ViT-based intelligent evaluation method and device for child humeroradial joint alignment relationship

Abstract

The invention relates to the technical field of medical image intelligent analysis, in particular to a ViT-based intelligent evaluation method and a ViT-based intelligent evaluation device for a child humeroradial joint alignment relationship. A child elbow joint image is segmented into a plurality of patches with the same size by applying a ViT-b-16 model; image characteristics of each patch and global context information of the image are captured, learned and analyzed by applying a multi-layer encoder and a multi-head self-attention mechanism, and image characteristics and spatial position characteristics of the capitulum humerus and the capitulum radius are identified emphatically, so that intelligent and accurate evaluation of the brachioradial joint alignment relationship of the child on the image is realized; according to the method, the limitation that the global image information is ignored due to the fact that only single-structure image features are concerned in the traditional convolution operation is improved, the uncertainty of the manual evaluation result of the child humeroradial joint alignment relation on the image is obviously reduced, the manual evaluation time and labor cost are reduced, and the method has remarkable clinical significance and wide application prospects.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, and in particular to an intelligent assessment method and device for the alignment relationship of the humeroradial joint in children based on ViT. Background Technology

[0002] In children, the alignment of the humeral radial joint is reflected in medical imaging as the degree of matching between the humeral head and the radial head, mainly divided into anatomical alignment and dislocation. Specifically, when the imaging shows that the longitudinal axis of the radius passes exactly through the center of the humeral head, it can be determined as anatomical alignment; if this condition is not met, it is considered humeral radial joint dislocation. The alignment of the humeral radial joint is a crucial anatomical landmark in the accurate diagnosis and treatment of various elbow joint diseases. Taking Monteggia fractures and elbow dislocations as examples, accurately determining the alignment of the humeral radial joint is of paramount importance for developing scientific and reasonable treatment plans and assessing prognostic effects. Once humeral radial joint dislocation is missed or misdiagnosed, it is very likely to cause a series of complications such as elbow pain, deformity, and functional impairment, seriously affecting the child's quality of life.

[0003] In current clinical practice, doctors primarily rely on X-rays to assess the alignment of the humeroradial joint in children. However, this method faces numerous challenges in practical application. Firstly, because children are in a stage of growth and development, the ossification centers of the radial and humeral heads are not yet visible on X-rays or are only partially visible, making it difficult for doctors to accurately determine the specific location and shape of the radial and humeral heads. Secondly, the irregular shape of the radial shaft makes confirming the radial longitudinal axis on X-rays extremely difficult. Furthermore, children often do not cooperate with the doctor's examination due to crying, fear, or other reasons, making it difficult for doctors to confirm whether there are signs of pain and limited movement at the humeroradial joint caused by joint dislocation. These three factors combined greatly increase the difficulty of accurately assessing the alignment of the humeroradial joint in children on imaging.

[0004] Related research results indicate that experienced physicians demonstrate higher accuracy in assessing the alignment of the humeroradial joint in children on imaging compared to less experienced physicians. However, due to the limited availability of high-quality medical resources, experienced physicians often cannot promptly guide less experienced physicians to accurately assess the alignment of the humeroradial joint on imaging, which can easily lead to misjudgments by less experienced physicians. This, in turn, can result in missed or misdiagnosed cases of related elbow joint diseases, hindering children from receiving timely and standardized diagnosis and treatment.

[0005] Furthermore, the current process for assessing the alignment of the humeroradial joint is quite cumbersome. Doctors need to manually draw the radial longitudinal axis on imaging software, a process that is not only time-consuming and labor-intensive but also prone to errors due to human factors. With the continuous development of technology, deep learning technology has provided new ideas and powerful tools for solving these problems. Deep learning technology can capture, learn, and analyze image features based on visual networks, thereby achieving intelligent, rapid, and accurate recognition of image features. Currently, researchers have successfully constructed intelligent diagnostic and treatment models for multiple diseases using image deep learning technology, demonstrating enormous application potential in the medical field.

