and treatment are enhanced.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically to a method for segmenting three-dimensional tooth point cloud instances based on a centroid-guided bi-branch geometric attention network, applicable to automated, high-precision instance segmentation of three-dimensional tooth models acquired by an oral scanner. Background Technology

[0002] With the rapid development of digital dental technology, accurate segmentation of 3D tooth models has become a crucial step in computer-aided treatment systems, significantly improving the accuracy of clinical diagnosis and treatment planning. Currently, 3D oral models are primarily acquired using intraoral scanners (IOS), and achieving automated and high-precision tooth instance segmentation based on these models has become a research focus. Accurate tooth instance segmentation has become a critical step in the digital dental workflow. Point cloud data is characterized by high noise, uneven distribution, and complex tooth structure, placing higher demands on automated segmentation methods.

[0003] Existing traditional tooth segmentation methods are mostly based on manual interaction, region growth, or contour-based geometric processing. These methods are prone to problems such as unclear boundaries or unstable segmentation in complex situations such as crowded, missing, rotated, or irregularly arranged teeth, making it difficult to meet the clinical needs for high precision and high robustness.

[0004] With the development of deep learning, convolutional neural networks (CNNs) and graph neural networks (GNNs) have been introduced into 3D tooth point cloud processing tasks, enabling the extraction of semantic features from point clouds through learning and improving the automation capabilities of the model. However, existing deep learning-based 3D tooth segmentation methods still have the following shortcomings: 1. Insufficient utilization of multi-source geometric features: Geometric information such as coordinates and normal vectors in point clouds are often processed by simple splicing or single branch in existing models, making it difficult to take into account both global structure and local boundary details.

[0005] 2. Lack of spatial prior modeling: The centroid layout and adjacency relationships of teeth in the oral cavity are usually not explicitly utilized, which makes the boundary areas between different teeth easy to be confused.

[0006] 3. Weak feature fusion mechanism: Some methods adopt single-branch or fixed fusion strategies, which are not conducive to the expression of complex geometric shapes.

[0007] 4. Inadequate loss design to balance localization and segmentation tasks: Existing methods generally use a single loss or fixed weights, which is not conducive to improving instance segmentation accuracy and spatial localization stability simultaneously in complex scenarios.

[0008] Therefore, existing technologies urgently need a three-dimensional tooth instance segmentation method that can fully integrate geometric features, introduce spatial structure priors, and possess strong robustness to adapt to complex clinical situations such as dense dentition, abnormal arrangement, and blurred boundaries in oral scan data. Summary of the Invention

[0009] To address the shortcomings of existing 3D tooth segmentation methods, such as insufficient geometric feature fusion, lack of spatial structure modeling, and poor adaptability to complex clinical cases, this invention proposes a centroid-guided bi-branch geometric attention network based on deep learning.

[0010] This invention fully utilizes multi-source geometric information such as coordinates and normal vectors in a 3D model. It extracts semantically expressive features by decoupling a dual-branch network structure and introduces the centroid distribution of teeth as a spatial prior, improving the model's robustness in identifying abnormally arranged cases such as crowding, missing teeth, and misalignment. Simultaneously, it employs a dynamic weighted fusion and joint loss mechanism to coordinate segmentation and localization tasks, significantly improving tooth boundary discrimination ability and instance segmentation accuracy. The method of this invention has a reasonable structure and efficient training, enabling accurate segmentation of teeth in 3D oral cavity scan data, providing reliable technical support for digital oral diagnosis and treatment.

[0011] This invention constructs a centroid-guided bi-branch geometric attention network (CenGNet) that extracts global structural and local detail information from the coordinates and normal vectors of tooth point clouds, respectively. It integrates graph convolutional attention mechanisms and multi-scale feature strategies, and introduces centroid priors to enhance boundary discrimination. Through dynamic weighting and joint loss, it collaboratively performs segmentation and localization tasks, improving robustness and accuracy in complex cases.

