Hallux valgus angle prediction and risk assessment research method based on multi-task network

Abstract

The invention provides a hallux valgus angle prediction and risk assessment research method based on a multi-task network, relates to the technical field of image processing, and collects actual clinical foot images, and constructs a hallux valgus angle measurement and damage determination data set for multi-task network training. The method comprises the following steps: firstly, acquiring a foot image, carrying out data labeling, and training a multi-task model by utilizing a labeled data set; then, the trained model is applied to a new foot image, and an angle line segment heat map of hallux valgus and a damage judgment result are obtained through model prediction. And finally, performing linear regression based on the line segment heat map to obtain two hallux valgus angles of HVA and IMA. The method provided by the invention can provide an accurate diagnosis basis for clinicians, help the clinicians to assist in diagnosis, significantly improve the accuracy and efficiency of hallux valgus treatment, and reduce the time and errors of manual operation. The method has wide application prospects in the fields of hallux valgus angle measurement and damage judgment.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to the measurement of the HVA (hallux valgus angle) and IMA (inter-metatarsal angle) of hallux valgus and the determination of the injury grade. Background Art

[0002] The HVA and IMA angles and injury grade of hallux valgus are particularly critical for the diagnosis and treatment of hallux valgus. If these two data can be provided as auxiliary diagnosis in clinical practice, the accuracy and effectiveness of clinical diagnosis will be greatly improved. Current angle measurement methods mainly rely on radiographic images and manual annotation, which are not only highly subjective but also have low diagnostic efficiency. In addition, relying solely on HVA and IMA angle measurements cannot accurately reflect the progression of the disease and the specific conditions of the patient, which further limits the accuracy and effectiveness of clinical diagnosis. Injury grade is based on the rule-based classification criteria of HVA and IMA. If the angles are not in the same range, it is difficult to give an accurate classification.

[0003] In recent years, multi-task learning has been widely used in medical image analysis. By sharing representational information between different tasks, it improves model generalization. Hallux valgus injury assessment and angle prediction are closely related tasks, and the hallux valgus angle largely determines the severity of injury assessment. Currently, there are no automated methods for measuring hallux valgus angles (HVA) and IMA (Inner Maneuvering Axis) and injury assessment. Integrating these two tasks into a multi-task learning framework can exploit the relationship between them, thereby improving the model's angle prediction capabilities and injury assessment efficiency. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a research method for hallux valgus angle prediction and risk assessment based on a multi-task network;

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A research method for hallux valgus angle prediction and risk assessment based on a multi-task network includes the following steps:

[0007] S1. Collect medical images of hallux valgus feet and perform data preprocessing;

[0008] First, the subject's foot CT images are collected and uploaded from the cloud server;

[0009] Then, according to the actual clinical foot images collected, the hallux valgus affected side is identified based on the foot images to obtain the hallux valgus affected side of the subject, and the hallux valgus injury is divided into three levels: mild, moderate and severe.

[0010] S2. Label the multi-task labels for hallux valgus angle prediction and injury determination, and perform data augmentation and dataset division;

[0011] First, the key points of the thumb joint are marked according to the two angles HVA and IMA, and a joint point image of the hallux valgus area is generated. This image is used as the label for hallux valgus angle prediction. There are a total of six key points, two points determine a line, and three lines determine the angles of the two angles HVA and IMA. According to the hallux valgus injury level, the corresponding three-category labels light, medium, and severe are generated, with light marked as 0, medium marked as 1, and severe marked as 2.

[0012] Then, various data enhancement methods are adopted to expand the number of samples in the multi-task dataset for hallux valgus angle prediction and injury assessment.

[0013] The data enhancement method includes: performing a vertical axis mirror flipping operation on the image, which can help the model learn the presentation characteristics of objects in different directions; by randomly rotating the image at an angle of -15° to +15°, the model's adaptability to object rotation changes can be enhanced, so that it can better handle the situation of objects at different angles; implementing a random scale scaling of 0.8 to 1.2 times, that is, randomly selecting a scaling ratio from this range, so that the model can cope with objects of different scales; performing scale scaling and shearing transformation of the image, the shearing transformation is to stretch a part of the image in the horizontal or vertical direction to form a parallelogram, so as to simulate the deformation of the object caused by the change of perspective; adding Gaussian noise, that is, introducing random noise that obeys Gaussian distribution into the image pixel values, so that the model can more easily learn the characteristics of real-world data, thereby improving the generalization ability; performing Gaussian blur processing, the degree of blurring of this operation is determined by the standard deviation parameter, which can reduce the details and noise in the image and help the model focus on the overall shape and structure of the object.

