Tooth segmentation method fusing slice attention and Transform semantic information

Abstract

The invention discloses a CBCT tooth segmentation method fusing a slice attention mechanism and Transform semantic information, and the method comprises the following steps: 1, collecting an original CBCT image, and carrying out the preprocessing of the original CBCT image, and obtaining a preprocessed CBCT image; 2, performing'global + local 'feature preliminary extraction on the preprocessed CBCT image to obtain a multi-scale initial feature map; 3, strengthening the spatial correlation and key region focusing of the multi-scale initial feature map to obtain an enhanced feature map; step 4, processing the enhanced feature map to obtain a noise suppression-multi-dimensional fusion feature map; and 5, recovering the resolution of the original CBCT image, and outputting a segmentation result of the three-dimensional tooth structure. According to the method, the precision of tooth segmentation in the CBCT image can be remarkably improved, segmentation errors caused by noise interference and structural complexity are reduced, the spatial consistency of a segmentation result and a real structure is improved, and the method has relatively strong advantages in segmentation accuracy and integrity.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence technology, and relates to a CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information. Background Technology

[0002] Oral health is an important component of overall health, and teeth, as the core carrier of oral function and morphology, require precise structural information for disease diagnosis, treatment planning, and surgical simulation. However, dental clinical skills training is lengthy and has a low margin for error; dentists need at least two years of systematic theoretical study and four years of clinical practice before they can independently perform complex oral surgeries. [2] Furthermore, oral imaging itself often presents challenges in accurately identifying lesions in some images due to factors such as noise, low contrast, and overlapping anatomical structures. Therefore, developing intelligent analysis tools to assist dentists in tooth segmentation and disease detection is of paramount importance.

[0003] Currently, cone-beam computed tomography (CBCT), as a high-resolution, low-radiation three-dimensional imaging technique, is widely used in the field of oral imaging. CBCT images can provide three-dimensional morphological information of structures such as teeth, jawbones, and skulls, and their high precision makes them an important tool in the diagnosis and treatment planning of oral diseases. Among these, accurately segmenting tooth structures from CBCT images has become one of the core tasks of digital dentistry applications. In recent years, many studies have focused on solving the problem of automatic segmentation of tooth structures in CBCT images. These studies can be broadly divided into two categories: CBCT tooth segmentation based on level set methods and CBCT tooth segmentation based on deep learning methods. However, existing tooth segmentation techniques based on level set methods still face the problem of insufficient automation. This limitation stems from the high sensitivity of level set methods to the initial contour, especially when dealing with complex oral anatomy structures, such as the contact relationships between adjacent teeth and the complexity of jawbone and alveolar bone structures. Choosing an appropriate method for determining the initial contour becomes particularly crucial. Despite significant progress in CBCT tooth segmentation based on deep learning methods, it still faces many challenges. For example, the density of tooth roots and alveolar bone is similar, resulting in low image contrast and difficulty in accurately distinguishing tooth root boundaries. Adjacent teeth in the crown region are often in close contact with blurred boundaries, which increases the difficulty of segmenting the crown region.

[0004] In summary, existing methods often struggle to balance the discovery of spatial relationships between slices and noise suppression when processing CBCT images, failing to meet clinical requirements for segmentation accuracy and robustness. Therefore, a new segmentation method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information, which solves the problem that existing tooth segmentation methods are difficult to simultaneously explore spatial correlation between slices and suppress noise, and are unable to meet clinical requirements for segmentation accuracy and robustness.

[0006] The technical solution adopted in this invention is a CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information. The specific steps are as follows: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; Step 4: Input the enhanced feature map into the dual fusion convolution module, and obtain the multi-dimensional fusion feature map through parallel dual-branch structure processing; Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

[0007] The invention is further characterized by: Step 1, the CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0008] Step 2 is as follows: The encoder is a hybrid architecture encoder, which includes a Transformer module and a convolution module to extract global semantic information and local feature information. It performs multi-scale feature extraction on the preprocessed image data through the following process: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed by an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module contains a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process From different depths of the hybrid architecture encoder, feature maps containing information at different scales are output, forming a multi-scale feature map.

[0009] Step 3's cross-slice attention module includes a position attention unit and a slice attention unit: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0010] The position attention unit generates the position attention map through the following operations: Step 3A.1: For the input multi-scale feature map, perform global max pooling and global average pooling along its channel dimensions respectively, to obtain two spatial dimensions. Feature descriptors; Step 3A.2: Concatenate the two feature descriptors to form a fused feature map; Step 3A.3: Perform convolution operation on the concatenated fused feature map, and apply the Sigmoid activation function to the result of the convolution operation to generate a positional attention map, specifically represented as follows: (1) in, The input is the multi-scale initial feature map. Note the diagram for output position. The semicolon indicates a convolution operation, and the semicolon indicates a concatenation operation.

[0011] The slice attention unit generates the slice attention map through the following operations: Step 3B.1: Perform global max pooling and global average pooling along the height, width and channel dimensions of the input multi-scale feature map to obtain two slice-level feature vectors. Step 3B.2: Input the two slice-level feature vectors into a multilayer perceptron with the same structure for nonlinear transformation; Step 3B.3: Sum the two output feature vectors processed by the multilayer perceptron, and reshape the summed vector V to obtain the feature tensor. ; Step 3B.4: For the feature tensor Three different linear transformations are performed in parallel to compute a low-rank Gaussian distribution. The parameters, including the mean vector Covariance factor and the diagonal components of the covariance Calculate the covariance matrix Σ of the Gaussian distribution based on the above parameters, and then proceed from the Gaussian distribution... The vector z is obtained by sampling from the middle, and z is activated by the Sigmoid function to generate a slice attention map; The following three linear operations are performed on tensors: (2) in, The mean parameter weight matrix, The weight matrix of the diagonal components of the covariance. This is the covariance factor weight matrix. yes Reshaping the tensor; Covariance matrix of the distribution The calculation formula is: (3) in, yes The reshaping tensor, It is a diagonal matrix, with [variables] on the diagonal. elements, For parameterized rank; The slice attention map is represented as follows: (4) Among them, the output It is the reshaped version This refers to a slice attention map.