[0012] A 3D tooth instance segmentation method based on a centroid-guided bi-branch geometric attention network includes the following steps: S1. Acquire 3D tooth point cloud data from an oral scanner, normalize the point cloud and estimate the normal vector, and construct a point cloud feature matrix containing coordinates and normal vectors; S2. Construct a dual-branch geometric attention network; the dual-branch geometric attention network includes a coordinate branch and a normal vector branch, wherein the coordinate branch adopts a multi-layer stacked graph convolution structure to capture the global spatial topological relationship of the tooth point cloud, and maintains the stability of deep feature transmission through cross-layer residual connections and batch normalization operations; the normal vector branch is based on local neighborhood convolution and graph attention mechanism to enhance the expression of subtle geometric differences at the tooth boundary, and also introduces residual and normalization modules to improve training convergence performance; the features output by the two branches are used in the subsequent fusion stage to construct complementary representations of global structural information and local boundary detail information; S3. Input the point cloud feature matrix into the coordinate branch and the normal vector branch respectively, wherein the coordinate branch extracts global geometric structure information and the normal vector branch extracts local boundary geometric details; S4. Calculate the centroid of the tooth based on the spatial distribution of each tooth instance in the training sample, and generate a centroid heatmap based on the Gaussian function, which is then used as a spatial prior guiding signal input into the network. S5. Introduce graph attention convolution mechanism in the backbone network and combine it with multi-scale feature fusion strategy to achieve adaptive fusion of dual-branch features, so as to simultaneously model global topology and local boundary information; S6. Construct a joint loss function consisting of segmentation loss and centroid prediction loss, and achieve task balance through dynamic weight adjustment, thereby outputting the tooth instance segmentation result and its corresponding centroid location.

[0013] The generation method of the centroid heatmap in step S4 is as follows: First, the centroid coordinates of each tooth instance in the training samples are calculated based on the point cloud distribution, and the centroids are mapped to a normalized three-dimensional spatial coordinate system; then, a three-dimensional Gaussian kernel is constructed with each centroid as the center to form a continuous spatial prior heatmap, the expression of which is:

[0014] Where p represents the position of any point. Let be the centroid coordinates of the i-th tooth, and N be the total number of teeth. The diffusion coefficient is used. The heatmap is input into the model as a spatial prior guiding signal during network training to enhance the boundary discrimination ability of adjacent tooth regions and the feature response of weak boundary regions.

[0015] In step S5, a graph attention convolution mechanism is introduced in the dual-branch feature fusion stage. The feature transfer weights are adaptively adjusted according to the geometric relationship between each point and its neighborhood in the point cloud. In addition, a multi-scale feature aggregation strategy is combined to integrate local geometric features and global topological structure at different scales to obtain a fused feature representation that takes into account both the overall tooth structure and boundary details.

[0016] In step S6, the joint loss function is used to simultaneously constrain the accuracy of tooth instance segmentation and centroid localization, and is determined by the segmentation loss. With centroid predicted loss Composition, which is defined as follows:

[0017] in, and These are weighting coefficients that are dynamically adjusted during the training process. They are used to adaptively balance segmentation accuracy and centroid heatmap prediction accuracy at different training stages, thereby improving the accuracy and stability of the overall segmentation results.

[0018] Furthermore, the input point cloud described in step S1 is constructed in the form of an N×6 dimensional feature matrix, where the three-dimensional coordinates of each point and the corresponding three-dimensional normal vector are concatenated by channel as network input and sent to the coordinate branch and the normal vector branch respectively, so as to provide global spatial structure information and local surface geometric information simultaneously.

[0019] This invention is applicable to complex clinical cases such as crowded, missing, and misaligned teeth. By using the spatial prior constraints provided by the centroid heatmap and the graph attention convolution mechanism to adaptively enhance the boundary region, it improves the ability to distinguish fuzzy boundaries between adjacent teeth, thereby maintaining high instance segmentation accuracy and robustness under the above-mentioned complex dentition conditions.