[0014] Finally, the training set and test set of the multi-task dataset for hallux valgus angle prediction and injury determination were randomly divided into two groups with a ratio of 8:2.

[0015] S3, convert key point labels into line segment labels and input them into the network;

[0016] First, before training, the two key points are connected into a straight line and input into the network. According to the hallux valgus key point label, the two key points at the beginning and end of the proximal phalanx midshaft are connected into a straight line, the two key point labels at the beginning and end of the first phalanx midshaft are connected into a line, and the two key point labels at the beginning and end of the second phalanx midshaft are connected into a line. The original six key point labels are connected into a straight line in groups of two to generate three line segment labels. In this way, the problem of single key point prediction is changed to the problem of foreground-background segmentation, and the final output is a line segment heat map instead of a key point heat map.

[0017] The sigmoid activation function is then used at the output layer to generate probabilities. Units with probabilities greater than 0.5 are considered foreground (corresponding to line segments), while others are considered background. This approach also makes the predictions robust by mitigating the imbalance between the foreground and background. More importantly, even if a few points are not predicted correctly, the remaining correctly predicted points can help fit the ideal linear equation through linear regression.

[0018] Finally, the neural network predicts three heatmaps from an input image and extracts the locations of points with a confidence score greater than 0.5. Linear regression is performed on these points to calculate the linear equation. With this linear equation, we can calculate the angles between the lines and visualize the results as the HVA and IMA angles.

[0019] S4. Build a multi-task network model for angle prediction and damage determination, and use labeled data to train the model.

[0020] First, a multi-task network is constructed, which contains two Hourglass network structures. The first Hourglass network performs downsampling. The downsampling stage includes four bottleneck blocks: F1, F2, F3, and F4. Through a series of convolutional layers and pooling layers, the spatial resolution of the feature map is gradually reduced. At the same time, the network increases the number of channels to achieve abstraction and extraction of image features, enabling the network to capture the key characteristics of hallux valgus.

[0021] The classification task is then completed by integrating a classification network at the end of the first hourglass network downsampling. The classification network consists of a CAF module, a FU module, and an MLP multi-layer perceptron. The CAF module uses a multi-head attention mechanism and a multi-scale convolution module to extract shape features for classification. First, a multi-scale convolution module uses three different convolution kernels to extract features at different scales. Three attention mechanisms are then applied to each of these modules to generate dynamic weights. A global average pooling layer and a concatenation layer are used to generate weights for each feature. These weights are then normalized by softmax to ensure their sum is 1. Features at each scale are weighted and then summed and fused to produce a dynamically weighted output. This allows the model to adaptively learn the importance of features at each scale, improving the flexibility of feature representation. Dynamic weighted fusion enables the model to better selectively focus on key features, which helps improve overall performance. This allows the model to adaptively adjust the importance of features at different scales. Finally, CBAM spatial attention is used to further extract spatial and channel features, achieving higher classification performance. The FU module uses residual connections and feature pyramid fusion to mitigate the vanishing gradient problem and improve classification stability. The FU module first performs a convolution on the input to reduce dimensionality and extract features. Next, batch normalization is applied to make the data distribution within each mini-batch consistent, thereby accelerating the convergence of network training. Then, the RELU activation function is applied to the batch normalized result to introduce nonlinear characteristics to enhance the model's expressiveness. In the convolution block of the second stage, the same operation is performed again to further improve the effect of feature extraction. Repeat this residual block four times. Finally, the convolution output is added element-by-element to the initial input through the residual connection. This residual connection mechanism can effectively alleviate the gradient vanishing problem in deep networks and accelerate the convergence of the model. The classification task is then completed through the MLP fully connected layer, and the hallux valgus injury judgment result is output. RELU is used as the activation function;

[0022] The RELU activation function is defined by formula (1) as:

[0023] f(x)=max(x,0) (1)

[0024] The first hourglass network then performs upsampling, including four bottleneck blocks: R1, R2, R3, and R4. This gradually increases the spatial resolution of the feature map, reduces the number of channels, and maps the abstract features extracted by the previous downsampling back to the original image size to generate a heatmap for angle prediction. Intermediate supervision is introduced between the two hourglass networks to further strengthen the relationship between the two tasks. The second hourglass network uses deformable convolutions in the upsampling stage to adapt to the shape characteristics of hallux valgus in different forms. Ultimately, at the end of the second hourglass, it outputs heatmap channels corresponding to the HVA and IMA angle predictions. Linear regression is then used to calculate the angle value to complete the angle prediction task.

