A scoliosis detection method and system based on multi-subdomain classification algorithm

Abstract

The present invention proposes a scoliosis detection method and system based on a multi-subdomain classification algorithm. Specifically, the following steps are: obtaining spinal image samples and preprocessing them, wherein the preprocessing includes encoding or quantifying clinical information; extracting comprehensive feature vectors of spinal image samples; clustering the spinal image samples and dividing them into multiple subdomains; constructing a classification model for each subdomain according to the number of samples in each subdomain and the distribution of comprehensive feature vectors; using the samples of each subdomain to train the corresponding classification model; extracting the comprehensive feature vector of the predicted sample, calculating the similarity between the comprehensive feature vector of the predicted sample and the comprehensive feature vector of each subdomain, selecting the classification model corresponding to the subdomain with the highest similarity to classify and predict the predicted sample, and predicting the classification of the predicted sample. The present invention improves the accuracy of classification in scoliosis detection, and the generalization ability is also significantly enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of scoliosis detection, and in particular to a scoliosis detection method and system based on a multi-subdomain classification algorithm. Background Art

[0002] Scoliosis is a three-dimensional deformity of the spine, including abnormalities in coronal, sagittal, and axial alignment. Early diagnosis and classification are essential to develop appropriate treatment plans and prevent progression of the disease.

[0003] In recent years, with the widespread application of digital medical imaging technology, a large amount of spinal imaging data (such as X-rays, CT scans, MRI, etc.) has been obtained, providing rich data resources for the automatic classification of scoliosis based on machine learning and computer vision technology.

[0004] At present, the classification methods of scoliosis mainly include those based on traditional machine learning algorithms and those based on deep learning algorithms. Methods based on traditional machine learning algorithms, such as support vector machines (SVM), require manual extraction of features, such as the curvature angle of the spine, the rotation angle of the vertebrae, etc., and then input these features into the classifier for classification. For example, a feature vector is constructed by measuring the angle of a specific vertebra on a spinal X-ray image, and then SVM is used to find an optimal hyperplane to separate normal, abnormal posture, and scoliosis samples. Methods based on deep learning algorithms, such as convolutional neural networks (CNN), can automatically learn features from image data. The imaging data of the spine (such as X-rays, CT scans, etc.) is directly input into CNN, and the network automatically extracts features from the image through multiple convolutional layers and pooling layers, and finally performs classification through a fully connected layer.

[0005] The shortcomings of existing classification methods:

[0006] Limitations of feature extraction: Traditional machine learning algorithms rely on manually designed features, which may not fully express the complex information of scoliosis, resulting in limited classification accuracy. For example, for some subtle changes in spinal morphology, manually extracted features are difficult to accurately capture.

[0007] Insufficient adaptation to data diversity: Both traditional machine learning algorithms and deep learning algorithms have poor generalization capabilities when faced with data of different types and degrees of scoliosis and large individual differences. A single model is difficult to adapt to the diverse feature space distribution and is prone to overfitting or underfitting.

[0008] High computing resource requirements: Although deep learning algorithms can automatically learn features, they require a large amount of training data and computing resources. For large-scale scoliosis imaging data, the training process takes a long time and has high requirements for hardware equipment. Summary of the invention

[0009] In order to overcome the defects in the above-mentioned prior art, the object of the present invention is to provide a scoliosis detection method and system based on a multi-subdomain classification algorithm.

[0010] In order to achieve the above-mentioned object of the present invention, the present invention provides a scoliosis detection method based on a multi-subdomain classification algorithm, comprising the following steps:

[0011] Acquire spinal image samples, and label and preprocess the spinal image samples;

[0012] Extracting comprehensive feature vectors of preprocessed spinal image samples;

[0013] Clustering the spinal image samples according to scoliosis categories to divide them into multiple subdomains;

[0014] A classification model is constructed for each subdomain based on the number of samples and the distribution of comprehensive feature vectors in each subdomain;

[0015] Use samples from each subdomain to train its corresponding classification model;

[0016] The comprehensive feature vector of the predicted sample is extracted, the similarity between the comprehensive feature vector of the predicted sample and the comprehensive feature vector of each subdomain is calculated, and the classification model corresponding to the subdomain with the highest similarity is selected to perform classification prediction on the predicted sample to obtain the classification of the predicted sample.

