A spine sagittal plane shape intelligent classification method based on multi-scale feature extraction

Abstract

The application discloses a kind of spine sagittal plane shape intelligent classification method based on multiscale feature extraction, including image preprocessing and feature extraction, key feature screening, data balancing and optimization and random forest classification;The present application can realize the automatic, high-precision classification of adult spine-pelvic sagittal plane X-ray image, provide scientific decision support for clinical diagnosis of spinal disease, large-scale screening, personalized prevention and surgical plan, and through the multiscale feature depth representation of spinal morphology biomechanical characteristics, combined with rigorous feature screening and model optimization strategy, effectively overcome the problem of strong subjectivity and single feature of traditional method.

Description

TECHNICAL FIELD

[0001] The present application relates to a spine sagittal morphology intelligent classification method, in particular to a spine sagittal morphology intelligent classification method based on multi-scale feature extraction, and belongs to the technical field of medical image processing and artificial intelligence. BACKGROUND

[0002] As the core support structure of the human body, the accurate classification of the sagittal morphology of the spine is a key basis for evaluating spinal health, diagnosing spinal diseases (such as disc herniation, scoliosis, spondylolisthesis, etc.) and developing treatment plans. The traditional classification method mainly relies on manual measurement of several static anatomical parameters on X-ray films by doctors, such as pelvic incidence, sacral slope, lumbar lordosis angle, etc., and is classified according to experience rules.

[0003] The traditional classification of the sagittal morphology of the spine relies on manual measurement of key parameters, which has obvious limitations:

[0004] 1. Large subjective error, existing literature:

[0005] The classification method proposed by P. Roussouly, S. Gollogly, E. Berthonnaud, et al., Classification of the normal variation in the sagittal alignment of the human lumbar spine and pelvis in the standing position. Spine (Phila Pa 1976), 2005, 30(3): 346-353, is based on only a few static parameters such as sacral slope (SS) and lumbar lordosis morphology, without considering the dynamic changes of the overall spinal curve and regional coordination. In the state of spinal degeneration, the decrease of lumbar lordosis will cause the distortion of SS value, and the classification accuracy will decrease significantly;

[0006] 2. Narrow scope, existing literature:

[0007] The grading system proposed by C. S. Lee, S. S. Chung, K. C. Kang, et al., Normal patterns of sagittal alignment of the spine in young adults radiological analysis in a Korean population. Spine (Phila Pa 1976), 2011, 36(25): E1648-1654, is only applicable to the single disease of lumbar spinal stenosis, and relies on high-resolution MRI images, which is limited by equipment conditions and cannot be extended to the classification of the overall sagittal morphology of the spine;

[0008] 3. Low degree of intelligence, existing literature:

[0009] S. Mohanty, F. M. Hassan, L. G. Lenke, et al., Machine learning clustering of adult spinal deformity patients identifies four prognostic phenotypes: a multicenter prospective cohort analysis with single surgeon external validation. Spine J, 2024, 24(6): 1095-1108; Although it reduces manual intervention, the selection of kernel function and parameter tuning is complex, and the generalization ability is weak when processing high-dimensional spinal image data, making it difficult to adapt to large-scale and diversified clinical data;

[0010] 4. Poor population adaptability: Most classification systems are based on foreign population data, ignoring the differences in sagittal spinal morphology development among different races and regions, and there is a lack of exclusive classification model for Chinese adults. SUMMARY

[0011] The purpose of the present application is to solve at least one of the above technical problems and provide a spinal sagittal plane morphology intelligent classification method based on multi-scale feature extraction, which can automatically and accurately process spinal sagittal plane images and extract multi-dimensional features reflecting the biomechanical essence from the images. The intelligent classification method has strong generalization ability and robustness to class imbalance, to overcome the technical defects of subjective error, incomplete feature representation, weak model generalization ability and poor population adaptability in the prior art.

[0012] The present application realizes the above-mentioned purpose through the following technical scheme: a spinal sagittal plane morphology intelligent classification method based on multi-scale feature extraction, which comprises the following steps:

[0013] Step 1, image preprocessing and feature extraction: for the spine-pelvis sagittal plane X-ray image, use the existing hollow space pyramid pooling module and multi-task deep learning method with cross-task attention mechanism to complete the key point detection of the four vertices of each spinal vertebra; After solving the centroid coordinates of the vertebrae by the vertex coordinates, the centroid coordinate sequence is fitted by a 12th order polynomial to generate a fitting curve; Based on the curve, multi-scale features that can represent the dynamic changes of spinal morphology are extracted to form an initial feature set;

[0014] Step 2, key feature selection: based on statistical correlation analysis method, the key feature subset strongly related to the spinal classification result is selected from the initial feature set;

[0015] Step 3, Data Balancing and Optimization: Divide the key feature subset into training and testing sets, perform class balancing, and optimize the hyperparameters of the random forest;

[0016] Step 4: Random Forest Classification: Based on the key feature subset, construct an optimized random forest model to classify the sagittal morphology of the spine, and output the classification results and performance indicators.

