CBCT tooth image segmentation method and device and electronic equipment

Abstract

The invention provides a CBCT tooth image segmentation method and device and electronic equipment, and the method comprises the steps: obtaining a to-be-segmented CBCT tooth image, inputting the CBCT tooth image into a tooth image segmentation model, and obtaining a tooth segmentation result, and the tooth image segmentation model is obtained through training a pre-constructed segmentation model through a tooth sample image. The pre-constructed segmentation model comprises an encoder deployed with a dual-scale collaborative attention module, a layered decoder deployed with a jump fusion gate module, and a multi-scale fusion head; according to the method, after the pre-constructed segmentation model is trained through the tooth sample image to obtain the tooth image segmentation model, the CBCT tooth image is segmented through the tooth image segmentation model, so that the adverse effects of metal artifacts, boundary blur and other problems on CBCT tooth image segmentation are reduced, the segmentation accuracy of the CBCT tooth image is improved, and the CBCT tooth image segmentation efficiency is improved. Dependence on annotated data is reduced, and generalization ability of a tooth image segmentation model is improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a CBCT tooth image segmentation method, apparatus, and electronic device. Background Technology

[0002] Cone beam computed tomography (CBCT) is an imaging technique that has been widely used in the field of dentistry in recent years. It can provide three-dimensional images of the oral and maxillofacial region, providing important information for the diagnosis and treatment of dental diseases.

[0003] However, in practical applications, due to the low contrast between teeth and surrounding soft tissues, coupled with metal stripe artifacts and the complexity of anatomical structures (such as root crowding or thin cortical bone), tooth segmentation methods in related technologies (e.g., the 3DU-Net architecture) often suffer from blurred boundaries and poor segmentation accuracy. Furthermore, the labor-intensive nature of manual annotation by dental experts results in limited labeled datasets, which not only exacerbates data scarcity but also limits the generalization ability of fully supervised deep learning models. This makes it difficult for them to adapt to the complex morphological structure, variable grayscale features, and similarity to surrounding tissues of teeth, leading to inaccurate segmentation results and affecting subsequent dental diagnosis and treatment planning.

[0004] Therefore, it is necessary to improve the tooth segmentation methods in related technologies to solve problems such as metal artifacts and blurred boundaries, so as to achieve high-precision segmentation of CBCT tooth images. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, this application provides a CBCT tooth image segmentation method, apparatus and electronic device to solve the above technical problems.

[0006] According to one aspect of the embodiments of this application, a CBCT tooth image segmentation method is provided, the method comprising: acquiring a CBCT tooth image to be segmented; inputting the CBCT tooth image into a tooth image segmentation model to obtain a tooth segmentation result; the tooth image segmentation model is obtained by training a pre-constructed segmentation model using tooth sample images; the pre-constructed segmentation model comprises: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head.

[0007] According to one aspect of the embodiments of this application, a CBCT tooth image segmentation apparatus is provided. The apparatus includes: an image acquisition module for acquiring CBCT tooth images to be segmented; an image segmentation module for inputting the CBCT tooth images into a tooth image segmentation model to obtain tooth segmentation results; the tooth image segmentation model is obtained by training a pre-constructed segmentation model using tooth sample images; the pre-constructed segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head.

[0008] According to one aspect of the embodiments of this application, an electronic device is provided, including: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the CBCT tooth image segmentation method as described above.

[0009] The beneficial effects of this application are as follows: This application acquires CBCT tooth images to be segmented, inputs the CBCT tooth images into a tooth image segmentation model, and obtains tooth segmentation results. The tooth image segmentation model is obtained by training a pre-built segmentation model using tooth sample images. The pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head. Through the above process, the pre-built segmentation model is obtained by the encoder with a dual-scale collaborative attention module, the hierarchical decoder with a skip fusion gate module, and the multi-scale fusion head, forming a Collaborative Adaptive Fusion Network (CAF-Net). After training the pre-built segmentation model with tooth sample images to obtain the tooth image segmentation model, the tooth image segmentation model is used to segment CBCT tooth images, reducing the adverse effects of metal artifacts and boundary blurring on CBCT tooth image segmentation, improving the segmentation accuracy of CBCT tooth images, reducing the dependence on labeled data, and improving the generalization ability of the tooth image segmentation model.

[0010] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 This is a schematic diagram of the 3D U-Net architecture in related technologies; Figure 2 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application; Figure 3 This is a flowchart illustrating a CBCT tooth image segmentation method in an exemplary embodiment of this application; Figure 4 This is a schematic diagram illustrating the structure of a pre-built segmentation model as shown in an exemplary embodiment of this application; Figure 5 This is a schematic diagram of the structure of the first dual-scale collaborative attention block shown in an exemplary embodiment of this application; Figure 6 This is a flowchart illustrating, in an exemplary embodiment of this application, the first skip fusion gate module processes the first scale fused feature map and the third upsampled feature map; Figure 7 This is a flowchart illustrating the training of a pre-built segmentation model, as shown in an exemplary embodiment of this application. Figure 8 This is a schematic diagram illustrating the tooth segmentation result output by a tooth image segmentation model, as shown in an exemplary embodiment of this application. Figure 9 This is a schematic diagram illustrating a visualization of the tooth segmentation result output by a tooth image segmentation model, as shown in an exemplary embodiment of this application. Figure 10 This is a block diagram illustrating a CBCT dental image segmentation apparatus as an exemplary embodiment of this application. Detailed Implementation

[0012] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0013] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0014] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0015] In the field of medical imaging, CBCT is widely used in dental diagnosis and preoperative planning due to its ability to provide high-quality 3D images, enabling the assessment of teeth and jaw structures. However, CBCT tooth segmentation faces multiple challenges: First, the low contrast between teeth and surrounding bone tissue and unclear boundaries significantly increase the difficulty of segmentation; second, metal implants such as restorations and implants introduce strong artifacts and noise interference, causing missegmentation and missed segmentation; third, the intricate and complex anatomical microstructures of teeth and roots place higher demands on the fine-grained representation and localization capabilities of tooth segmentation models; fourth, the voxel-by-voxel annotation of 3D volumetric data is extremely costly and time-consuming, directly leading to a scarcity of high-quality annotated data, making it difficult for purely supervised methods in related technologies to achieve stable and generalizable tooth segmentation performance.

[0016] Figure 1 This is a structural diagram of the 3D U-Net architecture in related technologies, such as... Figure 1 As shown, the 3D U-Net architecture includes an encoder and a decoder. The encoder includes a convolutional (conv(BN) + ReLU) layer and a max pooling (MaxPool) layer. The decoder includes an up-conv layer, skip connections, a concat layer, and an output layer (conv). When performing tooth image segmentation tasks, the above structure design often results in problems such as blurred boundaries, artifact interference, structural discontinuities, lack of global context information, and low segmentation accuracy.

[0017] Figure 2 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application.

[0018] Reference Figure 2As shown, the system architecture may include an acquisition device 201 and an electronic device 202. The electronic device 202 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, or a neural network computer. Those skilled in the art can use the electronic device 202 to acquire CBCT tooth images to be segmented, input the CBCT tooth images into a tooth image segmentation model, and obtain the tooth segmentation result. The tooth image segmentation model is trained using tooth sample images on a pre-built segmentation model, which includes: an encoder deployed with a dual-scale collaborative attention module, a hierarchical decoder deployed with a skip fusion gate module, and a multi-scale fusion head. The acquisition device 201 is used to acquire CBCT tooth images to be segmented and provide them to the electronic device 202 for processing.

