Temporomandibular joint magnetic resonance intelligent segmentation method, system and equipment based on deep learning, computer readable medium and program product

Abstract

The invention discloses a temporal-mandibular joint magnetic resonance segmentation method, system and device based on deep learning and a medium, and the method comprises the steps: building a data set of a mandible nuclear magnetic image, and dividing the data set into a training set and a test set; building a hybrid segmentation model, wherein the hybrid segmentation model comprises an encoder based on a global context vision Transform structure and a decoder based on a convolutional neural network; setting model training hyper-parameters, and training the hybrid segmentation model by using the training set to obtain segmentation model parameters; using the test set to test parameters of the hybrid segmentation model and the segmentation model; and using the adjusted segmentation model and the segmentation model parameters to segment the joint disc and the condyle process in other images. According to the method, the mixed segmentation model of the mixed model based on the deep learning convolutional neural network and the Transform is established, so that the pixel-level shape information of the joint disc and the condyle process position in the mandible joint nuclear magnetic resonance image is accurately extracted.

Description

Technical Field

[0001] This invention relates to computer image processing technology, specifically to a deep learning-based method, system, device, computer-readable medium, and program product for intelligent segmentation of the temporomandibular joint using magnetic resonance imaging. Background Technology

[0002] With the rapid development of computer technology, computer-aided disease diagnosis has been gradually applied to various disciplines in the medical field. Among them, image segmentation technology is also widely used in the field of medical image processing to extract organs and tissue structures from medical images, thereby assisting doctors in clinical diagnosis.

[0003] Computer vision techniques can assist physicians in interpreting MRI images, and have become a relatively objective diagnostic aid. Early methods for segmenting MRI images in medical imaging typically relied on traditional image processing methods such as thresholding, edge detection, graph cut-based methods, and shape-based methods. However, image features constructed based on human knowledge lose much of the image's internal information; furthermore, compared to natural images, medical images often suffer from blurriness, noise, and low contrast due to the different imaging equipment, making feature extraction more difficult.

[0004] Currently, deep learning methods are being used to detect regions of interest (ROIs) of key mandibular components in MRI images of the mandibular joint. Existing localization methods can only determine the approximate location of mandibular components, failing to accurately pinpoint the specific locations of the articular disc and condyles within the mandibular joint. Current network models are composed of pure convolutional neural networks, whose primary computational method is convolution operations. Limited by the receptive field size of the convolution kernel, they cannot effectively handle long-range dependencies, resulting in limited segmentation accuracy. For example, patent document CN113298828B proposes an automatic jawbone segmentation method based on convolutional neural networks, including the following steps: S1: Establish a dataset of maxillary and mandibular images, and randomly select 2 / 3 of the dataset as a training dataset and the other 1 / 3 as a test dataset; S2: Preprocess the image data of the maxillary and mandibular image dataset; S3: Build a convolutional neural network model and train the built convolutional neural network model using the training dataset to obtain a segmentation network model; S4: Test the segmentation network model using the test dataset; S5: Segment the maxillary and mandibular image data to be segmented using the segmentation network model. However, the automatic jawbone segmentation accuracy proposed in this document still cannot meet higher practical requirements. Summary of the Invention

[0005] To address existing problems, this invention aims to provide a deep learning-based method, system, device, and medium for segmenting temporomandibular joint MRI images, in order to handle long-range dependencies in images, accurately locate the specific positions of the articular disc and condyle in the mandibular joint, and segment out the specific shapes of the articular disc and condyle.

[0006] To achieve the above objectives, the method provided by this invention includes the following steps:

[0007] S1 creates a dataset of MRI images of the mandible and randomly divides it into a training set and a test set;

[0008] S2 preprocesses the dataset, including image processing methods such as image enhancement;

[0009] S3 builds a hybrid segmentation model, including an encoder based on a global context vision Transformer structure and a decoder based on a convolutional neural network; sets the model training hyperparameters, and trains the hybrid segmentation model using the preprocessed training set to obtain the segmentation model parameters;

[0010] S4 uses a test set to test the hybrid segmentation model and the segmentation model parameters;

[0011] S5 adjusts the structure and training hyperparameters of the hybrid segmentation model based on the evaluation results of the hybrid segmentation model and the segmentation model parameters on the test set; and evaluates the segmentation results of the articular disc and condyle.

[0012] S6 uses the adjusted segmentation model and segmentation model parameters to segment the articular discs and condyles in other MRI images to be segmented.

