Scoliosis Cobb angle automatic evaluation method
Through deep learning technology, medical image segmentation network and regression model are constructed to realize automatic evaluation of preoperative Cobb angles in scoliosis patients, solving the problem that traditional measurements rely on doctors' experience and improving diagnostic efficiency and accuracy.
Patent Information
- Application Number
- CN202411470217.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-21
- Publication Date
- 2025-01-17
AI Technical Summary
Cobb angle measurements of traditional scoliosis rely on physician experience, are time-consuming and limited in accuracy, making it difficult to achieve efficient diagnosis and treatment.
Using a deep learning technology-based method, the segmentation of spinal boundary lines and automatic evaluation of Cobb angles in spinal X-rays is realized by constructing a medical image segmentation network based on diffusion probability model and a DenseNet regression model integrating SENet module, and using data enhancement Strategy optimizes the training dataset and improves evaluation accuracy.
Automatic evaluation of preoperative Cobb angle in patients with scoliosis is achieved, which improves diagnostic efficiency and accuracy, reduces artificial errors, and improves the level of scoliosis treatment.
Smart Images
Figure CN119323558A_ABST
Abstract
Description
[Technical field]
[0001] The invention relates to a method for automatically evaluating the Cobb angle of scoliosis, and belongs to the field of medical image processing. [Background technology]
[0002] Scoliosis is a complex three-dimensional spinal deformity that is most common in adolescents. Accurate assessment of spinal curvature is crucial for the treatment of scoliosis. Currently, Cobb angle measurement is mainly used as a clinical evaluation standard. Scoliosis presents an "S" or "C" shaped curve on a standard full-length anteroposterior phase X-ray. Through X-rays, doctors manually measure the Cobb angle to quantitatively assess the severity of the patient's condition and formulate a treatment plan, but this method relies on personal experience, is time-consuming and has limited accuracy. With the development of deep learning technology, medical image processing has achieved remarkable results. Researchers have begun to focus on deep convolutional neural networks to achieve end-to-end automatic Cobb angle assessment, aiming to improve measurement accuracy and reduce manual errors. Therefore, how to achieve accurate assessment of Cobb in preoperative spinal X-rays based on deep learning technology is an important means to improve the level of scoliosis treatment. [Summary of the invention]
[0003] In response to the above technical problems, the present invention provides a method for automatic assessment of the Cobb angle of scoliosis. The method can realize automatic assessment of the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) based on the preoperative spinal X-ray of scoliosis patients. This method is of great significance for solving the problem that traditional Cobb angle measurement relies on the doctor's experience and the operation is time-consuming, and improving the efficiency of scoliosis diagnosis.
[0004] The present invention provides a method for automatically evaluating the Cobb angle of scoliosis, comprising:
[0005] Based on the public scoliosis X-ray dataset, rough annotation of the spine boundary line is obtained through data preprocessing;
[0006] Based on the constructed scoliosis X-ray dataset with annotated spine boundary lines, a medical image segmentation network based on the diffusion probability model is constructed, and the segmentation of the spine boundary lines in the spine X-rays is achieved through iterative optimization of network training;
[0007] According to the spine boundary lines obtained by training reasoning, a data enhancement strategy is designed to expand the training data set to obtain the spine boundary line data set after data enhancement;
[0008] Based on the data-enhanced spine boundary data set, a DenseNet regression model fused with the SENet module was constructed, and the automatic evaluation of the Cobb angle of scoliosis was achieved through network training.
[0009] Based on the Cobb angle gold standard annotated by doctors on spinal X-rays and the Cobb angle predicted by the network, the effectiveness of the Cobb angle evaluation method was verified by evaluating the symmetric mean absolute percentage error (SMAPE) and Manhattan distance (L1 distance).
[0010] The automatic evaluation method of the Cobb angle of scoliosis is preferably based on a public scoliosis X-ray data set, and a rough annotation of the spine boundary line is obtained through data preprocessing. The specific process is as follows:
[0011] According to the 2019 MICCAIAASCE competition public data set, we obtained the patient's spinal X-rays, the four corner points of each spine annotated by experts, and the corresponding Cobb angles. We performed data preprocessing steps such as spine boundary line connection, data enhancement, and data partitioning on the data set to reconstruct the scoliosis X-ray data set with annotated spine boundary lines.
