A CBCT-based tooth instance segmentation method based on boundary supervision and multiple attention.

Abstract

This invention discloses a CBCT tooth instance segmentation method based on boundary supervision and multiple attention mechanisms. The original CBCT image is input into a four-layer CBAM-based convolutional block feature encoding network to obtain the first and second layer feature maps generated after passing through the first two layers, and the feature map generated after passing through the fourth layer. The first two layer feature maps obtained in the previous step are then input into a boundary supervision module to obtain the boundary segmentation result. The fourth layer feature map is then input into a self-attention module to obtain a deep feature map, which is subsequently input into a four-layer CBAM-based convolutional block feature decoding network to obtain the region segmentation result. After fusing the boundary segmentation result and the region segmentation result, the final 3D tooth instance segmentation result is obtained. This invention significantly enhances the network's ability to extract boundary features and contextual information, enabling the model to achieve good generalization ability when processing complex CBCT images with varying morphologies.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a CBCT tooth instance segmentation method based on boundary supervision and multiple attention. Background Technology

[0002] With increasing demands for appearance and growing concern for oral health, orthodontics has become a major dental service. Orthodontics helps patients restore stable occlusion and improve facial aesthetics by treating complex tooth defects and severe craniofacial deformities. Compared to traditional panoramic radiographs, cone-beam computed tomography (CBCT) images avoid defects such as overlap, distortion, and deformation found in two-dimensional planes. Furthermore, CBCT images offer advantages such as faster scanning speed, lower radiation dose, higher image accuracy, and lower cost. Therefore, cone-beam computed tomography (CBCT) is increasingly used for diagnosis in dental practice.

[0003] Orthodontic treatment involves diagnosis, determining a treatment plan, appropriate treatment, and retention. The most crucial step in orthodontic treatment is diagnosis, as an incorrect diagnosis can lead to an incorrect treatment plan. A correct diagnosis requires understanding the patient's medical history, conducting a physical examination, examining radiographic imaging such as cephalometric images and CBCT images, analyzing diagnostic models, and fabricating an occlusal appliance. The most critical aspect of the diagnostic process is the dentist's ability to analyze the patient's dentofacial structure using radiographic images such as CBCT to determine the appropriate treatment plan, including implantation or extraction of teeth, correction of alignment angles, and postoperative evaluation.

[0004] However, this process requires dentists to mark and model each slice of the dental and maxillofacial region and perform cephalometric analysis to clearly understand the patient's condition. This process is time-consuming, labor-intensive, and highly dependent on the dentist's experience. Therefore, automated 3D tooth segmentation helps dentists simulate treatment and develop treatment plans, playing a crucial role in maintaining oral health and improving the diagnostic efficiency of dental hospitals. Thus, faster and more accurate finer tooth segmentation not only reduces the possibility of misdiagnosis but also shortens the overall diagnostic process.

[0005] The patent title is "A Method and System for Three-Dimensional Segmentation of Dental CBCT Images," publication number CN114241173A, publication date March 25, 2022. The patent describes a method and system for three-dimensional segmentation of dental CBCT images. This method first preprocesses the dental CBCT images using an anisotropic diffusion method; then, it uses a combination of thresholding and morphology to obtain the region of interest (ROI) of the teeth, thereby generating an initial level set function; based on this, a three-dimensional variable-scale region fitting segmentation model is used to extract the tooth region.

[0006] The patent title is "Dental CBCT 3D Tooth Segmentation Method Based on Deep Learning," application number CN113628223A, publication date 2021-11-09. The patent describes a dental CBCT 3D tooth segmentation method based on deep learning. This method first collects dental CBCT images and performs annotation and preprocessing. The preprocessed images are used to construct a model training set, which is then input into a newly constructed encoder-decoder structured network segmentation model for training and evaluation. The trained network segmentation model is then used to perform semantic segmentation of the teeth.