[0006] However, most existing deep learning technologies for image processing are based on convolutional neural networks (CNNs). While CNNs excel at extracting and learning local image features, they have significant limitations when dealing with complex image features involving the spatial locations of multiple anatomical landmarks. Intelligent assessment of the alignment of the humeral radial joint in children requires comprehensive consideration of the spatial relationship between the humeral and radial heads, a requirement that deep learning models based on CNNs clearly cannot meet. Therefore, CNNs are unsuitable for constructing intelligent image assessment models for the alignment of the humeral radial joint in children based on the spatial locations of the humeral and radial heads. Thus, improving upon the existing shortcomings of deep learning technologies for image processing to achieve accurate assessment of the alignment of the humeral radial joint in children has become a critical issue that urgently needs to be addressed.

[0007] In recent years, deep learning technology for medical imaging based on the Vision Transformer (ViT) model has attracted much attention, providing a novel strategy for solving this problem. The ViT model is primarily used for image classification tasks, and its unique multi-layer encoder and multi-head attention mechanism can effectively capture local features and global dependencies in images. Compared with traditional convolutional operations, the ViT model is no longer limited to focusing on the image features of a single structure, but can more comprehensively capture, learn, and analyze the features of different anatomical structures in images and their spatial relative positions. Given that the intelligent and accurate assessment of the alignment of the humeral and radial joints in children highly depends on the image features of the humeral and radial heads and their relative spatial positions, this invention aims to construct a ViT-based intelligent assessment method and device for the alignment of the humeral and radial joints in children, based on an image dataset constructed by senior physicians, with the application of the ViT model, in order to provide strong support for the accurate diagnosis and treatment of pediatric elbow joint diseases. Summary of the Invention

[0008] To address the aforementioned issues, this invention provides a method and apparatus for intelligent assessment of the alignment of the humeral and radial joint in children based on ViT. By applying the multi-layer encoder and multi-head self-attention mechanism of the ViT model, it fully captures the local features and global contextual information of children's elbow joint images, especially the spatial positional characteristics of the humeral and radial heads. This successfully achieves intelligent, rapid, and accurate assessment of the matching degree of the humeral and radial heads on children's elbow joint images, thereby reducing the probability of missed and misdiagnosed humeral and radial joint dislocations in children and reducing the time and manpower costs of manual assessment of the alignment of the humeral and radial joint on images.

[0009] The technical solution adopted in this invention is:

[0010] A ViT-based intelligent assessment method for pediatric humeroradial joint alignment includes the following steps:

[0011] S1. Preprocessing of anteroposterior and lateral X-ray images of the elbow joint in children with anatomical alignment or dislocation of the humeroradial joint, the preprocessing including noise reduction, normalization and image size adjustment;

[0012] S2. Perform data augmentation operations on the preprocessed images, including rotation, horizontal / vertical flipping, and scaling, and randomly divide the augmented image dataset into training and test sets;

[0013] S3. Input the training set into the image classification deep learning model based on the ViT-b-16 model;

[0014] S4. The image discrimination results of anatomical alignment or dislocation of the humeroradial joint in children are output through the feature extraction layer and fully connected classification layer of the model;

[0015] S5. Evaluate model performance based on the test set to test model stability and generalization ability;

[0016] S6. Visualize the model classification results.

[0017] Furthermore, the specific preprocessing steps in step S1 are as follows:

[0018] S11. Linearly normalize the image grayscale values ​​to a fixed range (0-255) to reduce differences between devices;

[0019] S12. Proportionally crop, scale, or fill the image dataset to ultimately adjust the images to a fixed size (224×224 pixels). The cropped images include the bony structures of the distal 1 / 3 of the humerus, the proximal 1 / 3 of the forearm, and the skin on both sides.

[0020] In this step, you can first adjust the image to the target size proportionally, and then crop (if the original image is larger than the target size) or fill (if the original image is smaller than the target size) as needed.