[0025] S5. Construct a multi-task loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters;

[0026] First, initialize the model, load the dataset and...

[0027] Then, a multi-task loss function is constructed, which is the sum of binary cross entropy loss and sparse category cross entropy loss, and is defined by formula (2):

[0028] L=λ1×L angle +λ2×L class (2)

[0029] In formula (2), L angle is the binary cross entropy loss for angle prediction, and the segment heatmap for supervised angle prediction; L class is the sparse category cross entropy loss for damage judgment, and λ1 and λ2 are task weight parameters.

[0030] Finally, the gradient is calculated through the back-propagation algorithm and the optimizer is used to update the parameters of the model;

[0031] The optimizer is the RMSprop optimizer, and the initial learning rate is set to 5×10 -4 , the training rounds are 200 rounds and the batch size is 2; input the training set data, update the network parameters through the back propagation algorithm, and the parameter update strategy of the optimizer is defined by formula (3):

[0032]

[0033] In formula (3), θ t is the model parameter, the number of iterations is t, and the gradient is g t , introduce the γ attenuation factor to calculate the mean value of the historical gradient square E[g 2 ] t , ∈ is a very small constant, η is the initial learning rate used to control the overall update step size.

[0034] S6. Load the weight parameters trained above, input the test set data into the model, and obtain the HVA and IMA angles and injury determination results of hallux valgus;

[0035] First, load the trained model weight parameters, input the test set images into the model, and obtain the angle predicted line segment heat map and key damage features.

[0036] Then, the positions of points with confidence greater than 0.5 were extracted from the angle prediction line segment heat map. Linear regression was performed on these points to fit the axis of the proximal metatarsal bone and the first and second phalanges to obtain a linear equation. The specific angle values ​​of HVA and IMA were calculated through linear regression.

[0037] According to the heat map of the key injury area and the HVA and IMA angle measurements, the final hallux valgus HVA and IMA measurement values ​​and injury grade are given.

[0038] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0039] (1) Based on actual clinical foot imaging data, a multi-task dataset was created that can be used for hallux valgus angle prediction and injury assessment.

[0040] (2) A multi-task learning network for hallux valgus angle prediction and injury assessment is proposed, which can simultaneously predict the angle and classification results.

[0041] (3) Convert the angle prediction key point labels into line segment labels to improve the robustness of the prediction.

[0042] (4) Two hourglass structures are adopted, and intermediate supervision is introduced between the hourglasses, which takes into account the connectivity of the hallux valgus joints and further enhances the relationship between the two tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the present invention;

[0044] Figure 2 This is a schematic diagram of the multi-task network model framework of the present invention;

[0045] Figure 3 It is the attention weighted fusion module (CAF block) in the classification network model of the present invention;

[0046] Figure 4 It is the feature fusion module (FU block) in the classification network model of the present invention;

[0047] Figure 5Figure 3 shows the use of a multi-task network model for hallux valgus angle prediction and injury determination in an embodiment of the present invention, wherein (a) is the input hallux valgus foot image test picture, (b) is the corresponding key point label for hallux valgus angle prediction, (c) is the mapping of the predicted corresponding line segment heat map to the original image, (d) is the corresponding injury determination key feature heat map, and (e) is the HVA and IMA angle measurement and injury determination result diagram after linear regression of the corresponding line segment heat map. DETAILED DESCRIPTION

[0048] Reference Figures 1 to 5 The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0049] A research method for hallux valgus angle prediction and risk assessment based on multi-task network, such as Figure 1 As shown, the following steps are included:

[0050] Step 1: Collect medical images of hallux valgus feet and perform data preprocessing;

[0051] Specifically, the following steps are performed: collecting CT images of the subject's foot uploaded from a cloud server;

[0052] According to the actual clinical foot images collected, the hallux valgus affected side is identified based on the foot images to obtain the hallux valgus affected side of the subject, and the hallux valgus injury is divided into three levels: mild, moderate and severe.

