A method for automatically segmenting masticatory muscles using cone-beam computed tomography images
Through a deep neural network of dual-graph inference, combined with feature enhancement methods of class diagrams and region diagrams, the accuracy problem of masticatory muscle segmentation in cone beam CT images is solved, and efficient masticatory muscle segmentation is achieved.
Patent Information
- Application Number
- CN202111676854.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-31
- Publication Date
- 2025-07-22
- Estimated Expiration
- 2041-12-31
AI Technical Summary
The prior art is difficult to effectively segment the masticatory muscle from cone beam CT images, especially due to the blurred soft tissue boundaries caused by low signal-to-noise ratio and imaging artifacts, and lacks an efficient masticatory muscle segmentation method.
A deep neural network based on dual-graph inference is adopted, combined with class diagram and regional map inference, muscle prior knowledge is obtained through a fully connected network and graph convolutional neural network to model the local and long-range dependencies between muscles, enhance feature representation, and achieve end-to-end masticatory muscle segmentation.
The accuracy of masticatory muscle segmentation is improved and the masticatory muscle segmentation task can be effectively handled in low-dose noisy cone beam CT images.
Smart Images
Figure CN116433682B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical computer vision, and particularly relates to a method for detecting and segmenting masticatory muscles using cone-beam computed tomography images (cone-beam CT images). Background Art
[0002] Automatic segmentation of masticatory muscles based on cone-beam CT images can be used to obtain the three-dimensional muscle morphology of a specific patient and perform quantitative analysis, and has a low radiation dose and high image resolution, and is widely used in computer-aided diagnosis and treatment in the fields of maxillofacial surgery and orthodontics. Considering problems such as low signal-to-noise ratio and imaging artifacts in cone-beam CT images, separating soft tissues from the attached muscles and simultaneously extracting masticatory muscles, such as the masseter muscle, medial pterygoid muscle, and lateral pterygoid muscle, etc., is still a very challenging task.
[0003] In recent years, graph-based visual reasoning has been widely used in computer vision tasks such as object detection and image classification. The graph structure is used to encode object associations and perform task-related feature enhancement. The graph-based regional association reasoning method can handle problems of low-quality features caused by occlusion and category ambiguity. However, current technologies are only used to handle object detection. Some existing deep learning-based research works apply cascaded convolution to extract image features to segment muscle tissues in three-dimensional medical images. However, due to the low signal-to-noise ratio of low-dose cone-beam CT images and blurred soft tissue boundaries, it is difficult to effectively obtain the boundaries of masticatory muscles. Transfer learning based on generative adversarial networks can also achieve muscle segmentation of cone-beam CT images through image style conversion. However, this method requires a large amount of cross-domain data for training the pixel-by-pixel transfer model, and it is difficult to be effectively applied to masticatory muscle image segmentation processing. Currently, there is a lack of an efficient method for segmenting masticatory muscles based on cone-beam CT images. Summary of the Invention
[0004] The present invention provides an automatic masticatory muscle segmentation method using cone-beam computed tomography images, aiming to overcome the above defects and perform efficient segmentation of masticatory muscles for three-dimensional cone-beam CT images.
[0005] To achieve the above object, the present invention proposes a deep neural network based on dual-graph reasoning for multi-class chewing muscle detection and segmentation of cone-beam CT images. The class graph and region graph reasoning are used to learn the high-level representations of multi-class muscles, overcoming the interference of local image artifacts in cone-beam CT images on image features. The class graph models the associations between multi-class chewing muscles, where a learnable weight matrix in the muscle classifier based on a fully connected network is used to obtain the prior knowledge representation of a specific class and enhance the image features of the specific class muscle. The region graph combines a graph convolutional neural network to model the local and long-range dependencies of multi-class muscle candidate regions in a single-layer image and enhance the region feature representation. By integrating the class graph and region graph reasoning, using the prior knowledge of multi-class chewing muscles and region associations, the feature representation learning is enhanced based on the symmetric and adjacent relationships of the anatomical structure in the local and long-range context, avoiding the perturbation of the chewing muscle feature representation caused by artifacts and surrounding soft tissue confusion in noisy cone-beam CT images. The present invention can effectively segment the chewing muscles from cone-beam CT images. In the test phase, the dual-graph reasoning network for multi-class chewing muscle segmentation proposed by the present invention can achieve end-to-end segmentation of multi-class chewing muscles.