[0007] The patent title is "A Method and Storage Medium for Tooth Segmentation and Reconstruction Based on CBCT Images," with authorization number CN114757960B and authorization date of 2022-09-09. The patent describes a method and storage medium for tooth segmentation and reconstruction based on CBCT images. This method first performs custom layer-by-layer annotation on CBCT data, automatically statistically analyzes the annotated tooth center points, and preprocesses the CBCT dataset. Then, based on the open-source deep learning platform TensorFlow and a fully convolutional neural network model, a CBCT tooth image segmentation network model is designed and trained. The probability value of each voxel is obtained through the output of the trained model. Optimization is achieved by separating and merging overlapping upper and lower teeth, followed by refinement of the root apex shape, and finally, a three-dimensional tooth model is reconstructed.

[0008] Existing technologies for CBCT 3D tooth segmentation using models can only perform semantic segmentation of teeth, failing to separate teeth individually based on their position, thus not meeting practical needs. While some technologies achieve instance segmentation, they do not employ targeted deep learning methods to improve tooth segmentation accuracy, do not fully consider various tooth features, and fail to address the issue of blurred tooth boundaries, resulting in inaccurate boundaries and proneness to incorrect tooth classification. Summary of the Invention

[0009] The purpose of this application is to propose a CBCT tooth instance segmentation method based on boundary supervision and multiple attention, which improves the problems of blurred tooth boundary segmentation and incorrect tooth category judgment in the existing technology when performing CBCT tooth segmentation on complex cases such as abnormal tooth number and abnormal tooth arrangement.

[0010] To achieve the above objectives, the technical solution of this application is as follows:

[0011] A CBCT tooth instance segmentation method based on boundary supervision and self-attention includes:

[0012] CBCT images are input into a feature coding network consisting of four layers of CBAM-based convolutional blocks, resulting in feature maps F1 and F2 generated by the first and second layers, and feature map F after passing through the fourth layer.deep_orig ;

[0013] The obtained feature maps F1 and F2 are input into the boundary supervision module to obtain the boundary segmentation result F. edge ;

[0014] The obtained feature map F deep_orig The input is fed into the self-attention module to obtain the feature map F. deep ;

[0015] Feature map F deep The input is fed into a feature decoding network consisting of four layers of CBAM-based convolutional blocks to obtain the region segmentation result F of the CBCT image. area ;

[0016] The boundary segmentation result F edge And the region segmentation result F area The fusion process yields the final three-dimensional tooth segmentation result F. out .

[0017] Furthermore, the CBCT image is input into the feature encoding network of the four-layer CBAM attention module to obtain feature maps F1 and F2 generated in the first and second layers, and feature map F generated after passing through the fourth layer. deep_orig ,include:

[0018] Step 1.1: Input the CBCT image into the first layer of the CBAM-based convolutional block to obtain the output feature map F1, which has dimensions of C1X D1X H1XW1;

[0019] Step 1.2: Max pool the feature map F1 and then input it into the second CBAM-based convolutional block to obtain the output feature map F2, which has dimensions of C2X D2X H2X W2.

[0020] Step 1.3: Max pool the feature map F2 and then input it into the third CBAM-based convolutional block to obtain the output feature map F3, which has dimensions of C3X D3X H3X W3.

[0021] Step 1.4: Perform max pooling on feature map F3, then input it into the fourth CBAM-based convolutional block to obtain the output feature map F. deep_orig Its dimensions are C4X D4X H4X W4.

[0022] Furthermore, the obtained feature maps F1 and F2 are input into the boundary supervision module to obtain the boundary segmentation result F. edge ,include:

[0023] Step 2.1: Concatenate feature map F1 and the upsampled feature map F2 along each channel to obtain feature map F.edge_orin Its dimensions are C5X D1X H1X W1;

[0024] Step 2.2: Transfer the feature map F edge_orin The input is fed into a convolutional layer with a kernel size of 1 x 1 x 1, and the boundary output F is obtained. edge Its dimensions are C6X D1X H1X W1, where C6 represents the number of categories of the boundaries of C6 different teeth.