[0053] Step 2: Label the multi-task labels corresponding to angle prediction and injury assessment, construct a multi-task dataset for hallux valgus angle prediction and injury assessment, and perform data enhancement and dataset partitioning.

[0054] The details are as follows:

[0055] Based on the HVA and IMA angles, the key points of the corresponding thumb joints are marked to generate a joint point image of the hallux valgus area, which is used as the label for hallux valgus angle prediction. Based on the hallux valgus injury grade, the corresponding three-category labels of light, medium, and severe are generated, with light labeled as 0, medium labeled as 1, and severe labeled as 2.

[0056] Multiple data augmentation methods were used to expand the sample size of the multi-task dataset for hallux valgus angle prediction and injury assessment. The data augmentation methods included mirror flipping along the vertical axis of the image, random rotation of -15° to +15°, random scaling of 0.8 to 1.2 times, shearing transformation, Gaussian noise, and Gaussian blurring.

[0057] The training set and test set of the multi-task dataset for angle prediction and damage determination are randomly divided into two parts with a ratio of 8:2.

[0058] Step 3: Convert key point labels into line segment labels and input them into the network;

[0059] First, before training, the two key points are connected into a straight line and input into the network. According to the hallux valgus key point label, the two key points at the beginning and end of the proximal phalanx midshaft are connected into a straight line, the two key point labels at the beginning and end of the first phalanx midshaft are connected into a line, and the two key point labels at the beginning and end of the second phalanx midshaft are connected into a line. The original six key point labels are connected into a straight line in groups of two to generate three line segment labels. In this way, the problem of single key point prediction is changed to the problem of foreground-background segmentation, and the final output is a line segment heat map instead of a key point heat map.

[0060] A sigmoid activation function is used at the output layer to generate probabilities. Units with probabilities greater than 0.5 are considered foreground (corresponding to line segments), while others are considered background. This approach also makes predictions robust by mitigating the imbalance between the foreground and background. More importantly, even if a few points are not predicted correctly, the remaining correctly predicted points can help fit the ideal linear equation through linear regression.

[0061] The neural network predicts three heatmaps from an input image and extracts the locations of points with a confidence score greater than 0.5. Linear regression is performed on these points to calculate the linear equation. With this linear equation, the angles between the lines can be calculated and visualized as the HVA and IMA angles.

[0062] Step 4: Build a multi-task network model for angle prediction and damage determination, and use labeled data to train the model;

[0063] The details are as follows:

[0064] Construct a multi-task network model for hallux valgus angle prediction and injury determination, such as Figure 2 As shown in the figure, it contains two Hourglass network structures. The first Hourglass performs downsampling first. The downsampling stage includes four bottleneck blocks: F1, F2, F3, and F4. Through a series of convolutional layers and pooling layers, the spatial resolution of the feature map is gradually reduced, while the number of channels is increased to achieve abstraction and extraction of image features, so that the network can capture the key features of hallux valgus at different scales.

[0065] At the end of the first hourglass network, the classification network is integrated to complete the classification task and output the damage judgment result. The classification network consists of the CAF module. Figure 3 As shown, and FU module as Figure 4The CAF module uses multi-scale convolution combined with a dynamic weighted attention mechanism to adaptively extract the shape features required for damage classification. The FU module is based on residual connection and feature pyramid fusion to alleviate the gradient vanishing problem and improve classification stability. MLP is a fully connected layer used to complete the classification task. The RELU formula (1) is used as the activation function.

[0066] f(x)=max(x,0) (1)

[0067] Upsampling is performed, including four bottleneck blocks (R1, R2, R3, and R4), to gradually increase the spatial resolution of the feature map while reducing the number of channels. The abstract features extracted by downsampling are then mapped back to the original image size to generate heatmaps for angle prediction. An intermediate supervision module is introduced between the two hourglass networks. This incorporation of intermediate supervision further enhances the underlying relationship between the two tasks. During the upsampling phase of the second hourglass network, deformable convolutions are added to the fusion process. These deformable convolutional layers provide offsets calculated by another parallel convolutional layer. This offset can also be learned end-to-end through gradient backpropagation. With this offset, the size and position of the deformable convolution kernel can be adjusted based on the requirements of dynamic image content recognition. The position of the convolution kernel varies based on the location of the sampling points, adapting to varying hallux valgus shapes as image content changes. Finally, at the end of the second hourglass network, heatmap channels corresponding to the HVA and IMA angle predictions are output. The angle value is then calculated through linear regression to complete the angle prediction task.