[0012] The method for automatically assessing the Cobb angle of scoliosis preferably constructs a medical image segmentation network based on a diffusion probability model based on a reconstructed scoliosis X-ray data set with annotated spine boundaries, and implements the segmentation of spine boundaries in spinal X-rays through network training. The specific process is as follows:
[0013] Based on the reconstructed scoliosis X-ray dataset with annotated spine boundaries, a medical image segmentation network based on a diffusion probability model is constructed. The generative model includes two stages: forward diffusion and reverse diffusion. In the forward diffusion stage, the segmentation label x0 introduces Gaussian noise through forward diffusion, which can be expressed as:
[0014] p α (x T )=N(x T ;0,I n×n )
[0015] Among them, x T is the segmentation label after adding Gaussian noise to the initial segmentation label x0 at time step T, I n×n is the original image.
[0016] In the reverse denoising stage, the neural network is trained to remove noise, thereby generating the segmentation result of the spine boundary line, which can be expressed as:
[0017]
[0018] Among them, α is the relevant parameter for constructing the neural network in the reverse process, that is, the weights and bias terms of iterative learning in network training, t is the time step, and T is the total number of time steps in which noise is gradually added during the forward diffusion process.
[0019] Through steps such as forward diffusion and reverse denoising, and through iterative optimization of network training, the trained network model parameters that can segment the spine boundary lines in spinal X-rays are obtained, which can generalize the segmentation of the spine boundary lines in spinal X-rays taken by patients with scoliosis.
[0020] The automatic evaluation method of the Cobb angle of scoliosis preferably designs a data enhancement strategy based on the spine boundary line obtained by training reasoning, expands the training data set, and obtains the spine boundary line data set after data enhancement. The specific process is as follows:
[0021] According to the spine boundary line image data segmented from the spine X-ray data set, two data enhancement strategies, random scaling and random rotation, are applied to amplify the spine boundary line image data, thereby obtaining a data-enhanced spine boundary line data set.
[0022] Random scaling involves scaling the image proportionally. Taking a point (x, y) on the image as an example, the scaled point (x', y') can be expressed as:
[0023] x′=s·x
[0024] y′=s·y
[0025] Where s is the scaling factor.
[0026] Random rotation involves rotating the image around the center point of the image by an angle θ. Taking the point (x, y) on the image as an example, assuming that the center point of the image is (x c ,y c ), then the rotated point (x',y') can be expressed as:
[0027]
[0028] The method for automatically evaluating the Cobb angle of scoliosis preferably constructs a DenseNet regression model fused with a SENet module based on the data-enhanced spine boundary line dataset, and realizes automatic evaluation of the Cobb angle of scoliosis through network training. The specific process is as follows:
[0029] Based on the data-enhanced spine boundary line dataset, a DenseNet regression model fused with the SENet module was constructed. After training and fitting parameters, the network model parameters of the Cobb angle can be directly regressed according to the spine boundary line after training, thereby realizing automatic evaluation of the Cobb angle of scoliosis. Therefore, the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) can be predicted and inferred based on the spine boundary line images of scoliosis X-rays.
[0030] The automatic evaluation method for the Cobb angle of scoliosis is preferably based on the Cobb angle gold standard annotated by the doctor on the spinal X-ray film and the Cobb angle predicted by the network, and is evaluated by two evaluation indicators, namely, the symmetric mean absolute percentage error (SMAPE) and the Manhattan distance (L1 distance), to verify the effectiveness of the Cobb angle evaluation method. The specific process is as follows:
[0031] The Cobb angle predicted by the regression network based on the test set is compared with the gold standard of the Cobb angle annotated by the corresponding doctor. The symmetric mean absolute percentage error (SMAPE) and L1 distance (also known as Manhattan distance) are used to evaluate the gap between the predicted result and the annotated Cobb, thereby verifying the overall effectiveness of the automatic Cobb angle evaluation method.
[0032] The calculation method of SMAPE is as follows:
[0033]
[0034] Among them, A is the Cobb angle marked by experts, B is the Cobb angle obtained by regression prediction, and N is the amount of test set data.
[0035] The calculation method of L1 distance is as follows:
[0036]
[0037] in, is the Cobb angle predicted by regression, y i is the Cobb angle annotated by the expert, and N is the number of test sets.