[0025] Furthermore, the obtained feature map F deep_orig The input is fed into the self-attention module to obtain the feature map F. deep ,include:

[0026] Step 3.1: Transfer the feature map F deep_orig The input is fed into a fully connected layer to obtain the feature map F. deep_1 Its dimensions are c7X D4X H4X W4;

[0027] Step 3.2: Transfer the feature map F deep_1 The input is fed into the location embedding layer to obtain the feature vector V1, which has a dimension of .

[0028] Step 3.3: Perform layer normalization on feature vector V1, then input it into the multi-head attention layer, and then add the output and feature vector V1 to obtain feature vector V2, whose dimension is the same as V1.

[0029] Step 3.4: Perform layer normalization on feature vector V2, then input it into the multilayer perceptron layer, and then add the output and feature vector V2 to obtain feature vector V3, whose dimension is the same as V1.

[0030] Step 3.5: Repeat steps 3.3 to 3.4 three times to obtain feature vector V4, whose dimension is the same as that of V1.

[0031] Step 3.6: Change the shape of the feature vector V4 to obtain the feature map F. deep_2 Its dimension is the same as F deep_1 Consistent.

[0032] Step 3.7: Transfer the feature map F deep_2 The input is fed into the convolutional block to obtain the feature map F. deep Its dimensions are C4X D4XH4X W4.

[0033] Furthermore, the feature map F deep The feature decoding network, fed into a four-layer CBAM attention module, yields the region segmentation result F of the CBCT image. area ,include:

[0034] Step 4.1: Transfer the feature map F generated by the self-attention module. deep Upsampling is performed, followed by concatenation with feature map F3 along channels to obtain a feature map with C4+C3 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up1 Its dimensions are C3X D3X H3X W3;

[0035] Step 4.2: Transfer the feature map F up1 Upsampling is performed, followed by concatenation with feature map F2 along channels to obtain a feature map with C3+C2 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up2 Its dimensions are C2X D2XH2X W2;

[0036] Step 4.3: Transfer the feature map F up2 Upsampling is performed, followed by concatenation with feature map F1 channel by channel to obtain a feature map with C2+C1 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up3 Its dimensions are C1X D1XH1X W1;

[0037] Step 4.4: Transfer the feature map F up3 The input is fed into a convolutional layer with a kernel size of 1 x 1 x 1 to obtain the region segmentation output F. area Its dimensions are C6 x D1 x H1 x W1, where C6 represents the number of different categories of C6 teeth.

[0038] Furthermore, the boundary segmentation result F edge And the region segmentation result F area The segments are then fused to obtain the final segmentation result F. out ,include:

[0039] The boundary segmentation result F edge And the region segmentation result F area The concatenation is performed along channels, with a channel size of C6+C6. This concatenation is then passed through a convolutional layer with a kernel size of 1x1x1 to obtain the output feature map F. out Its dimensions are C6 x D1 x H1 x W1, and F out Activate the pixels of each channel to obtain F. out The probability of each pixel belonging to the corresponding tooth category is calculated, and the channel value with the highest probability at the pixel position is taken as the category of the pixel at that position, thus obtaining the final tooth segmentation result.

[0040] Furthermore, the CBAM-based convolutional block includes two sets of cascaded convolutional layers with a kernel size of 3x3x3, a batch normalization layer and an activation layer, as well as a set of CBAM modules.

[0041] Furthermore, the CBAM module includes performing one average pooling operation, a convolutional layer with a kernel size of 1x1x1, an activation layer, and another convolutional layer with a kernel size of 1x1x1 on the input feature map, and one max pooling operation, a convolutional layer with a kernel size of 1x1x1, an activation layer, and another convolutional layer with a kernel size of 1x1x1 on the input feature map to obtain two channel weight vectors. Then, the two channel weight vectors are added together and activated according to the following formula, and then multiplied with the input feature map to obtain a channel feature map. The channel feature maps are then subjected to one average pooling operation and one max pooling operation to obtain two spatial feature maps. The two spatial feature maps are then concatenated by channel and passed through a convolutional layer with a kernel size of 3x3x3 to obtain a spatial attention weight map. The spatial attention weight map is then activated and multiplied with the channel feature maps to obtain an output feature map.