[0068] Step 5: Construct a multi-task loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters;

[0069] Specifically: initialize the model and load the dataset.

[0070] Construct a multi-task loss function, which is the sum of binary cross entropy loss and sparse category cross entropy loss, and is defined by formula (2):

[0071] L=λ1×L angle +λ2×L class (2)

[0072] In formula (2), L angle Binary cross entropy loss for angle prediction, supervised line segment label heatmap; L class is the sparse category cross entropy loss for damage judgment, and λ1 and λ2 are task weight parameters.

[0073] Calculate the gradient through the back-propagation algorithm and use the optimizer to update the model parameters;

[0074] The optimizer is the RMSprop optimizer, and the initial learning rate is set to 5×10 -4 , the training rounds are 200 rounds and the batch size is 2; input the training set data, update the network parameters through the back propagation algorithm, and the parameter update strategy of the optimizer is defined by formula (3):

[0075]

[0076] In formula (3), θ t is the model parameter, the number of iterations is t, and the gradient is g t , introduce the γ attenuation factor to calculate the mean value of the historical gradient square E[g 2 ] t ,∈ is a minimal constant, η is the initial learning rate used to control the overall update step size;

[0077] Step 6: Load the trained weight parameters, input the test set data into the model, and perform image post-processing on the output image to obtain the HVA and IMA angles of hallux valgus and the injury assessment results;

[0078] The details are as follows:

[0079] Load the trained model weight parameters, input the test set images into the model, and obtain the line segment heat map and injury judgment heat map corresponding to the two angles of hallux valgus HVA and IMA. Figure 5 As shown in the figure, (a) is the input hallux valgus foot image test picture, (b) is the corresponding key point label for hallux valgus angle prediction, (c) is the mapping of the predicted corresponding line segment heat map to the original image, (d) is the corresponding injury judgment key feature heat map, and (e) is the final hallux valgus HVA and IMA angle measurement and injury judgment result diagram after linear regression of the corresponding line segment heat map.

[0080] The angle prediction segment heat map was post-processed to extract the key points of the segment. The metatarsal and phalangeal axes were fitted by linear regression to calculate the specific angle values ​​of HVA and IMA.

[0081] Then, based on the predicted heat map of the key injury area, the HVA and IMA angles of hallux valgus and the injury judgment results are output.

[0082] The above description is merely a preferred embodiment of the present disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A research method for hallux valgus angle prediction and risk assessment based on multi-task network, characterized by: The following steps are involved: S1. Collect medical images of hallux valgus feet and classify them according to the injury grade; S2. Label the key points of hallux valgus HVA (hallux valgus angle) and IMA (inter-metatarsal angle) and injury grade labels, and perform data augmentation and data set partitioning on the dataset. S3, convert key point labels into line segment labels; S4. Construct a multi-task network model for angle prediction and damage determination; S5. Construct a multi-task loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters; S6. Load the weight parameters trained above, input the test set data into the model, and obtain the HVA and IMA angles and injury judgment results of hallux valgus.

2. The method for hallux valgus angle prediction and risk assessment based on a multi-task network according to claim 1, characterized in that: Step S1 includes the following: Collect the patient's foot CT images uploaded from the cloud server; According to the actual clinical foot images collected, the hallux valgus affected side is identified based on the foot images to obtain the hallux valgus affected side of the subject, and the hallux valgus injury is divided into three levels: mild, moderate and severe.