[0020] The method of the present invention includes: (1) Dual-branch geometric feature extraction module To address the different types of geometric information in a 3D tooth model, a decoupled dual-branch structure is designed to process point cloud coordinates and normal vectors separately. The coordinate branch captures global spatial structure information through stacked graph convolutional modules, while the normal vector branch focuses on extracting local geometric details. Residual connections and normalization operations are employed within each branch to improve the stability and depth representation of feature extraction. The two branches extract high-dimensional feature representations from different geometric perspectives in parallel, with structural symmetry and semantic complementarity, laying the foundation for subsequent fusion and segmentation.

[0021] (2) Generation of centroid heat map and spatial prior modeling

[0022] The model incorporates the centroid distribution of teeth as spatial prior knowledge to guide feature fusion and boundary discrimination. Specifically, it first calculates the centroid coordinates of each tooth instance in the training samples and maps these centroids to a standardized 3D space to generate a continuous heatmap. The value of each point in this heatmap is based on the distance between the point and each centroid, weighted by a Gaussian distribution function, thus forming a prior distribution that gradually decays with spatial distance. During training, this heatmap is used in conjunction with the feature fusion module as additional guiding information, enabling the network to pay more attention to the central region of the teeth and regions with blurred boundaries, thereby enhancing the ability to separate instances.

[0023] (3) Multi-scale feature fusion and geometric attention guidance

[0024] To enhance the expressive power of cross-scale and cross-modal features, the model constructs a multi-scale fusion module and introduces a centroid heatmap to adaptively weight and fuse the features extracted from the two branches along the channel dimension. Specifically, the centroid heatmap is first nonlinearly mapped to obtain weight coefficients, which are then used to control the proportion of coordinate branch and normal vector branch features in each channel. This allows the two features to be superimposed in different proportions to form fused features, thereby preserving both global structural information and local geometric details in the same representation.

[0025] In addition, the model integrates graph convolutional attention mechanism in the fusion stage, combining the inter-point connection weights and geometric similarity to adaptively adjust the propagation path of information in the neighborhood, thereby enhancing the ability to distinguish tooth instances with similar shapes but different internal structures.

[0026] (4) Joint optimization mechanism and loss function design

[0027] To collaboratively complete the tasks of tooth instance segmentation and centroid localization, the model employs a joint loss mechanism based on dynamic weight adjustment. Specifically, the segmentation subtask uses classification loss functions such as cross-entropy loss, while the centroid heatmap prediction subtask uses regression loss functions such as smoothing L1 loss. The two are then weighted and combined according to adjustable weights; these weight parameters can be adaptively adjusted according to the training stage or task complexity.

[0028] Through this joint optimization mechanism, the model can simultaneously take into account both tooth boundary recognition and central region localization during training, significantly improving the accuracy and robustness of segmentation results in complex clinical cases.

[0029] This invention introduces three key mechanisms at the engineering implementation level to enhance the geometric modeling capability and robustness of instance segmentation for 3D tooth point clouds: First, a dual-channel feature extraction path with coordinate-normal decoupling is designed, inputting the 3D coordinates and normal vectors of the tooth point cloud into independent encoding networks to model the global topology and local boundary details respectively, thereby improving the model's structural recognition capability in complex scenarios such as crowding and missing teeth; Second, a spatial heatmap module based on the centroid distribution of teeth is constructed, explicitly encoding prior location information as guiding signals to highlight the model's attention to blurred boundary regions, effectively improving the segmentation accuracy in abnormally arranged cases; Third, a graph attention convolution (GAC) module is introduced into the backbone network, and combined with a dynamic weighted fusion mechanism and a segmentation-localization collaborative loss function design, to achieve joint optimization of feature alignment and task coupling, thereby improving network training stability and prediction accuracy.

[0030] Based on the overall network architecture, the specific implementation process of this invention includes the following steps: (1) Dual-channel geometric feature encoding The input 3D tooth point cloud data is processed to extract coordinate and normal vector information, which are then fed into two independent encoding branches. The coordinate branch focuses on capturing the overall geometric shape of the point cloud, while the normal vector branch is used to model local surface details. The two branches are decoupled to improve the ability to identify boundary regions.

[0031] (2) Centroid Prior-Guided Modeling

[0032] A centroid heatmap is generated based on the spatial distribution of teeth and used as an additional spatial constraint input to the network. This module guides the model to focus on tooth edges and adjacent regions during feature learning, thereby improving segmentation stability in scenarios with dense, missing, or misaligned teeth.