[0038] The present invention adopts the above technical solution, which has the following advantages:
[0039] The Cobb automatic assessment method of the present invention can realize automatic assessment of the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) based on the preoperative X-ray images of scoliosis patients by constructing an end-to-end deep learning neural network model, which can effectively improve the level of scoliosis treatment. It is of great significance to solve the problem that traditional Cobb angle measurement relies on doctor's experience and is time-consuming to operate, and to improve the efficiency of scoliosis diagnosis.
Brief Description of the Drawings
[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It is obvious that the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creativity and labor.
[0041] Figure 1 It is a schematic diagram of a process flow of a method for automatically evaluating the Cobb angle of scoliosis provided by an example of the present invention;
[0042] Figure 2 : is a result diagram of the reconstructed spine boundary line dataset provided by an example of the present invention, wherein (a) is a schematic diagram of the three Cobb angles of the main thoracic (MT), the proximal thoracic (PT) and the thoracolumbar / lumbar (TL / L), (b) is the four corner points of the spine annotated in the original dataset, and (c) is a rough annotation of the spine boundary line obtained by connecting the four corner points;
[0043] Figure 3 It is a medical image segmentation network model constructed by segmenting the spine boundary line based on the spinal X-ray film of the present invention;
[0044] Figure 4 is a schematic diagram of a spine boundary line data set after data enhancement provided by an example of the present invention;
[0045] Figure 5 It is a schematic diagram of the dense layer of DenseNet integrated into SENet provided by the example of the present invention;
[0046] Figure 6 Schematic diagram of densely connected blocks of DenseNet provided by the example of the present invention;
[0047] Figure 7 It is a schematic diagram of the network structure of DenseNet provided by an example of the present invention; [Specific embodiment]
[0048] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work are within the scope of protection of the present invention.
[0049] At present, the Cobb angle measurement of traditional scoliosis mainly relies on the manual calculation of doctors, which leads to its extreme dependence on the doctor's experience, time-consuming operation, and limited accuracy of the measurement results. In recent years, with the development of deep learning technology, medical image processing has achieved remarkable results. In order to improve the calculation accuracy of Cobb and eliminate the influence of manual measurement, a lot of work in recent years has focused on the end-to-end automatic evaluation of Cobb angles based on deep convolutional neural networks. On this basis, the present invention provides an automatic evaluation method of Cobb in preoperative spinal X-rays based on deep learning technology, which can realize the automatic evaluation of the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) on the basis of the preoperative X-ray images of scoliosis patients, which can effectively improve the level of scoliosis treatment, and is of great significance for solving the problem that the traditional Cobb angle measurement depends on the doctor's experience and the operation is time-consuming, and improving the calculation accuracy.
[0050] like Figure 1 As shown, the automatic evaluation method of Cobb angle of scoliosis involved in the present invention comprises the following steps:
[0051] Step 101 : obtaining a rough annotation of the spine boundary line through data preprocessing based on a public spine X-ray dataset.
[0052] According to the 2019 MICCAIAASCE competition public data set, the patient's spinal X-ray, the Cobb angle calculated by experts, and the coordinates of the annotated spine corners were obtained. The three Cobb angle measurement methods of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) are as follows: Figure 2 As shown in (a), the four corner points of each spine are labeled as follows Figure 2 (b) shows the data set. The data is preprocessed by first converting the data and connecting the four corner points of each spine to obtain a rough annotation of the spine boundary line, as shown in Figure 2 (c) As shown; secondly, data cleaning is performed to remove low-quality data and normalize the data; finally, data division is performed, where the training set contains 431 X-rays, the validation set contains 50 X-rays, and the test set contains 128 X-rays.
[0053] Step 102, based on the constructed scoliosis X-ray film dataset with annotated spine boundary lines, a medical image segmentation network based on a diffusion probability model is constructed, and the segmentation of the spine boundary lines in the spine X-ray film is achieved through iterative optimization of network training.
[0054] According to the reconstructed scoliosis X-ray dataset with annotated spine boundaries, a medical image segmentation network based on MedSegDiff is constructed with rough annotations of spine boundaries as supervisory labels. The generative model includes two stages: forward diffusion and reverse diffusion. In the forward diffusion stage, the segmentation label x0 introduces Gaussian noise through forward diffusion, which can be expressed as:
[0055] p α (x T )=N(x T ;0,I n×n )
[0056] Among them, x T is the segmentation label after adding Gaussian noise to the initial segmentation label x0 at time step T, I n×n is the original image.
