Dental bracket point cloud segmentation method and system based on local feature fusion and parallel adaptive pooling attention mechanism

Abstract

The present invention relates to the field of oral medical image processing technology, and in particular to a method and system for dental bracket point cloud segmentation based on local feature fusion and parallel adaptive pooling attention mechanism, wherein the dental bracket point cloud data to be segmented is input into a pre-trained point cloud segmentation model, and the point cloud segmentation model is used to obtain the dental bracket point cloud data segmentation result; wherein the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining up-sampled features by interpolation and splicing, and a prediction head for class prediction output of the up-sampled features, and the encoder performs local area grouping and division with the sampling point as the center point, and uses the parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channel and spatial dimension to extract features from the grouped point cloud. The present invention enhances the segmentation effect of the dental bracket and effectively improves the efficiency and accuracy of the dental bracket segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of oral medical image processing, and in particular to a tooth bracket point cloud segmentation method and system based on local feature fusion and parallel adaptive pooling attention mechanism. Background Art

[0002] With the continuous progress of society, the development of dental technology is also changing with each passing day. Dental brackets play an important role in dental treatment, especially in oral restoration and orthodontic treatment. The retention mesh and macroconnector of dental brackets are of great significance to the treatment effect and comfort of patients. The retention mesh is a part of the bracket, which enhances the stability of the bracket in the mouth by providing a larger contact area. For patients, the retention mesh can significantly improve the retention effect of the bracket and reduce the risk of the bracket loosening or falling off, thereby improving the success rate of treatment and the comfort of patients. The macroconnector refers to the structure that connects the different parts of the bracket. It can evenly distribute the bite force and reduce the pressure on individual teeth or gums. For patients, the macroconnector can provide a more uniform and comfortable bite experience, while also helping to extend the service life of the bracket.

[0003] Point cloud segmentation techniques based on traditional methods have been widely used in the early stages of point cloud processing. Although these methods perform well in some cases, they also have some significant shortcomings. For example, region growing-based methods are sensitive to the selection of initial points, are easily affected by noise, and have poor results when processing complex shapes and irregular point clouds; K-means in clustering-based methods is sensitive to the selection of initial points and has difficulty processing non-spherical clusters; DBSCAN is sensitive to parameter settings and may miss some points in sparse areas. Point cloud segmentation methods based on deep learning have made significant progress in recent years and have solved many challenges faced by traditional methods. These methods can achieve high-precision point cloud segmentation in a variety of application scenarios by learning complex patterns and features in point cloud data. However, they also have some shortcomings and areas that need improvement. For example, PointNet has weak ability to capture local features and has difficulty processing complex geometric structures. Methods such as Spotr, DGCNN, PointNext, and Spotr rely on directional propagation to update parameters and do not select important features.

[0004] In different scanning or treatment processes in stomatology, the dental stents usually have complex shapes, including unevenly distributed retention meshes and complex connector structures. At the same time, the design of the stent depends on the missing part of the patient, and the position and shape of the missing parts of different patients are different, which makes the dental stent point cloud have a complex shape. The above problems make the existing segmentation algorithm unable to effectively segment the dental stent point cloud. Therefore, a dental stent point cloud segmentation solution is urgently needed to adapt to the processing of dental imaging data. Summary of the invention

[0005] To this end, the present invention provides a dental bracket point cloud segmentation method and system based on local feature fusion and parallel adaptive pooling attention mechanism, which fully utilizes the global and local information of the dental bracket point cloud based on local feature fusion and adaptive pooling attention mechanism to effectively segment the dental bracket point cloud.

[0006] According to the design scheme provided by the present invention, on the one hand, a tooth bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism is provided, comprising:

[0007] Acquire point cloud data of a tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors;

[0008] Input the tooth bracket point cloud data to be segmented into the pre-trained point cloud segmentation model, and use the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result;

[0009] Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features. The encoder groups local areas with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channels and spatial dimensions to extract features from the grouped point cloud.