[0013] In some embodiments of S2, preprocessing the dataset includes one or more of normalization and image enhancement.

[0014] In some embodiments of S3, the structure of the global context visual Transformer in the encoder is divided into three layers, each layer including alternating local attention modules and global self-attention modules; the layers are connected by downsampling modules, and the output of the encoder is three layers of image features after layer normalization.

[0015] In some embodiments, the number of channels in the three-layer global context visual Transformer is 64, 128, and 256, respectively.

[0016] In some embodiments, each downsampling layer module reduces the height and width of the feature map by half and doubles the number of feature map channels.

[0017] In some embodiments, in S3, the decoder consists of three convolutional attention layers, two attention skip connection layers, and two upsampling layers, with the convolutional attention layers connected to each other through the upsampling layers and the attention skip connection layers.

[0018] In some embodiments, each convolutional attention layer consists of a selective kernel convolutional attention layer, a spatial attention layer, and a convolutional block.

[0019] In some embodiments, the convolutional block consists of two sets of standard 3×3 convolutions, batch normalization, and nonlinear activation functions in sequence.

[0020] In some embodiments, the input to the attention skip connection layer is the output of the upsampled convolutional attention layer and the features of the encoder at the corresponding scale.

[0021] In some embodiments, the upsampling layer consists of double bilinear interpolation upsampling, standard 3×3 convolution, batch normalization, and a nonlinear activation function.

[0022] In some embodiments, the loss function for the image segmentation network is:

[0023] Loss = 0.4 × Loss CE +0.6×Loss Dice ;

[0024] Loss CE For cross-entropy loss, Loss Dice The Dice similarity coefficient loss is defined as follows:

[0025]

[0026] Where y is the ground truth segmentation value, p is the segmentation mask output by the segmentation model, and n is the number of pixels;

[0027] The loss function is used to calculate the difference between the true value and the model's output mask during retraining. Through iteration during training, the loss function decreases, thus optimizing the model parameters.

[0028] In some embodiments, S5 uses the average Dice similarity coefficient and the average Hausdorff distance as evaluation metrics for the segmentation results.

[0029] In some embodiments, the Dice similarity coefficient ranges from [0,1].

[0030] In some embodiments, the formulas for calculating the average Hausdorff distance are as follows:

[0031]

[0032] HD(P,T)=max{max p∈Pmin t∈T d(p,t),max t∈T min p∈P d(p,t)}

[0033] Where P is the segmentation mask of the model prediction output, T is the ground truth segmentation value, p and t represent each pixel in P and T, and d(p,t) is the distance between point p and point t.

[0034] The present invention also provides a temporomandibular joint MRI image segmentation system, the system comprising:

[0035] The data module is used to divide the MRI images of the mandible into training and testing sets;

[0036] The module is used to build a hybrid segmentation model, including an encoder based on a global context vision Transformer structure and a decoder based on a convolutional neural network.

[0037] The training module sets the model training hyperparameters, trains the hybrid segmentation model using the training set, and obtains the segmentation model parameters.

[0038] The testing module uses the test set to test the hybrid segmentation model and the parameters of the segmentation model.

[0039] The evaluation module is used to adjust the structure and training hyperparameters of the hybrid segmentation model based on the evaluation results of the hybrid segmentation model and its parameters on the test set, so that the adjusted segmentation model and its parameters can be used to segment articular discs and condyles in other images.

[0040] The present invention also provides a temporomandibular joint MRI image segmentation device, including a processor and a memory for storing executable instructions of the processor, wherein the processor executes any of the methods described herein when running. The present invention also provides a computer-readable medium storing computer program instructions thereon, characterized in that: when the computer program instructions are processed and executed, they implement any of the methods described herein.

[0041] Compared with existing technologies, this invention establishes a hybrid segmentation model based on a combination of deep learning convolutional neural networks and Transformers, which can accurately extract pixel-level shape information of the articular disc and condyle positions in mandibular joint MRI images. The encoder part effectively and efficiently models long-distance and short-distance spatial interactions, acquiring multi-scale feature information from the image; its decoder part processes the multi-scale features generated by the encoder through convolutional attention, promoting the effective fusion of multi-scale features. This effectively segments the articular disc and condyle in mandibular joint MRI images. Attached Figure Description

[0042] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example, not limitation, in which:

[0043] Figure 1 A schematic diagram of the structure of a hybrid segmentation model according to one embodiment of the present invention.

[0044] Figure 2 A flowchart of a temporomandibular joint segmentation method according to one embodiment of the present invention.