[0019] Indicatively, after acquiring the CBCT tooth image to be segmented from the acquisition device 201, the electronic device 202 inputs the CBCT tooth image into the tooth image segmentation model to obtain the tooth segmentation result. The tooth image segmentation model is obtained by training a pre-built segmentation model using tooth sample images. The pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head. Through the above process, the pre-built segmentation model is obtained by the encoder with a dual-scale collaborative attention module, the hierarchical decoder with a skip fusion gate module, and the multi-scale fusion head, forming a collaborative adaptive fusion network architecture. After training the pre-built segmentation model with tooth sample images to obtain the tooth image segmentation model, the tooth image segmentation model is used to segment the CBCT tooth image. This reduces the adverse effects of metal artifacts and boundary blurring on the CBCT tooth image segmentation, improves the segmentation accuracy of the CBCT tooth image, reduces the dependence on labeled data, and improves the generalization ability of the tooth image segmentation model.

[0020] The implementation details of the technical solutions in the embodiments of this application are described in detail below: Figure 3 This is a flowchart illustrating a CBCT tooth image segmentation method as a specific embodiment of this application. (Refer to...) Figure 3 As shown, this CBCT tooth image segmentation method includes at least steps S310 to S320, which are described in detail below: In step S310, the CBCT image of the tooth to be segmented is acquired. In one embodiment of this application, the CBCT image of the tooth to be segmented is acquired by a device such as a CBCT scanner. The CBCT scanner is a device used for CBCT imaging, which uses a cone X-ray beam and a two-dimensional detector to acquire three-dimensional volume data.

[0021] In step S220, the CBCT tooth image is input into the tooth image segmentation model to obtain the tooth segmentation result. In one embodiment of this application, the tooth image segmentation model is obtained by training a pre-built segmentation model with tooth sample images. The pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head; the tooth sample images include: labeled tooth sample images and unlabeled tooth sample images.

[0022] In one embodiment of this application, a pre-built segmentation model is obtained by deploying an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head, forming a collaborative adaptive fusion network architecture. After training the pre-built segmentation model with tooth sample images to obtain a tooth image segmentation model, the tooth image segmentation model is used to segment CBCT tooth images. This reduces the adverse effects of metal artifacts and boundary blurring on CBCT tooth image segmentation, improves the segmentation accuracy of CBCT tooth images, reduces the dependence on labeled data, and improves the generalization ability of the tooth image segmentation model.

[0023] In one embodiment of this application, the process of training a pre-built segmentation model using tooth sample images to obtain a tooth image segmentation model includes: A tooth sample image is input into an encoder equipped with a dual-scale collaborative attention module to obtain a multi-scale enhanced feature map. In one embodiment of this application, the encoder equipped with the dual-scale collaborative attention module includes multiple cascaded dual-scale collaborative attention blocks. After the tooth sample image is input into the encoder, each dual-scale collaborative attention block outputs an enhanced feature map of one scale, resulting in a multi-scale enhanced feature map.

[0024] The multi-scale enhanced feature map is input into a hierarchical decoder with a skip fusion gate module to obtain a decoded feature map. In one embodiment of this application, the hierarchical decoder with the skip fusion gate module includes: three hierarchical decoding branches and a third convolutional kernel combination module. Since the multi-scale fusion feature map includes four scale fusion feature maps, the process of inputting the multi-scale enhanced feature map into the hierarchical decoder with the skip fusion gate module to obtain the decoded feature map includes: inputting the fourth scale fusion feature map into the fourth convolutional kernel combination module to obtain a second extracted feature map; inputting the third scale fusion feature map and the first upsampled feature map into the third hierarchical decoding branch to obtain a third decoded feature map; the first upsampled feature map is obtained by upsampling the second extracted feature map; and the second scale fusion feature map and the second upsampled feature map are input into the second layer... The first decoding branch obtains the second decoded feature map; the second upsampled feature map is obtained by upsampling the third decoded feature map. The first scale fusion feature map and the third upsampled feature map are input into the first level decoding branch to obtain the first decoded feature map; the third upsampled feature map is obtained by upsampling the second decoded feature map. The first, second, and third decoded feature maps are input into the third convolutional kernel combination module to obtain the decoded feature map. In the above process, each level decoding branch selectively fuses multi-scale enhancement feature maps and decoded feature maps, which is beneficial for suppressing CBCT noise (such as metal artifacts) while maintaining the integrity of the microstructure, thereby promoting the reconstruction of multi-scale enhancement feature maps.

[0025] The decoded feature map is input into a multi-scale fusion head to obtain the predicted segmentation result. In one embodiment of this application, the multi-scale fusion head is used to capture and fuse multi-level features in the decoded feature map, thereby improving the segmentation accuracy of the predicted segmentation result and enhancing the robustness of the tooth image segmentation model.

[0026] With the goal of minimizing the difference between the predicted segmentation result and the tooth sample image, the parameters in the pre-built segmentation model are adjusted to obtain a tooth image segmentation model. In one embodiment of this application, if the difference between the predicted segmentation result and the tooth sample image is represented by a loss function, the expression of the loss function includes: Equation (1) in, Represents the loss function. Represents the supervised loss function. Represents the consistency loss function. This represents the consistency loss weighting function; The expression for the consistency loss weight function includes: Equation (2) in, This represents the consistency loss weight function. This represents the maximum weight. Indicates the current training round. The round parameter represents the number of rounds in the incremental process. This indicates an incremental behavior based on the sigmoid function, where the function value increases with the current training epoch. As Tramp gradually increases, a value of 60 indicates the rising phase based on the sigmoid function. Expressions based on the increasing behavior of the sigmoid function include: Equation (3) in, This indicates an increasing behavior based on the sigmoid function. Indicates the current training round. The round parameter represents the number of rounds in the incremental process; The expression for the supervised loss function includes: Equation (4) in, Represents the supervised loss function. Represents the Dice loss function. Represents the cross-entropy loss function; The expression for the Dice loss function includes: Equation (5) in, Represents the Dice loss function. This represents the predicted label in the predicted segmentation result. This indicates the category label corresponding to the tooth sample image. Indicates the sample index in the batch. Indicates category index, Indicates a depth index. Indicates height index, Indicates the width index. Represents a constant; The expression for the cross-entropy loss function includes: Equation (6) in, Represents the cross-entropy loss function. This represents the predicted label in the predicted segmentation result. This indicates the category label corresponding to the tooth sample image. Indicates the sample index in the batch. Indicates batch size, Indicates category index, Indicates the total number of categories. Indicates a depth index. Indicates the maximum depth. Indicates height index, Indicates the maximum height. Indicates the width index. Indicates the maximum width. Represents a constant; The expression for the consistency loss function includes: Equation (7) in, Represents the consistency loss function. Represents a mask matrix used to indicate the location Should consistency loss be calculated? Indicates the sample index in the batch. Indicates a depth index. Indicates height index, Indicates the width index. Indicates the teacher model in position Foreground probability distribution on, Indicates the student model at position Foreground probability distribution on, Represents a constant. This represents the Kullback-Leibler divergence, used to measure the teacher model's position. The foreground probability distribution on the student model at position The difference between the prospect probability distributions on the surface.