[0017] This method divides the spinal image samples into multiple subdomains and makes full use of the characteristic distribution information of the data, which effectively improves the classification accuracy of scoliosis detection. It has strong generalization ability in the prediction process and reduces the demand for computing resources.

[0018] Optionally, preprocessing includes coding, standardizing, or quantifying clinical information;

[0019] When extracting the comprehensive feature vector, the image data feature vector is obtained from the image data; the image data is input into a pre-trained deep learning model to extract a deep feature vector that can capture the high-level semantic information of the image; the encoded or quantized clinical information is combined to form an auxiliary feature vector; then the image data feature vector is spliced ​​with the deep feature vector, and further spliced ​​with the auxiliary feature vector to obtain a comprehensive feature vector for each sample.

[0020] This optional solution can effectively integrate imaging data and clinical information data, extract representative comprehensive feature vectors, fully reflect the various feature information of scoliosis, and solve the problem of feature extraction limitations in existing technologies.

[0021] Optionally, encoding or quantifying clinical information includes: encoding categorical variables; standardizing numerical variables; and quantifying text information.

[0022] Optionally, the image data feature vector obtained from the image data includes: statistical features based on image grayscale values, features based on edge detection, and features based on texture analysis.

[0023] Optionally, the steps for dividing the subdomains are:

[0024] At least two clustering algorithms are used to cluster the spinal image samples, and the clustering results of each clustering algorithm on the scoliosis data are compared, and the subdomains divided by the clustering algorithm in which the distances between different clusters are greater than a set distance threshold are selected as the subdomains for subsequent training.

[0025] This optional solution achieves a reasonable division of samples based on the intrinsic distribution characteristics of the data, ensuring the effectiveness of differences between subdomains and sample distribution within subdomains.

[0026] Optionally, a classification model can be built for each subdomain by:

[0027] For subdomains where the number of samples is greater than the set value and the feature distribution is simple, a linear classification model is selected;

[0028] For subdomains where the number of samples is not greater than the set value but the characteristic distribution is complex, a model with nonlinear modeling capabilities is selected.

[0029] This optional approach can improve model performance and training efficiency.

[0030] Optionally, the linear correlation between the features in the subdomain is calculated. When the linear correlation between the features in the subdomain is higher than a set requirement, the feature distribution of the subdomain is simple; otherwise, the feature distribution of the subdomain is complex.

[0031] Optionally, when the model with nonlinear modeling capability is a support vector machine or a deep learning model,

[0032] Training of support vector machines: In each subdomain, samples from the subdomain are used to train the support vector machine. For multi-classification problems, one-to-many or one-to-one strategies are used to convert them into multiple binary classification problems for training.

[0033] Training of deep learning models: Using the transfer learning method, we use a model pre-trained on a large-scale image dataset, freeze the underlying convolutional layers of the pre-trained model, and only train the top fully connected layer and the newly added classification layer.

[0034] Optionally, the steps for calculating the similarity between the comprehensive feature vector of the prediction sample and the comprehensive feature vector of each subdomain are:

[0035] At least two similarity measurement methods are used to respectively calculate the similarity between the feature vector of the prediction sample and each subdomain;

[0036] The similarities obtained by each similarity measurement method are weighted averaged to obtain the final similarity between the predicted sample feature vector and the feature vectors of each subdomain.

[0037] The present invention also proposes a scoliosis detection system, comprising: a processing module, a storage module and a data acquisition module for acquiring spinal image samples;

[0038] The data acquisition module is electrically connected to the processing module and sends the spinal image sample to the processing module; the processing module is electrically connected to the storage module, and the storage module is used to store at least one executable instruction, and the executable instruction enables the processing module to perform operations corresponding to the above-mentioned scoliosis detection method based on the multi-subdomain classification algorithm to detect the scoliosis condition.

[0039] The beneficial effects of the present invention are:

[0040] The present invention makes full use of the local characteristics and distribution rules of the training data through multi-subdomain division, so that the samples in each subdomain have more similar feature patterns, and the classification model can better learn the classification boundaries for different types of scoliosis, thereby improving the classification accuracy. Compared with the prior art, experiments on large-scale real data sets show that the classification accuracy of the present invention can be improved by more than 20%.

[0041] The present invention also considers the feature differences of different subdomains, so that the model can better adapt to the new data distribution and reduce the risk of overfitting. At the same time, combined with auxiliary features such as clinical information, the adaptability of the model to different scenarios and individual differences is further improved, and the generalization ability is significantly enhanced.