[0057] In the reverse denoising stage, the neural network is trained to remove noise, thereby generating the segmentation result of the spine boundary line, which can be expressed as:
[0058]
[0059] Among them, α is the relevant parameter of building the neural network in the reverse process, that is, the weight and bias term of iterative learning in network training, t is the time step, and T is the total number of time steps in which noise is gradually added during the forward diffusion process. The reverse process gradually restores the noisy image so that the latent variable distribution p α (x T ) is transformed into data distribution p α (x0), thus obtaining a clear segmentation.
[0060] Based on the basic structure of the diffusion probability model, UNet is used as the learning network. In order to solve the problem that the low contrast of medical images makes it difficult to learn features, dynamic conditional coding is used to fuse the current cloth segmentation information with the original image information at the feature level. Specifically, in the original image encoder, the current cloth segmentation map features are integrated into the intermediate features of the original image encoder, such as Figure 3 As shown, the class attention mechanism A is used to convert the conditional features With x T The encoding features The fusion can be expressed as:
[0061]
[0062] in, represents element-wise multiplication, LN represents layer normalization, and k represents the number of layers.
[0063] FF-Parser is connected to the path of feature integration, such as Figure 3As shown, it is used to constrain x T The noise-related components in the features eliminate the negative impact of high-frequency components.
[0064] The main architecture of MedSegDiff is ResNet encoder and UNet decoder, such as Figure 3 As shown, I n×n and x T It is encoded by two separate encoders, and after steps such as forward diffusion and reverse denoising, it is iteratively optimized through network training to obtain the trained network model parameters that can segment the spine boundary lines in spinal X-rays. It can generalize the segmentation of the spine boundary lines in spinal X-rays taken by patients with scoliosis.
[0065] Step 103, design a data enhancement strategy based on the spine boundary lines obtained by training reasoning, expand the training data set, and obtain a spine boundary line data set after data enhancement. The specific process is as follows:
[0066] First, for the preoperative spinal X-ray dataset, the spine boundary lines corresponding to the scoliosis patient X-ray dataset are obtained after inference and prediction by the trained MedSegDiff medical image segmentation network, such as Figure 4 (a) shows that the two data enhancement strategies of random scaling and random rotation are used to amplify the spine boundary line image data, thereby obtaining the data enhanced spine boundary line dataset.
[0067] Random scaling involves scaling the image proportionally, e.g. Figure 4 (b) As shown in Figure 2, taking the point (x, y) on the image as an example, the scaled point (x', y') can be expressed as:
[0068] x′=s·x
[0069] y′=s·y
[0070] Wherein, s is the scaling factor, and the scaling factor range used in the example of the present invention is (0.85 to 1.25).
[0071] Random rotation involves rotating the image by an angle θ around the center point of the image, such as Figure 4 (c) As shown in Figure 2. Taking the point (x, y) on the image as an example, assuming that the center point of the image is (x c ,y c ), then the rotated point (x',y') can be expressed as:
[0072]
[0073] The image rotation angle θ used in the embodiment of the present invention ranges from (-45° to 45°).
[0074] Step 104, based on the data-enhanced spine boundary line dataset, a DenseNet regression model fused with the SENet module is constructed, and the automatic evaluation of the Cobb angle of scoliosis is achieved through network training. The specific process is as follows:
[0075] According to the data-enhanced spine boundary data set, a regression network model based on DenseNet169 is constructed, and the SENet channel attention mechanism is inserted, such as Figure 5 As shown in the figure, SENet is inserted into the dense layer of DenseNet, thus serving as the dense connection layer SE_DenseLayer of DenseNet. Specifically, SENet uses global average pooling to compress the feature map of each channel into a single feature vector; then the FC1 fully connected layer is used to achieve dimensionality reduction, thereby reducing the number of parameters; then the Swish nonlinear activation function is used to activate the feature vector, thereby increasing the nonlinear expression ability of the model; then the feature vector dimension is restored through FC2 so that each channel has a corresponding weight; then the FC2 output is normalized using the Sigmoid activation function; finally, the weight vector output by the Sigmoid function is multiplied by the original input feature map channel by channel, thereby achieving feature recalibration, so that the network can adaptively adjust the importance of the channel in the feature map, which can be expressed as:
[0076] F scale =σ(W2δ(W1z))ΘF
[0077] Among them, F is the original input feature map, z is the feature vector after global average pooling, W1 and W2 are the weights of the two fully connected layers FC1 and FC2, δ is the nonlinear activation function, σ is the Sigmoid activation function, and Θ represents element-by-element multiplication. As a result, each channel in the original feature map is weighted by a weight ranging from [0,1], thereby enhancing important features and suppressing unimportant features.