[0010] As a dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the encoder includes a plurality of stacked SPoTr layers, the first two SPoTr layers constitute a PAPA-SPoTr layer for extracting key features of the input point cloud based on the parallel adaptive pooling attention mechanism, and the last SPoTr layer obtains the global features of the input point cloud based on the sampled point cloud features and the output features of the previous layer, wherein each SPoTr layer is composed of a sampling layer, a grouping layer and an SPoTrBlock layer, the sampling layer obtains the sampled point cloud from the input point cloud data using the farthest distance sampling method, the grouping layer takes the sampled point cloud as the center point and groups the spherical area of ​​the specified radius as the local area, and uses the SPoTrBlock layer to extract features from the grouped point cloud.

[0011] As a tooth bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the most relevant key features of point cloud channels and spatial dimensions are dynamically selected using the parallel adaptive pooling attention mechanism, including:

[0012] The first point cloud feature and the second point cloud feature are obtained from the center point and the neighbor point respectively using two attention pools;

[0013] The first point cloud feature and the second point cloud feature are concatenated, the input point cloud feature is compressed using a convolution function, and the input point cloud intermediate feature is obtained using a batch normalization operation and an activation function;

[0014] Decompose the intermediate features of the input point cloud into the first tensor and the second tensor along the channel and space directions, use convolution to convert the first tensor and the second tensor to the specified dimension, and obtain two attention weights through the activation function;

[0015] The input point cloud features are multiplied and weighted based on the attention weights to obtain the key features of the input point cloud.

[0016] As a dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the decoder is composed of multiple upsampling layers, each upsampling layer is upsampled by interpolation and connection splicing, wherein the interpolation performs a reverse interpolation operation based on the inverse distance weighted average of the nearest neighbor point cloud to obtain the upsampled point cloud features.

[0017] As a dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the interpolation process includes: for the point cloud in the current upsampling layer, obtaining the K neighbor points closest to the point cloud in the previous upsampling layer, performing weighted summation on the corresponding features of the point cloud and the neighbor points, and obtaining the point cloud features in the current upsampling layer, wherein the weight in the weighted summation is an inverse function of the distance.

[0018] As a dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the prediction head includes multiple MLP layers, and the MLP layers are used to predict the output point cloud category scores.

[0019] As a tooth bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism of the present invention, further, the point cloud segmentation model training process includes:

[0020] Constructing tooth bracket sample data, wherein the tooth bracket sample data is provided with a point cloud type label, and the point cloud type label includes a retention network point cloud and a connector point cloud;

[0021] A deep learning network model was constructed based on the U-Net network structure, and the deep learning network model was trained using dental bracket sample data. The trained deep learning network model was used as a point cloud segmentation model.

[0022] On the other hand, the present invention also provides a tooth bracket point cloud segmentation system based on local feature fusion and parallel adaptive pooling attention mechanism, comprising: a data acquisition module and a point cloud segmentation module, wherein:

[0023] A data acquisition module, used to acquire point cloud data of the tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors;

[0024] A point cloud segmentation module is used to input the tooth bracket point cloud data to be segmented into a pre-trained point cloud segmentation model, and use the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result;

[0025] Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features. The encoder groups local areas with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channels and spatial dimensions to extract features from the grouped point cloud.

[0026] Beneficial effects of the present invention:

[0027] The present invention utilizes a hierarchical feature fusion strategy to fuse high-dimensional features and low-dimensional features of point clouds, making full use of the global and local information of point clouds, thereby improving the segmentation effect of point clouds; through the parallel adaptive pooling attention mechanism, the input point cloud is modeled in space and channels, which can dynamically adjust the segmentation algorithm's attention to different features, and adaptively select the most relevant features for segmentation, thereby enhancing the segmentation effect of the segmentation algorithm on dental stents, effectively improving the efficiency and accuracy of dental stent segmentation, and has good application prospects in the field of stomatology. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the tooth bracket point cloud segmentation process based on local feature fusion and parallel adaptive pooling attention mechanism in the embodiment;

[0029] Figure 2 This is a schematic diagram of the point cloud segmentation model structure in the embodiment;

[0030] Figure 3 It is a schematic diagram of the local feature fusion strategy in the embodiment;

[0031] Figure 4 It is a schematic diagram of the principle of the parallel adaptive pooling attention mechanism in the embodiment;

[0032] Figure 5 This is a schematic diagram of the algorithm flow in the embodiment;

[0033] Figure 6 Schematic diagram of visualization results of different segmentation algorithms in the embodiments. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.