[0027] In one embodiment of this application, the Dice loss function and the cross-entropy loss function are used as supervised loss functions. The Dice loss function effectively solves the problem of imbalanced segmentation classes, while the cross-entropy loss provides a stable gradient signal during training, which is beneficial for the rapid convergence and performance improvement of the pre-built segmentation model. The consistency loss function is used to characterize the consistency between the student model and the teacher model in terms of prediction results. Both the student model and the teacher model have the same structure as the pre-built segmentation model, but the parameters of the student model are different from those of the teacher model. The teacher model is used to generate stable pseudo-labels to guide the training process of the student model.

[0028] In one embodiment of this application, during the training of a pre-built segmentation model, each batch of samples simultaneously includes labeled tooth sample images and unlabeled tooth sample images. For labeled tooth sample images, a supervised loss function is calculated and backpropagation is performed; for unlabeled tooth sample images, a consistency loss is calculated and backpropagation is performed. The supervised loss function and the consistency loss function are backpropagated independently. After each training step, the parameters of the teacher model are updated using EMA (Exponential Moving Average), and the optimizer parameters use the AdamW (Adam with Weight Decay) optimizer with a learning rate of... The weight decay rate is Furthermore, mixed-precision training is employed to accelerate training speed and reduce memory usage. Through the above training strategy, CAF-Net can be trained using both labeled and unlabeled tooth sample images simultaneously, solving the problem of missing labeled tooth sample images and significantly improving the segmentation performance of the tooth image segmentation model.

[0029] In one embodiment of this application, if the encoder deployed with a dual-scale collaborative attention module includes multiple cascaded dual-scale collaborative attention blocks, then the process of inputting a tooth sample image into the encoder deployed with the dual-scale collaborative attention module to obtain a multi-scale enhanced feature map includes: A tooth sample image is input into multiple cascaded dual-scale co-attention blocks to obtain an enhanced feature map output by each dual-scale co-attention block. In one embodiment of this application, the multiple cascaded dual-scale co-attention blocks include: a first dual-scale co-attention block, a second dual-scale co-attention block, a third dual-scale co-attention block, and a fourth dual-scale co-attention block. Each dual-scale co-attention block includes: a convolution kernel combination module, a convolution kernel module, a similarity-based attention module, and an efficient channel attention mechanism module. When the size of the tooth sample image is... At that time, the first dual-scale collaborative attention block output Scale-enhanced feature maps, second dual-scale co-attention block output Scale-enhanced feature maps, third dual-scale co-attention block output Scale-enhanced feature maps, fourth dual-scale co-attention block output Scale-enhanced feature maps.

[0030] The enhanced feature maps output from each dual-scale co-attention block are combined to obtain a multi-scale enhanced feature map. In one embodiment of this application, the multi-scale enhanced feature map includes: Scale-enhanced feature maps Scale-enhanced feature maps Scale-enhanced feature maps The number of scale-enhanced feature maps corresponds to the number of dual-scale collaborative attention blocks.

[0031] In one embodiment of this application, if a plurality of cascaded dual-scale co-attention blocks include a first dual-scale co-attention block, then the process of inputting a tooth sample image into the first dual-scale co-attention block includes: The tooth sample image is input into the convolution kernel combination module to obtain the first extracted feature map. In one embodiment of this application, the convolution kernel combination module includes: a first convolution kernel combination module and a second convolution kernel combination module. The first convolution kernel combination module includes: a convolution kernel, a batch normalization layer, and an activation function layer; the second convolution kernel combination module includes: a convolution kernel and a batch normalization layer. The calculation formula of the first convolution kernel combination module is as follows: Equation (8) in, This indicates the extracted feature map from the middle. This represents the activation function layer (i.e., the modified linear unit). Indicates the batch normalization layer. Represents a three-dimensional convolution kernel. This represents an image of a tooth sample.

[0032] The calculation formula for the second convolution kernel combination module is as follows: Equation (9) in, This represents the first extracted feature map. Indicates the batch normalization layer. Represents a three-dimensional convolution kernel. This represents the extracted feature map from the middle.

[0033] The tooth sample image data is input into the convolution kernel module to obtain a texture feature map. In one embodiment of this application, the calculation formula for the texture feature map is as follows: Equation (10) in, Represents texture feature map, Represents a tooth sample image. This represents a 1×1 convolution kernel.

[0034] The first extracted feature map and the texture feature map are concatenated to obtain the basic feature map. In one embodiment of this application, the calculation formula for the basic feature map is as follows: Equation (11) in, Represents the basic feature map. This represents the first extracted feature map. Represents a texture feature map.

[0035] The basic feature map is input into a similarity-based attention module to obtain a weighted feature map. In one embodiment of this application, the calculation formula for the weighted feature map is as follows: Equation (12) in, This represents a weighted feature map. Represents the basic feature map. This represents the attention weight matrix generated by the similarity-based attention module.

[0036] In one embodiment of this application, the attention weight matrix consists of multiple attention weight elements, and the expressions for the attention weight elements are as follows: Equation (13) in, Indicates the first In the attention weight matrix of each channel, the position Attention weight elements at the location, Indicates the first In each channel, the location The energy at the location, This represents the Sigmoid function.

[0037] In one embodiment of this application, in the first In each channel, the location The formula for calculating the energy at a given location is as follows: Equation (14) in, Indicates the first In each channel, the location The energy at the location, Indicates the first In each channel, the location Basic feature map of the location. The basic feature map is represented in the th... The mean of each channel, The basic feature map is represented in the th... Variance on each channel It represents a positive number.

[0038] The basic feature map in the first place The formula for calculating the mean across all channels is shown below: Equation (15) in, The basic feature map is represented in the th... The mean of each channel, Indicates the first In each channel, the location Basic feature map of the location. Indicates the channel index. Indicates a depth index. Indicates the maximum depth. Indicates height index, Indicates the maximum height. Indicates the width index. This indicates the maximum width.

[0039] The basic feature map in the first place The formula for calculating the variance of each channel is as follows: Equation (16) in, The basic feature map is represented in the th... Variance on each channel Indicates the first In each channel, the location Basic feature map of the location. Indicates the channel index. Indicates a depth index. Indicates the maximum depth. Indicates height index, Indicates the maximum height. Indicates the width index. Indicates the maximum width. The basic feature map is represented in the th... The mean value across all channels.

[0040] In one embodiment of this application, the similarity-based attention module (SimAM3D, Similarity-based Attention Module for 3D data) calculates the attention weight of each tensor relative to its spatial neighborhood by statistical difference, calculates the attention weight matrix, and by multiplying the basic feature map with the attention weight matrix element by element, it can assign higher weights to low-contrast regions, effectively highlighting the features of key regions (such as the root tip or thin cortical bone).

[0041] The basic feature map is input into the efficient channel attention mechanism module to obtain the spatial-channel feature map. In one embodiment of this application, the calculation formula for the spatial-channel feature map is as follows: Equation (17) in, Represents spatial-channel feature maps. Represents the basic feature map. Represents the spatial-channel attention weight matrix.

[0042] In one embodiment of this application, the spatial-channel attention weight matrix includes: normalized attention weight values ​​for multiple channels, and the calculation formula for the normalized attention weight value of each channel is as follows: Equation (18) in, Indicates the first Normalized attention weights for each channel. This represents the Sigmoid function. Indicates the first Attention weight values ​​for each channel.