[0021] Furthermore, the specific image data enhancement operation steps in step S2 are as follows:

[0022] S21. Randomly rotate (±10°) and horizontally / vertically flip the image dataset to exclude individual differences in X-ray imaging position. The adjusted images show the anatomical structures of the distal humerus and proximal forearm from top to bottom.

[0023] S22. Randomly divide the dataset into a training set and a test set, and keep the proportion of anatomical alignment and dislocation of the humeral radial joint in the training set roughly the same as that in the test set.

[0024] Furthermore, the model construction step described in step S3 includes:

[0025] S31. Based on the ViT-b-16 deep learning model, construct the intelligent evaluation model of the pediatric humeroradial joint alignment image of the present invention;

[0026] S32. After inputting the image dataset into the model, the input image is divided into 16×16 pixel non-overlapping patches, each patch having a dimension of 16×16×3 (corresponding to RGB three channels).

[0027] S33. Flatten each 16×16×3 patch into a 768-dimensional vector. Project the 768-dimensional vector of each patch onto the model's hidden dimension through a fully connected layer, generating a total of 196 768-dimensional embedding vectors. Assign an independent 768-dimensional position vector to each patch position so that the model can learn the spatial positional relationship of each patch later. Add the patch embedding vector and the position embedding vector element by element to form an input sequence with positional information. Add a learnable CLS tag vector to the beginning of the sequence. Finally, the model input is a sequence of length 197, including 196 patches and 1 CLS tag, with each element being a 768-dimensional vector.

[0028] S34. By aggregating local features and global contextual information of images through a multi-layer Transformer encoder and a multi-head self-attention mechanism, the structural relationship of the humeral head and radial head and the surrounding anatomical region on the images is enhanced.

[0029] S35. Use pre-trained weights to fine-tune the model, and employ weight decay and early stopping strategies to prevent overfitting.

[0030] Furthermore, in step S4, by connecting a fully connected layer after the global features marked by CLS, the Softmax function is used to output the probability distribution of "anatomical alignment / dislocation of the humeroradial joint in children on imaging".

[0031] Furthermore, the specific steps in step S5 are as follows:

[0032] S51. Calculate the model's accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Plot the ROC curve and calculate the area under the curve (AUC) to evaluate the model's performance.

[0033] Among them, accuracy represents the proportion of correctly predicted samples out of the total number of samples; sensitivity represents the proportion of samples that were actually positive but were correctly predicted as positive; specificity represents the proportion of samples that were actually negative but were correctly predicted as negative; positive predictive value represents the proportion of samples that were predicted as positive but were actually positive; negative predictive value represents the proportion of samples that were predicted as negative but were actually negative; and F1 score represents the harmonic mean of sensitivity and positive predictive value.

[0034] S52. Cross-domain evaluation is performed using independent datasets to test the model's generalization ability, while misjudgment analysis is conducted to statistically analyze failure modes under different perspectives and imaging conditions.

[0035] Furthermore, in step S6, a heatmap overlay is performed on the regions of interest of the ViT model during training, including the heatmap after image patch embedding and Transformer encoding, to evaluate whether the model focuses on key anatomical regions of the elbow joint, including the head of the humerus and the head of the radius.

[0036] Based on the same inventive concept, this application also provides a ViT-based intelligent assessment device for the alignment of the humeroradial joint in children, including a processor and a memory. The memory stores executable instructions, and when the instructions are executed by the processor, the above-mentioned method steps are implemented.

[0037] Furthermore, the device employs a high-performance GPU to accelerate training and is configured with a parameter optimization module to automatically adjust the learning rate, momentum, and batch size. Specifically, the training epochs are set to 150 epochs to ensure the model converges with sufficient iterations; the initial learning rate is set to 1×10⁻⁶. -3 The step size for parameter updates is set to 0.9 to accelerate gradient descent and suppress oscillations; the batch size is set to 32 to strike a balance between computational efficiency and model stability; and the cross-entropy loss function is used to measure the difference between the predicted results and the true labels in classification tasks.