[0078] The dense connection layer SE_DenseLayer is an important part of the dense block SE_DenseBlock, such as Figure 6 As shown, dense connections are used between layers to achieve feature reuse, which can be expressed as:
[0079] X i =H i ([X0,X1,X2,···,X i-1 ])
[0080] Among them, H i(·) represents a series of nonlinear transformations of the densely connected layer SE_DenseLayer. From the above formula, we can see that the dense block realizes the direct connection from any layer to the subsequent layer. The current feature map is the result of the concatenation of the feature maps of all previous layers. Due to its dense connection, the image features can be deeply extracted and effectively reused.
[0081] The main structure of the constructed regression network model is the DenseNet network architecture. Take a DenseNet containing two dense blocks (SE_DenseBlock) as an example. Figure 7 As shown in the figure, after the original image initially obtains the feature map through convolution, it undergoes multiple dense block feature extractions and transition layer convolution pooling to achieve deep feature extraction, and finally inputs into a fully connected layer through global average pooling to obtain the Cobb angle regression result.
[0082] According to the constructed regression network model, the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) are predicted and inferred based on the spinal boundary line image of scoliosis X-ray.
[0083] Step 105, based on the Cobb angle gold standard annotated by the doctor on the spinal X-ray film and the Cobb angle predicted by the network, the effectiveness of the Cobb angle evaluation method is verified by evaluating the two evaluation indicators of symmetric mean absolute percentage error (SMAPE) and Manhattan distance (L1 distance).
[0084] The Cobb angle predicted by the regression network based on the test set was compared with the gold standard of Cobb angle annotated by the corresponding doctor. The symmetric mean absolute percentage error (SMAPE) and Manhattan distance (L1 distance) were used to evaluate the gap between the predicted result and the annotated Cobb. The final experimental evaluation SMAPE was 30.89, and the L1 distances were 4.12, 5.15, and 5.54, which verified that the automatic Cobb angle evaluation method designed in this project is effective.
[0085] The calculation method of SMAPE is as follows:
[0086]
[0087] Among them, A is the Cobb angle marked by the expert, B is the Cobb angle obtained by regression prediction, and N is the number of test sets.
[0088] The calculation method of L1 distance is as follows:
[0089]
[0090] in, is the Cobb angle predicted by regression, y i is the Cobb angle annotated by the expert, and N is the number of test sets.
[0091] The technical solution provided by the present invention can realize the automatic evaluation of the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) based on the preoperative spinal X-ray of scoliosis patients. It is of great significance to solve the problem that traditional Cobb angle measurement relies on doctor's experience and the operation is time-consuming, and to improve the efficiency of scoliosis diagnosis.
[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
[0093] The contents not described in detail in the specification of the present invention belong to the common knowledge of those skilled in the art.
Claims
1. A method for automatically evaluating the Cobb angle of scoliosis, characterized in that: include: Based on the public scoliosis X-ray dataset, rough annotation of the spine boundary line is obtained through data preprocessing; Based on the constructed scoliosis X-ray dataset with annotated spine boundary lines, a medical image segmentation network based on the diffusion probability model is constructed, and the segmentation of the spine boundary lines in the spine X-rays is achieved through iterative optimization of network training; According to the spine boundary lines obtained by training reasoning, a data enhancement strategy is designed to expand the training data set to obtain the spine boundary line data set after data enhancement; Based on the data-enhanced spine boundary data set, a DenseNet regression model fused with the SENet module was constructed, and the automatic evaluation of the Cobb angle of scoliosis was achieved through network training. According to the gold standard of Cobb angle annotated by doctors on spinal X-rays and the Cobb angle predicted by the network, the effectiveness of the Cobb angle evaluation method was verified by evaluating the symmetric mean absolute percentage error (SMAPE) and Manhattan distance (L1 distance).
2. The method for automatically evaluating the Cobb angle of scoliosis according to claim 1, characterized in that: Based on the public scoliosis X-ray dataset, rough annotation of the spine boundary line is obtained through data preprocessing. The specific process is as follows: According to the 2019 MICCAIAASCE competition public data set, obtain the patient's spinal X-ray, the four corner points of each vertebra annotated by experts, and the corresponding Cobb angle; The dataset was subjected to data preprocessing steps such as spine boundary line connection, data enhancement, and data partitioning to reconstruct the scoliosis X-ray dataset with annotated spine boundary lines.