[0012] The specific steps for processing the dual-branch structure in step 4 are as follows: Step 4.1: Use the pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0013] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0014] Step 4.2's dual-fusion convolution module includes parallel 2D and 3D convolution branches: The specific steps for capturing the three-dimensional spatial structural features of teeth using 3D convolutional branches are as follows: Step A.1: Expand the dimension of the input feature map F by adding an extra channel, transforming it into a 4D real tensor, denoted as . ,in, This refers to the expanded feature tensor prepared for input to the 3D convolutional layer. The shape of this 4D real tensor is ( P,P,S, 1); Step A.2: Perform a 3D convolution operation on the 4D real tensor to obtain... ,in, The feature tensor output from the 3D convolutional layer, and zero-padding is used to ensure the output's... Space dimensions and input The matching process, in which the three-dimensional convolution operation is calculated, is shown in equation (5): (5) in, and These are the weights and bias parameters of the convolution, respectively. For three-dimensional convolution operators, For activation functions; Step A.3: Output of the 3D convolution operation Perform the extrusion operation to Converting back from 4D space to the original 3D space yields... The feature map F is then converted back to a three-dimensional space that matches the feature map F to obtain the three-dimensional spatial structural features of the teeth. The specific steps for extracting the two-dimensional planar semantic information of the feature map F using a two-dimensional convolutional branch are as follows: Step B.1: Zero-padding is performed on the input feature map F to ensure that the spatial dimensions of the output feature map after the convolution operation are consistent with those of the feature map F. Step B.2, using Two-dimensional convolution kernels are used to perform convolution operations on the padded feature map to obtain the convolution result; Step B.3: After normalizing the result of the convolution operation, apply a non-linear activation function to obtain the enhanced two-dimensional planar semantic features. ; The calculation process of the two-dimensional convolution branch is shown in equation (6): (6) in, Representing semantic features of a two-dimensional plane, and These are the weights and bias parameters of the convolution, respectively. It is a two-dimensional convolution operator, and BN is a normalization layer. This is the activation function.

[0015] Three-dimensional spatial structural features in step 4.3 semantic features of two-dimensional plane The splicing process is shown in equation (7): (7) in, This indicates a splicing operation. This is the spliced ​​feature map. It is a three-dimensional spatial structure feature. It represents two-dimensional planar semantic features.

[0016] The specific process of step 5 is as follows: Step 5.1, Upsampling Initialization: Input the multi-dimensional fused feature map into the decoding network, and perform upsampling through interpolation or transpose convolution operations to improve the spatial resolution of the multi-dimensional fused feature map and obtain the upsampled feature map; Step 5.2, Skip Connection Fusion: Introducing a skip connection module to fuse the high-resolution local feature map output by the encoder at the corresponding scale with the upsampled feature map of the current decoder stage; the fusion method is channel splicing or element-wise addition to supplement the spatial detail information lost during the encoder downsampling process and obtain the fused feature map; Step 5.3, Feature Recovery and Output: Perform convolution operation on the fused feature map to eliminate the checkerboard effect caused by upsampling, and repeat steps 5.1-5.3 to restore the spatial resolution to the original CBCT image step by step, and finally output the three-dimensional tooth segmentation result.

[0017] The beneficial effects of this invention are as follows: The method of this invention designs a cross-slice attention mechanism, which effectively mines the spatial dependencies between slices by enhancing the vertical spatial correlation between slices, thus solving the problem of not being able to fully utilize the information between slices in the prior art. A dual-fusion convolution module is designed, which superimposes two-dimensional planar semantic information on the basis of three-dimensional spatial features, and a denoising mechanism is used to effectively improve segmentation accuracy, especially when processing complex structures and low-contrast images. The entire model can significantly improve the accuracy of tooth segmentation in CBCT images, reducing segmentation errors caused by noise interference and structural complexity. Compared with the prior art, this invention reduces Hausdorff distance and average symmetric surface distance by 3.48 mm and 0.35 mm, respectively, improving the spatial consistency between the segmentation results and the real structure; the intersection-union ratio, similarity overlap, and DICE similarity coefficient are improved by 1.1%, 0.69%, and 0.24%, respectively, demonstrating the advantages of this model in segmentation accuracy and completeness. Attached Figure Description

[0018] Figure 1 This is a network structure diagram of the CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information according to the present invention; Figure 2 This is a structural diagram of the cross-slice attention module of the CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information according to the present invention; Figure 3 This is a structural diagram of the dual-fusion convolutional module of the CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information in this invention; Figure 4 This is a comparison of the segmentation results of the CBCT tooth segmentation method using the fusion of slice attention mechanism and Transformer semantic information in this invention with other models; Figure 5 This is a comparison of the 3D reconstruction results of the CBCT tooth segmentation method using the fusion of slice attention mechanism and Transformer semantic information in this invention with other models. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0020] Example 1 The CBCT tooth segmentation method, which integrates slice attention mechanism and Transformer semantic information, has the following specific network structure: Figure 1 As shown, the specific segmentation steps are as follows: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; The CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0021] Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; The encoder is a hybrid architecture encoder, containing a Transformer module and a convolutional module for extracting global semantic information and local feature information. In step 2, the convolutional module of the encoder works in conjunction with the Transformer module to achieve complementary feature extraction of "global semantics + local details". Structurally, it adopts an embedded design that alternates with the Transformer encoding layer: first, the preprocessed CBCT image is converted into a hidden feature representation through linear projection, and then multiple sets of concatenated Transformer layers are used for the extraction. Between or after modules, local feature extraction units centered on 3×3 two-dimensional convolutions are inserted. After the convolutional layers, batch normalization (BN) layers and ReLU activation functions are used. The former is used to stabilize feature value fluctuations during training, while the latter introduces nonlinearity to adapt to the learning of complex local morphology of teeth (such as crown margins and fine root structures). At the same time, this module adopts a dense feature connection method similar to DenseNet, allowing the local features output from the previous convolutional layer to be directly passed to the next layer, enhancing the reuse of detailed information and alleviating the gradient vanishing problem. Finally, feature maps are extracted from different depths of the encoder (i.e., the outputs of "Transformer + convolution" units in different rounds) to form multi-scale feature maps containing shallow fine-grained details (such as pixel differences at tooth boundaries) and deep global associations (such as the context of tooth alignment), providing basic feature inputs with both global semantics and local accuracy for subsequent modules.