[0033] (3) Graph structure attention and multi-scale fusion

[0034] A Global Area Combination (GAC) module is introduced into the backbone network, combined with a multi-scale feature aggregation mechanism, to achieve joint modeling of global and local features. Through adaptive weight allocation, the complementarity of features in terms of spatial structure and detail representation is enhanced.

[0035] (4) Dynamic fusion and joint optimization

[0036] In the decoding stage, a dynamic weighted fusion strategy is employed, combined with a joint loss function for instance segmentation and centroid localization, to achieve feature alignment and task collaboration. This mechanism improves segmentation accuracy while ensuring the stability of the training process and overall inference efficiency.

[0037] In summary, the centroid-guided bi-branch geometric attention network proposed in this invention achieves significant performance improvements in 3D tooth point cloud segmentation tasks. By decoupling coordinate and normal vector features, and combining centroid prior spatial constraints with graph structure attention mechanisms, the model demonstrates excellent advantages in tooth boundary clarity, instance segmentation accuracy, and adaptability to complex cases. Experimental results show that the invention not only achieves high-precision segmentation results in conventional dentition samples but also exhibits strong robustness in abnormal scenarios such as crowding, missing teeth, and misalignment. Furthermore, the dynamic fusion and joint loss optimization strategies employed make the training process more stable and inference efficiency higher, demonstrating good potential for engineering implementation. Overall, the method of this invention can provide reliable technical support for 3D digital oral diagnosis and treatment, possessing significant clinical application value and promising prospects for widespread adoption.

[0038] The innovation of this invention is: (1) This invention proposes a spatial prior guidance mechanism based on centroid heatmap, which differs from traditional segmentation methods that rely solely on geometric features or spatial location. It utilizes the center coordinates of tooth instances to construct a continuous heatmap through a Gaussian kernel function, which is then embedded into the network encoder as a guiding feature. This effectively focuses on the blurred areas of instance boundaries, improving boundary recognition capability and segmentation accuracy.

[0039] (2) The present invention constructs a dual-branch geometric coding structure with coordinate and normal decoupling. Unlike the traditional point cloud feature extraction method that directly splices the input, the two branches learn global topology and local boundary geometric information respectively. Feature compensation and stable fusion are achieved through residual connection and normalization module, which improves the robustness of the model to complex cases such as crowding and missing data.

[0040] (3) This invention introduces a GAC ​​module and a separable point attention mechanism into the backbone network, and realizes non-local interaction between features through a multi-scale graph structure. While improving the boundary discrimination ability, it keeps the model lightweight and efficient, and is suitable for clinical real-time segmentation scenarios.

[0041] (4) This invention proposes a loss function design that combines segmentation and center point localization, constructs a weighted combination of boundary segmentation loss and center point heatmap supervision, dynamically adjusts the weights during training to adaptively optimize the target, improves the spatial recognition integrity and positional accuracy of tooth instances, and enhances the clinical applicability and engineering deployment feasibility of the method.

[0042] The working principle of this invention is: By introducing centroid prior and a bi-branch coding structure, this invention achieves high-precision segmentation and localization of tooth instances.

[0043] (1) Feature extraction stage: The normalized 3D point cloud data is input into a dual-branch encoding structure. The coordinate branch focuses on extracting spatial distribution and global geometric shape, while the normal vector branch focuses on details such as boundary normal abrupt changes and local curvature. Both branches use graph neural network modules for layer-by-layer encoding, and feature stability is improved through residual connections and normalization mechanisms.

[0044] (2) Spatial prior generation stage: Based on the point cloud distribution of tooth instances, calculate the geometric center coordinates of each instance, construct a continuous centroid heat map using Gaussian kernel function, and embed it as spatial attention guidance information into the coding features to enhance the network's response to the blurred regions of the target boundary.