[0035] Existing attention mechanisms have the following shortcomings: (1) Insufficient information coupling. Lack of comprehensive information: Focusing only on one dimension in the channel or space may lead to ignoring important information in the other dimension and fail to fully capture the complex relationship between global and local features. Order dependency: The design of the serial method usually relies on processing one dimension first (such as space first and then channel or vice versa). This order dependency may lead to insufficient information coupling and fail to fully utilize the synergistic effect of multi-dimensional features. (2) Limited feature expression ability. Single dimension limitation: Processing channel or spatial attention alone may limit the model's ability to express complex features, especially when processing complex shapes and structures (such as dental bracket point clouds), which may miss important feature information. Feature redundancy: The serial design is prone to feature redundancy. Because information of different dimensions cannot fully interact during the processing process, redundant features may be generated, reducing the efficiency and performance of the model. (3) Model complexity and computational overhead. Increasing model complexity: The serial hybrid attention mechanism may increase the complexity and computational overhead of the model because it is necessary to calculate the attention of the channel and spatial dimensions separately and perform multiple feature transformations. Low computational efficiency: The serial processing method has more processing steps and relatively low computational efficiency, especially in the processing of high-dimensional data (such as 3D point clouds), which may significantly increase the computational burden. (4) Poor dynamic adaptability. Lack of dynamics: The attention mechanism of serial design is usually static and lacks dynamic adaptability, making it difficult to flexibly adjust according to the different characteristics of the input data. Slow response speed: When it is necessary to quickly respond to changes in the importance of different features, the serial attention mechanism may respond slowly and fail to adjust the feature weights in time. During different scanning or treatment processes, the stent usually has a complex shape, including an unevenly distributed retention network and a complex connector structure. At the same time, the design of the stent depends on the missing part of the patient, and the position and shape of the missing part of different patients are different, which makes the tooth stent point cloud have a complex shape. The above problems make it impossible for existing segmentation algorithms to effectively segment tooth stent point clouds. To this end, the embodiments of the present invention, see Figure 1 As shown, a tooth bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism is provided, comprising:

[0036] S101, obtaining point cloud data of a tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors;

[0037] S102, inputting the tooth bracket point cloud data to be segmented into a pre-trained point cloud segmentation model, and using the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result,

[0038] Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features. The encoder groups local areas with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channels and spatial dimensions to extract features from the grouped point cloud.

[0039] See also Figure 2 As shown in Figure 1, the encoder consists of multiple SPoTr layers. The SPoTr layer consists of a sampling layer, a grouping layer, and a SPoTrBlock layer. The sampling layer uses the farthest distance sampling method to obtain N points from the input. s In the grouping layer, N s The sampling points are taken as the center point, and a spherical area with a radius of d is divided as the local area, and each local area is guaranteed to contain K neighboring points. When the number of neighboring points in the local area is less than K, the coordinates of the first neighboring point are repeated until the set requirements are met. SPoTrBlock extracts features from the grouped features. The adaptive pooling coordinate attention designed in the first two SPoTr layers constitutes the PAPA-SPoTr layer, which is used to improve the model's ability to extract important features. In addition, due to the feature fusion strategy, in the last SPoTr, all corresponding features of the 64 sampling points in the previous layer are connected, and the features are expanded to 975 dimensions as global features of the point cloud.