[0017] As a further aspect of the present invention: in step one, the key points include 24 key points of the spinal vertebrae; the centroid coordinates of the spinal vertebrae are calculated using the polygon centroid formula; the 12th-order polynomial fitting employs the least squares method, with the sum of squared errors... The objective is to minimize the coefficients of the 12th-order polynomial, and the coefficients are solved using the Vandermonde matrix. The ordinate of the feature point is... Predict the ordinate of the fitted curve to ensure a good fit between the fitted curve and the centroid coordinates;

[0018] Multiscale features include the first derivative, second derivative, and curvature of the centroids of the 24 vertebral bodies corresponding to the fitted curves.

[0019] As a further aspect of the present invention: the first derivative reflects the vertebral displacement rate, the second derivative reflects the degree of joint curvature, and the curvature reflects the degree of spinal load concentration.

[0020] As a further aspect of the present invention: In step two, the statistical correlation analysis method is Spearman rank correlation analysis, and features with an absolute value of correlation coefficient not less than 0.6 are selected; the key feature subset includes L4 first derivative, L3 first derivative, T12 second derivative, L2 first derivative, L1 second derivative, L3 second derivative, L2 second derivative, T11 second derivative, T10 first derivative, T9 first derivative, L3 curvature, and T11 curvature.

[0021] As a further aspect of this invention: the Spearman rank correlation analysis calculation formula is as follows:

[0022] ;

[0023] in, The rank correlation coefficient has a value range of [-1, 1]. Let n be the rank difference between paired samples, and n be the sample size.

[0024] As a further scheme of the present application: in step three, the class balance processing adopts the SMOTE-ENN algorithm, wherein the SMOTE oversampling ratio is 1:3, and the ENN algorithm removes noise samples based on 3 nearest neighbor samples; the random forest hyperparameters are searched by the GridSearchCV method, and the random forest hyperparameters include at least one of the following: the number of decision trees (100 / 200 / 300), the maximum depth of the decision tree (None / 10 / 20 / 30), the maximum number of features (sqrt, log2, None), the minimum number of samples for node splitting (2 / 5 / 10), the minimum number of samples for leaf nodes (1 / 2 / 4), autonomous sampling (True, False), and class weight (balanced, None).

[0025] As a further scheme of the present application: the optimal hyperparameter combination of the random forest classification model is: the number of decision trees n_estimators=200, the maximum number of features max_features=sqrt, the minimum number of samples for node splitting min_samples_split=2, autonomous sampling bootstrap=False, the class weight class_weight=balanced, the maximum depth of the decision tree max_depth=None, and the minimum number of samples for leaf nodes min_samples_leaf=2.

[0026] As a further scheme of the present application: in step four, the sagittal spinal morphology is divided into four types according to the position of the horizontal lumbar line relative to the lumbar vertebral body and the intervertebral space:

[0027] Type I, HLL is located below the L4 / 5 intervertebral space; Type II, HLL passes through the L4 vertebral body and the L4 / 5 intervertebral space; Type III, HLL passes through the L3 vertebral body and the L3 / 4 intervertebral space; Type IV, HLL is located above the L3 upper endplate.

[0028] As a further scheme of the present application: the training process of the random forest model adopts a weighted cross-entropy loss function.

[0029] A sagittal spinal morphology intelligent classification system based on multi-scale feature extraction, comprising an image preprocessing module, a feature extraction module, a feature screening module, a data optimization module, and a random forest classification module connected in sequence and cooperatively completing the sagittal spinal morphology classification.

[0030] The image preprocessing module and the feature extraction module are used to perform image preprocessing and feature extraction of step one; the feature screening module is used to perform key feature screening of step two; the data optimization module is used to perform data balance and optimization of step three; and the random forest classification module is used to perform random forest classification of step four.