[0006] The technical solution provided by the present invention is as follows:
[0007] The present invention proposes a method for automatically segmenting chewing muscles using cone-beam computed tomography images, creating a deep neural network with dual-graph reasoning for multi-class chewing muscle segmentation of cone-beam CT images. The dual-graph reasoning network models the associations between different class muscle regions in cone-beam computed tomography images, uses class graph and region graph reasoning to mine the prior representations of muscles and the local and long-range dependencies between muscles, and efficiently extracts and learns features from low-dose noisy cone-beam CT images, improving the chewing muscle segmentation accuracy, including the following steps:
[0008] 1) Construction of the dual-graph reasoning network.
[0009] The dual-graph reasoning network includes a chewing muscle classification network and a chewing muscle segmentation network, as well as a reasoning module based on the class graph and a reasoning module based on the region graph; the chewing muscle classification network consists of a feature extraction network based on a convolutional neural network and a classification network; the classification network generates the class labels of the regions belonging to the muscles; the chewing muscle segmentation network is used to predict the pixel labels. Feature representation learning is performed through the reasoning module based on the class graph and the reasoning module based on the region graph, and feature enhancement is carried out; the enhanced features are used as the input of the second chewing muscle classification network and the chewing muscle segmentation network to obtain the final classification and segmentation results; specifically as follows:
[0010] 1a) The reasoning module based on the class graph outputs the enhanced features based on the class graph.
[0011] First, a class diagram is established. The weight matrix R of the muscle classification model based on the fully connected network is used as the prior knowledge representation of the high-level muscle categories. The muscle prior knowledge matrix can be updated by optimizing the muscle classifier (muscle classification network) during the training process. The inter-class association matrix A c is defined as:
[0012]
[0013] where A c is the inter-class association matrix; A c,ij is the inter-class association matrix between the i-th muscle category and the j-th muscle category; R i and R j represent the prior knowledge of the i-th and j-th muscle categories respectively, K is the total number of masticatory muscle categories; l represents the masticatory muscle category.
[0014] The nodes in the class diagram represent muscle categories; the edges in the class diagram represent the associations between muscle categories; the class diagram reasoning based on the masticatory muscle prior knowledge alleviates the low-quality feature problems caused by local image artifacts and enhances the image features of the muscle region. The enhanced features based on the class diagram are defined as follows:
[0015] F c = R(A c + I)c⊙s
[0016] where R is the prior knowledge of the muscle, A c is the inter-class association matrix, I is the identity matrix, c is the muscle classification probability distribution output by the muscle classifier, s is the probability that the pixel predicted by the segmentation network belongs to the foreground muscle; ⊙ represents the element-wise product.
[0017] 1b) The inference module based on the region graph outputs the enhanced features based on the region graph inference.
[0018] Construct a region graph G r (N r , E r ) to model the associations between muscle region images. Among them, the graph node set N r is defined by the multi-class masticatory muscle regions output by the region extraction network, and E r represents the edge set of pairwise connections between regions; the inference learning based on the region graph enhances the region features using the muscle region associations and models the local and long-range dependencies between regions, including symbiotic, symmetric, and adjacency relationships, using the graph convolutional neural network.
[0019] The region association matrix A r is calculated as follows:
[0020]
[0021] where Ar,ij is the regional association matrix of the i-th and j-th muscle regions; F i associated with F j are the features of the i-th and j-th muscle regions, and exp is the exponential function.
[0022] The graph Laplacian matrix L is defined as: L = D - A r , and the graph convolution operation is implemented using the graph Laplacian matrix L, and regional feature enhancement is performed using regional association. D is the degree matrix.
[0023] The regional graph can model the implicit correlations of muscle regions, avoiding the problem of feature degradation caused by low local image signal-to-noise ratio. The output of the graph convolutional neural network is used as the enhanced feature F based on regional graph reasoning r .
[0024] 1c) Feature fusion.
[0025] The enhanced features based on the class graph and the enhanced features based on the regional graph are concatenated to obtain the enhanced features based on the dual-graph reasoning model
[0026]
[0027] where is the feature concatenation operation, F c represents the enhanced features of class graph reasoning, F r represents the enhanced features of regional graph reasoning.
[0028] The concatenated features are then used as the input to the masticatory muscle classification network and the masticatory muscle segmentation network for further fine classification and segmentation.
[0029] 2) The reasoning module based on the class graph and the reasoning module based on the regional graph are both dynamically learnable modules and are optimized during the training phase to obtain a trained dual-graph reasoning network model with enhanced features.