3. The method for automatically evaluating the Cobb angle of scoliosis according to claim 1, characterized in that: Based on the reconstructed scoliosis X-ray dataset with annotated spine boundary lines, a medical image segmentation network based on the diffusion probability model is constructed. After network training, the segmentation of spine boundary lines in spinal X-rays is achieved. The specific process is as follows: Based on the reconstructed scoliosis X-ray dataset with annotated spine boundaries, a medical image segmentation network based on a diffusion probability model was constructed. The generative model includes two stages: forward diffusion and reverse diffusion. In the forward diffusion stage, the segmentation label x0 introduces Gaussian noise through forward diffusion, which can be expressed as: p α (x T )=N(x T ;0,I n×n ) Among them, x T is the segmentation label after adding Gaussian noise to the initial segmentation label x0 at time step T, I n×n is the original image; In the reverse denoising stage, the neural network is trained to remove noise, thereby generating the segmentation result of the spine boundary line, which can be expressed as: Among them, α is the relevant parameter for constructing the neural network in the reverse process, that is, the weight and bias term of iterative learning in network training, t is the time step, and T is the total number of time steps in which noise is gradually added during the forward diffusion process; Through steps such as forward diffusion and reverse denoising, and through iterative optimization of network training, the trained network model parameters that can segment the spine boundary lines in spinal X-rays are obtained, which can generalize the segmentation of the spine boundary lines in spinal X-rays taken by patients with scoliosis.
4. The method for automatically evaluating the Cobb angle of scoliosis according to claim 1, characterized in that: According to the spine boundary lines obtained by training reasoning, a data enhancement strategy is designed to expand the training data set and obtain the spine boundary line data set after data enhancement. The specific process is as follows: According to the spine boundary line image data segmented from the spine X-ray data set, two data enhancement strategies, random scaling and random rotation, are applied to amplify the spine boundary line image data, thereby obtaining a data-enhanced spine boundary line data set; Random scaling involves scaling the image proportionally. Taking a point (x, y) on the image as an example, the scaled point (x', y') can be expressed as: x′=s·x y′=s·y Where s is the scaling factor; Random rotation involves rotating the image around the center point of the image by an angle θ. Taking the point (x, y) on the image as an example, assuming that the center point of the image is (x c ,y c ), then the rotated point (x',y') can be expressed as:
5. The method for automatically evaluating the Cobb angle of scoliosis according to claim 1, characterized in that: Based on the data-enhanced spine boundary data set, a DenseNet regression model fused with the SENet module was constructed. After network training, the automatic evaluation of the Cobb angle of scoliosis was achieved. The specific process is as follows: Based on the data-enhanced spine boundary line dataset, a DenseNet regression model fused with the SENet module was constructed. After training and fitting parameters, the network model parameters of the Cobb angle can be directly regressed according to the spine boundary line after training, thereby realizing automatic evaluation of the Cobb angle of scoliosis. Therefore, the three Cobb angles of main thoracic (MT), proximal thoracic (PT) and thoracolumbar / lumbar (TL / L) can be predicted and inferred based on the spine boundary line images of scoliosis X-rays.
6. The method for automatically evaluating the Cobb angle of scoliosis according to claim 1, characterized in that: According to the Cobb angle gold standard annotated by doctors on spinal X-rays and the Cobb angle predicted by the network, the effectiveness of the Cobb angle evaluation method was verified by evaluating the symmetric mean absolute percentage error (SMAPE) and Manhattan distance (L1 distance). The specific process is as follows: The Cobb angle predicted by the regression network based on the test set was compared with the gold standard Cobb angle annotated by the corresponding doctor. The Symmetric Mean Absolute Percentage Error (SMAPE) and Manhattan Distance (L1 Distance) were used to evaluate the difference between the predicted result and the annotated Cobb, thereby verifying the overall effectiveness of the Cobb angle automatic evaluation method. The calculation method of SMAPE is as follows: Among them, A is the Cobb angle annotated by the expert, B is the Cobb angle obtained by regression prediction, and N is the amount of test set data; The calculation method of L1 distance is as follows: in, is the Cobb angle predicted by regression, y i is the Cobb angle annotated by the expert, and N is the number of test sets.