Cervical vertebra bone age staging method based on combined model

Abstract

The invention relates to a cervical vertebra bone age staging identification method. The method comprises the following steps: S1, acquiring a skull side film of a cervical vertebra bone age staging object; s2, inputting the skull side position film into a cervical vertebra region target detection model to remove irrelevant backgrounds, and only retaining target regions containing C2, C3 and C4 joints; s3, inputting the area detected in the S2 into an instance segmentation model to segment C2, C3 and C4 joints, and labeling category labels; s4, inputting the area detected in the S2 into a mark point detection model to extract mark points with detail information at the edges of the bone joints C2, C3 and C4; and S5, inputting the features extracted in S3 and S4 into a cervical vertebra staging estimation module to obtain a final staging result. Compared with the prior art, the method has the advantages that the traditional manual skull side position film evaluation is converted into automatic detection, segmentation and staging tasks by using deep learning, so that the diagnosis efficiency is greatly improved, the staging accuracy is improved, and a doctor is effectively assisted in making an orthodontic plan.

Description

Technical Field

[0001] The invention relates to a cervical vertebrae bone age staging method based on a joint model, and belongs to the technical field of medical image processing and computer vision. Background Art

[0002] In the field of oral medicine, skeletal maturity assessment is crucial for treatment planning and determining the best time for treatment (for example, planning the appropriate time for dentofacial plastic surgery, orthognathic surgery, and implants). Doctors usually perform cervical maturity assessment, namely CVM (Cervical Vertebral Maturation), on lateral skull radiographs for orthodontic diagnosis and treatment planning. However, CVM staging is usually assessed by naked eye, which requires experienced and well-trained radiologists or orthodontists, and the repeatability of non-expert examiners is poor, which cannot meet existing clinical needs. It is necessary to use artificial intelligence to assist doctors in diagnosis.

[0003] However, the existing artificial intelligence methods still have many defects. The existing methods use a single model, such as the instance segmentation model U-Net, which predicts the cervical vertebrae bone age by classifying each pixel and extracting fine-grained features and object edge information. Although U-Net can capture fine pixel-level features, it focuses on local detail information and does not adequately process the overall bone morphology or relative position information. It is easy to ignore the global features required for the bone age classification task. Some people also use the target detection model Yolo to quickly and globally locate the cervical vertebrae joint area in the image and extract the overall features of the joint. Although it can quickly locate the entire cervical vertebrae area, it has limited processing capabilities for subtle anatomical changes, such as the shape and structural details of the cervical vertebrae joints, and cannot fully capture the local features required for bone age judgment. These factors make the cervical vertebrae bone age prediction results under a single model have poor stability and low accuracy, and cannot effectively help doctors judge bone maturity.

[0004] Therefore, we proposed a cervical vertebrae bone age staging method based on a joint model, which combines target detection, instance segmentation and landmark point detection, and can effectively solve the problems encountered by current artificial intelligence methods in cervical vertebrae bone age staging.

[0005] After searching, the application publication number CN113570577B, a method for determining the cervical vertebrae bone age, is based on a lateral skull X-ray film, and a neural network cervical vertebrae detection model, a neural network cervical vertebrae landmark point positioning model and a neural network cervical vertebrae segmentation model are established; the cervical vertebrae detection model is used to detect the cervical vertebrae position in the lateral skull X-ray film; the cervical vertebrae landmark point positioning model is used to locate the cervical vertebrae shape landmark points; the cervical vertebrae are segmented using the cervical vertebrae segmentation model; the cervical vertebrae are extracted from the segmented cervical vertebrae; the relationship between the landmark points is quantified and interpreted, and then the current growth and development stage is obtained.