[0022] The specific steps for multi-scale feature extraction from preprocessed image data are as follows: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed by an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module contains a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process From different depths of the hybrid architecture encoder, feature maps containing information at different scales are output, forming a multi-scale feature map.

[0023] Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map. The structure diagram of the cross-slice attention module is as follows: Figure 2 As shown; The cross-slice attention module includes position attention units and slice attention units: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0024] The position attention unit generates the position attention map through the following operations: Step 3A.1: For the input multi-scale feature map, perform global max pooling and global average pooling along its channel dimensions respectively, to obtain two spatial dimensions. Feature descriptors; Step 3A.2: Concatenate the two feature descriptors to form a fused feature map; Step 3A.3: Perform convolution operation on the concatenated fused feature map, and apply the Sigmoid activation function to the result of the convolution operation to generate a positional attention map, specifically represented as follows: (1) in, The input is the multi-scale initial feature map. Note the diagram for output position. The semicolon indicates a convolution operation, and the semicolon indicates a concatenation operation.

[0025] The slice attention unit generates the slice attention map through the following operations: Step 3B.1: Perform global max pooling and global average pooling along the height, width and channel dimensions of the input multi-scale feature map to obtain two slice-level feature vectors. Step 3B.2: Input the two slice-level feature vectors into a multilayer perceptron with the same structure for nonlinear transformation; Step 3B.3: Sum the two output feature vectors processed by the multilayer perceptron, and reshape the summed vector V to obtain the feature tensor. ; Step 3B.4: For the feature tensor Three different linear transformations are performed in parallel to compute a low-rank Gaussian distribution. The parameters, including the mean vector Covariance factor and the diagonal components of the covariance Calculate the covariance matrix Σ of the Gaussian distribution based on the above parameters, and then proceed from the Gaussian distribution... The vector z is obtained by sampling from the middle, and z is activated by the Sigmoid function to generate a slice attention map; The following three linear operations are performed on tensors: (2) in, The mean parameter weight matrix, The weight matrix of the diagonal components of the covariance. This is the covariance factor weight matrix. yes Reshaping the tensor; Covariance matrix of the distribution The calculation formula is: (3) in, yes The reshaping tensor, It is a diagonal matrix, with [variables] on the diagonal. elements, For parameterized rank; The slice attention map is represented as follows: (4) Among them, the output It is the reshaped version This refers to a slice attention map.

[0026] Step 4: Input the enhanced feature map into the dual fusion convolution module, and obtain the multi-dimensional fusion feature map through parallel dual-branch structure processing; The structure diagram of the dual-fusion convolutional module is as follows: Figure 3As shown, the dual-fusion convolution module uses "parallel dual-branch + feature concatenation" as its core structure to achieve collaborative extraction of two-dimensional planar semantics and three-dimensional spatial structure from the enhanced feature map: First, a single-layer pointwise convolution is used to reduce the channel dimensionality of the input feature map (preserving core spatial information while reducing computational cost), resulting in a unified-dimensional feature map F. Then, this feature map is input into two parallel branches—the two-dimensional convolution branch first performs zero-padding on the feature map F (ensuring the spatial size remains unchanged after convolution), and then a single-layer 3×3 two-dimensional convolution is used to extract planar semantic features within the slice (focusing on details such as local tooth boundaries and textures). After convolution, the features are then processed by batch normalization. The Batch Normalization (BN) and ReLU activation functions are used to optimize feature representation. The 3D convolution branch first expands the dimension (adapting the 3D convolution to the 4D tensor format with a shape of (P,P,S,1)), then uses a 3×3×3 3D convolution layer to capture the 3D spatial structural features of the teeth (such as the three-dimensional positional relationship between the tooth root and alveolar bone), and finally restores the 3D feature dimension through compression. Finally, the output feature maps of the two branches are directly spliced ​​along the channel dimension to form a multi-dimensional fusion feature map that combines two-dimensional semantic details with three-dimensional spatial correlation. At the same time, the complementarity of the two branches is used to suppress noise interference in CBCT images, thereby improving the integrity and robustness of feature representation.