[0030] Specifically, during implementation, the training method and loss function of the classification and segmentation model MaskRCNN are used for model training; the model is programmed and implemented using the PyTorch deep learning library, and the first-order optimization algorithm Adam is used for backpropagation iterative update of model parameters. The loss function specifically includes bounding box localization loss, classification loss, and segmentation loss. The bounding box localization loss measures the positional difference between the network-predicted muscle bounding box and the ground truth bounding box, the classification loss measures the consistency between the network-predicted muscle class and the ground truth, and the segmentation loss measures the similarity between the network-predicted muscle segmentation result and the ground truth.
[0031] The present invention enhances muscle features using class diagrams and regional diagrams, and passes through two muscle classification and segmentation networks. The parameters of the inference module based on the class diagram and the inference module based on the regional diagram are learnable and are dynamically updated through backpropagation. The feature extraction network, regional classification, segmentation, and the optimization of the dual-diagram inference module use the loss function of Mask R-CNN. The difference between the present invention and the existing Mask R-CNN training is that in the second fine classification and segmentation of the present invention, enhanced muscle features are used. As the input, the parameters of the dual-diagram inference module are updated through the gradient backpropagation of the loss function. In the existing Mask R-CNN training, since there is no dual-diagram inference module and only one prediction is made using muscle feature F without the second fine classification and segmentation step.
[0032] 3) Use the trained dual-diagram inference feature enhancement network to perform muscle segmentation of the masticatory muscles on the cone beam CT images.
[0033] Input the cone beam CT images, and use the feature extraction network based on cascaded convolution to map the input images to a high-dimensional feature space. This regional feature is used for preliminary muscle classification and segmentation. The muscle classification model based on the fully connected network is used to obtain the prior knowledge of multi-class muscles, and the muscle segmentation model is used to obtain the probability distribution of belonging to the foreground muscles by pixel. The dual-diagram inference module combines class diagram inference and regional diagram inference to enhance the regional features, and this enhanced feature is used to further improve muscle classification and segmentation. Use the supervised learning mechanism to train the feature extraction network, regional classification, segmentation, and the dual-diagram inference module.
[0034] When the method of the present invention is specifically implemented, head cone beam CT images collected clinically are used to test the classification and segmentation processing of the masticatory muscles in the cone beam CT images, verifying that the method of the present invention can achieve the segmentation of multi-class masticatory muscles including the masseter muscle, the medial pterygoid muscle, and the lateral pterygoid muscle based on the cone beam CT images, and can effectively improve the segmentation accuracy of the masticatory muscles.
[0035] Compared with the prior art, the beneficial effects of the present invention include:
[0036] The present invention effectively realizes the automatic segmentation of the masticatory muscles using cone beam computed tomography images, and performs efficient segmentation of the masticatory muscles for three-dimensional cone beam CT images. The dual-diagram inference network for multi-class masticatory muscle segmentation proposed by the present invention models the association of different class muscle regions in the cone beam computed tomography images, uses class diagram and regional diagram inferences to mine the prior representation of muscles and the local and long-range dependencies between muscles, and efficiently extracts and learns features from low-dose noisy cone beam CT images, and can achieve end-to-end segmentation of multi-class masticatory muscles, and can effectively improve the segmentation accuracy of the masticatory muscles. Description of the Drawings
[0037] Figure 1Flow chart of the method for automatically segmenting masticatory muscles using cone-beam computed tomography images provided by the present invention.
[0038] Figure 2 Figure showing the masticatory muscle segmentation result obtained by automatically segmenting masticatory muscles from cone-beam CT images using the method of the present invention in the embodiment;
[0039] Among them, the segmentation mask represents the muscle category; the left figure is the cone-beam CT image; the right figure is the masticatory muscle segmentation result. Detailed implementation manners
[0040] The present invention will be further described below with reference to the accompanying drawings through embodiments, but it is not limited to the scope of the present invention in any way.
[0041] The present invention proposes a method for automatically segmenting masticatory muscles using cone-beam computed tomography images, creating a deep neural network for dual-graph reasoning for multi-class masticatory muscle segmentation of cone-beam CT images. The dual-graph reasoning network models the association of different category muscle regions in the cone-beam computed tomography images, uses class graph and region graph reasoning to mine the prior representation of muscles and the local and long-range dependencies between muscles, and efficiently performs feature extraction and learning from low-dose noisy cone-beam CT images to improve the accuracy of masticatory muscle segmentation.