[0045] Figure 3 A visual comparison of the effects of different segmentation methods according to one embodiment of the present invention. Detailed Implementation

[0046] Example 1

[0047] See Figure 2 , Figure 2 This illustration shows one embodiment of the present invention. The first embodiment of the present invention provides a method for segmenting MRI images of the mandibular joint, including:

[0048] S1: Organize the dataset of mandibular MRI images and randomly divide it into training and test sets.

[0049] As a preferred approach, the training set and the test set are divided in an 8:2 ratio.

[0050] S2: Preprocess the training and test sets in the mandibular dataset, including image processing methods such as normalization and image enhancement.

[0051] S3. Construct an initial convolutional neural network and Transformer hybrid image segmentation model, including an encoder, decoder, and skip connections. This model structure can effectively extract long-range and short-range spatial dependencies between image pixels, fully utilize the extracted features, and achieve feature fusion to obtain better segmentation results. Set the model training hyperparameters and train the constructed hybrid segmentation model using the preprocessed training set described above to obtain the segmentation model parameters.

[0052] Preferably, the training optimizer uses an SGD optimizer with a momentum of 0.9 and a weight decay of 1e-4. The batch size used for training is set to 10, the learning rate is 0.01, and the number of training epochs is set to 300.

[0053] S4. Use the above test set to test the hybrid segmentation model and the segmentation model parameters.

[0054] S5. Based on the evaluation results of the hybrid segmentation model and the segmentation model parameters on the test set, adjust the segmentation model structure and training hyperparameters.

[0055] The average Dice similarity coefficient and the average Hausdorff distance are preferred as evaluation metrics for assessing the segmentation results of the articular disc and condyle.

[0056] S6. Use the finalized segmentation model and segmentation model parameters to segment the articular discs and condyles in other MRI images to be segmented.

[0057] Among them, see Figure 1 As shown, the hybrid segmentation network for mandibular joint images input from MRI images consists of an encoder and a decoder. The encoder primarily comprises three global contextual vision Transformer layers, effectively modeling long-range and short-range dependencies in images at 1 / 4, 1 / 8, and 1 / 16 scales to obtain encoded features. The global contextual vision Transformer layers are sequentially composed of local and global attention modules. The decoder consists of three convolutional attention layers, two skip connection layers, and two upsampling layers. Encoded features are input to the decoder and decoded through the convolutional attention modules. These decoded features are then upsampled and interleaved with encoder features at larger scales via skip connections in the attention skip connection layers. This is followed by convolutional attention at even larger image feature scales to obtain the next scale of decoded features, until the feature scale is restored to 1 / 4 of the original input image. Finally, an image restoration upsampling layer upsamples the image by a factor of 4 to restore the original image size. The segmentation head converts the feature channel count to 3, including three categories: background, articular disc, and condyle. The softmax function is used to obtain the score for each class, and the class with the highest score for each pixel is taken as the classification result.

[0058] The encoder's specific process involves dividing the image into blocks to reduce the image size to 1 / 4 of its original size, increasing the number of channels to 64. Then, a preliminary feature map at this scale is obtained through local and global attention modules. A downsampling layer downsamples the feature map by a factor of two, making the scale half the size of the input downsampling layer, and simultaneously doubling the number of channels. The preliminary feature maps at each scale are then normalized through layers to obtain the encoder's output feature maps, with 64, 128, and 256 channels, corresponding to sizes of 1 / 4, 1 / 8, and 1 / 16, respectively.

[0059] The decoder's specific process first acquires decoding features through a convolutional attention module, then upsamples the input decoding features by a factor of two through an upsampling layer, effectively doubling the image feature scale. After upsampling, the decoding features interact with encoded features at the same scale in an attention-skip connection layer. Through an attention gate mechanism, the encoder features are weighted based on the upsampled decoding features to extract effective information from the encoded features and suppress useless information. Finally, the weighted encoded features are concatenated with the decoding features. The upsampling layer in the decoder contains a 2x bilinear interpolation upsampling layer, a 3×3 standard 2D convolution, a batch normalization layer, and a ReLU activation function. The image reconstruction upsampling layer consists of a 4x bilinear interpolation upsampling layer, a 3×3 standard 2D convolution, a batch normalization layer, and a ReLU activation function. The segmentation head is a 3×3 standard 2D convolution.