[0027] The specific steps for processing a dual-branch structure are as follows: Step 4.1: Use a single pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain a feature map of uniform dimension. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0028] The dual-fusion convolution module includes parallel 2D and 3D convolution branches: The specific steps for capturing the three-dimensional spatial structural features of teeth using 3D convolutional branches are as follows: Step A.1: Expand the dimension of the input feature map F by adding an extra channel, transforming it into a 4D real tensor, denoted as . ,in, This refers to the expanded feature tensor prepared for input to the 3D convolutional layer. The shape of this 4D real tensor is ( P,P,S, 1); Step A.2: Perform a 3D convolution operation on the 4D real tensor to obtain... (such as the three-dimensional positional relationship between the tooth root and the alveolar bone), among which, The feature tensor output from the 3D convolutional layer, and zero-padding is used to ensure the output's... Space dimensions and input The matching process, in which the three-dimensional convolution operation is calculated, is shown in equation (5): (5) in, and These are the weights and bias parameters of the convolution, respectively. For three-dimensional convolution operators, For activation functions; Step A.3: Output of the 3D convolution operation Perform the extrusion operation to Converting back from 4D space to the original 3D space yields... The feature map F is then converted back to a three-dimensional space that matches the feature map F to obtain the three-dimensional spatial structural features of the teeth. The specific steps for extracting the two-dimensional planar semantic information of the feature map F using a two-dimensional convolutional branch are as follows: Step B.1: Zero-padding is performed on the input feature map F to ensure that the spatial dimensions of the output feature map after the convolution operation are consistent with those of the feature map F. Step B.2, use one layer Two-dimensional convolution kernels are used to perform convolution operations on the padded feature map to obtain the convolution result; Step B.3: After normalizing the result of the convolution operation, apply a non-linear activation function to obtain the enhanced two-dimensional planar semantic features. (Focusing on details such as the local boundaries and texture of teeth); The calculation process of the two-dimensional convolution branch is shown in equation (6): (6) in, Representing semantic features of a two-dimensional plane, and These are the weights and bias parameters of the convolution, respectively. It is a two-dimensional convolution operator, and BN is a normalization layer. This is the ReLU activation function.

[0029] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0030] Three-dimensional spatial structural features semantic features of two-dimensional plane The splicing process is shown in equation (7): (7) in, This indicates a splicing operation. This is the spliced ​​feature map. It is a three-dimensional spatial structure feature. It represents two-dimensional planar semantic features.

[0031] Step 5: Upsample the multi-dimensional fused feature map using a decoding network, and perform step-by-step recovery by combining the skip connection features from the encoder, finally outputting a 3D tooth segmentation result with the same resolution as the input image. The specific process is as follows: Step 5.1, Upsampling Initialization: Input the multi-dimensional fused feature map into the decoding network, and perform upsampling through interpolation or transpose convolution operations to improve the spatial resolution of the multi-dimensional fused feature map and obtain the upsampled feature map; Step 5.2, Skip Connection Fusion: Introducing a skip connection module to fuse the high-resolution local feature map output by the encoder at the corresponding scale with the upsampled feature map of the current decoder stage; the fusion method is channel stitching to supplement the spatial detail information lost during the encoder downsampling process, resulting in a fused feature map; Step 5.3, Feature Recovery and Output: Perform convolution operation on the fused feature map to eliminate the checkerboard effect caused by upsampling, and repeat steps 5.1-5.3 to restore the spatial resolution to the original CBCT image step by step, and finally output the three-dimensional tooth segmentation result.

[0032] Example 2 The difference between this embodiment and Embodiment 1 is that in step 5.2, the fusion method in the skip connection fusion process is to use element-wise addition.

[0033] Example 3 The CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information has the following specific steps: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; The CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0034] Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; The encoder is a hybrid architecture encoder, containing a Transformer module and a convolutional module, used to extract global semantic information and local feature information. It performs multi-scale feature extraction on preprocessed image data through the following process: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed by an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module contains a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process From different depths of the hybrid architecture encoder, feature maps containing information at different scales are output, forming a multi-scale feature map.

[0035] Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; The cross-slice attention module includes position attention units and slice attention units: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0036] The position attention unit generates the position attention map through the following operations: Step 3A.1: For the input multi-scale feature map, perform global max pooling and global average pooling along its channel dimensions respectively, to obtain two spatial dimensions. Feature descriptors; Step 3A.2: Concatenate the two feature descriptors to form a fused feature map; Step 3A.3: Perform convolution operation on the concatenated fused feature map, and apply the Sigmoid activation function to the result of the convolution operation to generate a positional attention map, specifically represented as follows: (1) in, The input is the multi-scale initial feature map. Note the diagram for output position. The semicolon indicates a convolution operation, and the semicolon indicates a concatenation operation.

[0037] The slice attention unit generates the slice attention map through the following operations: Step 3B.1: Perform global max pooling and global average pooling along the height, width and channel dimensions of the input multi-scale feature map to obtain two slice-level feature vectors. Step 3B.2: Input the two slice-level feature vectors into a multilayer perceptron with the same structure for nonlinear transformation; Step 3B.3: Sum the two output feature vectors processed by the multilayer perceptron, and reshape the summed vector V to obtain the feature tensor. ; Step 3B.4: For the feature tensor Three different linear transformations are performed in parallel to compute a low-rank Gaussian distribution. The parameters, including the mean vector Covariance factor and the diagonal components of the covariance Calculate the covariance matrix Σ of the Gaussian distribution based on the above parameters, and then proceed from the Gaussian distribution... The vector z is obtained by sampling from the middle, and z is activated by the Sigmoid function to generate a slice attention map; The following three linear operations are performed on tensors: (2) in, The mean parameter weight matrix, The weight matrix of the diagonal components of the covariance. This is the covariance factor weight matrix. yes Reshaping the tensor; Covariance matrix of the distribution The calculation formula is: (3) in, yes The reshaping tensor, It is a diagonal matrix, with [variables] on the diagonal. elements, For parameterized rank; The slice attention map is represented as follows: (4) Among them, the output It is the reshaped version This refers to a slice attention map.