[0043] In one embodiment of this application, the first The expression for the attention weight value of each channel is as follows: Equation (19) in, Indicates the first Attention weight values ​​for each channel, This represents a one-dimensional convolution operation. This represents the base feature map after global average pooling. This indicates the size of the convolution kernel, which is usually set to 3.

[0044] In one embodiment of this application, the expression for the base feature map after global average pooling is as follows: Equation (20) in, This represents the base feature map after global average pooling. Indicates the first In each channel, the location Basic feature map of the location. Indicates a depth index. Indicates the maximum depth. Indicates height index, Indicates the maximum height. Indicates the width index. This indicates the maximum width.

[0045] In one embodiment of this application, the Efficient Channel Attention for 3D Convolutional Networks (ECA3D) module can dynamically adjust the attention weights between different feature categories, thereby achieving better feature category discrimination capability.

[0046] The weighted feature map and the spatial-channel feature map are weighted and summed to obtain the first-scale enhanced feature map. In one embodiment of this application, the calculation formula for the first-scale enhanced feature map is as follows: Equation (21) in, This represents the first-scale enhanced feature map. This represents a weighted feature map. Represents spatial-channel feature maps. This represents the weight coefficient of the weighted feature map item. This represents the weight coefficients of the spatial-channel feature map items. This represents the PReLU activation function.

[0047] In one embodiment of this application, the setting of the first dual-scale collaborative attention block (DSCA) can ensure the preservation of residual information in the basic feature map and strengthen the attention mechanism in the spatial and channel dimensions. At the same time, the first dual-scale collaborative attention block (DSCA) balances the contribution of different categories of attention weights by adjusting the fusion mechanism. In the CBCT tooth segmentation task, it can significantly improve the segmentation performance of low-contrast and complex structural regions.

[0048] In one embodiment of this application, the multiple cascaded dual-scale collaborative attention blocks further include: a second dual-scale collaborative attention block, a third dual-scale collaborative attention block, and a fourth dual-scale collaborative attention block. The first, second, third, and fourth dual-scale collaborative attention blocks have the same structure and are cascaded. During the process of inputting a tooth sample image into an encoder equipped with a dual-scale collaborative attention module, the tooth sample image is processed by the first dual-scale collaborative attention block to obtain a first-scale enhanced feature map. The first-scale enhanced feature map is then processed by the second dual-scale collaborative attention block to obtain a second-scale enhanced feature map. The second-scale enhanced feature map is then processed by the third dual-scale collaborative attention block to obtain a third-scale enhanced feature map. The third-scale enhanced feature map is then processed by the fourth dual-scale collaborative attention block to obtain a fourth-scale enhanced feature map.

[0049] In one embodiment of this application, if the hierarchical decoder with a skip fusion gate module includes: three hierarchical decoding branches, a third convolutional kernel combination module, and the multi-scale fusion feature map includes four-scale fusion feature maps, then the process of inputting the multi-scale enhanced feature map into the hierarchical decoder with the skip fusion gate module to obtain the decoded feature map includes: The fourth-scale fused feature map is input into the fourth convolutional kernel combination module to obtain the second extracted feature map. In one embodiment of this application, the fourth convolutional kernel combination module includes: a convolutional kernel, a batch normalization layer, and an activation function layer. The fourth-scale fused feature map is sequentially input into the convolutional kernel, the batch normalization layer, and the activation function layer. The convolutional kernel is used to perform feature extraction on the fourth-scale fused feature map, the batch normalization layer is used to normalize the extracted features to accelerate training speed and improve model stability, and the activation function layer is used to introduce nonlinear factors to enhance the expressive power of the pre-built segmentation model.

[0050] The third-scale fused feature map and the first upsampled feature map are input into the third-level decoding branch to obtain the third decoded feature map. In one embodiment of this application, the first upsampled feature map is obtained by upsampling the second extracted feature map. Upsampling the second extracted feature map helps to restore the spatial resolution of the feature map, making the first upsampled feature map match the third-scale fused feature map, thereby enabling effective feature fusion. The third-level decoding branch includes: a third jump fusion gate module, a third multi-scale feature aggregation module, a second dual-scale collaborative attention block, and a third feature concatenation module. The third jump fusion gate module dynamically fuses the third-scale fused feature map and the first upsampled feature map to obtain the third enhanced fused feature map. The third jump fusion gate module can automatically adjust the fusion weights according to the importance of different features to ensure the effective transmission of key features. The third multi-scale feature aggregation module performs multi-scale feature aggregation on the third enhanced fusion feature map and the third upsampled feature map to obtain the third aggregated feature map, further enriching the feature information. The second dual-scale collaborative attention block in the third-level decoding branch is used to enhance the attention mechanism of the third aggregated feature map, improving the feature capture capability of low-contrast regions, resulting in the third enhanced aggregated feature map. The third feature concatenation module concatenates the third enhanced aggregated feature map and the first upsampled feature map to obtain the third decoded feature map. The upsampling process of the second extracted feature map includes: performing trilinear interpolation calculation on the second extracted feature map, and performing a 1×1×1 convolution transformation on the calculated feature data.

[0051] The second-scale fused feature map and the second upsampled feature map are input into the second-level decoding branch to obtain the second decoded feature map. In one embodiment of this application, the second upsampled feature map is obtained by upsampling the third decoded feature map. Upsampling the third decoded feature map helps to further recover the spatial detail information of the feature map, making the second upsampled feature map and the second-scale fused feature map better aligned in the spatial dimension, providing a good foundation for subsequent fusion operations. The second-level decoding branch includes a second skip fusion gate module, a second multi-scale feature aggregation module, a second dual-scale collaborative attention block, and a second feature stitching module. The second skip fusion gate module dynamically allocates fusion weights based on the different feature attributes of the second-scale fused feature map and the second upsampled feature map, realizing the effective fusion of the second-scale fused feature map and the second upsampled feature map, and generating a second enhanced fused feature map. The second multi-scale feature aggregation module performs multi-scale feature aggregation operations on the second enhanced fused feature map and the second upsampled feature map, mining feature information at different scales to obtain a second aggregated feature map. The second dual-scale collaborative attention block in the second-level decoding branch strengthens the attention mechanism for the second aggregated feature map from both spatial and channel dimensions, highlighting key features and generating a second enhanced aggregated feature map. The second feature concatenation module concatenates the second enhanced aggregated feature map and the second upsampled feature map to obtain the second decoded feature map. The process of upsampling the third decoded feature map is the same as the process of upsampling the second extracted feature map.