[0006] The patent uses Yolov3 for target detection, which is fast but relatively low in accuracy, especially in processing the details of complex targets (such as the cervical vertebra), which may result in a high false detection rate and missed detection rate. The present invention uses Yolov7 to improve the detection accuracy while maintaining high speed, especially in complex scenes, Yolov7 can better identify small objects (such as the cervical vertebra). The present invention uses Blendmask for instance segmentation, which can finely capture the boundary information of the skeleton on multi-scale features, and the segmentation result is more accurate. Compared with the U2Net used in the invention, Blendmask has higher segmentation accuracy and can better handle local complex structures. The invention uses a common convolutional network combined with a fully connected layer to predict landmarks, which performs poorly on complex geometric structures. For complex bone structures such as the cervical vertebra, it is difficult to capture accurate landmark information and the BP algorithm used converges slowly, the training efficiency is very low, and it is easy to fall into a local optimal solution. SimCC in the present invention significantly improves the stability and accuracy of landmark detection by converting the regression problem into a classification problem, and can provide more robust and smooth results, especially when processing structures with irregular shapes. Summary of the invention

[0007] The present invention aims to solve the above problems of the prior art. A cervical vertebrae bone age staging method based on a joint model is proposed. The technical solution of the present invention is as follows:

[0008] A cervical vertebrae bone age staging method based on a joint model comprises the following steps:

[0009] S1: Collect and preprocess lateral skull films, remove abnormal data such as mobile phone photos, dark or excessive images to ensure the quality of the data set, assign categories to each pixel or area of ​​the image for deep learning training, and divide all annotated images into training and test sets in a ratio of 8:2;

[0010] S2: Use the training set to train the cervical vertebra area target detection network to generate a cervical vertebra area target detection network model, and use the test set to test the model performance;

[0011] S3: Use cervical spine images to train the instance segmentation network and generate C2, C3, and C4 joint instance segmentation models. Use the test set to test the model performance and check whether the C2, C3, and C4 joints can be accurately segmented and labeled.

[0012] S4: Use the cervical spine area images to train the landmark detection network, generate the C2, C3, C4 joint edge landmark detection model, and use the test set to test the model performance to detect the C2, C3, C4 joint edge landmarks;

[0013] S5: Design a cervical vertebrae bone age staging estimation module based on the cervical vertebrae maturity staging standard, and input the C2, C3, and C4 joint feature information extracted in steps S3 and S4 for automated staging.

[0014] Furthermore, the step S1 preprocesses the lateral cranial radiograph and divides the data set, specifically including:

[0015] Collect lateral skull films that meet medical requirements, clean the collected lateral skull films to exclude low-quality images, ensure that all images have a unified standard, scale the images, adjust the images to a uniform resolution, normalize the images to ensure that the pixel values ​​of the images are consistent, and divide the preprocessed lateral skull films into training and test sets in a ratio of 8:2.

[0016] Furthermore, the step S2 generates a cervical vertebra area target detection network model and performs a test, specifically including:

[0017] Yolov7 is selected as the cervical vertebra area target detection network. Multiple parallel branches are introduced into the network through the multi-branch stacking module. The information of different scales and feature layers is extracted and stacked and fused, which helps to capture diverse features. The feature layers of different resolutions are connected together through the improved transition module to reduce information loss and ensure that the features are fully utilized at different scales. The target detection model is generated through training and tested on the test set to verify the model performance, that is, whether it can detect the cervical vertebra area where the C2, C3, and C4 joints are located.

[0018] Furthermore, the step S3 generates the C2, C3, and C4 condyle instance segmentation models, specifically including:

[0019] The training generation model is based on the BlendMask instance segmentation network, which consists of three main parts: the bottom module is used to predict the score map Base; the top layer predicts the weight value of each pixel in the score map Base; the fusion module Blender module fuses the score map Base and the weight value of each pixel and superimposes them by channel to obtain the final mask.

[0020] Furthermore, the Bottommodule prediction score map Base specifically includes:

[0021] The bottom module uses a pre-trained backbone convolutional neural network to extract features from the input image to obtain a set of multi-scale feature maps, which are further processed using convolutional layers to generate a relatively rough score map Base with position and scale information, that is, an image containing the approximate area probability of the target.

[0022] The top layer predicts the weight value of each pixel in the base image, including:

[0023] The top layer obtains more detailed features from the bottom feature map, processes these features through multiple layers of convolution and activation functions to predict the weight value of each pixel in the base map. This weight indicates the relative importance of the pixel in generating the final segmentation mask.