[0045] (3) Feature fusion and decoding stage: In the fusion stage, the GAC module and separable point attention mechanism are introduced. Through multi-scale graph structure modeling and attention weight calculation, the feature adaptive weighted aggregation of the coordinate-normal two branches is realized, and the instance segmentation result and corresponding centroid localization information are output in a layer-by-layer upsampling manner.

[0046] (4) Joint optimization training mechanism: Construct a joint loss function, combine the boundary segmentation loss with the center point localization supervision signal in a weighted manner, and improve the localization accuracy and boundary clarity of the model in complex case scenarios through dynamic weight adjustment.

[0047] The core principle of this invention is to decompose, model, and optimize the global topology, boundary curvature, and centroid position in the point cloud spatial structure, and combine spatial prior guidance and attention mechanism to break through the recognition bottleneck of traditional methods in scenarios such as overlapping, missing, and unclear boundaries of adjacent teeth, thereby improving the overall accuracy, stability, and engineering deployment feasibility of the 3D tooth segmentation task.

[0048] The advantages of this invention are: while maintaining a lightweight design, it significantly improves the segmentation accuracy and robustness in complex cases, and enhances the automation and clinical applicability of oral diagnosis and treatment. Attached Figure Description

[0049] Figure 1 This is a framework diagram of the centroid-guided bi-branch geometric attention network structure in an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the tooth centroid prediction subnetwork structure.

[0051] Figure 3 A schematic diagram of a separable point self-attention module structure guided by the centroid.

[0052] Figure 4 This is a schematic diagram of the graph attention convolution (GAC) module.

[0053] Figure 5 This is a visual comparison of the segmentation results of the method of the present invention and existing technologies on a typical three-dimensional tooth model. Detailed Implementation

[0054] To make the technical solution of the present invention clearer, the present invention will be further described below with reference to specific embodiments. It should be understood that those skilled in the art can make various modifications or equivalent substitutions to the following embodiments without departing from the principles and spirit of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention. These embodiments are only used to illustrate the present invention and do not constitute a limitation on the protection scope of the present invention.

[0055] The present invention provides a three-dimensional tooth instance segmentation method based on a centroid-guided bi-branch geometric attention network, comprising the following steps: (1) Overall process of centroid-guided bi-branch geometric attention network To improve the instance segmentation accuracy and robustness of 3D tooth point clouds, this embodiment provides a centroid-guided bi-branch geometric attention network method. Its overall process is as follows: Figure 1 As shown, the process mainly includes stages such as input data preparation, geometric feature encoding, centroid guidance mechanism, multi-scale feature fusion, and joint decoding optimization. Combined with an end-to-end training strategy, it achieves high-precision instance segmentation of complex oral cavity point clouds.

[0056] First, the original tooth point cloud data undergoes normal vector estimation, uniform sampling, and normalization to obtain two types of input features: coordinates and normal vectors, at a uniform scale. To enhance the modeling ability for geometric details, a coordinate-normal decoupled dual-channel encoder is constructed to extract global structural information and local surface features from the tooth point cloud, respectively. The coordinate channel focuses on topological structure modeling, while the normal channel emphasizes surface normal variations, providing crucial support for boundary discrimination.

[0057] Secondly, based on the prior knowledge of the spatial distribution of tooth centroids in clinical experience, this embodiment introduces a centroid guidance module. By embedding the labeled centroids into the backbone network in the form of heatmaps, the model is guided to focus on areas with blurred boundaries or complex structures during training, which significantly improves the segmentation stability of abnormal morphologies such as misalignment, missing teeth, and dense dentition.

[0058] Furthermore, the model introduces a GAC ​​module during the feature fusion stage to dynamically aggregate spatially correlated node information within the local neighborhood, improving the consistency and recognizability of structural representation. Simultaneously, to enhance inter-branch complementarity and fusion efficiency, this embodiment employs a dynamic weighted fusion mechanism, adaptively fusing two types of geometric features according to task relevance. It also jointly designs instance segmentation loss and centroid localization loss, collaboratively optimizing the two tasks during training to improve the overall accuracy and stability of the model.