[0038] The beneficial effects of this invention are as follows:

[0039] This invention provides a ViT-based intelligent assessment method and device for the alignment of the humeroradial joint in children. By applying the ViT-b-16 model, it segments pediatric elbow joint images into multiple patches of the same size. A multi-layer encoder and multi-head self-attention mechanism are then used to capture, learn, and analyze the image characteristics and global contextual information of each patch, including the image features and spatial positional characteristics of the humeral and radial heads. This enables intelligent and accurate assessment of the alignment of the humeroradial joint in children's images, overcoming the limitations of traditional convolution operations that focus only on single structural image features and ignore global image information. This provides a new method to improve the accuracy of physicians' assessment of the alignment of the humeroradial joint in children's images, reducing the probability of missed and misdiagnosed related pediatric elbow joint diseases. Furthermore, the method and device provided by this invention can quickly and accurately output the assessment results of the alignment of the humeroradial joint in pediatric images, significantly reducing the time and labor costs of manual assessment, and has significant clinical significance and broad application prospects. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the model construction process in an embodiment of the present invention;

[0041] Figure 2 This is a visual schematic diagram of the model building process in an embodiment of the present invention, including X-ray film input (A), image block embedding (B), and Transformer encoding process (C).

[0042] Figure 3 The ROC plot shown in the embodiments of the present invention reflects the model performance based on the test set;

[0043] Figure 4 This is a confusion matrix diagram in an embodiment of the present invention, reflecting the model performance based on the test set. In the diagram, A-HRJA is the abbreviation for Anatomical Humeroradial Joint Alignment; NA-HRJA is the abbreviation for Non-anatomical Humeroradial Joint Alignment. Detailed Implementation

[0044] To facilitate understanding of the present invention, it will be described more fully below through embodiments, and preferred embodiments are given below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Any other implementation schemes obtained by modifying or equivalently substituting the technical solutions of the present invention without inventive step are all within the protection scope of the present invention.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0046] The numerical values ​​disclosed in the embodiments of this invention are approximate values, not definitive values. Where error or experimental conditions permit, all values ​​within the error range may be included, and the specific numerical values ​​disclosed in the embodiments of this invention are not limited to those specified.

[0047] Unless otherwise stated, the various datasets, parameter settings, model training and testing methods used in the following embodiments are all datasets, parameter settings and model training and testing methods commonly used in the art.

[0048] See Figure 1 As shown, this embodiment provides a ViT-based intelligent assessment method for the alignment of the humeroradial joint in children, including the following steps:

[0049] S1. Preprocessing of anteroposterior and lateral X-ray images of the elbow joint in children with anatomical alignment or dislocation of the humeroradial joint. The preprocessing includes noise reduction, normalization, and image size adjustment. The specific steps are as follows:

[0050] S11. Linearly normalize the image grayscale values ​​to a fixed range (0-255) to reduce differences between devices;

[0051] S12. Proportionally crop, scale, or fill the image dataset, and finally adjust the images to a fixed size (224×224 pixels). The cropped images include the bony structures of the distal 1 / 3 of the humerus and the proximal 1 / 3 of the forearm, as well as the skin on both sides.

[0052] S2. Perform data augmentation operations on the preprocessed images, including rotation, horizontal / vertical flipping, and scaling, and randomly divide the augmented image dataset into training and test sets. The specific steps are as follows:

[0053] S21. Rotate (±10°), horizontally / vertically flip the image dataset to exclude individual differences in X-ray imaging position;

[0054] S22. The X-ray dataset (n=5118, including 2932 anatomical alignment images of the humeral radial joint and 2186 images of humeral radial joint dislocation) was randomly divided into a training set (n=4095, including 2346 anatomical alignment images of the humeral radial joint and 1749 images of humeral radial joint dislocation) and a test set (n=1023, including 586 anatomical alignment images of the humeral radial joint and 437 images of humeral radial joint dislocation) at a ratio of 4:1, so that the ratio of anatomical alignment and dislocation images of the humeral radial joint in children in the training set was approximately the same as that in the test set.