[0038] Step 4: Input the enhanced feature map into the dual-fusion convolution module, and obtain the multi-dimensional fused feature map through parallel dual-branch structure processing; the specific steps of the dual-branch structure processing are as follows: Step 4.1: Use the pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0039] The dual-fusion convolution module includes parallel 2D and 3D convolution branches: The specific steps for capturing the three-dimensional spatial structural features of teeth using 3D convolutional branches are as follows: Step A.1: Expand the dimension of the input feature map F by adding an extra channel, transforming it into a 4D real tensor, denoted as . ,in, This refers to the expanded feature tensor prepared for input to the 3D convolutional layer. The shape of this 4D real tensor is ( P,P,S, 1); Step A.2: Perform a 3D convolution operation on the 4D real tensor to obtain... ,in, The feature tensor output from the 3D convolutional layer, and zero-padding is used to ensure the output's... Space dimensions and input The matching process, in which the three-dimensional convolution operation is calculated, is shown in equation (5): (5) in, and These are the weights and bias parameters of the convolution, respectively. For three-dimensional convolution operators, For activation functions; Step A.3: Output of the 3D convolution operation Perform the extrusion operation to Converting back from 4D space to the original 3D space yields... The feature map F is then converted back to a three-dimensional space that matches the feature map F to obtain the three-dimensional spatial structural features of the teeth. The specific steps for extracting the two-dimensional planar semantic information of the feature map F using a two-dimensional convolutional branch are as follows: Step B.1: Zero-padding is performed on the input feature map F to ensure that the spatial dimensions of the output feature map after the convolution operation are consistent with those of the feature map F. Step B.2, using Two-dimensional convolution kernels are used to perform convolution operations on the padded feature map to obtain the convolution result; Step B.3: After normalizing the result of the convolution operation, apply a non-linear activation function to obtain the enhanced two-dimensional planar semantic features. ; The calculation process of the two-dimensional convolution branch is shown in equation (6): (6) in, Representing semantic features of a two-dimensional plane, and These are the weights and bias parameters of the convolution, respectively. It is a two-dimensional convolution operator, and BN is a normalization layer. This is the activation function.

[0040] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0041] Three-dimensional spatial structural features semantic features of two-dimensional plane The splicing process is shown in equation (7): (7) in, This indicates a splicing operation. This is the spliced ​​feature map. It is a three-dimensional spatial structure feature. It represents two-dimensional planar semantic features.

[0042] Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

[0043] Example 4 The CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information has the following specific steps: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; The CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0044] Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; The encoder is a hybrid architecture encoder, containing a Transformer module and a convolutional module, used to extract global semantic information and local feature information. It performs multi-scale feature extraction on preprocessed image data through the following process: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed by an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module contains a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process From different depths of the hybrid architecture encoder, feature maps containing information at different scales are output, forming a multi-scale feature map.

[0045] Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; The cross-slice attention module includes position attention units and slice attention units: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0046] The position attention unit generates the position attention map through the following operations: Step 3A.1: For the input multi-scale feature map, perform global max pooling and global average pooling along its channel dimensions respectively, to obtain two spatial dimensions. Feature descriptors; Step 3A.2: Concatenate the two feature descriptors to form a fused feature map; Step 3A.3: Perform convolution operation on the concatenated fused feature map, and apply the Sigmoid activation function to the result of the convolution operation to generate a positional attention map, specifically represented as follows: (1) in, The input is the multi-scale initial feature map. Note the diagram for output position. The semicolon indicates a convolution operation, and the semicolon indicates a concatenation operation.

[0047] The slice attention unit generates the slice attention map through the following operations: Step 3B.1: Perform global max pooling and global average pooling along the height, width and channel dimensions of the input multi-scale feature map to obtain two slice-level feature vectors. Step 3B.2: Input the two slice-level feature vectors into a multilayer perceptron with the same structure for nonlinear transformation; Step 3B.3: Sum the two output feature vectors processed by the multilayer perceptron, and reshape the summed vector V to obtain the feature tensor. ; Step 3B.4: For the feature tensor Three different linear transformations are performed in parallel to compute a low-rank Gaussian distribution. The parameters, including the mean vector Covariance factor and the diagonal components of the covariance Calculate the covariance matrix Σ of the Gaussian distribution based on the above parameters, and then proceed from the Gaussian distribution... The vector z is obtained by sampling from the middle, and z is activated by the Sigmoid function to generate a slice attention map; The following three linear operations are performed on tensors: (2) in, The mean parameter weight matrix, The weight matrix of the diagonal components of the covariance. This is the covariance factor weight matrix. yes Reshaping the tensor; Covariance matrix of the distribution The calculation formula is: (3) in, yes The reshaping tensor, It is a diagonal matrix, with [variables] on the diagonal. elements, For parameterized rank; The slice attention map is represented as follows: (4) Among them, the output It is the reshaped version This refers to a slice attention map.

[0048] Step 4: Input the enhanced feature map into the dual-fusion convolution module, and obtain the multi-dimensional fused feature map through parallel dual-branch structure processing; the specific steps of the dual-branch structure processing are as follows: Step 4.1: Use the pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0049] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0050] Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

[0051] Example 5 The CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information has the following specific steps: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; The CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0052] Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; The encoder is a hybrid architecture encoder, containing a Transformer module and a convolutional module, used to extract global semantic information and local feature information. It performs multi-scale feature extraction on preprocessed image data through the following process: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed by an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module contains a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process From different depths of the hybrid architecture encoder, feature maps containing information at different scales are output, forming a multi-scale feature map.

[0053] Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; The cross-slice attention module includes position attention units and slice attention units: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0054] Step 4: Input the enhanced feature map into the dual-fusion convolution module, and obtain the multi-dimensional fused feature map through parallel dual-branch structure processing; the specific steps of the dual-branch structure processing are as follows: Step 4.1: Use the pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0055] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0056] Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

[0057] Example 6 The CBCT tooth segmentation method that integrates slice attention mechanism and Transformer semantic information has the following specific steps: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; The CBCT image preprocessing process includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise and other data enhancement operations, and finally normalizing the image to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and uniform format.