[0042] The present invention calculates the similarity between the comprehensive feature vector of the prediction sample and the comprehensive feature vector of each subdomain, and based on the model selection strategy, ensures that the prediction sample can be matched to the most appropriate subdomain model for classification prediction.

[0043] The present invention comprehensively utilizes imaging data and clinical information data, fully mines multimodal information related to scoliosis, provides a more comprehensive basis for classification, and improves the reliability and stability of classification.

[0044] The present invention can provide patients with personalized classification results and diagnostic suggestions based on the characteristics of different subdomains. In generating diagnostic suggestions, it will combine factors such as the patient's age and living habits to formulate exclusive strategies for the classification of each subdomain. For example, for teenagers with mild scoliosis caused by bad posture, it is recommended to correct habits and exercise; for middle-aged and elderly people with degenerative scoliosis, medication is used and attention is paid to physical function. In addition, the method can also dynamically monitor and adjust individuals. Regularly resample the image analysis subdomain. If changes in the feature space are found, such as worsening scoliosis or the appearance of new abnormalities, the diagnosis and treatment plan will be optimized immediately, which will help doctors formulate more accurate treatment plans and improve medical quality and efficiency.

[0045] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 It is a schematic diagram of the feature extraction process;

[0049] Figure 3 It is a flowchart of using DBSCAN clustering algorithm to perform subdomain division;

[0050] Figure 4 It is a schematic diagram of the model training process;

[0051] Figure 5 It is a schematic diagram of the classification prediction process. DETAILED DESCRIPTION

[0052] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0053] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0054] Embodiment 1

[0055] like Figure 1 As shown, the present invention provides a scoliosis detection method based on a multi-subdomain classification algorithm, comprising the following steps:

[0056] Obtain spinal imaging samples. Spinal imaging samples can be collected from medical institutions. Spinal imaging samples include spinal imaging data (such as X-rays, CT scans, etc.) and basic patient information (such as age, gender, medical history, etc.). For imaging data, it is necessary to ensure that the image is clear and complete, covering the key parts of the spine; for the basic information of the patient, each item must be accurately recorded for subsequent analysis.

[0057] Professional doctors annotate the collected spinal image samples. They determine whether the samples are normal, abnormal or scoliotic based on the shape and curvature of the spine.

[0058] Then the spinal image samples are preprocessed. The preprocessing includes the image data preprocessing and clinical information preprocessing of the spinal image samples.

[0059] Image data preprocessing includes grayscale processing of image data (for color images) and converting them into single-channel grayscale images to reduce the amount of data and highlight the spinal structure information.

[0060] Image data preprocessing also includes normalizing the image data to map pixel values ​​to a specific interval (such as [0,1]). The normalization formula is:

[0061]

[0062] Where x is the original pixel value, x normalized is the normalized pixel value, min(x) and max(x) are the minimum and maximum pixel values ​​in the image data, respectively.

[0063] Image data preprocessing also includes image data enhancement operations, such as rotating, flipping, scaling and other transformations to increase data diversity and improve the robustness of the model.

[0064] Clinical information preprocessing includes coding, standardizing or quantifying clinical information. Specifically:

[0065] Encode categorical variables (such as gender), such as encoding male as 1 and female as 0.

[0066] For numerical variables (such as age, height, weight, etc.), standardization is performed to make them have zero mean and unit variance. The standardization formula is: Where μ is the mean, σ is the standard deviation, and x is the numerical variable to be standardized.

[0067] Text information such as family medical history is quantified, for example, having a family medical history is recorded as 1 and having no family medical history is recorded as 0; information such as symptom duration is segmented, such as 0-1 year is recorded as 1, 1-3 years is recorded as 2, and more than 3 years is recorded as 3, etc.

[0068] Then the comprehensive feature vector of the preprocessed spine image samples is extracted.