[0024] The Blender module fuses the score map Base and the weight value of each pixel and superimposes them by channel to get the final mask, including:

[0025] The Blender module multiplies the Base score map with the corresponding weight map pixel by pixel to ensure that the score of each position in the score map is adjusted by its corresponding weight. The weight values ​​of different convolutional layers are applied to the score map of the corresponding channel by channel superposition to obtain masks of different instances. The activation function is used to process the superposition result to generate an accurate instance segmentation mask to indicate the probability that each pixel belongs to a specific target instance.

[0026] Furthermore, the step S4 generates a C2, C3, C4 condyle edge landmark point detection model, which is as follows:

[0027] The simple coordinate classification strategy SimCC (Simple Coordinate Classification) is used to detect key points. SimCC only needs two classifier heads for coordinate classification. For the x-coordinate and y-coordinate, two independent classifier heads are created respectively. Each classifier head will output a category probability distribution. The x-coordinate classifier is responsible for classifying the horizontal coordinates (x-axis) of the image, and the y-coordinate classifier is responsible for classifying the vertical coordinates (y-axis) of the image. Each classifier head maps the feature map to a category space through a convolutional layer. For the x-coordinate classifier, it generates a probability distribution, indicating the possibility of the x-coordinate corresponding to each pixel. The y-coordinate classifier also generates a probability distribution representing the y-coordinate. The probability of each coordinate category is obtained by normalization through the softmax function, and the category with the largest probability is selected as the coordinate prediction value of the key point. By combining these two classifiers, the precise position of all key points is finally obtained.

[0028] Furthermore, the step S5: designing a cervical vertebrae bone age staging estimation module based on the cervical vertebrae maturity staging standard and inputting the feature information extracted by S3 and S4 into the module to realize automatic staging, specifically includes:

[0029] The clinical maturity stages of the cervical spine have the following criteria:

[0030] CS1: The lower edge of the C2 to C4 vertebrae is flat (sometimes slightly convex), the third and fourth vertebrae are trapezoidal in shape, the posterior edge of the vertebrae is higher than the anterior edge, and the upper surface is inclined downward and forward;

[0031] CS2: There is a visible notch at the lower edge of C2, the lower edges of the third and fourth vertebrae remain flat, and both C3 and C4 retain a trapezoidal shape;

[0032] CS3: The lower borders of C2 and C3 are clearly notched, and the lower edge of C4 remains flat;

[0033] CS4: All vertebrae have obvious concavities on their lower surfaces. In the CS4 period, both C3 and C4 vertebrae have rectangular horizontal shapes instead of trapezoidal shapes.

[0034] CS5: At least one of the C3 and C4 vertebrae is square; if not square, the vertebral body of the cervical spine is a horizontal rectangle with obvious depressions at the lower borders of all three vertebrae;

[0035] CS6: At least one of the C3 and C4 vertebrae has a rectangular vertical shape, i.e., the posterior border is longer than the inferior border. If the vertebral body is not rectangular vertically shaped, the vertebral body of the other cervical vertebra is square;

[0036] According to the segmentation results and landmark point prediction results, the manual features of C2, C3, and C4, including the edge curvature and the vertebral length-width ratio, were extracted. These manual features were input into the estimation module designed according to the cervical maturity staging standard to obtain the final staging results.

[0037] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a cervical vertebrae bone age staging method based on a joint model as described in any one of the above is implemented.

[0038] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the combined model-based cervical vertebrae bone age staging methods described.

[0039] The advantages and beneficial effects of the present invention are as follows:

[0040] After adopting the above technical solution, compared with the prior art, the beneficial effect of the present invention is that the bone age staging method proposed in the present invention can automatically estimate the maturity of the cervical spine based on the lateral skull radiograph data using deep learning to assist doctors in diagnosis and improve clinical efficiency. This method combines the yolov7 target detection network, the Blendmask instance segmentation network and the SimCC landmark detection network. Yolov7, as the target detection network, can quickly and accurately identify and locate the cervical vertebrae. It provides the necessary context information for subsequent segmentation and key point detection, ensuring that only the area of ​​interest is processed, reducing the waste of computing resources. Blendmask receives the detection frame output by yolov7, and refines the contour of the cervical spine through instance segmentation. Its mask fusion technology can effectively capture complex shapes and improve detail processing capabilities, ensuring that the boundaries of each vertebra are clear and accurate. SimCC avoids the coordinate regression error in the traditional method and provides a smoother and more stable landmark prediction, which not only improves the robustness of the detection, but also effectively copes with cervical structures with different postures and deformations. Blendmask provides the global shape contour of the cervical vertebrae, while SimCC focuses on local geometric landmarks. The combination of the two enables the model to grasp the overall structure of the cervical vertebrae and accurately locate the local landmarks of the cervical vertebrae. This combination of global and local features provides comprehensive cervical vertebrae information, which helps to make more accurate bone age staging judgments. By combining these models to effectively extract the different detailed features of the C2, C3, and C4 joints in the area of ​​interest and designing a cervical bone age staging module based on the CVM standard, the automatic staging results are more robust, achieving a staging accuracy of 76% on the ISBI2015 public data set, further improving the efficiency of doctors' diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the overall process of the cervical vertebrae bone age staging method based on the joint model according to a preferred embodiment of the present invention;

[0042] Figure 2 This is a network framework diagram of the cervical vertebrae bone age staging method based on the joint model. DETAILED DESCRIPTION

[0043] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0044] The technical solution of the present invention to solve the above technical problems is:

[0045] refer to Figure 1 , which is a schematic diagram of the steps of this method, specifically comprising the following steps:

[0046] S1 data acquisition, preprocessing and data set division:

[0047] Collect lateral skull films that meet medical ethics, clean the collected lateral skull films to exclude low-quality images, such as images taken with mobile phones, ensure that all images have certain standards, such as file suffixes, etc., scale the images, adjust the images to a uniform resolution, normalize the images, ensure the consistency of the pixel values ​​of the images, and divide the preprocessed lateral skull films into training and test sets in a ratio of 8:2.

[0048] S2 uses the training set to train the cervical vertebrae area target detection network and generate a target detection model:

[0049] Yolov7 is selected as the cervical spine area target detection network. The network focuses on lightweight and high efficiency. Through multi-branch stacking modules and innovative transition modules, it can achieve good detection effects on small data medical image datasets. The target detection model is generated through training and tested on the test set to verify the model performance, that is, whether it can detect the cervical spine area where the C2, C3, and C4 joints are located.

[0050] S3 uses the training set to train the instance segmentation network and generate an instance segmentation model:

[0051] The training is based on the BlendMask instance segmentation network, which consists of three main parts: using the Bottom module to predict the score map Base; the Top layer predicts the weight value of each pixel in the Base map; the Blender module fuses the score map Base and the weight value of each pixel and superimposes them by channel to obtain the final mask. The C2, C3, and C4 condyle edges segmented by this network are more accurate, providing more detailed feature information for the subsequent calculation of the manual features of the condyle.

[0052] S4 uses the training set to train the landmark detection network and generate a landmark model:

[0053] A novel simple coordinate classification strategy, SimCC, is used to detect landmark points. The traditional landmark point model predicts landmark points through heat maps. Since continuous coordinate values ​​need to be mapped to a discretized two-dimensional reduced heat map, there is a serious quantization error problem. Expensive computing resources are required to improve the resolution of the image to alleviate this problem, and additional post-processing is required to improve the prediction accuracy. SimCC simplifies the prediction process, requiring only two classifier heads for coordinate classification, and eliminates expensive refinement post-processing and additional upsampling layers. The landmark points predicted by this strategy are more accurate, and the manual features calculated through the coordinate information between landmark points are more effective in improving the accuracy of staging.

[0054] S5 designs a cervical vertebrae bone age staging estimation module based on the cervical vertebrae maturity staging standard, and inputs the C2, C3, and C4 joint feature information extracted by S3 and S4 for automatic staging:

[0055] The clinical maturity stages of the cervical spine have the following criteria:

[0056] CS1: The lower edge of the vertebrae from C2 to C4 is flat (sometimes slightly convex), the third and fourth vertebrae are trapezoidal in shape, the posterior edge of the vertebrae is higher than the anterior edge, and the upper surface is inclined downward and forward.