[0040] The decoder consists of multiple FP layers. The FP layer mainly uses interpolation and concatenation to perform upsampling. Among them, interpolation is based on the inverse distance weighted average of the K nearest neighbor points to perform reverse interpolation operations to obtain the features of the upsampled points. Specifically, for a point p in the upsampling layer i , find the corresponding features of the K points closest to it in the previous layer and perform weighted summation, where the weight is the inverse function of the corresponding distance. In the decoder part, the skip link concatenation method is used to splice the features corresponding to the points in the encoder part.

[0041] The prediction head consists of three MLP layers. MLP is mainly used to output the prediction score for each category.

[0042] Spotr is based on a hierarchical framework and can effectively capture local features of point clouds to complete classification and segmentation tasks. The point cloud is sampled through a downsampling algorithm, and the number of points in each layer continues to decrease, but the feature dimension continues to increase, so the information contained in each point is richer. However, Spotr feature extraction only uses the highest dimensional features corresponding to the sampling points in the last layer, while ignoring the low-dimensional features corresponding to the points in other SPoTr layers. The high-dimensional features of the point cloud can better reflect the global information of the point cloud, while the low-dimensional features can better characterize the local features and reflect more detailed information.

[0043] Therefore, in order to make full use of the feature information of each SPoTr layer, in this embodiment, the local feature fusion strategy (LFF) is used to splice the low-dimensional and high-dimensional features of the point cloud to improve the accuracy of the model. Figure 3 As shown in the figure, the green, red and blue points represent the points in the original point cloud, the sampling points and the remaining points after sampling, respectively. M and N represent the number of points in the point cloud and the number of sampling points, respectively. The features of the sampling points are increased from the original D1 dimension to the D2 dimension through the SPoTr layer. The hierarchical feature fusion strategy used in the embodiment of this case combines the features of the sampling points with the features before the point correspondence, thereby expanding the features to (D1+D2) dimensions.

[0044] Among them, the parallel adaptive pooling attention mechanism is used to dynamically select the most relevant key features of the point cloud channel and spatial dimension, which can be designed to include the following:

[0045] The first point cloud feature and the second point cloud feature are obtained from the center point and the neighbor point respectively using two attention pools;

[0046] The first point cloud feature and the second point cloud feature are concatenated, the input point cloud feature is compressed using a convolution function, and the input point cloud intermediate feature is obtained using a batch normalization operation and an activation function;

[0047] Decompose the intermediate features of the input point cloud into the first tensor and the second tensor along the channel and space directions, use convolution to convert the first tensor and the second tensor to the specified dimension, and obtain two attention weights through the activation function;

[0048] The input point cloud features are multiplied and weighted based on the attention weights to obtain the key features of the input point cloud.

[0049] There are many types of attention mechanisms, including compression and excitation attention mechanisms, convolutional attention mechanism modules, etc. However, most existing attention mechanisms only focus on channels or spaces and often design mixed attention in a serial manner. This means that only focusing on one dimension in the channel or space may lead to ignoring important information in another dimension and fail to fully capture the complex relationship between global and local features. In addition, the design of the serial method usually relies on processing one dimension first, such as space first and then channel or vice versa. This order dependency may lead to insufficient information coupling and easily lead to feature redundancy, because information of different dimensions cannot fully interact during the processing process, and the synergistic effect of multi-dimensional features cannot be fully utilized.

[0050] The above problems make it impossible for the existing attention mechanism to effectively segment the dental bracket point cloud. Therefore, in the embodiment of this case, the parallel adaptive pooling attention mechanism (Parallel Adaptive Pooling Attention, PAPA) is used to model the attention in the channel and spatial dimensions of the point cloud in parallel, so that the global structure and local detail information can be captured at the same time, making full use of the multi-dimensional features of the point cloud data. At the same time, the feature weights are adaptively adjusted, and the most relevant features are dynamically selected for segmentation to enhance the ability to capture complex shapes and details. In addition, compared with the serial attention mechanism, the parallel adaptive pooling attention mechanism can focus computing resources on important features, optimize the allocation of computing resources, and improve overall computing efficiency.