[0042] This application proposes a CBCT tooth instance segmentation method based on boundary supervision and multiple attention mechanisms. Firstly, by incorporating a boundary supervision module, the model can extract boundary information of the target object, significantly improving the accuracy of tooth boundary segmentation. Even when CBCT morphology varies considerably, valuable boundary information can still be extracted. Secondly, by adding a CBAM module and a self-attention module, the model can extract richer global contextual information and more effective local feature information from the feature map. This allows the model to identify target objects of different locations and sizes, enhancing its generalization ability when handling complex CBCTs, ultimately resulting in more accurate 3D tooth instance segmentation. Attached Figure Description

[0043] Figure 1 This is a flowchart of the CBCT tooth instance segmentation method based on boundary supervision and attention proposed in this application;

[0044] Figure 2 This is a schematic diagram of the network structure in an embodiment of this application;

[0045] Figure 3 This is a schematic diagram of the boundary monitoring module of this application;

[0046] Figure 4 This is a schematic diagram of the self-attention module of this application;

[0047] Figure 5 This is a schematic diagram of a CBAM-based convolutional block in this application;

[0048] Figure 6This is a schematic diagram of the CBAM module of this application. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one embodiment, such as Figure 1 As shown, a CBCT tooth instance segmentation method based on boundary supervision and multiple attention is provided, including:

[0051] Step S1: Input the CBCT image into a feature coding network consisting of four layers of CBAM-based convolutional blocks to obtain feature maps F1 and F2 generated by the first and second layers, and feature map F generated after passing through the fourth layer. deep_orig .

[0052] like Figure 2 As shown in the figure, the CBCT image is input into a feature coding network comprising four layers of CBAM-based convolutional blocks to obtain feature map F. deep_orig ,include:

[0053] Step 1.1: Input the CBCT image into a CBAM-based convolutional block to obtain the output feature map F1, which has dimensions of C1X D1 X H1 X W1;

[0054] Step 1.2: Max pool the feature map F1 and then input it into a CBAM-based convolutional block to obtain the output feature map F2, which has dimensions of C2X D2 X H2X W2.

[0055] Step 1.3: Max pool the feature map F2 and then input it into a CBAM-based convolutional block to obtain the output feature map F3, which has dimensions of C3XD3XH3XW3.

[0056] Step 1.4: Perform max pooling on feature map F3, then input it into a CBAM-based convolutional block to obtain the output feature map F. deep_orig Its dimensions are C4 x D4 x H4 x W4.

[0057] The CBAM-based convolutional block, such as Figure 5 As shown, it includes two sets of cascaded convolutional layers with kernel size of 3x3x3, batch normalization layers and activation layers, as well as a set of CBAM modules.

[0058] The CBAM module, such as Figure 6As shown, the process includes performing one flat pooling operation, a convolutional layer with a kernel size of 1×1×1, an activation layer, and another convolutional layer with a kernel size of 1×1×1 on the input feature map, and one max pooling operation, a convolutional layer with a kernel size of 1×1×1, an activation layer, and another convolutional layer with a kernel size of 1×1×1 to obtain two channel weight vectors. These two channel weight vectors are then added together, activated according to the following formula, and multiplied with the input feature map to obtain a channel feature map. One flat pooling operation and one max pooling operation are then performed on each channel feature map to obtain two spatial feature maps. These two spatial feature maps are then concatenated by channel and passed through a convolutional layer with a kernel size of 3×3×3 to obtain a spatial attention weight map. Finally, the spatial attention weight map is activated according to the following formula and multiplied with the channel feature map to obtain the output feature map.

[0059]

[0060] Where x is the spatial attention weight map.

[0061] Step S2: Input the feature maps F1 and F2 obtained in the previous step into the boundary supervision module to obtain the edge segmentation result F. edge .