[0060] To effectively integrate information, reduce information loss caused by downsampling in the encoder, and improve network performance, skip connections are employed for interaction between the encoder and decoder. Through skip connections, feature maps extracted by the encoder are fused and superimposed with feature maps from the decoder to minimize information loss. In the decoder, the convolutional attention module fully utilizes image feature space and channel information to effectively achieve feature fusion, thereby improving the accuracy of the segmentation network. This design strategy helps optimize model performance, making it more efficient in information transfer and feature fusion.

[0061] Furthermore, the convolutional attention module in this embodiment captures the spatial and channel feature relationships of the input features, promoting feature fusion and obtaining decoded features. This module includes a kernel selection attention module, a spatial attention module, and a convolutional block. The specific process of the convolutional attention module is to sequentially pass through the kernel selection attention module, the spatial attention module, and the convolutional block.

[0062] The kernel attention module comprises four standard convolutions, four batch normalization layers, four ReLU activation functions, five linear layers, and one sigmoid activation function. Specifically, the input features are passed through a convolutional layer, a batch normalization layer, and a ReLU activation function to obtain four sets of features, with kernel sizes of 1, 3, 5, and 7 respectively. These four sets of features are then summed and averaged before passing through a linear layer. Next, they are passed through four linear layers, each using a softmax function to obtain four sets of weights. These four weights are then applied to the four sets of features using a Hadamard product, and finally summed to obtain the module's output.

[0063] Spatial attention consists of a max pooling, an average pooling, a 7×7 standard 2D convolution, and a sigmoid activation function. The specific process is as follows: the input features are first processed through channel-dimension max pooling and average pooling, then concatenated and passed through a 7×7 convolution to obtain the spatial attention weights. Finally, a Hadamard product is applied to the input features.

[0064] The convolutional block consists of two 3×3 standard two-dimensional convolutions, two batch normalization layers, and two ReLU activation functions. The specific process is as follows: sequentially pass through 3×3 convolution, batch normalization, ReLU activation function, 3×3 convolution, batch normalization, and ReLU activation function.

[0065] Furthermore, in this embodiment, the attention gate consists of three 1×1 standard two-dimensional convolutions, three batch normalization layers, a sigmoid activation function, and a ReLU activation function.

[0066] The encoded and decoded features are each passed through a 1×1 convolution and a batch normalization layer, respectively. The results are summed, passed through a ReLU activation function, then through another 1×1 convolution, a batch normalization layer, and a sigmoid activation function to obtain the attention weights. These weights are then applied to the encoded features via a Hadamard product to obtain the attention gate output.

[0067] Furthermore, in this embodiment, the loss function used for the image segmentation network is:

[0068] Loss = 0.4 × Loss CE +0.6×Loss Dice ;

[0069] Loss CE For cross-entropy loss, Loss Dice The Dice similarity coefficient loss is defined as follows:

[0070]

[0071] Where y is the ground truth segmentation value, p is the segmentation mask output by the segmentation model, and n is the number of pixels.

[0072] The loss function is used to calculate the difference between the true value and the model output mask during retraining. Through iteration during training, the loss function decreases, and the model parameters are optimized.

[0073] The mandibular MRI dataset in this embodiment contains 725 images from 79 patients. The original sample includes images from patients with normal articular discs and those with disc displacement. The MRI images used in this embodiment were primarily acquired using two MRI scanners, both manufactured by Philips, Netherlands. This embodiment uses only proton density-weighted sagittal images (TR 2000ms, TE 20ms). All joint MRIs were evaluated by temporomandibular joint specialists, and the data were manually labeled using 3D slicer software. The manually segmented images were reviewed by three experts skilled in interpreting the articular disc and condylar region on MRI, and the images unanimously agreed upon by the experts were used as the original dataset. The original TMJ dataset was randomly divided into a training set (584 images) and a test set (141 images) in an 8:2 ratio. Note: MRI imaging standards should refer to the "Expert Consensus on Routine MRI Examination Standards for Temporomandibular Joint and Diagnostic Criteria for Articular Disc Displacement" published in 2020.

[0074] In this embodiment, the Dice similarity coefficient (DSC) and Hausdorff distance (HD), commonly used to evaluate the quality of medical image segmentation, are selected as evaluation metrics. DSC is a measure of set similarity, with a value ranging from [0,1]. A value closer to 1 indicates greater similarity, with 1 representing the best segmentation result. HD calculates the distance between two sets; a smaller value indicates greater similarity. DSC is more sensitive to internal padding during segmentation, while HD is more sensitive to segmentation boundaries. Their calculation formulas are as follows:

[0075]

[0076] HD(P,T)=max{max p∈P min t∈T d(p,t),max t∈T min p∈P d(p,t)}

[0077] Where P is the segmentation mask of the model prediction output, T is the ground truth segmentation value, p and t represent each pixel in P and T, and d(p,t) is the distance between point p and point t.