[0069] like Figure 2 As shown in the figure, a variety of imaging data feature vectors are obtained from imaging data using traditional feature extraction methods, including but not limited to: statistical features based on image grayscale values, including mean, variance, skewness, kurtosis, etc.; features based on edge detection, including length, curvature, angle change, etc. of the spine edge; features based on texture analysis, such as grayscale co-occurrence matrix features, local binary pattern features, etc. The imaging data is input into a pre-trained deep learning model (such as VGGNet, ResNet, etc.), and the intermediate layer feature vector that can capture the high-level semantic information of the image is extracted as a deep feature vector. The encoded, standardized or quantized clinical information is combined to form an auxiliary feature vector. Finally, the image feature vector extracted by the traditional method is spliced ​​with the deep feature vector, and further spliced ​​with the auxiliary feature vector to obtain the comprehensive feature vector of each sample, which is used as the basis for subsequent analysis or model training.

[0070] At the same time, the pre-processed spinal image samples are clustered according to the scoliosis categories and divided into multiple sub-domains. In this embodiment, the scoliosis categories include normal, abnormal posture, and scoliosis.

[0071] The clustering method here may adopt one of the DBSCAN algorithm, the spectral clustering algorithm, the fuzzy C-means clustering algorithm or any combination thereof, or may adopt other clustering algorithms.

[0072] When using the DBSCAN algorithm, such as Figure 3 As shown in the figure, the DBSCAN algorithm is based on the density of sample points. If the density of sample points in an area exceeds the set threshold, these sample points are divided into a cluster, i.e., a subdomain. Specifically, its core concepts include the neighborhood radius ∈ and the minimum number of sample points MinPts. For a sample point p, the number of sample points in its ∈-neighborhood (a spatial area with sample point p as the center and ∈ as the radius) is recorded as |N∈(p)|. If |N∈(p)|≥Minpts, then p is a core point. The algorithm starts from a core point and continuously expands the sample points in its neighborhood to form a cluster. If a sample point is neither a core point nor belongs to any cluster, it is regarded as a noise point. By adjusting the values ​​of ∈ and MinPts, the number and shape of subdomains can be controlled to adapt to different data distributions, so that the subdomains of scoliosis data can be better divided.

[0073] When the spectral clustering algorithm is used, clustering is performed based on the similarity matrix between sample points, and clustering structures of arbitrary shapes can be discovered.

[0074] When the fuzzy C-means clustering algorithm is used, sample points are allowed to belong to multiple clusters, which has a certain degree of fuzziness.

[0075] When at least two clustering algorithms are used for subdomain division, the clustering results of each clustering algorithm on the scoliosis data are compared, and the subdomains divided by the clustering algorithm whose distances between different clusters are greater than the set distance threshold are selected as the subdomains for subsequent training. The distance between clusters can be represented by the distance between two cluster centers. For example, for two clusters C1 and C2, whose cluster centers are a1 and a2 respectively, the distance between clusters can be calculated by d=||a1-a2||. Where ||·|| represents a distance metric (such as Euclidean distance).

[0076] After the subdomain division is completed, it is necessary to evaluate the scoliosis category distribution of samples in each subdomain to ensure that there are obvious differences in the feature space distribution between different subdomains (large inter-domain distances), and that the distribution of samples of each category in each subdomain is relatively concentrated, which is convenient for the subsequent learning of the classification model.

[0077] Then, a classification model is constructed for each subdomain based on the number of samples and the distribution of comprehensive feature vectors in each subdomain. Figure 4 As shown in the figure, for subdomains with a sample size greater than a set value (for example, 1000, which can be adjusted according to specific problems and data characteristics) and relatively simple feature distribution, a linear classification model, such as a logistic regression model, is selected; for subdomains with a sample size not greater than a set value (for example, 500, which can also be adjusted according to specific problems and data characteristics) and complex feature distribution, a model with strong nonlinear modeling capabilities is selected, such as support vector machines, deep learning models, random forests, gradient boosting decision trees, classification models based on graph neural networks (GNNs), and so on.

[0078] In this embodiment, the measurement method for whether the feature distribution in the subdomain is simple or not is:

[0079] When a subdomain has multiple features, the correlation coefficient matrix between the features is calculated to understand the linear relationship between the features. Calculation method: Assuming there are m features, calculate the correlation coefficient matrix R between the features to obtain an m×m matrix. Among them, the matrix element r in the correlation coefficient matrix R ij Represents feature x i and x j The present embodiment preferably but not limited to uses the Pearson correlation coefficient to calculate the feature x i and x jThe correlation coefficient between them is taken as the matrix element r ij . The correlation coefficient range is [-1,1]. If the correlation coefficient is close to 1, it means that the two features have a strong positive linear correlation; if it is close to -1, it means that the negative linear correlation is strong; if it is close to 0, it means that the linear relationship is weak. When the linear correlation between the features in a subdomain is higher than the set requirement, the correlation between the features in the subdomain is strong (for example, the absolute value of most elements in the correlation coefficient matrix is ​​higher than 0.3), which means that the distribution of features in the subdomain is relatively simple; when the linear correlation between the features in the subdomain is lower than the set requirement, it means that there is a strong nonlinear relationship between the features in the subdomain, and the distribution of features in the subdomain is complex.