[0052] The first-scale fused feature map and the third-scale upsampled feature map are input into the first-level decoding branch to obtain the first decoded feature map. In one embodiment of this application, the third-scale upsampled feature map is obtained by upsampling the second-scale decoded feature map. By upsampling the second-scale decoded feature map, the spatial information of the tooth structure can be recovered more meticulously, so that the third-scale upsampled feature map and the first-scale fused feature map achieve a high degree of matching in spatial resolution and feature detail, laying the foundation for subsequent accurate segmentation. The first-level decoding branch includes a first skip fusion gate module, a second dual-scale collaborative attention block, and a first feature splicing module. The first skip fusion gate module dynamically adjusts the fusion ratio according to the feature differences between the first-scale fused feature map and the third-scale upsampled feature map to ensure that key anatomical structural features are not lost, and generates a first enhanced fused feature map. The first multi-scale feature aggregation module performs cross-scale feature interaction on the first enhanced fused feature map and the third-scale upsampled feature map through multi-scale convolution kernel combination, extracting multi-level features from coarse to fine, to obtain the first aggregated feature map. In the first-level decoding branch, the second dual-scale collaborative attention block simultaneously applies spatial and channel-dimensional attention mechanisms to the first aggregated feature map, enhancing the feature response to low-contrast regions such as the tooth root and enamel, generating the first enhanced aggregated feature map. The first feature concatenation module concatenates the first enhanced aggregated feature map and the third upsampled feature map along the channel dimension, fusing shallow positional information and deep semantic information to obtain a first decoded feature map rich in multi-scale contextual information. The upsampling process for the second decoded feature map is the same as the upsampling process for the second extracted feature map.

[0053] The first, second, and third decoded feature maps are input into the third convolutional kernel combination module to obtain the decoded feature map. In one embodiment of this application, the third convolutional kernel combination module includes: a convolutional kernel, a batch normalization layer, and an activation function layer. The third convolutional kernel combination module extracts and batch normalizes key feature information from the first, second, and third decoded feature maps, improving the stability of the pre-built segmentation model during training and enhancing its generalization ability. The activation function layer introduces nonlinear transformations to enhance the pre-built segmentation model's ability to express complex tooth structures.

[0054] In one embodiment of this application, if the first-level decoding branch includes: a first skip fusion gate module, a second dual-scale collaborative attention block, and a first feature concatenation module, then the process of inputting the first-scale fused feature map and the third upsampled feature map into the first-level decoding branch to obtain the first decoded feature map includes: The flowchart illustrates how the first-scale fused feature map and the third-upsampled feature map are input into the first skip fusion gate module for processing, resulting in a first enhanced fused feature map. In one embodiment of this application, the first skip fusion gate module dynamically adjusts the fusion ratio based on the feature differences between the first-scale fused feature map and the third-upsampled feature map to ensure that key anatomical structural features are not lost, thereby generating the first enhanced fused feature map.

[0055] The first enhanced fusion feature map and the third upsampled feature map are input into the first multi-scale feature aggregation module to obtain the first aggregated feature map. In one embodiment of this application, the first multi-scale feature aggregation module is used to perform cross-scale feature interaction on the first enhanced fusion feature map and the third upsampled feature map by combining multi-scale convolution kernels, extracting multi-level features from coarse to fine, and obtaining the first aggregated feature map.

[0056] The first aggregated feature map is input into the second dual-scale collaborative attention block to obtain the first enhanced aggregated feature map. In one embodiment of this application, the second dual-scale collaborative attention block in the first-level decoding branch is used to simultaneously apply spatial and channel-dimensional attention mechanisms to the first aggregated feature map, thereby enhancing the feature response to low-contrast regions such as tooth roots and enamel, and generating the first enhanced aggregated feature map.

[0057] The first enhanced aggregated feature map and the third upsampled feature map are input into the first feature concatenation module to obtain the first decoded feature map. In one embodiment of this application, the first feature concatenation module is used to concatenate the first enhanced aggregated feature map and the third upsampled feature map along the channel dimension, fusing shallow positional information and deep semantic information to obtain a first decoded feature map rich in multi-scale contextual information.

[0058] In one embodiment of this application, the process of inputting a first-scale fused feature map and a third upsampled feature map into a first skip fusion gate module to obtain a first enhanced fused feature map includes: A first-scale fused feature map and a third upsampled feature map are concatenated to obtain a first stitched feature map. In one embodiment of this application, the calculation formula for the first stitched feature map is as follows: Equation (22) in, This represents the first concatenated feature map. This represents the first-scale fused feature map. This represents the third upsampled feature map. This represents the join operation function.

[0059] A convolution operation is performed on the first concatenated feature map to obtain an intermediate feature map. In one embodiment of this application, the calculation formula for the intermediate feature map is as follows: Equation (23) in, This represents the intermediate feature map. This represents the first concatenated feature map. express The convolution kernel.

[0060] Global average pooling, fully connected operations, and activation operations are performed on the intermediate feature map to obtain the fused attention weight values. In one embodiment of this application, the calculation formula for the global average pooling operation on the intermediate feature map is shown in formula (20), which obtains the feature vector of each channel of the intermediate feature map. Fully connected operations and ReLU activation operations are performed on the feature vector of each channel of the intermediate feature map to obtain the fused attention weight values ​​before normalization. The calculation formula for the fused attention weight values ​​before normalization is as follows: Equation (24) in, This represents the fusion attention weights before normalization. The first feature map of the inter-region Feature vectors of each channel Indicates the total number of channels. Represents the normalization factor. express convolution kernel, This represents the ReLU activation function.

[0061] In one embodiment of this application, the formula for calculating the fusion attention weight value is as follows: Equation (25) in, This indicates the fusion attention weight value. This represents the Sigmoid function. This represents the fusion attention weights before normalization. Indicates the total number of channels. Represents the normalization factor. express The convolution kernel.

[0062] By fusing attention weight values, a weighted fusion of the first-scale fused feature map and the third upsampled feature map is performed to obtain a first enhanced fused feature map. In one embodiment of this application, the calculation formula for the first enhanced fused feature map is as follows: Equation (26) in, This represents the first enhanced fusion feature map. This indicates the fusion attention weight value. This represents the first-scale fused feature map. This represents the third upsampled feature map.

[0063] In one embodiment of this application, the first Skip Fusion Gate Module The flowchart of SFG3D's processing of the first-scale fused feature map and the third upsampled feature map adopts a 1×1×1 fusion operation, effectively alleviating the problem caused by the inconsistency between encoder and decoder feature forms. Secondly, the fusion process is regulated through a global average pooling mechanism, dynamically adjusting the contribution weights of the encoder and decoder. When encoder features contain a large amount of noise, the weight of decoder features increases accordingly, thus achieving accurate fusion based on real features. Simultaneously, SFG3D effectively preserves microstructural information. When encoder features contain important structural components (such as root boundaries and thin cortical bone), SFG3D significantly improves feature fusion quality by establishing a strong correlation with decoder features. When the fusion attention weight value is large, encoder features are preferentially preserved, thus effectively maintaining microstructural features. SFG3D's structural design, while suppressing noise, accurately preserves key structural information, achieving a more stable feature fusion effect. This significantly improves the performance of the pre-established segmentation module in CBCT tooth segmentation tasks, especially in noise reduction of low-contrast, noisy, and complex feature regions.

[0064] Figure 4 This is a schematic diagram illustrating the structure of a pre-built segmentation model as shown in an exemplary embodiment of this application, such as... Figure 4 As shown, the pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head. The encoder with the dual-scale collaborative attention module includes: a first dual-scale collaborative attention block, a second dual-scale collaborative attention block, a third dual-scale collaborative attention block, and a fourth dual-scale collaborative attention block. The hierarchical decoder with the skip fusion gate module includes: three hierarchical decoding branches and a third convolutional kernel combination module. Each hierarchical decoding branch includes: a skip fusion gate module, a multi-scale feature aggregation module, a second dual-scale collaborative attention block, and a feature concatenation module.