[0058] Step 2: Use the encoder to extract features from the preprocessed image to obtain a multi-scale feature map; The encoder is a hybrid architecture encoder, containing a Transformer module and a convolutional module, used to extract global semantic information and local feature information. It performs multi-scale feature extraction on preprocessed image data through the following process: Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; The cross-slice attention module includes position attention units and slice attention units: The positional attention unit is used to process the input multi-scale feature map to generate a positional attention map representing the key spatial regions within a single slice; The slice attention unit is used to process the input multi-scale feature map and generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

[0059] Step 4: Input the enhanced feature map into the dual-fusion convolution module, and obtain the multi-dimensional fused feature map through parallel dual-branch structure processing; the specific steps of the dual-branch structure processing are as follows: Step 4.1: Use the pointwise convolutional layer of the dual-fusion convolutional module to perform channel dimensionality reduction on the input enhanced feature map (reducing computation while preserving core spatial information) to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. The three-dimensional convolution branch is used to extract features in the three-dimensional space plane. It expands the dimension of the feature map F after dimensionality reduction, then performs convolution operation using a three-dimensional convolution kernel, and finally compresses the result back into the three-dimensional space that matches F to obtain the three-dimensional spatial structure features. The two-dimensional convolution branch is used to extract features in the two-dimensional space plane. It performs convolution operation on the feature map F and performs activation function and normalization processing in sequence to obtain two-dimensional planar semantic features.

[0060] Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

[0061] Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

[0062] The decoder used in this application adopts an L-level hierarchical structure, which is completely symmetrical to the downsampling levels of the encoder. Each level corresponds to one upsampling + feature fusion + feature extraction operation, ultimately resulting in... Restore to .

[0063] Table 1 Decoder L-level Hierarchical Structure

[0064] Note: Hk=2×Hk+1, Wk=2×Wk+1. Each upsampling step increases the spatial resolution by 2 times. The number of channels is usually halved proportionally (e.g., Ck=Ck+1 / 2) to ensure a balanced computational load.

[0065] Comparative Example 1 The technical advantages of this application will be illustrated below with a specific comparative example; This experiment used 129 scans, divided into a training set of 103 scans and a test set of 26 scans. The original image data format was NIFTI, the voxel spacing was 0.25 mm, and the resolutions were as follows: voxels, voxels and Voxel.

[0066] The server is configured as follows: it uses an Intel(R) Xeon(R) Platinum 8352v processor, equipped with a single NVIDIA GeForce RTX 4090 GPU, 32GB of memory, the operating system is Ubuntu 20.04, the programming language is Python 3.8, and the deep learning framework is PyTorch 2.0.0.

[0067] For the CBCT dental dataset, each 3D image is resampled to a uniform voxel spacing of [value missing]. mm, then randomly crop each image to The dimensions are in mm. Data augmentation methods such as random elastic deformation, Gaussian noise addition, random flipping, random scaling, random rotation, and random shifting were also employed. All images were standardized and normalized to fit within the intensity range of [0,1]. Gradient accumulation techniques were also used to improve training stability. The Adam optimizer was used with an initial learning rate of 0.0005 and weight decay of 0.00005. All networks were trained for 20 epochs with a batch size of 1.

[0068] The model is trained using the Binary Cross Entropy Loss (BCE Loss) function, a commonly used loss function in binary classification problems such as image semantic segmentation. Its mathematical expression is shown in equation (8):

[0069] in, express Medium pixel labels, This is the predicted segmentation map. BCE Loss can be used to evaluate the quality of predictions from a binary classification network model. For example, for binary labels... If the predicted value is 1, If the predicted value approaches 1, then the value of the loss function approaches 0. Conversely, if the predicted value... If the value approaches 0, then the value of the loss function approaches 1.

[0070] The performance of the model on the CBCT dental dataset was evaluated by comprehensively comparing this application with seven representative 3D networks, and the results are as follows: These networks include UNet3D based on the traditional UNet architecture, DenseVNe based on dense feature stacking, DenseVoxelNet based on attention mechanism, and UNETR, SwinUNETR, nnFormer, 3DUX-Net, and nnU-Net RL based on Transformer. The comparison results are shown in Table 2.

[0071] Table 2 Comparison results of different methods on the dataset

[0072] Experimental results show that the proposed method outperforms other comparative methods in tooth segmentation tasks. As shown in Table 2, in terms of specific evaluation metrics, the proposed method achieves a Hausdorff distance (HD) of 4.73 mm, an average symmetric surface distance (ASSD) of 0.93 mm, an intersection-over-union ratio (IoU) of 85.67%, a similarity overlap (SO) of 95.57%, and a DICE similarity coefficient (DICE) of 92.39%. While UNet3D has a simple structure, it lacks an effective feature reuse mechanism and is prone to gradient vanishing as network depth increases, leading to insufficient segmentation boundary accuracy. Compared to UNet3D, the proposed method improves the IoU metric by 15.05%.

[0073] DenseVNet employs dense connections to enhance feature reuse, but this also introduces redundant features, increasing model complexity and limiting further improvements in segmentation accuracy. Compared to DenseVNet, this application achieves a 1.10% improvement in IoU. While DenseVoxelNet enhances feature reuse in 3D space with its dense connection strategy, its overly complex structure leads to feature redundancy, limiting the model's ability to segment fine structures. Compared to DenseVoxelNet, this application achieves a 4.16% improvement in IoU. The UNETR model combines a Transformer structure to capture long-range dependencies, but lacks fine-grained characterization of local spatial features, resulting in decreased boundary recognition performance. Compared to UNETR, this method achieves an 11.37% improvement in IoU.