[0080] Use samples from each subdomain to train its corresponding classification model;

[0081] Taking the support vector machine as an example, the samples in each subdomain are used to train the support vector machine. For multi-classification problems (normal, abnormal posture, scoliosis), a one-vs-rest or one-vs-one strategy can be used to convert them into multiple binary classification problems for training. During the training process, appropriate kernel functions (such as linear kernels, polynomial kernels, radial basis function kernels, etc.) and model parameters (such as penalty parameter C) are selected through cross-validation and other methods to optimize model performance.

[0082] For the subdomains that select deep learning models, transfer learning methods are used to fine-tune models pre-trained on large-scale image datasets (such as ResNet models pre-trained on ImageNet datasets) on subdomain samples. During fine-tuning, some of the underlying convolutional layers of the pre-trained model are frozen, and only the top fully connected layers and newly added classification layers are trained to speed up training and reduce the risk of overfitting. During training, optimization algorithms such as stochastic gradient descent (SGD) or Adam are used to adjust parameters such as learning rate and momentum to minimize the loss function (such as the cross entropy loss function).

[0083] For the subdomains where random forest and gradient boosted decision tree are selected, random forest and gradient boosted decision tree improve the classification performance by combining multiple weak classifiers, and have good robustness and generalization ability.

[0084] The classification model of graph neural network (GNN) uses the relationship information between samples to make classification decisions and is more suitable for processing data with complex structures.

[0085] After the training is completed, the prediction samples are predicted, such as Figure 5As shown, first, the comprehensive feature vector of the prediction sample is extracted, and then the similarity between the comprehensive feature vector of the prediction sample and the comprehensive feature vector of each subdomain is calculated, and the trained classification model corresponding to the subdomain with the highest similarity is selected to perform classification prediction on the prediction sample.

[0086] In this embodiment, at least two similarity measurement methods are used to respectively calculate the similarity between the feature vector of the prediction sample and each subdomain, such as Euclidean distance, cosine similarity, Mahalanobis distance, etc., and the similarity obtained by each similarity measurement method is weighted averaged to obtain the final similarity between the feature vector of the prediction sample and the feature vector of each subdomain, that is, the final similarity between the prediction sample and the i-th subdomain is:

[0087] S i =w1s i1 +w2s i2 +…+w j s ij +…+w n s in

[0088] Among them, n is the number of similarity measurement methods used, w j is the weight of the j-th similarity measurement method, s ij is the similarity between the predicted sample and the i-th subdomain calculated using the j-th similarity measurement method.

[0089] Take Euclidean distance and cosine similarity as examples:

[0090] For example, let the predicted sample feature vector be x test , the feature center of the i-th subdomain is μ i , then the Euclidean distance formula is:

[0091]

[0092] Where n is the feature vector dimension.

[0093] The formula for cosine similarity is:

[0094] Then the final similarity between the predicted sample and the i-th subdomain is obtained by weighted average: S i =w1d i +w2s i , where w1 and w2 are weight coefficients.

[0095] The trained classification model corresponding to the subdomain with the highest final similarity is selected to perform classification prediction on the prediction sample to obtain the classification label of the prediction sample, such as normal, abnormal posture or scoliosis.

[0096] Embodiment 2

[0097] The present invention also provides a scoliosis detection system, comprising: a processing module, a storage module and a data acquisition module for acquiring spinal image samples;

[0098] The data acquisition module is electrically connected to the processing module and sends the spinal image sample to the processing module; the processing module is electrically connected to the storage module, and the storage module is used to store at least one executable instruction, and the executable instruction enables the processing module to perform operations corresponding to the scoliosis detection method based on the multi-subdomain classification algorithm as described in Example 1, and detect the scoliosis condition.