[0065] exist Figure 4 In the encoder, from top to bottom, there are the first dual-scale co-attention block, the second dual-scale co-attention block, the third dual-scale co-attention block, and the fourth dual-scale co-attention block. In the layered decoder, from top to bottom, there are the first-level decoding branch, the second-level decoding branch, and the third-level decoding branch.

[0066] Figure 5This is a schematic diagram of the structure of the first dual-scale cooperative attention block shown in an exemplary embodiment of this application, as follows: Figure 5 As shown, the first dual-scale collaborative attention block includes: a convolutional kernel combination module, a convolutional kernel module, a similarity-based attention module, and an efficient channel attention mechanism module. The process of inputting a tooth sample image into the first dual-scale collaborative attention block includes: inputting the tooth sample image into the convolutional kernel combination module to obtain a first extracted feature map; the convolutional kernel combination module includes: a first convolutional kernel combination module and a second convolutional kernel combination module; the first convolutional kernel combination module includes: a convolutional kernel, a batch normalization layer, and an activation function layer; the second convolutional kernel combination module includes: a convolutional kernel and a batch normalization layer; inputting the tooth sample image data into the convolutional kernel module to obtain a texture feature map; The first extracted feature map and the texture feature map are concatenated to obtain the basic feature map; the basic feature map is input into the similarity-based attention module to obtain the weighted feature map; the basic feature map is input into the efficient channel attention mechanism module to obtain the spatial-channel feature map; the weighted feature map and the spatial-channel feature map are weighted and summed to obtain the first scale-enhanced feature map.

[0067] Figure 6 This is a flowchart illustrating, in an exemplary embodiment of this application, the processing of a first-scale fusion gate module on a first-scale fusion feature map and a third-scale upsampled feature map by the first skip fusion gate module, as shown below. Figure 6 As shown, the flowchart of the first skip fusion gate module processing the first-scale fused feature map and the third upsampled feature map includes: performing a concatenation operation on the first-scale fused feature map and the third upsampled feature map to obtain a first concatenated feature map; performing a convolution operation on the first concatenated feature map to obtain an intermediate feature map; performing global average pooling, fully connected operation, and activation operation on the intermediate feature map to obtain fusion attention weight values; and performing weighted fusion on the first-scale fused feature map and the third upsampled feature map using the fusion attention weight values ​​to obtain a first enhanced fused feature map.

[0068] Figure 7 This is a flowchart illustrating the training of a pre-built segmentation model, as shown in an exemplary embodiment of this application. Figure 7As shown, the framework for training the pre-built segmentation model is the average teacher semi-supervised segmentation training framework. The process of training the pre-built segmentation model through the average teacher semi-supervised segmentation training framework includes: processing the tooth sample image (tooth sample image in CBCT format) with strong augmentation technology and inputting it into the student CAF-Net model to obtain the student model segmentation result (Student's Prediction); processing the tooth sample image with weak augmentation technology and inputting it into the teacher CAF-Net model to obtain the teacher model segmentation result (Teacher's Prediction); and calculating the consistency loss function value (KL) between the student model segmentation result and the teacher model segmentation result according to formula (7). Divergence), calculate the Dice loss function value and cross-entropy (CE) loss function value between the student model segmentation result and the real label corresponding to the tooth sample image according to formula (4)-(6), and update the parameters of the teacher CAF-Net model and the student CAF-Net model according to the consistency loss function value, Dice loss function value and cross-entropy loss function value. The parameters of the student CAF-Net model are updated by gradient descent, and the parameters of the teacher CAF-Net model are updated by the moving average of the student CAF-Net model.

[0069] In one embodiment of this application, the sample images of each training batch simultaneously include labeled tooth sample images and unlabeled tooth sample images. For labeled tooth sample images, the supervision loss function value is calculated and backpropagation is performed; for unlabeled tooth sample images, the consistency loss function value is calculated and backpropagation is performed. This is beneficial to improve the robustness and stability of the pre-built segmentation model, and the supervision loss and consistency loss are backpropagated independently.

[0070] In one embodiment of this application, the student CAF-Net model and the teacher CAF-Net model have the same structure. Both the student CAF-Net model and the teacher CAF-Net model include: an encoder with a dual-scale collaborative attention module (Encoderwith DSCA), a neck skip fusion gate module (Bottleneck SFG3D), and a hierarchical decoder.

[0071] In one embodiment of this application, strong augmentation is used to process the tooth sample images, increasing the difficulty of recognition and forcing the student CAF-Net model to make robust predictions even under interference. The teacher CAF-Net model, acting as a guide for the student CAF-Net model, performs weak augmentation on the tooth sample images, which helps the teacher CAF-Net model generate a relatively accurate and stable prediction result, serving as a "pseudo-label" to guide the student CAF-Net model.

[0072] Figure 8 This is a schematic diagram illustrating the tooth segmentation result output by a tooth image segmentation model, as shown in an exemplary embodiment of this application. Figure 8 The paper presents three cases comparing tooth sample images (Ground Truth) with tooth segmentation images (Predicted Value). Odd-numbered rows show the actual data of the tooth sample images, while even-numbered rows show the tooth segmentation results output by CAF-Net. Each column displays axial, coronal, sagittal views and 3D reconstructed images, respectively. As can be seen from the figures, the tooth segmentation results output by CAF-Net in this application have highly consistent and accurate boundary delineation and high anatomical fidelity with manual annotation.

[0073] Figure 9 This is a schematic diagram illustrating a visualization of the tooth segmentation result output by a tooth image segmentation model, as shown in an exemplary embodiment of this application. Figure 9 The visualization includes three sets of tooth segmentation results. Each set includes 2D cross-sectional views in the axial, coronal, and sagittal planes, as well as corresponding 3D surface reconstruction renderings. In the 2D cross-sectional views, different colors represent different tooth categories (corresponding to 49 semantic categories). The visualizations show that the tooth image segmentation model can accurately fit the tooth edges, generating clear and smooth segmentation masks even in low-contrast root apical regions or crowded areas, effectively separating adjacent tooth instances. In the 3D reconstruction view, the tooth segmentation results exhibit extremely high anatomical fidelity. The generated tooth model has a smooth surface and complete topological structure, accurately reproducing not only the crown morphology but also precisely capturing the slender root structure without obvious fractures or adhesions. This fully demonstrates the robustness of the tooth image segmentation model in voxel-level multi-class segmentation tasks, meeting the high-precision requirements of clinical applications such as orthodontic planning and implant guide design.