[0059] (2) Dual-branch geometric coding module

[0060] This embodiment addresses the heterogeneity of coordinates and normal vectors in the geometric representation of tooth point cloud data by designing a decoupled dual-channel encoding structure to model the global topology and local surface features respectively, thereby enhancing the semantic representation capability of the model under complex boundaries and abnormal structures.

[0061] During the input phase, the point cloud data is first split into two parts: 3D coordinates (x, y, z) and normal vectors (nx, ny, nz), which are then input into two parallel feature extraction paths. The coordinate branch employs a multi-layer graph convolutional structure to construct spatial topological relationships based on a K-nearest neighbor graph, and models the structural relationships between points through edge feature aggregation. The normal vector branch uses a lightweight convolutional stacking structure and introduces direction normalization to capture local differential changes in surface morphology. Both branches use residual connections to improve gradient propagation stability and scale alignment is achieved through a shared parameter layer.

[0062] In the intermediate fusion stage, the outputs of the two branches are concatenated through feature connection operations and then fed into the subsequent fusion module for processing, ensuring geometric and semantic alignment driven by the collaborative efforts of global and local features. This structure exhibits superior robustness during training, especially in abnormal scenarios such as broken, tilted, or tightly packed teeth, effectively extracting structural features while maintaining semantic consistency.

[0063] (3) Centroid Heatmap Module Design

[0064] To further enhance the model's ability to perceive critical boundaries between teeth, this embodiment designs a spatial heatmap encoding mechanism based on the prior of the tooth centroid, which serves as a guide for embedding information into the backbone network feature fusion process, significantly strengthening the model's ability to distinguish and locate complex instances.

[0065] Specifically, firstly, the centroid coordinates of the 3D point cloud for each tooth in the training set are statistically analyzed to calculate the average centroid position of each type of tooth in the standard oral coordinate system, thus constructing a category-level spatial prior distribution map. This distribution map is then diffused onto each centroid using a Gaussian kernel, forming a continuously differentiable heatmap template. For any point... The thermal value can be defined as:

[0066] Where p represents the position of any point. Let be the centroid coordinates of the i-th tooth, and N be the total number of teeth. The degree of heat distribution diffusion is controlled. The heatmap serves as a guiding signal during training, inputting along with the backbone features into subsequent modules to guide the network to focus on potential instance boundary regions.

[0067] During the inference phase, after the model performs coarse classification prediction on the input point cloud, it selects a heatmap template corresponding to the category and aligns it with the normalized spatial coordinates of the current input point cloud to generate a guiding heatmap with the same number of points. This heatmap serves as an additional geometric feature channel and is fused with the main branch features to guide the network to focus on possible instance boundary regions.

[0068] In addition, to enhance the module's adaptability to actual clinical variations (such as missing teeth, misalignment, etc.), learnable scaling factors and offset parameters can be introduced during the heatmap generation process to achieve fine-tuning of the template shape and center position, so that the heatmap can be dynamically adjusted during training to adapt to the spatial layout of different samples.

[0069] Experiments have shown that the heatmap module has significant effects in enhancing model boundary perception and improving segmentation accuracy, especially in clinical scenarios with densely packed multiple instances or blurred boundaries, where it is more stable and reliable.

[0070] (4) Design of graph attention fusion module

[0071] To achieve effective integration of cross-branch features and enhance the model's ability to jointly model local neighborhoods and global structures, this embodiment introduces a GAC ​​module into the backbone network and combines a multi-scale feature aggregation strategy to dynamically weight and fuse the outputs of the two branches.

[0072] Specifically, the GAC module first constructs adjacency relationships based on the K-nearest neighbor graph of the point cloud, measuring the geometric similarity between each point and its neighbors. During feature propagation, a learnable attention weight function is used to adaptively allocate the contributions of different neighbors, enabling the model to highlight features more semantically relevant to the target point and suppress redundancy or noise interference. Compared to traditional convolution operations, this mechanism can effectively capture long-range dependencies while preserving local details.

[0073] In the feature fusion stage, the weights generated by the centroid heatmap can be used to weight and superimpose the features of the coordinate branch and the normal branch to obtain the fused features. Specifically, it can be defined as:

[0074] in, and The feature maps representing the coordinate and normal vector branches are respectively. For heatmaps, This is the Sigmoid function.