[0057] CS2: There is a visible notch at the lower edge of C2. The lower edges of the third and fourth vertebrae remain flat. Both C3 and C4 retain a trapezoidal shape.

[0058] CS3: The lower borders of C2 and C3 are notched, and the lower border of C4 remains flat.

[0059] CS4: All vertebrae have a distinct concavity on their lower surfaces. In the CS4 period, both C3 and C4 vertebrae have a rectangular horizontal shape instead of a trapezoidal shape.

[0060] CS5: At least one of the C3 and C4 vertebrae is square. If not square, the vertebral bodies of the cervical spine are horizontally rectangular with obvious concavities at the lower borders of all three vertebrae.

[0061] CS6: At least one of the C3 and C4 vertebral bodies has a rectangular vertical shape, i.e., the posterior border is longer than the inferior border. If the vertebral body is not rectangular vertically shaped, the other cervical vertebral body is square.

[0062] According to the segmentation results and landmark point prediction results, manual features of C2, C3, and C4 were extracted, such as the degree of lower edge curvature, vertebral length-to-width ratio, etc. These manual features were input into the estimation module designed according to the cervical maturity staging standard to obtain the final staging results.

[0063] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0064] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0065] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0066] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A cervical vertebrae bone age staging method based on a joint model, characterized in that: The following steps are involved: S1: Collect and preprocess lateral skull films, remove abnormal data such as mobile phone photos, dark or excessive images to ensure the quality of the data set, assign categories to each pixel or area of ​​the image for deep learning training, and divide all annotated images into training and test sets in a ratio of 8:2; S2: Use the training set to train the cervical vertebra area target detection network to generate a cervical vertebra area target detection network model, and use the test set to test the model performance; S3: Use cervical spine images to train the instance segmentation network and generate C2, C3, and C4 joint instance segmentation models. Use the test set to test the model performance and check whether the C2, C3, and C4 joints can be accurately segmented and labeled. S4: Use the cervical spine area images to train the landmark detection network, generate the C2, C3, C4 joint edge landmark detection model, and use the test set to test the model performance to detect the C2, C3, C4 joint edge landmarks; S5: Design a cervical vertebrae bone age staging estimation module based on the cervical vertebrae maturity staging standard, and input the C2, C3, and C4 joint feature information extracted in steps S3 and S4 for automated staging.

2. A cervical vertebrae bone age staging method based on a joint model according to claim 1, characterized in that: The step S1 preprocesses the lateral cranial radiograph and divides the data set, specifically including: Collect lateral skull films that meet medical requirements, clean the collected lateral skull films to exclude low-quality images, ensure that all images have a unified standard, scale the images, adjust the images to a uniform resolution, normalize the images to ensure that the pixel values ​​of the images are consistent, and divide the preprocessed lateral skull films into training and test sets in a ratio of 8:

2.

3. The cervical vertebrae bone age staging method based on a joint model according to claim 1, characterized in that: The step S2 generates a cervical vertebra area target detection network model and performs a test, specifically including: Yolov7 is selected as the cervical vertebra area target detection network. Multiple parallel branches are introduced into the network through the multi-branch stacking module. The information of different scales and feature layers is extracted and stacked and fused, which helps to capture diverse features. The feature layers of different resolutions are connected together through the improved transition module to reduce information loss and ensure that the features are fully utilized at different scales. The target detection model is generated through training and tested on the test set to verify the model performance, that is, whether it can detect the cervical vertebra area where the C2, C3, and C4 joints are located.

4. The cervical vertebrae bone age staging method based on a joint model according to claim 1, characterized in that: The step S3 generates the C2, C3, and C4 condyle instance segmentation models, specifically including: The training generation model is based on the BlendMask instance segmentation network, which consists of three main parts: the bottom module is used to predict the score map Base; the top layer predicts the weight value of each pixel in the score map Base; the fusion module Blender module fuses the score map Base and the weight value of each pixel and superimposes them by channel to obtain the final mask.