3. The method for hallux valgus angle prediction and wind direction assessment based on a multi-task network according to claim 2, characterized in that: Wherein step S2 comprises the following: First, the key points of the thumb joint are marked based on the HVA and IMA angles to generate a joint point image of the hallux valgus area. This image is used as the label for hallux valgus angle prediction. Based on the injury grade, mild is marked as 0, moderate is marked as 1, and severe is marked as 2 to generate a label for hallux valgus injury determination. Then, a variety of data augmentation methods were used to expand the number of samples in the multi-task dataset for hallux valgus angle prediction and injury assessment. The data augmentation methods included mirror flipping along the vertical axis of the image, random rotation of -15° to +15°, random scaling of 0.8 to 1.2 times, shearing transformation, Gaussian noise, and Gaussian blur. Finally, the training set and test set of the multi-task dataset for angle prediction and damage determination are randomly divided in the ratio of 8:2 in the number of images.

4. The method for hallux valgus angle prediction and risk assessment based on a multi-task network according to claim 3, characterized in that: Wherein step S3 comprises the following: According to the hallux valgus key point labels, the two key points at the beginning and end of the midshaft of the proximal phalanx are connected into a straight line, the two key point labels at the beginning and end of the midshaft of the first phalanx are connected into a line, and the two key point labels at the beginning and end of the midshaft of the second phalanx are connected into a line. The original six key point labels are connected into a straight line in groups of two to generate three line segment labels. In this way, the problem of single key point prediction is changed to the problem of foreground and background segmentation, and a line segment heat map is predicted instead of a key point heat map.

5. The method for hallux valgus angle prediction and risk assessment based on a multi-task network according to claim 4, characterized in that: Wherein step S4 comprises the following: First, a multi-task network is constructed, which consists of two Hourglass networks. The first Hourglass network performs downsampling. During the downsampling phase, a series of convolutional and pooling layers are used to gradually reduce the spatial resolution of the feature map while increasing the number of channels. This allows the network to abstract and extract image features, enabling it to capture the key characteristics of hallux valgus at different scales. Then, the classification network is integrated at the end of the first hourglass network downsampling to complete the classification task. The classification network consists of a CAF module and a FU module. The CAF module uses multi-scale convolution combined with a dynamic weighted attention mechanism to adaptively extract the shape features required for damage classification. The FU module is based on residual connection and feature pyramid fusion to alleviate the gradient vanishing problem and improve classification stability. It then outputs the corresponding damage judgment result, using the RELU formula (1) as the activation function. f(x)=max(x,0) (1) Then upsampling is performed to gradually expand the spatial resolution of the feature map, reduce the number of channels, and map the abstract features extracted by downsampling back to the original image size to generate a heat map for angle prediction; Intermediate supervision is introduced between the two Hourglass networks to enhance the relationship between the two tasks. A deformable convolution module is used in the sampling stage of the second Hourglass network to adapt to the shape features of hallux valgus of different forms.

6. The method for hallux valgus angle prediction and risk assessment based on a multi-task network according to claim 5, characterized in that: Wherein step S5 comprises the following: First, initialize the model and load the dataset; Then, a multi-task loss function is constructed, which is the sum of binary cross entropy loss and sparse category cross entropy loss, and is defined by formula (2): L=λ1×L angle +λ2×L class (2) In formula (2), L angle is the binary cross entropy loss for angle prediction, and the segment heatmap for supervised angle prediction; L class is the sparse category cross entropy loss for damage judgment, λ1 and λ2 are task weight parameters; The optimizer is the RMSprop optimizer, and the initial learning rate is set to 5×10 -4 , the training rounds are 200 rounds and the batch size is 2; input the training set data, update the network parameters through the back propagation algorithm, and the parameter update strategy of the optimizer is defined by formula (3): In formula (3), θ t is the model parameter, the number of iterations is t, and the gradient is g t , introduce the γ attenuation factor to calculate the mean value of the historical gradient square E[g 2 ] t , ∈ is a very small constant, η is the initial learning rate used to control the overall update step size.

7. The method for hallux valgus angle prediction and risk assessment based on a multi-task network according to claim 6, characterized in that: Wherein step S6 comprises the following: First, load the trained model weight parameters, input the test set images into the model, and obtain the angle predicted line segment heat map and key damage features; Then, the positions of points with a confidence level greater than 0.5 are extracted from the angle prediction line segment heat map. These points are then fitted with the axis of the proximal metatarsal bone and the first and second phalanges through linear regression to obtain a linear equation. The specific angle values ​​of HVA and IMA are calculated through linear regression. Finally, based on the characteristics of the key injury area and the angle measurement results, the final hallux valgus HVA and IMA measurement values ​​and injury judgment level are given.

Images