[0075] During the fusion phase, the feature tensors output by the coordinate branch and the normal branch can be input into independent GAC units to obtain attention-enhanced representations of the two types of features through parallel computation. Subsequently, a channel-level weighting mechanism is used to jointly model the two types of features, and the feature fusion module completes the step-by-step merging of features at different resolutions to ensure the coordinated expression of local geometry and global topology.

[0076] Furthermore, this embodiment introduces residual connections and normalization operations in the GAC module to ensure gradient stability and feature distribution consistency during training. During inference, this module significantly improves the model's ability to distinguish boundary regions between teeth, especially demonstrating higher segmentation accuracy between adjacent teeth with similar morphology.

[0077] In summary, the graph attention fusion module effectively improves the robustness and boundary recognition performance of the model in complex oral point cloud scenarios by explicitly modeling the correlation between points and the interaction of cross-branch features.

[0078] (5) Joint optimization strategy design

[0079] To achieve simultaneous optimization of localization accuracy and contour details in the 3D tooth point cloud segmentation task, this embodiment designs a joint optimization strategy during the training phase, integrating multiple task objectives and loss functions to achieve efficient end-to-end learning in the overall architecture.

[0080] First, at the level of supervision signals, the model simultaneously outputs a point-level classification probability map and a centroid-guided instance mask map. This dual supervision mechanism guides the network to learn structural features and instance separation information. Specifically, the point-level classification task uses cross-entropy loss to ensure correct allocation of basic semantics; instance mask prediction employs a weighted structure combining Dice loss and boundary smoothing loss to enhance the model's ability to perceive boundaries and small structures.

[0081] Overall, the segmentation loss and the centroid heatmap prediction loss can be combined by weights into a joint loss function, which takes the form:

[0082] in, The cross-entropy loss used for splitting subtasks This is the smooth L1 loss for the centroid heatmap. and It is a dynamically adjustable coefficient that can be adaptively adjusted according to the training stage or task complexity.

[0083] Secondly, a dynamic loss weight mechanism is introduced into the optimization process, which adaptively adjusts the contribution ratio of each loss term based on the convergence rate and stability of different tasks during the training phase. This mechanism avoids the training bias problem caused by fixed weights and ensures that the model achieves co-evolution between global and local tasks.

[0084] Furthermore, to enhance the effectiveness of the centroid guidance module, a spatial thermal guidance loss term can be introduced to constrain the structural consistency between the instance mask boundary and the centroid heatmap. This loss guides the network to maintain high consistency output in irregular or blurred-boundary regions by comparing the response distribution of the instance mask with that of the corresponding centroid region.

[0085] Finally, this embodiment uses the Adam optimizer, sets a stepped learning rate decay strategy, and combines BatchNorm and Dropout techniques to further improve the training stability and generalization ability of the model.

[0086] In summary, this invention proposes a centroid-guided bi-branch feature fusion method for 3D tooth segmentation. By constructing a five-stage technical process consisting of image preprocessing, bi-path encoding, centroid heatmap, graph attention fusion, and joint optimization, accurate segmentation of 3D tooth target structures is achieved.

[0087] The modules of this invention possess excellent collaborative characteristics, combining spatial structure perception and local texture recognition capabilities, effectively improving segmentation accuracy and robustness in complex scenes. The above embodiments are merely preferred descriptions of the technical solution of this invention and do not constitute a limitation on the scope of protection of this invention. Without departing from the core idea of ​​this invention, those skilled in the art can replace, adjust, or reconstruct some of the modules, all of which should be considered equivalent variations and extensions of this invention.