[0055] S3. Input the training set into the image classification deep learning model based on the ViT-b-16 model. The specific steps are as follows:

[0056] S31. Based on the ViT-b-16 deep learning model, construct the intelligent evaluation model of the pediatric humeroradial joint alignment image of the present invention;

[0057] S32. After inputting the image dataset into the model, the input image is divided into 16×16 pixel non-overlapping patches, each patch having a dimension of 16×16×3 (corresponding to RGB three channels).

[0058] S33. Flatten each 16×16×3 patch into a 768-dimensional vector. Project the 768-dimensional vector of each patch onto the model's hidden dimension through a fully connected layer, generating a total of 196 768-dimensional embedding vectors. Assign an independent 768-dimensional position vector to each patch position so that the model can learn the spatial positional relationship of each patch later. Add the patch embedding vector and the position embedding vector element by element to form an input sequence with positional information. Add a learnable CLS tag vector to the beginning of the sequence. Finally, the model input is a sequence of length 197, including 196 patches and 1 CLS tag, with each element being a 768-dimensional vector.

[0059] S34. By aggregating local features and global contextual information of images through a multi-layer Transformer encoder and a multi-head self-attention mechanism, the structural relationship of the humeral head and radial head and the surrounding anatomical region on the images is enhanced.

[0060] S35. Use pre-trained weights to fine-tune the model, and employ weight decay and early stopping strategies to prevent overfitting;

[0061] S4. The image discrimination results of anatomical alignment or dislocation of the humeral radial joint in children are output through the feature extraction layer and the fully connected classification layer of the model. Specifically, the fully connected layer is connected after the global features labeled by CLS, and the Softmax function is used to output the probability distribution of "anatomical alignment / dislocation of the humeral radial joint in children on the image".

[0062] S5. Evaluate model performance based on the test set to test model stability and generalization ability. The specific steps are as follows:

[0063] S51. Calculate the model's accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Plot the ROC curve and calculate the area under the curve (AUC) to evaluate the model's performance.

[0064] Among them, accuracy represents the proportion of correctly predicted samples out of the total number of samples; sensitivity represents the proportion of samples that were actually positive but were correctly predicted as positive; specificity represents the proportion of samples that were actually negative but were correctly predicted as negative; positive predictive value represents the proportion of samples that were predicted as positive but were actually positive; negative predictive value represents the proportion of samples that were predicted as negative but were actually negative; and F1 score represents the harmonic mean of sensitivity and positive predictive value.

[0065] S52. Use independent datasets for cross-domain evaluation to test the model's generalization ability, and perform misjudgment analysis to statistically analyze failure modes under different perspectives and imaging conditions.

[0066] S6. Visualize the model classification results and display the heatmap overlay of the model's focus areas during ViT model training, including the heatmap after image patch embedding and Transformer encoding, to evaluate whether the model focuses on key anatomical areas of the elbow joint, including the head of the humerus and the head of the radius.

[0067] In this embodiment, five commonly used traditional convolutional neural network models (ResNet18, ResNet50, DenseNet121, VGG16, and RegNetY2) were compared with the ViT-b-16 model constructed in this invention to evaluate the experimental results. Specifically, comparative experiments were conducted on the test set for the six models. The accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score of the models were calculated to evaluate the performance of the ViT-b-16 model in recognizing the alignment of the humeroradial joint in images of children. The specific comparison results are shown in Table 1.

[0068] Table 1. Comparison results of the present invention with five commonly used traditional convolutional neural network models.