[0078] Table 1

[0079]

[0080] Table 1 shows the comparison results of the method of this invention with other medical image segmentation methods in recent years on the test set. The comparison algorithms are as follows: UNeXt is the medical image segmentation method proposed in "UNeXt: MLP-based Rapid Medical Image Segmentation Network" published at the MICCAI conference in 2022; MISSFormer is the medical image segmentation method proposed in "MISSFormer: An Effective Medical Image Segmentation Transformer" published in 2021; and Swin U-net is the medical image segmentation method proposed in "Swin-unet: Unet-like pure transformer for medical image segmentation" published at ECCV 2022 Workshops. As can be seen from Table 1, this method achieved the best results in both average DSC and HD, and also achieved the best DSC results in articular disc and condyle segmentation. Figure 3 The visualization of the segmentation results of this model is compared with those of the above segmentation models. It can be seen from the figure that the segmentation results of this method are most similar to the true values.

[0081] Example 2

[0082] The second embodiment of the present invention provides a temporomandibular joint MRI image segmentation system, including: a data module for dividing mandibular MRI images into a training set and a test set;

[0083] The module is used to build a hybrid segmentation model, including an encoder based on a global context vision Transformer structure and a decoder based on a convolutional neural network.

[0084] The training module sets the model training hyperparameters, trains the hybrid segmentation model using the training set, and obtains the segmentation model parameters.

[0085] The testing module uses the test set to test the hybrid segmentation model and the parameters of the segmentation model.

[0086] The evaluation module is used to adjust the structure and training hyperparameters of the hybrid segmentation model based on the evaluation results of the hybrid segmentation model and its parameters on the test set, so that the adjusted segmentation model and its parameters can be used to segment articular discs and condyles in other images.

[0087] Example 3

[0088] A third embodiment of the present invention provides a temporomandibular joint magnetic resonance imaging segmentation device, including a processor and a memory for storing executable instructions of the processor, wherein the processor executes any of the methods described above when running.

[0089] Memory may include random access memory, flash memory, read-only memory, programmable read-only memory, non-volatile memory, or registers, etc. The processor may be a central processing unit (CPU), or a graphics processing unit (GPU). Memory can store executable instructions. The processor can execute the executable instructions stored in memory to implement the various processes described herein.

[0090] It is understood that the memory in this embodiment can be volatile memory or non-volatile memory, or may include both. The non-volatile memory can be ROM (Read-Only Memory), PROM (Programmable ROM), EPROM (Erasable PROM), EEPROM (Electrically Erasable EPROM), or flash memory. The volatile memory can be RAM (Random Access Memory), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as SRAM (Static RAM), DRAM (Dynamic RAM), SDRAM (Synchronous DRAM), DDR SDRAM (Double Data Rate SDRAM), ESDRAM (Enhanced SDRAM), SLDRAM (Synchlink DRAM), and DRRAM (Direct Rambus RAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0091] In some implementations, the memory stores elements such as upgrade packages, executable units, or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0092] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications used to implement various application functions. Programs implementing the methods of this invention can be included within these application programs.

[0093] In this embodiment of the invention, the processor executes the method steps provided in the first aspect by calling a program or instruction stored in the memory, specifically a program or instruction stored in an application program.

[0094] Example 4

[0095] A fourth embodiment of the present invention also provides a chip for performing the method in the first aspect described above. Specifically, the chip includes a processor for calling and running a computer program from a memory, such that a device having the chip installed performs the method in the first aspect described above.

[0096] Example 5

[0097] A fifth embodiment of the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method of the present invention.

[0098] For example, machine-readable storage media can include, but are not limited to, various known and unknown types of non-volatile memory.

[0099] Example 6

[0100] This invention also provides a computer program product, including computer program instructions that cause a computer to perform the method described in the first aspect above.