[0074] In one embodiment of this application, the comparison results of experimental data of the CAF-Net proposed in this application and experimental data of related technologies such as VNet (Volumetric Medical Image Segmentation Network), nnUNet (neural network Universal Network), Mean Teacher, UA-MT (Uncertainty-aware Self-ensembling Model for Semi-supervised 3D Left Atrium Segmentation), SASSNet (Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images), DTC (DualTask ​​Consistency), SS-Net (Smoothness and Separation-aware Network for Semi-supervised Medical Image Segmentation), and MC-Net+ (Mutual Consistency Learning Network Plus) using the STS 2024 dataset are shown in Table 1: Table 1 As shown in Table 1, all semi-supervised methods (e.g., nnUNet, Mean Teacher, UA-MT, SASSNet, DTC, SS-Net, MC-Net+, and CAF-Net models) consistently outperform fully supervised baseline models (e.g., VNet), confirming that integrating unlabeled tooth sample images through consistency regularization can significantly improve the feature representation learning effect when labeled tooth sample images are scarce. Among the competing semi-supervised methods, CAF-Net performs best on all evaluation metrics. Specifically, the CAF-Net model in this application achieves a Dice score of 89.60% and an IoU value of 84.15%, significantly outperforming the classic Mean Teacher baseline by 3.48% on Dice. Furthermore, the CAF-Net in this application outperforms the second-place MC-Net+ by 1.15% on Dice and 1.35% on IoU. In terms of boundary delineation accuracy, the CAF-Net model in this application has an HD95 value of 6.45 mm (the highest) that is superior to SS-Net (6.62 mm) and MC-Net+ (6.65 mm), thus verifying the effectiveness of the DSCA module settings in the CAF-Net model.

[0075] In one embodiment of this application, ablation experimental data on the progressive contributions of each component in the CAF-Net proposed in this application using the STS 2024 dataset are shown in Table 2: Table 2 As shown in Table 2, the baseline U-Net model achieved a Dice score of 85.12% and an HD95 of 2.35 mm. Integrating the proposed Dual-Scale Collaborative Attention (DSCA) module significantly improved performance, raising the Dice score to 86.85% and the IoU to 80.20%. This improvement demonstrates that the dual-scale attention mechanism effectively optimizes feature representation by suppressing background noise and highlighting tooth structure (especially in low-contrast regions). Subsequently, introducing Skip Fusion Gate 3D (SFG3D) further increased the Dice score to 88.10% and significantly reduced the HD95 to 1.68 mm, validating that the adaptive gating mechanism successfully generated clearer boundary predictions by filtering semantic ambiguity in skip connections. Finally, by equipping the multi-scale fusion head, the complete CAF-Net achieved optimal performance, with a Dice score of 89.60% and an IoU of 84.15%. The sustained performance improvement across all metrics confirms that aggregating multi-level semantic information is crucial for capturing overall anatomical morphology and fine geometric details, demonstrating that all proposed components are indispensable complements to the framework.

[0076] In one embodiment of this application, a plain residual block without an attention mechanism was first constructed as a benchmark. Then, SimAM (using only spatial attention), ECA (using only channel attention), and a fixed summation dual-scale combination scheme were compared. Ablation experiments were conducted on the plain residual block, SimAM, ECA, and the dual-scale collaborative attention (DSCA) module in the proposed CAF-Net using the STS 2024 dataset. The obtained ablation experiment data are shown in Table 3. Table 3 As shown in Table 3, the Plain Residual Block, as the baseline model, achieves a Dice score of 86.10%. Introducing single-stream attention mechanisms (both SimAM and ECA) brings moderate improvements, indicating a significant advantage in emphasizing informational features. Furthermore, combining SimAM, ECA, and fixed weights further increases the Dice score to 87.55%, confirming the complementarity between spatial and channel features. However, the Dual-Scale Cooperative Attention (DSCA) module in the proposed CAF-Net, after using an adaptive fusion strategy, achieves a Dice score of 89.60% and an IoU value of 84.15%, representing a 2.05% improvement in Dice compared to the SimAM, ECA, and fixed weights variant. This result demonstrates that learnable parameters (α, β) play a crucial role in dynamically balancing voxel-level and channel-level information contributions, enabling tooth image segmentation models to segment tooth features based on the specific anatomical background of dental structures.

[0077] This application introduces a dual-scale collaborative attention (DSCA) module for adaptive feature optimization, employs skip fusion gate 3D (SFG3D) technology to suppress noise propagation in skip connections, and combines a multi-level decoder with a robust consistency regularization strategy, enabling CAF-Net to effectively utilize unlabeled sample images to improve generalization ability.

[0078] The Collaborative Adaptive Fusion Network CAF-Net in this application is constructed based on a semi-supervised learning framework. Since semi-supervised learning (SSL) can make full use of a large amount of unlabeled CBCT volume data with a small amount of labeled volume data, it can improve the robustness and generalization ability of the tooth image segmentation model through consistency constraints and pseudo-label learning, thereby alleviating the core contradiction of "expensive 3D annotation - insufficient data - limited performance". Therefore, the Collaborative Adaptive Fusion Network CAF-Net in this application performs feature fusion and enhancement for problems such as low contrast, metal artifacts and complex microstructures under the semi-supervised framework, thereby achieving more reliable and clinically usable CBCT tooth segmentation.

[0079] The following describes an embodiment of the apparatus described in this application, which can be used to execute the CBCT tooth image segmentation method described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the CBCT tooth image segmentation method described above in this application.

[0080] Figure 10 This is a block diagram illustrating a CBCT dental image segmentation apparatus as an exemplary embodiment of this application.

[0081] like Figure 10As shown, the exemplary CBCT dental image segmentation apparatus 1000 includes: The image acquisition module 1001 is used to acquire CBCT images of teeth to be segmented.

[0082] The image segmentation module 1002 is used to input CBCT tooth images into the tooth image segmentation model to obtain tooth segmentation results. The tooth image segmentation model is obtained by training a pre-built segmentation model with tooth sample images. The pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head.

[0083] In one embodiment of this application, the CBCT tooth image to be segmented is acquired by a device such as a CBCT scanner. The CBCT scanner is a device used for CBCT imaging, which uses a cone X-ray beam and a two-dimensional detector to acquire three-dimensional volume data.

[0084] In one embodiment of this application, the tooth image segmentation model is obtained by training a pre-built segmentation model with tooth sample images. The pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head; the tooth sample images include: labeled tooth sample images and unlabeled tooth sample images.

[0085] In one embodiment of this application, a pre-built segmentation model is obtained by deploying an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head, forming a collaborative adaptive fusion network architecture. After training the pre-built segmentation model with tooth sample images to obtain a tooth image segmentation model, the tooth image segmentation model is used to segment CBCT tooth images. This reduces the adverse effects of metal artifacts and boundary blurring on CBCT tooth image segmentation, improves the segmentation accuracy of CBCT tooth images, reduces the dependence on labeled data, and improves the generalization ability of the tooth image segmentation model.

[0086] It should be noted that the CBCT dental image segmentation device and the CBCT dental image segmentation method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the CBCT dental image segmentation device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0087] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to implement the CBCT tooth image segmentation method provided in the above embodiments.

[0088] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A CBCT tooth image segmentation method, characterized in that, The method includes: Obtain the CBCT image of the tooth to be segmented; The CBCT tooth image is input into the tooth image segmentation model to obtain the tooth segmentation result; the tooth image segmentation model is obtained by training a pre-built segmentation model with tooth sample images; the pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head.

2. The CBCT tooth image segmentation method according to claim 1, characterized in that, The process of training a pre-built segmentation model using tooth sample images to obtain the tooth image segmentation model includes: The tooth sample image is input into the encoder equipped with a dual-scale collaborative attention module to obtain a multi-scale enhanced feature map; The multi-scale enhanced feature map is input into the hierarchical decoder with the skip fusion gate module to obtain the decoded feature map; The decoded feature map is input into the multi-scale fusion head to obtain the predicted segmentation result; With the goal of minimizing the difference between the predicted segmentation result and the tooth sample image, the parameters in the pre-built segmentation model are adjusted to obtain the tooth image segmentation model.