[0099] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0100] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A scoliosis detection method based on a multi-subdomain classification algorithm, characterized in that: The following steps are involved: Acquire spinal image samples, and label and preprocess the spinal image samples; Extracting comprehensive feature vectors of preprocessed spinal image samples; Clustering the spinal image samples according to scoliosis categories to divide them into multiple subdomains; A classification model is constructed for each subdomain based on the number of samples and the distribution of comprehensive feature vectors in each subdomain; Use samples from each subdomain to train its corresponding classification model; The comprehensive feature vector of the predicted sample is extracted, the similarity between the comprehensive feature vector of the predicted sample and the comprehensive feature vector of each subdomain is calculated, and the classification model corresponding to the subdomain with the highest similarity is selected to perform classification prediction on the predicted sample to obtain the classification of the predicted sample.

2. The scoliosis detection method based on the multi-subdomain classification algorithm according to claim 1, characterized in that: Preprocessing involves coding or quantifying clinical information; When extracting the comprehensive feature vector, the image data feature vector is obtained from the image data; the image data is input into a pre-trained deep learning model to extract a deep feature vector that can capture the high-level semantic information of the image; the encoded or quantized clinical information is combined to form an auxiliary feature vector; then the image data feature vector is spliced ​​with the deep feature vector, and further spliced ​​with the auxiliary feature vector to obtain a comprehensive feature vector for each sample.

3. The scoliosis detection method based on the multi-subdomain classification algorithm according to claim 2 is characterized in that: Coding, standardizing or quantifying clinical information includes: encoding categorical variables; standardizing numerical variables; and quantifying text information.

4. The scoliosis detection method based on the multi-subdomain classification algorithm according to claim 2, characterized in that: The image data feature vectors obtained from the image data include: statistical features based on image grayscale values, features based on edge detection, and features based on texture analysis.

5. The scoliosis detection method based on multi-subdomain classification algorithm according to claim 1, characterized in that: The steps for dividing the subdomains are: At least two clustering algorithms are used to cluster the spinal image samples, and the clustering results of each clustering algorithm on the scoliosis data are compared, and the subdomains divided by the clustering algorithm in which the distances between different clusters are greater than a set distance threshold are selected as the subdomains for subsequent training.

6. The scoliosis detection method based on multi-subdomain classification algorithm according to claim 1, characterized in that: The method to build a classification model for each subdomain is: For subdomains where the number of samples is greater than the set value and the feature distribution is simple, a linear classification model is selected; For subdomains where the number of samples is not greater than the set value but the characteristic distribution is complex, a model with nonlinear modeling capabilities is selected.

7. The scoliosis detection method based on multi-subdomain classification algorithm according to claim 6, characterized in that: The linear correlation between the features in the subdomain is calculated. When the linear correlation between the features in the subdomain is higher than the set requirement, the feature distribution of the subdomain is simple; otherwise, the feature distribution of the subdomain is complex.

8. The scoliosis detection method based on multi-subdomain classification algorithm according to claim 6, characterized in that: When the model with nonlinear modeling capability is a support vector machine or a deep learning model, Training of support vector machines: In each subdomain, samples from the subdomain are used to train the support vector machine. For multi-classification problems, one-to-many or one-to-one strategies are used to convert them into multiple binary classification problems for training. Training of deep learning models: Using the transfer learning method, we use a model pre-trained on a large-scale image dataset, freeze the underlying convolutional layers of the pre-trained model, and only train the top fully connected layer and the newly added classification layer.

9. The scoliosis detection method based on multi-subdomain classification algorithm according to claim 1, characterized in that: The calculation steps of the similarity between the comprehensive feature vector of the predicted sample and the comprehensive feature vector of each subdomain are as follows: At least two similarity measurement methods are used to respectively calculate the similarity between the feature vector of the prediction sample and each subdomain; The similarities obtained by each similarity measurement method are weighted averaged to obtain the final similarity between the predicted sample feature vector and the feature vectors of each subdomain.

10. A scoliosis detection system, characterized in that: include: A processing module, a storage module and a data acquisition module for acquiring spinal image samples; The data acquisition module is electrically connected to the processing module, and sends the spinal image sample to the processing module; the processing module is electrically connected to the storage module, and the storage module is used to store at least one executable instruction, and the executable instruction enables the processing module to perform operations corresponding to the scoliosis detection method based on the multi-subdomain classification algorithm as described in any one of claims 1-9, and detect the scoliosis condition.

Images