5. The cervical vertebrae bone age staging method based on the joint model according to claim 4, characterized in that: The bottom module prediction score map Base specifically includes: Use the pre-trained backbone convolutional neural network to extract features from the input image to obtain a set of multi-scale feature maps, which are further processed using convolutional layers to generate a relatively rough score map Base with position and scale information, that is, an image containing the approximate area probability of the target. The top layer predicts the weight value of each pixel in the base image, including: The top layer obtains more detailed features from the bottom feature map, processes these features through multiple layers of convolution and activation functions to predict the weight value of each pixel in the base map. This weight indicates the relative importance of the pixel in generating the final segmentation mask. The Blender module fuses the score map Base and the weight value of each pixel and superimposes them by channel to get the final mask, including: The Blender module multiplies the Base score map with the corresponding weight map pixel by pixel to ensure that the score of each position in the score map is adjusted by its corresponding weight. The weight values ​​of different convolutional layers are applied to the score map of the corresponding channel by channel superposition to obtain masks of different instances. The activation function is used to process the superposition result to generate an accurate instance segmentation mask to indicate the probability that each pixel belongs to a specific target instance.

6. The cervical vertebrae bone age staging method based on a joint model according to claim 1, characterized in that: The step S4 generates a C2, C3, C4 condyle edge landmark point detection model, which is as follows: The simple coordinate classification strategy SimCC (Simple Coordinate Classification) is used to detect landmark points. SimCC only needs two classifier heads for coordinate classification. For the x-coordinate and y-coordinate, two independent classifier heads are created respectively. Each classifier head will output a category probability distribution. The x-coordinate classifier is responsible for classifying the horizontal coordinates (x-axis) of the image, and the y-coordinate classifier is responsible for classifying the vertical coordinates (y-axis) of the image. Each classifier head maps the feature map to a category space through a convolutional layer. For the x-coordinate classifier, it generates a probability distribution, indicating the possibility of the x-coordinate corresponding to each pixel. The y-coordinate classifier also generates a probability distribution representing the y-coordinate. The probability of each coordinate category is obtained by normalization through the softmax function, and the category with the largest probability is selected as the coordinate prediction value of the landmark point. By combining these two classifiers, the precise position of all landmark points is finally obtained.

7. The cervical vertebrae bone age staging method based on the joint model according to claim 6, characterized in that: The step S5: designing a cervical vertebrae bone age staging estimation module based on the cervical vertebrae maturity staging standard and inputting the feature information extracted in S3 and S4 into the module to realize automatic staging, specifically includes: The clinical maturity stages of the cervical spine have the following criteria: CS1: The lower edge of the C2 to C4 vertebrae is flat (sometimes slightly convex), the third and fourth vertebrae are trapezoidal in shape, the posterior edge of the vertebrae is higher than the anterior edge, and the upper surface is inclined downward and forward; CS2: There is a visible notch at the lower edge of C2, the lower edges of the third and fourth vertebrae remain flat, and both C3 and C4 retain a trapezoidal shape; CS3: The lower borders of C2 and C3 are clearly notched, and the lower edge of C4 remains flat; CS4: All vertebrae have obvious concavities on their lower surfaces. In the CS4 period, both C3 and C4 vertebrae have rectangular horizontal shapes instead of trapezoidal shapes. CS5: At least one of the C3 and C4 vertebrae is square; if not square, the vertebral body of the cervical spine is a horizontal rectangle with obvious depressions at the lower borders of all three vertebrae; CS6: At least one of the C3 and C4 vertebrae has a rectangular vertical shape, i.e., the posterior border is longer than the inferior border. If the vertebral body is not rectangular vertically shaped, the vertebral body of the other cervical vertebra is square; According to the segmentation results and landmark point prediction results, the manual features of C2, C3, and C4, including the degree of edge curvature and the aspect ratio of the vertebral body, were extracted. These manual features were input into the estimation module designed according to the cervical maturity staging standard to obtain the final staging results.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the cervical vertebrae bone age staging method based on the joint model as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the cervical vertebrae bone age staging method based on a joint model as described in any one of claims 1 to 7 is implemented.

Images