[0042] The input image of the method of the present invention is the slice of each frame of the CT image, and the output result is a muscle classification and segmentation map of the same size as the CT image. The specific implementation process steps are as Figure 1 shown, including:
[0043] 1) Construction of the dual-graph reasoning network. The dual-graph reasoning network includes a masticatory muscle classification network and a masticatory muscle segmentation network. The masticatory muscle classification network consists of a feature extraction network based on a convolutional neural network and a classification network. The classification network generates the category label of the region belonging to the muscle; the masticatory muscle segmentation network is used to predict the pixel label. The dual-graph reasoning network consists of a reasoning module based on the class graph and a reasoning module based on the region graph for feature representation learning, specifically as follows:
[0044] a) Reasoning module based on the class graph.
[0045] First, a class graph is established. The weight matrix R of the muscle classification model based on the fully connected network is used as the prior knowledge representation of the high-level muscle categories. The muscle prior knowledge matrix can be updated by optimizing the muscle classifier during the training process. The inter-class association matrix A c is defined as:
[0046]
[0047] where R i and R jIt represents the prior knowledge of the i-th and j-th types of muscles, and K is the number of masticatory muscle categories. The edges in the class graph represent the associations between muscle categories. The class graph reasoning based on muscle prior knowledge alleviates the problem of low-quality features caused by local image artifacts and enhances the regional image features. The enhanced features based on the class graph are defined as follows:
[0048] F c = R(A c + I)c⊙s
[0049] where R is the prior knowledge of the muscle, A c is the inter-class association matrix, I is the identity matrix, c is the muscle classification probability distribution output by the muscle classifier, and s is the probability that the pixel predicted by the segmentation network belongs to the foreground muscle.
[0050] b) Inference module based on the region graph.
[0051] Construct a region graph G r (N r , E r ) to model the associations between muscle region images. The set of graph nodes N r is defined by the multi-class masticatory muscle regions output by the region extraction network, and E r represents the set of edges connecting pairs of regions. The inference learning based on the region graph enhances the regional features using the muscle region associations and models the local and long-range dependencies between regions, including co-occurrence, symmetry, and adjacency relationships, using a graph convolutional neural network. The regional association matrix A r is calculated as follows:
[0052]
[0053] where F i and F j are the features of the i-th and j-th muscle regions, and exp is the exponential function. The graph Laplacian matrix is defined as: L = D - A r , and the graph convolution operation is implemented using the graph Laplacian matrix L to enhance the regional features using the regional associations. D is the degree matrix. The region graph can model the implicit correlations of muscle regions and avoid the problem of feature degradation caused by low local image signal-to-noise ratio. The output of the graph convolutional neural network is used as the enhanced feature F r based on the region graph inference.
[0054] c) Feature fusion. Concatenate the features based on the class graph and the enhanced features based on the region graph to obtain the enhanced features after the double-graph inference model
[0055]
[0056] For feature concatenation operation, F c Represents the enhanced features of class graph reasoning, F r Represents the enhanced features of region graph reasoning. The concatenated features Are used as the input to the masticatory muscle classification network and the masticatory muscle segmentation network.
[0057] The reasoning module based on the class graph and the reasoning module based on the region graph are both dynamically learnable modules and are optimized during the training phase.
[0058] In specific implementation, the training method and loss function of the classification and segmentation model MaskRCNN are used for model training; the model is programmed and implemented using the PyTorch deep learning library, and the first-order optimization algorithm Adam is used for backpropagation iteration to update the model parameters. The loss function specifically includes bounding box localization loss, classification loss, and segmentation loss. The bounding box localization loss measures the position difference between the network-predicted muscle bounding box and the ground truth bounding box, the classification loss measures the consistency between the network-predicted muscle category and the ground truth, and the segmentation loss measures the similarity between the network-predicted muscle segmentation result and the ground truth.
[0059] The present invention enhances muscle features using class graphs and region graphs and passes through two muscle classification and segmentation networks. The parameters of the reasoning module based on the class graph and the reasoning module based on the region graph are learnable and are dynamically updated through backpropagation. The feature extraction network, region classification, segmentation, and the optimization of the dual-graph reasoning module use the loss function of MaskRCNN. The difference between the present invention and the existing MaskRCNN training is that in the second fine classification and segmentation of the present invention, the enhanced muscle features Are used as the input, and the parameters of the dual-graph reasoning module are updated through the gradient backpropagation of the loss function. The existing MaskRCNN training does not have a dual-graph reasoning module and only uses the muscle feature F for one prediction without a second fine classification and segmentation step.