3. The CBCT tooth image segmentation method according to claim 2, characterized in that, If the encoder equipped with the dual-scale collaborative attention module comprises multiple cascaded dual-scale collaborative attention blocks, then the process of inputting the tooth sample image into the encoder equipped with the dual-scale collaborative attention module to obtain a multi-scale enhanced feature map includes: The tooth sample image is input into multiple cascaded dual-scale collaborative attention blocks to obtain the enhanced feature map output by each dual-scale collaborative attention block; the multiple cascaded dual-scale collaborative attention blocks include: a convolution kernel combination module, a convolution kernel module, a similarity-based attention module, and an efficient channel attention mechanism module; The enhanced feature map is obtained by combining the enhanced feature maps output by each dual-scale co-attention block.

4. The CBCT tooth image segmentation method according to claim 3, characterized in that, If multiple cascaded dual-scale collaborative attention blocks include a first dual-scale collaborative attention block, then the process of inputting the tooth sample image into the first dual-scale collaborative attention block includes: The tooth sample image is input into the convolution kernel combination module to obtain a first extracted feature map; the convolution kernel combination module includes: a first convolution kernel combination module and a second convolution kernel combination module, the first convolution kernel combination module includes: a convolution kernel, a batch normalization layer, and an activation function layer; the second convolution kernel combination module includes: a convolution kernel and a batch normalization layer. The tooth sample image data is input into the convolution kernel module to obtain a texture feature map; The first extracted feature map and the texture feature map are concatenated to obtain a basic feature map; The basic feature map is input into the similarity-based attention module to obtain a weighted feature map; The basic feature map is input into the efficient channel attention mechanism module to obtain the spatial-channel feature map; The weighted feature map and the spatial-channel feature map are weighted and summed to obtain the first-scale enhanced feature map.

5. The CBCT tooth image segmentation method according to claim 2, characterized in that, If the hierarchical decoder with the skip fusion gate module includes three hierarchical decoding branches and a third convolutional kernel combination module, and the multi-scale fusion feature map includes four-scale fusion feature maps, then the process of inputting the multi-scale enhanced feature map into the hierarchical decoder with the skip fusion gate module to obtain the decoded feature map includes: The fourth-scale fused feature map is input into the fourth convolutional kernel combination module to obtain the second extracted feature map; The third-scale fused feature map and the first upsampled feature map are input into the third-level decoding branch to obtain the third decoded feature map; the first upsampled feature map is obtained by upsampling the second extracted feature map; The second-scale fused feature map and the second upsampled feature map are input into the second-level decoding branch to obtain the second decoded feature map; the second upsampled feature map is obtained by upsampling the third decoded feature map; The first scale fused feature map and the third upsampled feature map are input into the first level decoding branch to obtain the first decoded feature map; the third upsampled feature map is obtained by upsampling the second decoded feature map; The first decoded feature map, the second decoded feature map, and the third decoded feature map are input into the third convolutional kernel combination module to obtain the decoded feature map.

6. The CBCT tooth image segmentation method according to claim 5, characterized in that, If the first-level decoding branch includes: a first skip fusion gate module, a second dual-scale collaborative attention block, and a first feature concatenation module, then the process of inputting the first-scale fused feature map and the third upsampled feature map into the first-level decoding branch to obtain the first decoded feature map includes: The first scale fusion feature map and the third upsampled feature map are input into the first skip fusion gate module to obtain the first enhanced fusion feature map; The first enhanced fusion feature map and the third upsampled feature map are input into the first multi-scale feature aggregation module to obtain the first aggregated feature map; The first aggregated feature map is input into the second dual-scale collaborative attention block to obtain the first enhanced aggregated feature map; The first enhanced aggregated feature map and the third upsampled feature map are input into the first feature concatenation module to obtain the first decoded feature map.

7. The CBCT tooth image segmentation method according to claim 6, characterized in that, The process of inputting the first scale-fused feature map and the third upsampled feature map into the first skip fusion gate module to obtain the first enhanced fused feature map includes: A concatenation operation is performed on the first scale fused feature map and the third upsampled feature map to obtain a first stitched feature map; Perform a convolution operation on the first spliced ​​feature map to obtain an intermediate feature map; Global average pooling, fully connected operations, and activation operations are performed on the intermediate feature map to obtain the fused attention weight values; The first scale fusion feature map and the third upsampled feature map are weighted and fused using the fusion attention weight value to obtain the first enhanced fusion feature map.

8. The CBCT tooth image segmentation method according to any one of claims 2-7, characterized in that, If the difference between the predicted segmentation result and the tooth sample image is represented by a loss function, then the expression of the loss function includes: ， in, Represents the loss function. Represents the supervised loss function. Represents the consistency loss function. This represents the consistency loss weighting function; The expression for the consistency loss weight function includes: ， in, This represents the consistency loss weight function. This represents the maximum weight. Indicates the current training round. The round parameter represents the number of rounds in the incremental process. This indicates an incremental behavior based on the sigmoid function, where the function value increases with the current training epoch. As Tramp gradually increases, a value of 60 indicates the rising phase based on the sigmoid function. Expressions based on the increasing behavior of the sigmoid function include: ， in, This indicates an increasing behavior based on the sigmoid function. Indicates the current training round. The round parameter represents the number of rounds in the incremental process; The expression for the supervised loss function includes: ， in, Represents the supervised loss function. Represents the Dice loss function. Represents the cross-entropy loss function; The expression for the Dice loss function includes: ， in, Represents the Dice loss function. This represents the predicted label in the predicted segmentation result. This indicates the category label corresponding to the tooth sample image. Indicates the sample index in the batch. Indicates category index, Indicates a depth index. Indicates height index, Indicates the width index. Represents a constant; The expression for the cross-entropy loss function includes: ， in, Represents the cross-entropy loss function. This represents the predicted label in the predicted segmentation result. This indicates the category label corresponding to the tooth sample image. Indicates the sample index in the batch. Indicates batch size, Indicates category index, Indicates the total number of categories. Indicates a depth index. Indicates the maximum depth. Indicates height index, Indicates the maximum height. Indicates the width index. Indicates the maximum width. Represents a constant; The expression for the consistency loss function includes: ， in, Represents the consistency loss function. Represents a mask matrix used to indicate the location Should consistency loss be calculated? Indicates the sample index in the batch. Indicates a depth index. Indicates height index, Indicates the width index. Indicates the teacher model in position Foreground probability distribution on, Indicates the student model at position Foreground probability distribution on, Represents a constant. This represents the Kullback-Leibler divergence, used to measure the teacher model's position. The foreground probability distribution on the student model at position The difference between the prospect probability distributions on the surface.

9. A CBCT tooth image segmentation device, characterized in that, The device includes: The image acquisition module is used to acquire CBCT images of teeth to be segmented; The image segmentation module is used to input the CBCT tooth image into the tooth image segmentation model to obtain the tooth segmentation result; the tooth image segmentation model is obtained by training a pre-built segmentation model with tooth sample images; the pre-built segmentation model includes: an encoder with a dual-scale collaborative attention module, a hierarchical decoder with a skip fusion gate module, and a multi-scale fusion head.

10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the CBCT tooth image segmentation method as described in any one of claims 1 to 8.

Images