[0062] The feature maps F1 and F2 described in this application are input into the boundary supervision module to obtain the edge segmentation result F. edge ,include:

[0063] Step 2.1: Concatenate feature map F1 and the upsampled feature map F2 along each channel to obtain feature map F. edge_orin Its dimensions are C5 x D1 x H1 x W1;

[0064] Step 2.2: Transfer the feature map F edge_orin The input is fed into a convolutional layer with a kernel size of 1 x 1 x 1, and the boundary output F is obtained. edge Its dimensions are C6X D1 X H1X W1, where C6 represents the number of categories of the boundaries of C6 different teeth.

[0065] The boundary monitoring module, such as Figure 3 As shown, it is used to fuse the feature maps generated by the first and second layer encoders to extract the boundary information obtained after the convolution operation.

[0066] Step S3: Obtain the feature map F deep_orig The input is fed into the self-attention module to obtain the feature map F. deep .

[0067] The feature map F obtained as described in this application deep_orig The input is fed into the self-attention module to obtain the feature map F.deep ,include:

[0068] Step 3.1: Transfer the feature map F deep_orig The input is fed into a fully connected layer to obtain the feature map F. deep_1 Its dimensions are C7 x D4 x H4 x W4;

[0069] Step 3.2: Transfer the feature map F deep_1 The input is fed into the location embedding layer to obtain the feature vector V1, which has a dimension of .

[0070] Step 3.3: Perform layer normalization on feature vector V1, then input it into the multi-head attention layer, and then add the output and feature vector V1 to obtain feature vector V2, whose dimension is the same as V1.

[0071] Step 3.4: Perform layer normalization on feature vector V2, then input it into the multilayer perceptron layer, and then add the output and feature vector V2 to obtain feature vector V3, whose dimension is the same as V1.

[0072] Step 3.5: Repeat steps 3.2 to 3.4 three times to obtain a feature vector V4 repeated four times, with the same dimension as V1.

[0073] Step 3.6: Change the shape of the feature vector V4 to obtain the feature map F. deep_2 Its dimension is the same as F deep_1 Consistent.

[0074] Step 3.7: Transfer the feature map F deep_2 The input is fed into a channel-space attention-based convolutional block to obtain the feature map F. deep Its dimensions are C4 x D4 x H4 x W4.

[0075] The self-attention module, such as Figure 4 As shown, this is used to extract richer contextual information from the input feature map to better identify targets of varying sizes and dimensions.

[0076] It should be noted that in the attention module of this application, the number of repetitions can be adjusted according to the graphics card memory, for example, it can be adjusted to 2 times or 6 times.

[0077] Step S4: Transfer feature map F deep The input is fed into a feature decoding network consisting of four layers of CBAM-based convolutional blocks to obtain the region segmentation result F of the CBCT image. area .

[0078] like Figure 2 As shown, the feature map F described in this application deepThe input is fed into a feature decoding network consisting of four layers of CBAM-based convolutional blocks to obtain the region segmentation result F of the CBCT image. area ,include:

[0079] Step 4.1: Transfer the feature map F deep Upsampling is performed, followed by concatenation with feature map F3 along channels to obtain a feature map with C4+C3 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up1 Its dimensions are C3 x D3 x H3 x W3;

[0080] Step 4.2: Transfer the feature map F up1 Upsampling is performed, followed by concatenation with feature map F2 along channels to obtain a feature map with C3+C2 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up2 Its dimensions are C2X D2 XH2X W2;

[0081] Step 4.3: Transfer the feature map F up2 Upsampling is performed, followed by concatenation with feature map F1 channel by channel to obtain a feature map with C2+C1 channels. This feature map is then input into a CBAM-based convolutional block to obtain feature map F. up3 Its dimensions are C1X D1 XH1X W1;

[0082] Step 4.4: Transfer the feature map F up3 The input is fed into a convolutional layer with a kernel size of 1 x 1 x 1 to obtain the region segmentation output F. area Its dimensions are C6 x D1 x H1 x W1, where C6 represents the number of different categories of C6 teeth.