[0051] The specific modeling process of the parallel adaptive pooling attention mechanism is as follows Figure 4 As shown, it can be summarized as including the following contents:

[0052] (1) Assume that the given input point cloud feature is F∈R C×N×K , where C, N, and K represent the feature dimension of the point in the point cloud, the number of center points, and the number of neighbor points in the local area. Different from SE that only generates single channel attention, first, the present invention uses two attention pools to obtain two features F from the center point and the neighbor point respectively. N and F K , the corresponding dimensions are C×N×1 and C×1×K. Modeling method, formulas are shown in (1) and (2).

[0053] F N =BN(Conv(meanpool(softmax(Conv(F))·F,dim=2)+maxpool(softmax(Conv(F))·F,dim=2)))(1)

[0054] F K=BN(Conv(meanpoll(softmax(Conv(F))·F,dim=1)+maxpoll(softmax(Conv(F))·F,dim=1))) (2)

[0055] In formulas (1) and (2), Conv represents the convolution operation; softmax represents the normalization operation; maxpool represents the maximum pooling; meanpool represents the average pooling; and BN represents batch normalization.

[0056] (2) Secondly, in the attention generation stage, the two features are concatenated, and the original C-dimensional features are compressed into C / r dimensions using the convolution function. Then, the batch normalization operation and the swish activation function are used to obtain the intermediate feature F m , the modeling method is shown in formula (3).

[0057] F m =δ(BN(Conv[F N ,F K ])) (3)

[0058] In formula (3), [·,·] represents the concatenation operation; Conv represents the convolution operation; BN represents the batch normalization operation; δ represents the swish activation function.

[0059] (3) Next, the intermediate feature F m Decompose into different tensors as well as Use different convolutions to convert the features into C dimensions, obtain two attention weights through the sigmoid activation function, and perform a weighted multiplication operation with the input features to obtain the output feature F of the module. out .

[0060]

[0061] In formula (4), Conv represents the convolution transformation operation; BN represents the batch normalization operation; δ represents the swish activation function.

[0062] Among them, the point cloud segmentation model training process can be designed to include the following contents:

[0063] Constructing tooth bracket sample data, wherein the tooth bracket sample data is provided with a point cloud type label, and the point cloud type label includes a retention network point cloud and a connector point cloud;

[0064] A deep learning network model was constructed based on the U-Net network structure, and the deep learning network model was trained using dental bracket sample data. The trained deep learning network model was used as a point cloud segmentation model.

[0065] Furthermore, based on the above method, an embodiment of the present invention also provides a tooth bracket point cloud segmentation system based on local feature fusion and parallel adaptive pooling attention mechanism, comprising: a data acquisition module and a point cloud segmentation module, wherein:

[0066] A data acquisition module, used to acquire point cloud data of the tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors;

[0067] A point cloud segmentation module is used to input the tooth bracket point cloud data to be segmented into a pre-trained point cloud segmentation model, and use the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result;

[0068] Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features. The encoder groups local areas with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channels and spatial dimensions to extract features from the grouped point cloud.

[0069] By utilizing the local feature fusion strategy, the channels of the low-dimensional feature point cloud and the high-dimensional feature point cloud are merged, so that the low-dimensional and high-dimensional features of the point cloud are correspondingly spliced, and the local information of the low-dimensional features and the high-dimensional global information are fully utilized, thereby improving the segmentation accuracy of the model; by utilizing the parallel adaptive pooling attention mechanism, the input point cloud is spatially and channel-wise pooled modeling is performed, the feature weights are adaptively adjusted, and important features are adaptively selected and enhanced through the attention mechanism, especially in detail capture and heterogeneous data processing; the local feature fusion strategy is combined with the parallel adaptive pooling attention mechanism and Spotr to enhance the segmentation effect of Spotr on dental brackets.