[0074] SwinUNETR extracts features using a sliding window approach combined with a Transformer module, but it still suffers from the loss of local detail information, leading to reduced segmentation accuracy in complex regions. Compared to SwinUNETR, our method improves the IoU by 2.57%. nnFormer uses all 3D convolutions, effectively capturing 3D spatial features, but lacks a 2D feature fusion mechanism, resulting in insufficient planar feature parsing and limiting overall segmentation accuracy. Compared to nnFormer, our method improves the IoU by 2.13%. While the 3D-UX-Net structure effectively captures multi-scale information in 3D images, its complex structure and insufficient feature aggregation lead to unclear feature representation, thus affecting segmentation performance. Compared to 3D-UX-Net, our method improves the IoU by 10.27%.

[0075] While nnU-Net RL enhances feature extraction capabilities through residual coding layers, its inter-layer fusion mechanism lacks flexibility and its context modeling ability is limited, leading to insufficient multi-scale feature interaction and missegmentation of regions with blurred boundaries, thus restricting overall segmentation accuracy. Compared to nnU-Net RL, our method improves the IoU metric by 1.54%.

[0076] Figure 4 The qualitative study in the paper presents a performance comparison of different networks on the CBCT dental dataset. Figure 4 The image in (a) shows the segmentation results of different networks on the hemidentate. The method in this paper can accurately segment small regions, while the contrasting methods often lead to oversegmentation. Figure 4 The image in (b) shows the segmentation results of different networks in cases of missing teeth and interference from other bone tissues. The method of this application demonstrates good segmentation performance and does not over-segment other bone tissues or the edentulous area. Figure 4 The image in (c) shows the segmentation results of different networks on the uneven surface of the crown. Compared with other networks, the method in this application preserves more details of the irregular surface contour. Figure 4 The image in (d) shows the segmentation results of teeth with implants using different networks. Compared with other networks, the method in this application preserves the complete details of the implant shape surface contour.

[0077] Figure 5 This paper presents a comparison of the 3D reconstruction performance of different networks on the CBCT dental dataset. Figure 5 The image in (a) shows the segmentation results of different networks for the hemidentate region. The method proposed in this application can accurately segment the hemidentate region, while the comparison methods often result in undersegmentation. Figure 5 The image in (b) shows that in cases of missing teeth or involvement of other bone tissue, the method proposed in this application does not oversegment adjacent bone tissue or incorrectly segment the missing tooth area, while other methods exhibit oversegmentation or incorrect segmentation. Figure 5 (c) shows the segmentation results of different networks when processing irregular shapes on the crown surface. The method proposed in this application can better preserve the details of irregular shapes on the crown surface, while other networks often fail to accurately preserve these irregular surface features. Figure 5 (d) shows the segmentation results of different networks under the influence of missing teeth and other bone tissues. Compared with other methods, the method proposed in this application can maintain good segmentation results and avoid over-segmentation of other bone tissues or edentulous areas.

Claims

1. A CBCT tooth segmentation method integrating slice attention mechanism and Transformer semantic information, characterized in that, The specific steps are as follows: Step 1: Acquire raw CBCT images and preprocess them to obtain preprocessed images; Step 2: Use an encoder to extract features from the preprocessed image to obtain a multi-scale feature map; Step 3: Input the multi-scale feature map into the cross-slice attention module and output the enhanced feature map; Step 4: Input the enhanced feature map into the dual fusion convolution module, and obtain a multi-dimensional fusion feature map through parallel dual-branch structure processing; Step 5: Upsample the multi-dimensional fused feature map through the decoding network, and combine it with the skip connection features from the encoder for stepwise recovery, finally outputting a three-dimensional tooth segmentation result with the same resolution as the input image.

2. The CBCT tooth segmentation method according to claim 1, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The CBCT image preprocessing process described in step 1 includes resampling the original CBCT image to a uniform voxel spacing, randomly cropping it to a fixed size, performing random elastic deformation, adding Gaussian noise, and other data enhancement operations. Finally, the image is normalized to the intensity range of [0,1] to obtain a preprocessed image with no redundant interference and a uniform format.

3. The CBCT tooth segmentation method according to claim 1, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The specific process of step 2 is as follows: The encoder is a hybrid architecture encoder, which includes a Transformer module and a convolution module, used to extract global semantic information and local feature information. It performs multi-scale feature extraction on the preprocessed image data through the following process: Step 2.1, Embedding Process The preprocessed image is converted into a hidden feature representation through linear projection; Step 2.2, Global Context Encoding Process The hidden feature representation is processed using an encoding layer consisting of multiple cascaded Transformer modules, where each Transformer module includes a multi-head self-attention mechanism and a feedforward neural network to capture the global semantic information of the image. Step 2.3: Local detail enhancement process Convolutional operations are introduced between or after the global context encoding process to enhance the capture of local detailed features; Step 2.4, Multi-scale output process The hybrid architecture encoder outputs feature maps containing information at different scales from different depths, forming the multi-scale feature map.

4. The CBCT tooth segmentation method according to claim 1, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The cross-slice attention module mentioned in step 3 includes a position attention unit and a slice attention unit: The position attention unit is used to process the input multi-scale feature map to generate a position attention map representing the key spatial regions within a single slice. The slice attention unit is used to process the input multi-scale feature map to generate a slice attention map that represents the importance between different slices; The location attention map and the slice attention map work together on the multi-scale feature map to enhance its feature representation in the spatial and slice dimensions, and output the spatial-slice association enhanced feature map.