[0101] It should be understood that in the embodiments of the present invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0102] It is worth noting that although the spirit and principles of this invention have been described with reference to several specific embodiments, it should be understood that this invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that the features in these aspects cannot be combined; such division is merely for the convenience of description. This invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A deep learning-based magnetic resonance imaging segmentation method for the temporomandibular joint, characterized in that: S1. Establish a dataset of MRI images of the mandible and randomly divide it into a training set and a test set; S2 preprocesses the dataset, including image processing methods such as image enhancement. S3. Build a hybrid segmentation model, including an encoder based on a global context vision Transformer structure and a decoder based on a convolutional neural network; set the model training hyperparameters, and train the hybrid segmentation model using the preprocessed training set to obtain the segmentation model parameters; S4, use the test set to test the hybrid segmentation model and the segmentation model parameters; S5. Based on the evaluation results of the hybrid segmentation model and the segmentation model parameters on the test set, adjust the structure and training hyperparameters of the hybrid segmentation model; and evaluate the segmentation results of the articular disc and condyle. S6. Use the adjusted segmentation model and segmentation model parameters to segment the articular discs and condyles in other MRI images to be segmented.

2. The method according to claim 1, characterized in that: It also includes step S2, which preprocesses the dataset, including one or more of the following: normalization and image enhancement.

3. The method according to claim 1, characterized in that: In step S3, the structure of the global context vision Transformer in the encoder is divided into three layers. Each layer includes alternating local attention modules and global self-attention modules. The layers are connected by downsampling modules, and the output of the encoder is the three-layer image features after layer normalization.

4. The method according to claim 3, characterized in that: The number of channels in the three-layer global context visual Transformer are 64, 128, and 256, respectively.

5. The method according to claim 3, characterized in that: Each downsampling layer module reduces the height and width of the feature map by half and doubles the number of channels in the feature map.

6. The method according to claim 1, characterized in that: In step S3, the decoder consists of three convolutional attention layers, two attention skip connection layers, and two upsampling layers. The convolutional attention layers are connected to each other through the upsampling layers and the attention skip connection layers.

7. The method according to claim 3, characterized in that: Each convolutional attention layer consists of a selective kernel convolutional attention layer, a spatial attention layer, and a convolutional block.

8. The method according to claim 7, characterized in that: The convolutional block consists of two sets of standard 3×3 convolutions, batch normalization, and nonlinear activation functions in sequence.

9. The method according to claim 7, characterized in that: The input to the attention jump connection layer is the output of the upsampled convolutional attention layer and the encoder features at the corresponding scale.

10. The method according to claim 7, characterized in that: The upsampling layer consists of double bilinear interpolation upsampling, standard 3×3 convolution, batch normalization, and a nonlinear activation function.

11. The method according to any one of claims 7 to 10, characterized in that: The loss function used for image segmentation networks is: Loss=0.4×Loss CE +0.6×Loss Dice ; Loss CE For cross-entropy loss, Loss Dice The Dice similarity coefficient loss is defined as follows: Where y is the ground truth segmentation value, p is the segmentation mask output by the segmentation model, and n is the number of pixels; The loss function is used to calculate the difference between the true value and the model output mask during retraining. Through iteration during training, the loss function decreases, and the model parameters are optimized.

12. The method according to claim 1, characterized in that: In step S5, the average Dice similarity coefficient and the average Hausdorff distance are used as evaluation metrics for the segmentation results.

13. The method according to claim 12, characterized in that: The Dice similarity coefficient ranges from [0,1]; and / or the formulas for calculating the average Hausdorff distance are as follows: HD(P,T)=max{max p∈P min t∈T d(p,t),max t∈T min p∈P d(p,t)} Where P is the segmentation mask of the model prediction output, T is the ground truth segmentation value, p and t represent each pixel in P and T, and d(p,t) is the distance between point p and point t.

14. A temporomandibular joint magnetic resonance segmentation system, characterized in that... include: The data module is used to divide the MRI images of the mandible into training and testing sets; The module is used to build a hybrid segmentation model, including an encoder based on a global context vision Transformer structure and a decoder based on a convolutional neural network. The training module sets the model training hyperparameters, trains the hybrid segmentation model using the training set, and obtains the segmentation model parameters. The testing module uses the test set to test the hybrid segmentation model and the parameters of the segmentation model; The evaluation module is used to adjust the structure and training hyperparameters of the hybrid segmentation model based on the evaluation results of the hybrid segmentation model and the parameters of the segmentation model on the test set. The adjusted segmentation model and its parameters can be used to segment articular discs and condyles in other images.

15. A temporomandibular joint magnetic resonance imaging (MRI) segmentation device, characterized in that: It includes a processor and a memory for storing executable instructions of the processor, wherein the processor, when running, performs the method of any one of claims 1-13.

16. A computer-readable medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are processed and executed, they implement the method described in any one of claims 1-13.

17. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor to implement the method of any one of claims 1-13.

Images