[0069] Model Accuracy (%) Sensitivity (%) Specificity (%) Positive predictive value (%) Negative predictive value (%) F1 score (%) ResNet18 80.64 81.30 81.41 80.68 81.22 80.99 ResNet50 83.97 88.57 77.80 84.25 83.54 86.36 DenseNet121 88.33 89.48 85.80 88.67 87.87 89.08 VGG16 87.55 89.56 83.85 87.33 87.85 88.44 RegNetY2 86.73 86.40 86.47 86.53 86.54 84.42 ViT-b-16 92.29 92.56 90.18 91.35 89.17 91.45

[0070] Table 1 shows that the ViT-b-16 model provided by this invention outperforms traditional convolutional neural network models such as ResNet18, ResNet50, DenseNet121, VGG16, and RegNetY2 in multiple evaluation metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Furthermore, the model heatmap overlay results indicate that the model constructed in this invention can focus on the anatomical positions of the humeral and radial heads when detecting the alignment of the humeral and radial joints in images of children. Figure 2 This fully demonstrates the model's good detection performance. Furthermore, the ROC analysis results ( Figure 3 The results show that the AUC value of the ViT-b-16 model constructed in this invention is 0.944, and the confusion matrix diagram ( Figure 4The results show that the model of the present invention has a high accuracy rate, indicating that the model has excellent performance in distinguishing the alignment relationship of the humeroradial joint in children.

[0071] The above results confirm that the method and apparatus provided by this invention have significant innovation and feasibility. Since the assessment of the alignment of the humeral radial joint in children's images is mainly based on the spatial positional characteristics of the humeral and radial heads, traditional convolutional neural networks focus primarily on local image features, resulting in poor performance in capturing and distinguishing global image information. To address this, the method and apparatus provided by this invention innovatively introduce a global self-attention mechanism. This mechanism not only effectively learns and distinguishes the local image features of the humeral and radial heads, but more importantly, it also captures global image information, including the accurate differentiation of the spatial positions of the humeral and radial heads, thereby achieving precise assessment of the alignment of the humeral radial joint in children's images.

[0072] In summary, the ViT-based intelligent assessment method and device for pediatric humeroradial joint alignment provided by this invention segments pediatric elbow joint images into multiple patches of the same size using the ViT-b-16 model. It then employs a multi-layer encoder and multi-head self-attention mechanism to capture, learn, and analyze the image characteristics and global contextual information of each patch, including the image features of the humeral and radial heads and their spatial positional characteristics. This achieves intelligent and accurate assessment of the humeroradial joint alignment in images, overcoming the limitations of traditional convolution operations that focus only on single structural image features and ignore global image information. This provides a new method to improve the accuracy of physicians' assessment of pediatric humeroradial joint alignment in images, reducing the probability of missed and misdiagnosed related pediatric elbow joint diseases. Furthermore, the method and device provided by this invention can quickly and accurately output the assessment results of pediatric humeroradial joint alignment, significantly reducing the time and labor costs of manual assessment of pediatric humeroradial joint alignment in images, and has significant clinical significance and broad application prospects.

[0073] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A ViT-based intelligent assessment method for the alignment relationship of the humeroradial joint of children, characterized in that, Comprising the following steps: S1. Preprocessing the anteroposterior and lateral X-ray images of the elbow joint of children with anatomical alignment or dislocation of the humeroradial joint, the preprocessing including denoising, normalization and image size adjustment; S2. Data enhancement operation is performed on the preprocessed images, including rotation, horizontal / vertical flipping and scaling, and the enhanced image dataset is randomly divided into a training set and a test set; S3. Input the training set into the image classification deep learning model based on the ViT-b-16 model; S4. Output the image discrimination result of the anatomical alignment or dislocation of the humeroradial joint of children through the feature extraction layer and the fully connected classification layer of the model; S5. Evaluate the model performance based on the test set to test the model stability and generalization ability; S6. Visualize the model classification result.

2. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, The specific preprocessing steps in step S1 are: S11. Linearly normalize the image gray value to a fixed range (0-255) to reduce the difference between devices; S12. Crop, scale or pad the image dataset proportionally, and finally adjust the image to a fixed size (224x224 pixels), and the cropped image contains the bony structure of the distal 1 / 3 of the humerus and the proximal 1 / 3 of the forearm and the skin on both sides.

3. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, The specific image data enhancement operation steps in step S2 are: S21. Randomly rotate, horizontally / vertically flip the image dataset to exclude individual differences in X-ray shooting position, and the adjusted image shows the anatomical structure of the distal humerus and the proximal forearm from top to bottom; S22. Randomly divide the dataset into a training set and a test set, and keep the proportion of image data of children with anatomical alignment and dislocation of the humeroradial joint in the training set basically consistent with that in the test set.

4. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, The model construction steps in step S3 include: S31. Constructing the image intelligent evaluation model of the anatomical alignment of the humeroradial joint of children in the present application based on the ViT-b-16 deep learning model; S32. After inputting the image dataset into the model, the input image is divided into 16x16 non-overlapping patches, each patch has a dimension of 16x16x3 (corresponding to RGB three channels); S33. Flatten each 16x16x3 patch into a 768-dimensional vector, project each patch's 768-dimensional vector into the model's hidden dimension through a fully connected layer, and generate a total of 196 768-dimensional embedding vectors; assign an independent 768-dimensional position vector to each patch position to allow the model to learn the spatial position relationship of each patch in the later stage, and add the patch embedding vector and the position embedding vector element by element to form an input sequence with position information; add a learnable CLS marker vector at the beginning of the sequence; finally, the model input is a sequence of length 197, including 196 patches and 1 CLS marker, and each element is a 768-dimensional vector; S34. Aggregate local features and global context information through multi-layer Transformer encoder and multi-head self-attention mechanism to strengthen the representation of the structure relationship of the humeral capitulum, radial capitulum and surrounding anatomical area in the image; S35. Fine-tune the model using pre-trained weights, and use weight decay and early stopping strategy to prevent model overfitting.

5. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, In step S4, the probability distribution of "child elbow radiocapitellar joint anatomical alignment / dislocation on the image" is output by connecting the global features marked by CLS with a fully connected layer and using a Softmax function.

6. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, The specific steps in step S5 are: S51. Calculate the accuracy, sensitivity, specificity, positive predictive value, negative predictive value and F1 score of the model, draw the ROC curve and calculate the area under the curve to evaluate the model performance; S52. Cross-domain evaluation is performed on an independent data set to test the model generalization, and misjudgment analysis is performed to count the failure modes under different viewing angles and imaging conditions.

7. The ViT-based intelligent assessment method for child humeroradial joint alignment relationship according to claim 1, characterized in that, In step S6, the heat map of the model attention area in the ViT model training is superimposed and displayed, including the heat map after image block embedding and Transformer encoding, so as to evaluate whether the model focuses on the key anatomical region of the elbow joint including the capitulum humeri and capitulum radii.

8. A ViT-based intelligent assessment device for the alignment relationship of the humeroradial joint of children, characterized in that The device includes a processor and a memory, and the memory stores executable instructions which, when executed by the processor, implement the method steps of any one of claims 1 to 7.

9. The ViT-based child humeroradial joint alignment relationship intelligent assessment device according to claim 8, characterized in that, The device uses high-performance GPU acceleration training and configures a parameter optimization module to automatically adjust the learning rate, momentum and batch size. The specific steps are: setting the training epoch to 150 epochs to ensure that the model converges under sufficient iteration; setting the initial learning rate to 1×10⁻³ to control the step size of parameter update; setting the momentum parameter to 0.9 to accelerate gradient descent and suppress oscillation; setting the batch size to 32 to balance between computational efficiency and model stability; using a cross-entropy loss function to measure the difference between the predicted results and the true labels in the classification task.

Images