5. The CBCT tooth segmentation method according to claim 4, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The location attention unit generates a location attention map through the following operations: Step 3A.1: For the input multi-scale feature map, perform global max pooling and global average pooling along its channel dimensions respectively, to obtain two spatial dimensions. Feature descriptors; Step 3A.2: Concatenate the two feature descriptors to form a fused feature map; Step 3A.3: Perform a convolution operation on the concatenated fused feature map, and apply the Sigmoid activation function to the result of the convolution operation to generate the positional attention map, specifically as follows: (1) in, The input is the multi-scale initial feature map. Note the diagram for output position. The semicolon indicates a convolution operation, and the semicolon indicates a concatenation operation.

6. The CBCT tooth segmentation method according to claim 4, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The slice attention unit generates the slice attention map through the following operations: Step 3B.1: Perform global max pooling and global average pooling along the height, width and channel dimensions of the input multi-scale feature map to obtain two slice-level feature vectors. Step 3B.2: Input the two slice-level feature vectors into a multilayer perceptron with the same structure for nonlinear transformation; Step 3B.3: Sum the two output feature vectors processed by the multilayer perceptron, and reshape the summed vector V to obtain the feature tensor. ; Step 3B.4: For the feature tensor Three different linear transformations are performed in parallel to compute a low-rank Gaussian distribution. The parameters, including the mean vector Covariance factor and the diagonal components of the covariance Calculate the covariance matrix Σ of the Gaussian distribution based on the above parameters, and then proceed from the Gaussian distribution... The vector z is obtained by sampling from the middle, and z is activated by the Sigmoid function to generate a slice attention map; The following three linear operations are performed on tensors: (2) in, The mean parameter weight matrix, The weight matrix for the diagonal components of the covariance. This is the covariance factor weight matrix. yes Reshaping the tensor; Covariance matrix of the distribution The calculation formula is: (3) in, yes The reshaping tensor, It is a diagonal matrix, with [variables] on the diagonal. elements, For parameterized rank; The slice attention map is represented as follows: (4) Among them, the output It is the reshaped version This refers to a slice attention map.

7. The CBCT tooth segmentation method according to claim 1, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The steps for processing the dual-branch structure in step 4 are as follows: Step 4.1: Use the pointwise convolutional layer of the dual fusion convolution module to perform channel dimensionality reduction on the input enhanced feature map to obtain the feature map. S represents the number of channels obtained after point-wise convolution. yes Reshaping the tensor; Step 4.2: Input the feature map F into one two-dimensional convolutional branch and one three-dimensional convolutional branch of the dual fusion convolutional module simultaneously to obtain two-dimensional planar semantic features and three-dimensional spatial structural features, respectively. Step 4.3: Concatenate the two-dimensional planar semantic features and the three-dimensional spatial structural features to output a multi-dimensional fusion feature map with noise suppression.

8. The CBCT tooth segmentation method according to claim 7, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The dual-fusion convolution module described in step 4.2 includes parallel two-dimensional convolution branches and three-dimensional convolution branches: The specific steps for capturing the three-dimensional spatial structural features of teeth using 3D convolutional branches are as follows: Step A.1: Expand the dimension of the input feature map F by adding an extra channel, transforming it into a 4D real tensor, denoted as . ,in, This refers to the expanded feature tensor prepared for input to the 3D convolutional layer. The shape of this 4D real tensor is ( P,P,S, 1); Step A.2: Perform a three-dimensional convolution operation on the 4D real tensor to obtain... ,in, The feature tensor output from the 3D convolutional layer, and zero-padding is used to ensure the output's... Space dimensions and input The matching process, in which the three-dimensional convolution operation is calculated, is shown in equation (5): (5) in, and These are the weights and bias parameters of the convolution, respectively. For three-dimensional convolution operators, For activation functions; Step A.3: Output of the 3D convolution operation Perform the extrusion operation to Converting back from 4D space to the original 3D space yields... The feature map is then converted back to a three-dimensional space that matches the feature map F to obtain the three-dimensional spatial structural features of the tooth. The specific steps for extracting the two-dimensional planar semantic information of the feature map F using the two-dimensional convolutional branch are as follows: Step B.1: Zero-padding is performed on the input feature map F to ensure that the spatial dimensions of the output feature map after the convolution operation are consistent with those of the feature map F. Step B.2, using Two-dimensional convolution kernels are used to perform convolution operations on the padded feature map to obtain the convolution result; Step B.3: After normalizing the result of the convolution operation, apply a non-linear activation function to obtain the enhanced two-dimensional planar semantic features. ; The calculation process of the two-dimensional convolution branch is shown in equation (6): (6) in, Representing semantic features of a two-dimensional plane, and These are the weights and bias parameters of the convolution, respectively. It is a two-dimensional convolution operator, and BN is a normalization layer. This is the activation function.

9. The CBCT tooth segmentation method according to claim 7, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The three-dimensional spatial structural features described in step 4.3 semantic features of two-dimensional plane The splicing process is shown in equation (7): (7) in, This indicates a splicing operation. This is the spliced ​​feature map. It is a three-dimensional spatial structure feature. It represents two-dimensional planar semantic features.

10. The CBCT tooth segmentation method according to claim 1, which integrates slice attention mechanism and Transformer semantic information, is characterized in that... The specific process of step 5 is as follows: Step 5.1, Upsampling Initialization: Input the multi-dimensional fused feature map into the decoding network, and perform upsampling through interpolation or transpose convolution to obtain the upsampled feature map; Step 5.2, Skip Connection Fusion: Introducing a skip connection module to fuse the high-resolution local feature map output by the encoder at the corresponding scale with the upsampled feature map of the current decoder stage; the fusion method is channel splicing or element-wise addition to supplement the spatial detail information lost during the encoder downsampling process and obtain the fused feature map; Step 5.3, Feature Recovery and Output: Perform convolution operation on the fused feature map to eliminate the checkerboard effect caused by upsampling. Repeat steps 5.1-5.3 to restore the spatial resolution to the original CBCT image step by step, and finally output the three-dimensional tooth segmentation result.

Images