Claims

1. A three-dimensional tooth segmentation method based on a centroid-guided bi-branch geometric attention network, characterized in that, Includes the following steps: S1. Acquire 3D tooth point cloud data from an oral scanner, normalize the point cloud and estimate the normal vector, and construct a point cloud feature matrix containing coordinates and normal vectors; S2. Construct a dual-branch geometric attention network; the dual-branch geometric attention network includes a coordinate branch and a normal vector branch, wherein the coordinate branch adopts a multi-layer stacked graph convolution structure to capture the global spatial topological relationship of the tooth point cloud, and maintains the stability of deep feature transmission through cross-layer residual connections and batch normalization operations; the normal vector branch is based on local neighborhood convolution and graph attention mechanism to enhance the expression of subtle geometric differences at the tooth boundary, and also introduces residual and normalization modules to improve training convergence performance; the features output by the two branches are used in the subsequent fusion stage to construct complementary representations of global structural information and local boundary detail information; S3. Input the point cloud feature matrix into the coordinate branch and the normal vector branch respectively, wherein the coordinate branch extracts global geometric structure information and the normal vector branch extracts local boundary geometric details; S4. Calculate the centroid of the tooth based on the spatial distribution of each tooth instance in the training sample, and generate a centroid heatmap based on the Gaussian function, which is then used as a spatial prior guiding signal input into the network. S5. Introduce graph attention convolution mechanism in the backbone network and combine it with multi-scale feature fusion strategy to achieve adaptive fusion of dual-branch features, so as to simultaneously model global topology and local boundary information; S6. Construct a joint loss function consisting of segmentation loss and centroid prediction loss, and achieve task balance through dynamic weight adjustment, thereby outputting the tooth instance segmentation result and its corresponding centroid location.

2. The three-dimensional tooth instance segmentation method based on centroid-guided bi-branch geometric attention network according to claim 1, characterized in that: The generation method of the centroid heatmap in step S4 is as follows: First, the centroid coordinates of each tooth instance are calculated based on the point cloud distribution in the training samples, and the centroids are mapped to a normalized three-dimensional spatial coordinate system; then, a three-dimensional Gaussian kernel is constructed with each centroid as the center to form a continuous spatial prior heatmap, the expression of which is: Where p represents the position of any point. Let be the centroid coordinates of the i-th tooth, and N be the total number of teeth. The diffusion coefficient is used. The heatmap is input into the model as a spatial prior guiding signal during network training to enhance the boundary discrimination ability of adjacent tooth regions and the feature response of weak boundary regions.

3. The three-dimensional tooth instance segmentation method based on centroid-guided bi-branch geometric attention network according to claim 1, characterized in that: In step S5, a graph attention convolution mechanism is introduced in the dual-branch feature fusion stage. The feature transfer weights are adaptively adjusted according to the geometric relationship between each point and its neighborhood in the point cloud. Combined with a multi-scale feature aggregation strategy, local geometric features at different scales and global topological structure are integrated in a unified manner to obtain a fused feature representation that takes into account both the overall tooth structure and boundary details.

4. The three-dimensional tooth instance segmentation method based on centroid-guided bi-branch geometric attention network according to claim 1, characterized in that: The joint loss function described in step S6 is used to simultaneously constrain the accuracy of tooth instance segmentation and centroid localization, and is determined by the segmentation loss. With centroid predicted loss Composition, which is defined as follows: in, and These are weighting coefficients that are dynamically adjusted during the training process. They are used to adaptively balance segmentation accuracy and centroid heatmap prediction accuracy at different training stages, thereby improving the overall accuracy and stability of the segmentation results.

5. The three-dimensional tooth instance segmentation method based on centroid-guided bi-branch geometric attention network according to claim 1, characterized in that: The input point cloud described in step S1 is constructed in the form of an N×6 dimensional feature matrix, where the three-dimensional coordinates of each point and the corresponding three-dimensional normal vector are concatenated by channel as network input and fed into the coordinate branch and the normal vector branch respectively, so as to provide global spatial structure information and local surface geometric information at the same time.

6. The three-dimensional tooth instance segmentation method based on centroid-guided bi-branch geometric attention network according to claim 1, characterized in that: It is suitable for complex clinical cases such as crowded, missing, and misaligned teeth. By using the spatial prior constraints provided by the centroid heatmap and the graph attention convolution mechanism to adaptively enhance the boundary region, it improves the ability to distinguish fuzzy boundaries between adjacent teeth, thus maintaining high instance segmentation accuracy and robustness under the above complex dentition conditions.

Images