[0083] Step S5: Divide the boundary segmentation result F edge And the region segmentation result F area The segments are then fused to obtain the final segmentation result F. out .

[0084] The boundary segmentation result F described in this application is... edge And the region segmentation result F area The segments are then fused to obtain the final segmentation result F. out ,include:

[0085] The boundary segmentation result F edge And the region segmentation result F area The concatenation is performed along channels, with a channel size of C6+C6. This concatenation is then passed through a convolutional layer with a kernel size of 1x1x1 to obtain the output feature map F. out Its dimensions are C6 x D1 x H1 x W1, and F outEach pixel in each channel is activated using the following formula to obtain F. out The probability of each pixel belonging to the corresponding tooth category is calculated, and the channel value with the highest probability at the pixel position is taken as the category of the pixel at that position, thus obtaining the final tooth segmentation result.

[0086]

[0087] Where, x i F represents out The pixels of each channel.

[0088] The method and technical framework described in this application, such as Figure 2 As shown, in other embodiments, the number of layers in the feature encoding network and the feature decoding network can be modified.

[0089] It should be noted that in this application, C is the number of channels, D is the depth of the image, H is the height of the image, W is the width of the image, L is the vector length, and the subscripts of the letters indicate the sequence number, which is used to distinguish the dimensions of different feature maps.

[0090] This application improves the model's segmentation ability when faced with complex CBCT images by directly learning the features of the target object's boundary through boundary supervision. It can more effectively find the segmentation boundary, thereby improving the accuracy of tooth boundary segmentation. Furthermore, by incorporating the CBAM module and the self-attention module, the model can extract richer global contextual information and more effective local feature information from the feature map. This allows the model to identify target objects of different locations and sizes, enhancing its generalization ability when processing complex CBCT images.

[0091] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A CBCT tooth instance segmentation method based on boundary supervision and multiple attention, characterized in that, The CBCT tooth instance segmentation method based on boundary supervision and multiple attention includes: CBCT images are input into a feature coding network consisting of four layers of CBAM-based convolutional blocks to obtain feature maps generated by the first and second layers. , and the feature map generated after passing through four layers. ; The obtained feature map , The input is fed into the boundary supervision module to obtain the boundary segmentation result. ; The obtained feature map The input is fed into the self-attention module to obtain the feature map. ; feature map The input is fed into a feature decoding network consisting of four layers of CBAM-based convolutional blocks to obtain the region segmentation results of the CBCT image. ; Boundary segmentation results and region segmentation results The data is then fused to obtain the final three-dimensional tooth segmentation result. ; The CBAM-based convolutional block comprises two sets of cascaded convolutional kernels with a size of [missing value]. It consists of convolutional layers, batch normalization layers, activation layers, and a set of CBAM modules; The CBAM module includes performing average pooling on the input feature map once, and setting the convolution kernel size to be... The size of the convolutional layer, activation layer, and convolutional kernel is The operations of the convolutional layer and one max pooling, with a convolutional kernel size of The size of the convolutional layer, activation layer, and convolutional kernel is The convolutional layer operations yield two channel weight vectors. These two channel weight vectors are then summed, activated, and multiplied with the input feature map to obtain a channel feature map. One flat pooling operation and one max pooling operation are performed on each channel feature map to obtain two spatial feature maps. These two spatial feature maps are then concatenated by channel and passed through a convolutional kernel of size [size missing]. The convolutional layer is used to obtain a spatial attention weight map, which is then activated and multiplied with the channel feature map to obtain the output feature map.

2. The CBCT tooth instance segmentation method based on boundary supervision and multiple attention as described in claim 1, characterized in that, The CBCT images are input into a feature coding network consisting of four layers of CBAM-based convolutional blocks to obtain feature maps generated in the first and second layers. , and the feature map generated after passing through four layers. ,include: Step 1.1: Input the CBCT image into the first layer of the CBAM-based convolutional block to obtain the output feature map. Its dimension size is ; Step 1.2: Transfer the feature map Max pooling is performed, and then the result is fed into the second CBAM-based convolutional block to obtain the output feature map. Its dimension size is ; Step 1.3: Transfer the feature map Max pooling is performed, and then the result is fed into the third CBAM-based convolutional block to obtain the output feature map. Its dimension size is ; Step 1.4: Transfer the feature map Max pooling is performed, and then the result is fed into the fourth CBAM-based convolutional block to obtain the output feature map. Its dimension size is .