[0070] In order to verify the effectiveness of this solution, the following is a further explanation based on experimental data:

[0071] Due to the hierarchical structure, the local area divided by the network model gradually expands, and higher-dimensional local features of the point cloud can be obtained. The input of the model is a set of N×C point clouds. In all experiments, only the six-dimensional coordinate information (x, y, z, n) of the input point cloud is used. x ,n y ,n z ), so the feature dimension C of the input point cloud is set to 6. The bracket point cloud segmentation algorithm based on hierarchical feature fusion and adaptive pooling coordinate attention mechanism is as follows Figure 5As shown in the figure, the overall content is as follows: use the farthest sampling method to downsample the point cloud; set a threshold with the sampled point as the center to find the nearest 32 or 64 domain points for combination; use PAPA-SPoTr to extract features from the grouped points; use LFF-SPoTr to extract features for the last layer of the encoder; extract features from the point cloud of the last layer of the encoder and continuously upsample until the number is the same as the original input point cloud; use the segmentation head to predict the segmentation of the upsampled last layer.

[0072] Compared with the existing point cloud segmentation algorithms Spotr, PointMLP, DGCNN, and PointNext, the results are shown in Table 1. Compared with Spotr and PointNext, the mIou of this algorithm is increased by 5.46% and 4.15%, and the mAc is increased by 5.79% and 4.42%, respectively. Compared with the compared point cloud segmentation methods, the mIou and mAc of this algorithm are as high as 85.58% and 96.88%, respectively, which effectively improves the segmentation effect of the segmentation algorithm on the bracket point cloud.

[0073] Table 1 Comparison of different segmentation algorithms

[0074]

[0075] Draw the visualization results of different segmentation algorithms such as Figure 6 As shown, (a) represents the original annotated point cloud, blue represents the background, red represents the large connector, and green represents the retention network; (b) represents the point cloud segmentation result of Spotr; (c) represents the point cloud segmentation result of PointMLP; (d) represents the point cloud segmentation result of PointNext; (e) represents the point cloud segmentation result of the algorithm in this case.

[0076] As can be seen from the yellow boxes in the first row of images (b) and (c), Spotr and PointMLP mistakenly identify the retention network as a large connector; as can be seen from the yellow boxes in (d) and (e), PointNext and the solution in this case can better segment the large connector and the retention network. As can be seen from the orange and red boxes in the first row of images (b) and (c), Spotr, PointMLP, and PointNext mistakenly identify the retention network as the background; as can be seen from the orange and red boxes in (e), the solution in this case can better segment the large connector and the retention network.

[0077] From the yellow boxes in the second row of images (c) and (d), it can be seen that PointMLP and PointNext will mistakenly identify the retention network as a large connector or background; from the yellow boxes in (b) and (e), it can be seen that Spotr and the solution in this case can better segment the large connector and the retention network. From the orange and red boxes in the second row of images (b) and (c), it can be seen that Spotr and PointMLP mistakenly identify the retention network as the background; from the orange and red boxes in (d) and (e), it can be seen that PointNext and the solution in this case can better segment the retention network and the segmented area of ​​the solution in this case is closer to the real area.

[0078] From the yellow boxes in the third row of images (b), (c), and (d), it can be seen that Spotr, PointMLP, and PointNext will mistakenly identify some large connectors as background; from the yellow box in (e), it can be seen that the solution in this case can better segment the retention network. From the orange and red boxes in the second row of images (c) and (d), it can be seen that PointMLP and PointNext will mistakenly identify the retention network as background; from the orange and red boxes in (a) and (e), it can be seen that Spotr and the solution in this case can better segment the retention network and the segmented area of ​​the solution in this case is closer to the real area.

[0079] The above experimental data show that this solution can achieve effective segmentation of dental brackets, improve efficiency and segmentation accuracy, and has good application prospects in the field of oral medical image processing.