[0031] The beneficial effects of the present application are:

[0032] 1) The present application extracts multiple vertebral centroid coordinates and performs high-order polynomial fitting, and then calculates the first derivative, the second derivative and the curvature, innovatively constructs a multi-scale feature set from the geometric and biomechanical dual angles, can comprehensively and deeply represent the dynamic changes, bending shape and load distribution characteristics of the spinal curve, overcomes the limitations of traditional methods relying on a few static parameters, and lays a rich and physically meaningful feature foundation for accurate classification;

[0033] 2) The present application uses strict statistical correlation analysis to automatically select a key feature subset from high-dimensional initial features, effectively eliminates redundant information, reduces model complexity and enhances interpretability, while avoiding overfitting risk, combines SMOTE oversampling and ENN denoising hybrid algorithm to process the common class imbalance problem in clinical data, and combines grid search to fine-tune the random forest model hyperparameters, significantly improving the model's recognition ability for minority classes and overall generalization performance;

[0034] 3) The present application realizes accurate four-classification of the sagittal shape of Chinese adult spine based on the optimized random forest model, with high classification accuracy and good stability; The method is fully automated, greatly reducing manual intervention and subjective errors, and is specially trained and optimized for Chinese population data, with excellent population adaptability, and has important clinical application value for auxiliary diagnosis, large-scale screening and individualized treatment planning of spinal diseases;

[0035] 4) The present application can realize automatic and high-precision classification of adult spine-pelvis sagittal X-ray images, provide scientific decision support for clinical diagnosis, large-scale screening, individualized prevention and surgical planning of spinal diseases, and represent the biomechanical characteristics of spinal shape through multi-scale features, combined with rigorous feature selection and model optimization strategies, effectively overcoming the problem of strong subjectivity and single feature of traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a schematic diagram of the overall flow structure of the present application;

[0037] Figure 2 is a schematic diagram of the present application vertebral centroid calculation;

[0038] Figure 3 is a schematic diagram of the 12th order polynomial curve fitting result structure of the present application;

[0039] Figure 4 is a schematic diagram of the physical meaning of the first derivative of the present application;

[0040] Figure 5A second derivative physical meaning diagram of the present application;

[0041] Figure 6 A curvature physical meaning diagram of the present application;

[0042] Figure 7 A random forest model training process structure diagram of the present application;

[0043] Figure 8 A four-type spine type diagram of the improved Lee classification of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0045] Embodiment one, as shown in the drawing, the present embodiment provides a spine sagittal plane shape intelligent classification method based on multi-scale feature extraction, the spine sagittal plane shape intelligent classification method comprises the following steps: Figures 1 to 8

[0046] Step one, image preprocessing and feature extraction: for the spine-pelvic sagittal plane X-ray image, the key point detection of the four vertices of each spine vertebra is completed by using the existing cavity space pyramid pooling module and the multi-task deep learning method with cross-task attention mechanism; after solving the centroid coordinates of the vertebra through the vertex coordinates, the centroid coordinate sequence is fitted by using the 12th order polynomial to generate a fitting curve; based on the curve, multi-scale features that can represent the dynamic changes of the spine shape are extracted to form an initial feature set;

[0047] Step two, key feature screening: based on statistical correlation analysis method, the key feature subset strongly related to the spine classification result is screened from the initial feature set;

[0048] Step three, data balancing and optimization, the key feature subset is divided into training set and test set, and the class balancing processing and the optimization of random forest hyperparameters are carried out;

[0049] Step four, random forest classification, based on the key feature subset, the optimized random forest model is constructed to classify the spine sagittal plane shape, and the classification result and performance index are output.

[0050] Embodiment two, in addition to including all the technical features in embodiment one, the present embodiment also includes:

[0051] In step one: the key points include 24 spine vertebra key points;​

[0052] As shown in Figure 2 , the centroid coordinates of the spinal vertebrae are calculated by the polygon centroid formula;

[0053] As shown in Figure 3 , the 12-order polynomial fitting adopts the least square method to minimize the sum of squared errors , and the 12-order polynomial coefficients are solved in combination with the Vandermonde matrix, where is the longitudinal coordinate of the feature point, is the predicted longitudinal coordinate of the fitting curve to ensure the fitting degree of the fitting curve and the centroid coordinates;

[0054] As shown in Figure 4 , Figure 5 and Figure 6 , the multi-scale features include the first derivative, the second derivative and the curvature of the fitting curve corresponding to the 24 centroid of the spinal vertebrae, wherein the first derivative reflects the displacement rate of the spinal vertebrae, the second derivative reflects the degree of joint bending, and the curvature reflects the degree of spinal load concentration.