3. The CBCT tooth instance segmentation method based on boundary supervision and multiple attention as described in claim 1, characterized in that, The feature map to be obtained , The input is fed into the boundary supervision module to obtain the boundary segmentation result. ,include: Step 2.1: Transfer the feature map and the feature map after upsampling By splicing the data along the channels, we obtain the feature map. Its dimension size is ; Step 2.2: Transfer the feature map The input to the convolution kernel size is The convolutional layer yields the boundary output. Its dimension size is ,in represent The number of categories of different tooth boundaries.

4. The CBCT tooth instance segmentation method based on boundary supervision and multiple attention as described in claim 1, characterized in that, The feature map to be obtained The input is fed into the self-attention module to obtain the feature map. ,include: Step 3.1: Transfer the feature map The input is fed into a fully connected layer to obtain a feature map. Its dimension size is ; Step 3.2: Transfer the feature map The input is fed into the location embedding layer to obtain the feature vector. Its dimension size is ; Step 3.3: Transfer the feature vector Perform layer normalization, then input it into a multi-head attention layer, and then combine the output with the feature vector. Add them together to obtain the feature vector. Its dimensional size is Consistent; Step 3.4: Transfer the feature vector Perform layer normalization, then input it into a multilayer perceptron layer, and then combine the output with the feature vector. Add them together to obtain the feature vector. Its dimensional size is Consistent; Step 3.5: Repeat steps 3.3 to 3.4 three times to obtain the feature vector. Its dimensional size is Consistent; Step 3.6: Transfer the feature vector By changing the shape, a feature map is obtained. Its dimensional size is Consistent; Step 3.7: Transfer the feature map The input is fed into a convolutional block to obtain a feature map. Its dimension size is .

5. The CBCT tooth instance segmentation method based on boundary supervision and multiple attention as described in claim 1, characterized in that, The feature map The input is fed into a feature decoding network consisting of four layers of CBAM-based convolutional blocks to obtain the region segmentation results of the CBCT image. ,include: Step 4.1: Convert the feature map generated by the self-attention module. Upsampling is performed, followed by comparison with feature maps. By concatenating the channels, the number of channels is obtained as follows: The feature map is input into a CBAM-based convolutional block to obtain the feature map. Its dimension size is ; Step 4.2: Transfer the feature map Upsampling is performed, followed by comparison with feature maps. By concatenating the channels, the number of channels is obtained as follows: The feature map is input into a CBAM-based convolutional block to obtain the feature map. Its dimension size is ; Step 4.3: Transfer the feature map Upsampling is performed, followed by comparison with feature maps. By concatenating the channels, the number of channels is obtained as follows: The feature map is input into a CBAM-based convolutional block to obtain the feature map. Its dimension size is ; Step 4.4: Transfer the feature map The input to the convolution kernel size is The convolutional layer produces the region segmentation output. Its dimension size is ,in represent The number of different categories of teeth.

6. The CBCT tooth instance segmentation method based on boundary supervision and multiple attention as described in claim 1, characterized in that, The boundary segmentation result and region segmentation results The data is then fused to obtain the final segmentation result. ,include: Boundary segmentation results and region segmentation results The components are joined by channels, with each channel having a size of [value missing]. After passing through a convolution kernel of size The convolutional layer produces the output feature map. Its dimensions are ,Will Activate the pixels in each channel to obtain The probability of each pixel belonging to the corresponding tooth category is calculated, and the channel value with the highest probability at the pixel position is taken as the category of the pixel at that position, thus obtaining the final tooth segmentation result.

Images