[0080] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0081] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0082] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0083] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0084] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism, characterized in that: Include: Acquire point cloud data of a tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors; Input the tooth bracket point cloud data to be segmented into the pre-trained point cloud segmentation model, and use the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result; Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features, and the encoder groups and divides the local area with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channel and spatial dimension to extract features from the grouped point cloud; using a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channel and spatial dimension, including: obtaining the first point cloud feature and the second point cloud feature respectively from the center point and the neighbor point and using two attention pooling; splicing the first point cloud feature and the second point cloud feature, compressing the input point cloud feature using a convolution function, and obtaining the input point cloud intermediate feature using a batch normalization operation and an activation function; decomposing the input point cloud intermediate feature into a first tensor and a second tensor along the channel and space directions, and using convolution to convert the first tensor and the second tensor into a specified dimension, and obtaining two attention weights through an activation function; multiplying and weighting the input point cloud features based on the attention weights to obtain the key features of the input point cloud.

2. The dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism according to claim 1 is characterized in that: The encoder includes a plurality of SPoTr layers stacked together, wherein the first two SPoTr layers constitute a PAPA-SPoTr layer for extracting key features of an input point cloud based on a parallel adaptive pooling attention mechanism, and the last SPoTr layer obtains global features of the input point cloud based on the features of the sampled point cloud and the output features of the previous layer, wherein each SPoTr layer is composed of a sampling layer, a grouping layer and an SPoTrBlock layer, wherein the sampling layer obtains a sampled point cloud from the input point cloud data using the farthest distance sampling method, the grouping layer takes the sampled point cloud as the center point and groups a spherical area of ​​a specified radius as a local area, and uses the SPoTrBlock layer to extract features from the grouped point cloud.

3. The dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism according to claim 1 is characterized in that: The decoder is composed of multiple upsampling layers, each of which is upsampled by interpolation and concatenation, wherein the interpolation performs a reverse interpolation operation based on an inverse distance weighted average of the nearest neighbor point cloud to obtain the upsampled point cloud features.

4. The method for tooth bracket point cloud segmentation based on local feature fusion and parallel adaptive pooling attention mechanism according to claim 3 is characterized in that: The interpolation process includes: for the point cloud in the current upsampling layer, obtaining the K neighboring points in the previous upsampling layer that are closest to the point cloud, performing weighted summation on the corresponding features of the point cloud and the neighboring points, and obtaining the point cloud features in the current upsampling layer, where the weight in the weighted summation is the inverse function of the distance.

5. The dental bracket point cloud segmentation method based on local feature fusion and parallel adaptive pooling attention mechanism according to claim 1, characterized in that: The prediction head includes multiple MLP layers, and the MLP layers are used to predict the output point cloud category scores.

6. The method for dental bracket point cloud segmentation based on local feature fusion and parallel adaptive pooling attention mechanism according to claim 1, characterized in that: The point cloud segmentation model training process includes: Constructing tooth bracket sample data, wherein the tooth bracket sample data is provided with a point cloud type label, and the point cloud type label includes a retention network point cloud and a connector point cloud; A deep learning network model was constructed based on the U-Net network structure, and the deep learning network model was trained using dental bracket sample data. The trained deep learning network model was used as a point cloud segmentation model.

7. A dental bracket point cloud segmentation system based on local feature fusion and parallel adaptive pooling attention mechanism, characterized in that: The method according to claim 1 is implemented, comprising: a data acquisition module and a point cloud segmentation module, wherein: A data acquisition module, used to acquire point cloud data of the tooth bracket to be segmented, wherein the point cloud data of the tooth bracket includes three-dimensional coordinates of the point cloud and corresponding normal vectors; A point cloud segmentation module is used to input the tooth bracket point cloud data to be segmented into a pre-trained point cloud segmentation model, and use the point cloud segmentation model to obtain the tooth bracket point cloud data segmentation result; Among them, the point cloud segmentation model includes an encoder for sampling, grouping and feature extraction of input data, a decoder for obtaining upsampled features through interpolation and splicing, and a prediction head for class prediction output of upsampled features. The encoder groups local areas with the sampling point as the center point, and uses a parallel adaptive pooling attention mechanism to dynamically select the most relevant key features of the point cloud channels and spatial dimensions to extract features from the grouped point cloud.

8. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 6 can be implemented.

Images