[0060] 2) Muscle segmentation based on feature enhancement of the dual-graph reasoning network.
[0061] Input cone-beam CT images, and use the feature extraction network based on cascaded convolution to map the input images to a high-dimensional feature space. This regional feature is used for preliminary muscle classification and segmentation. The muscle classification model based on the fully connected network is used to obtain the prior knowledge of multi-class muscles, and the muscle segmentation model is used to obtain the probability distribution of foreground muscles by pixel. The dual-graph reasoning module combines class graph reasoning and region graph reasoning to enhance the regional features, and this enhanced feature is used to further improve muscle classification and segmentation. The feature extraction network, region classification, segmentation, and the dual-graph reasoning module are trained using the supervised learning mechanism.
[0062] Such as Figure 2As shown, in order to verify the effectiveness of the method for classifying and segmenting masticatory muscles in cone-beam CT images based on a dual-graph inference network, experiments were conducted on clinically acquired head cone-beam CT images to test the classification and segmentation of masticatory muscles in cone-beam CT images. The segmentation of multiple types of masticatory muscles based on cone-beam CT images, including the masseter muscle, medial pterygoid muscle, and lateral pterygoid muscle, can be achieved, and the segmentation accuracy of masticatory muscles is effectively improved. Figure 2 The left figure in the middle is the cone-beam CT image; the right figure is the segmentation result of the masticatory muscles. The small figures in the first row need to be deleted and replaced with the right figure in the second row. The left figure in the second row is the cone-beam CT image; the right figure is the segmentation result of the masticatory muscles, and the same gray level of the segmentation mask represents the same muscle category.
[0063] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can also make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention shall be defined by the claims.
Claims
1. A method for automatically segmenting masticatory muscles using cone-beam computed tomography images, creating a deep neural network based on dual-graph reasoning, applying class-graph and region-graph reasoning to learn high-level representations of multi-class muscles, and modeling the association of different class muscle regions within cone-beam computed tomography images; Among them, the learnable weight matrix in the muscle classifier based on the fully connected network is utilized by the class graph to obtain the prior knowledge representation of a specific class and enhance the muscle image features of the specific class; The class graph and the region graph are used for reasoning to mine the prior representation of muscles and the local and long-range dependencies between muscles, and feature extraction and learning are efficiently performed from low-dose noisy cone-beam CT images to improve the segmentation accuracy of the masticatory muscles and effectively achieve the end-to-end segmentation of multiple classes of masticatory muscles. The method includes the following steps: 1) Build a dual-graph reasoning network: including a masticatory muscle classification network and a masticatory muscle segmentation network, as well as a reasoning module based on the class graph and a reasoning module based on the region graph; The masticatory muscle classification network includes a feature extraction network based on a convolutional neural network for extracting features and a classification network for generating class labels indicating that a region belongs to a muscle; the masticatory muscle segmentation network is used to predict pixel labels; Feature representation learning and feature enhancement are performed through the reasoning module based on the class graph and the reasoning module based on the region graph of the dual-graph reasoning network; the enhanced features are then used as the inputs of the masticatory muscle classification network and the masticatory muscle segmentation network to obtain the final classification and segmentation results; 1a) The reasoning module based on the class graph outputs the enhanced features based on the class graph; Build a class graph, where the nodes in the class graph represent muscle classes; the edges in the class graph represent the associations between muscle classes; Define the inter-class association matrix A c ; The weight matrix R of the muscle classification model based on the fully connected network is used as the prior knowledge representation of the high-level muscle classes, and the muscle prior knowledge matrix is updated by optimizing the muscle classifier during the training process; Define the enhanced features based on the class graph; the class graph reasoning based on the prior knowledge of the masticatory muscles alleviates the problem of low-quality features caused by local image artifacts and enhances the muscle region image features; 1b) The reasoning module based on the region graph outputs the enhanced features based on the region graph reasoning; Construct the regional graph G r (N r , E r ), model the associations between muscle regional images, where the set of graph nodes N r is defined by the multi-class masticatory muscle regions output by the region extraction network, and E r represents the set of edges connecting pairs of regions; Define the region association matrix A r ; The inference learning based on the region graph enhances the region features by using the muscle region association, models the local and long-range dependence relationships between regions by using the graph convolutional neural network; enhances the region features by using the region association; 1c) Feature fusion: concatenate the enhanced features based on the class graph and the enhanced features based on the region graph to obtain the enhanced features based on the dual-graph reasoning model; Use the concatenated features as the inputs of the masticatory muscle classification network and the masticatory muscle segmentation network, and further perform the fine classification and segmentation of the masticatory muscles; 2) Both the reasoning module based on the class graph and the reasoning module based on the region graph are dynamically learnable modules and are optimized during the training phase to obtain a trained dual-graph reasoning network model with enhanced features; 3) Use the trained dual-graph reasoning feature enhancement network to perform muscle segmentation of the masticatory muscles on the cone-beam CT images; Input the cone-beam CT images, and use the feature extraction network based on cascaded convolution to map the input images to a high-dimensional feature space. The regional features are used for preliminary muscle classification and segmentation; use the muscle classification model based on the fully connected network to obtain the prior knowledge of multiple classes of muscles; the muscle segmentation model is used to obtain the probability distribution of the foreground muscles by pixel. The regional features are enhanced through the dual-graph reasoning module by combining class graph reasoning and region graph reasoning, and the enhanced features are used to further improve muscle classification and segmentation; Through the above steps, the segmentation of multiple classes of masticatory muscles based on cone-beam CT images is achieved.
2. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images according to claim 1, characterized in that, In step 1a), the inter-class association matrix A c is defined as: Among them, A c is the inter-class association matrix; A c,ij is the inter-class association matrix between the i-th type of muscle and the j-th type of muscle; R i and R j respectively represent the prior knowledge of the i-th type and the j-th type of muscle, K is the total number of masticatory muscle categories; l represents the masticatory muscle category.
3. The method for automatically segmenting the masticatory muscles using cone-beam computed tomography images according to claim 2, wherein, In step 1a), the enhanced features based on the class graph are defined as: F c =R(A c +I)cos where R is the prior knowledge of the muscle, A c is the inter-class correlation matrix, I is the identity matrix, c is the muscle classification probability distribution output by the muscle classifier, and s is the probability that the pixel predicted by the segmentation network belongs to the foreground muscle; ⊙ represents the per-pixel product.
4. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images according to claim 2, characterized in that, In step 1b), the region graph-based inference module uses a graph convolutional neural network to model the local and long-range dependencies between regions, where the local and long-range dependencies include symbiotic relationships, symmetric relationships, and adjacency relationships; the output of the graph convolutional neural network is used as the enhanced feature F for region graph-based inference r ; The regional association matrix A r Adopts the following calculation method: Among them, A r,ij is the regional association matrix of the i-th and j-th muscle regions; F i and F j are the characteristics of the i-th and j-th muscle regions, and exp is the exponential function.
5. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images as claimed in claim 4, wherein The graph convolution is specifically implemented using the graph Laplacian matrix L; the graph Laplacian matrix L is defined as: L = D - A r , where D is the degree matrix.
6. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images according to claim 4, wherein Obtain the enhanced features based on the dual-graph inference model through the following concatenation of enhancement features Among them, is the feature concatenation operation, F c represents the enhanced feature of class diagram reasoning, F r represents the enhanced feature of region diagram reasoning.
7. The method for automatically segmenting the masticatory muscles using cone beam computed tomography images according to claim 6, wherein The training method and loss function of the classification and segmentation model MaskRCNN are used for model training, and the inference module based on the class graph and the inference module based on the region graph are optimized in the training stage to obtain a trained dual-graph inference network model with enhanced features.
8. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images as claimed in claim 7, wherein, Specifically, the dual-graph inference model is implemented by programming with the PyTorch deep learning library, and the first-order optimization algorithm Adam is used for backpropagation iterative update of the model parameters.
9. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images according to claim 7, wherein, The loss function for model training specifically includes bounding box localization loss, classification loss, and segmentation loss; the bounding box localization loss is used to measure the position difference between the muscle bounding box predicted by the network and the ground truth bounding box; the classification loss is used to measure the consistency between the muscle category predicted by the network and the ground truth; the segmentation loss is used to measure the similarity between the muscle segmentation result predicted by the network and the ground truth.
10. The method for automatically segmenting masticatory muscles using cone-beam computed tomography images according to claim 1, characterized in that, The multi-class masticatory muscles segmented based on cone-beam CT images include the masseter muscle, the medial pterygoid muscle, and the lateral pterygoid muscle.