[0055] In step two, the correlation analysis method is Spearman rank correlation analysis, and the features with an absolute correlation coefficient not less than 0.6 are selected; the key feature subset includes the first derivative of L4, the first derivative of L3, the second derivative of T12, the first derivative of L2, the second derivative of L1, the second derivative of L3, the second derivative of L2, the second derivative of T11, the first derivative of T10, the first derivative of T9, the curvature of L3 and the curvature of T11.

[0056] Further, the calculation formula of Spearman rank correlation analysis is:

[0057] ;

[0058] Wherein, is the rank correlation coefficient (value range [-1, 1]), is the rank difference of the paired samples, and n is the sample size.

[0059] In step three, the class balance processing adopts the SMOTE-ENN algorithm, wherein the SMOTE oversampling ratio is 1:3, and the ENN algorithm removes noise samples based on 3 nearest neighbor samples;

[0060] The random forest hyperparameters are searched by the GridSearchCV method, and the random forest hyperparameters include at least one of the following: the number of decision trees (100 / 200 / 300), the maximum depth of the decision tree (None / 10 / 20 / 30), the maximum number of features (sqrt, log2, None), the minimum number of samples for node splitting (2 / 5 / 10), the minimum number of samples for leaf nodes (1 / 2 / 4), self-sampling (True, False), and class weight (balanced, None).

[0061] In step four, the random forest hyperparameter combination of the random forest classification model is: the number of decision trees n_estimators=200, the maximum number of features max_features=sqrt, the minimum number of samples for node splitting min_samples_split=2, self-sampling bootstrap=False, the class weight class_weight=balanced, the maximum depth of the decision tree max_depth=None, and the minimum number of samples for leaf nodes min_samples_leaf=2.

[0062] The sagittal spinal morphology is divided into four types according to the position of the horizontal lumbar line relative to the lumbar vertebral body and intervertebral space:

[0063] Type I (HLL is located below the L4 / 5 intervertebral space);

[0064] Type II (HLL passes through the L4 vertebral body and L4 / 5 intervertebral space);

[0065] Type III (HLL passes through the L3 vertebral body and L3 / 4 intervertebral space);

[0066] Type IV (HLL is located above the L3 endplate).

[0067] As shown in Figure 7 , the complete process of data preprocessing, SMOTE-ENN balancing, GridSearchCV hyperparameter optimization, and cross-validation is shown; the training process of the random forest model uses a weighted cross-entropy loss function, and the Type I and Type IV minority class samples are given higher weights.

[0068] In embodiment three, a sagittal spinal morphology intelligent classification system based on multi-scale feature extraction is provided, which is used to realize the sagittal spinal morphology intelligent classification method in embodiment one. The sagittal spinal morphology intelligent classification system includes an image preprocessing module, a feature extraction module, a feature screening module, a data optimization module, and a random forest classification module connected in sequence and cooperatively completing the sagittal spinal morphology classification.

[0069] Wherein, the image preprocessing module and the feature extraction module are used for performing image preprocessing and feature extraction; the feature screening module is used for performing key feature screening; the data optimization module is used for performing data balancing and optimization; and the random forest classification module is used for performing random forest classification.

[0070] Embodiment four is a spine sagittal plane shape intelligent classification method based on multi-scale feature extraction, which completes the spine sagittal plane shape classification through the cooperation of four core modules of image preprocessing and feature extraction module, key feature screening module, data balancing and optimization module, and random forest classification module, and the process is as follows:

[0071] First, image preprocessing and feature extraction: pre-process the spine-pelvis sagittal plane X-ray image, extract the key points of the four vertices of the spine vertebrae and calculate the vertebrae centroid, and generate multi-scale features (first derivative, second derivative, curvature) through 12-order polynomial curve fitting.

[0072] The image preprocessing and feature extraction module specifically includes:

[0073] 1) Image preprocessing: receiving adult spine-pelvis sagittal plane X-ray image, extracting four key point coordinates.

[0074] 2) Key point detection and centroid calculation: using the existing hollow space pyramid pooling module and multi-task deep learning method with cross-task attention mechanism, completing the key point detection of the four vertices of the 24 vertebrae C1-C7, T1-T12, L1-L5; calculating the centroid coordinates of each vertebra (quadrilateral structure) according to the polygon centroid formula as the feature point of curve fitting, the formula is as follows:

[0075] ;

[0076] Wherein, is the area of the vertebra quadrilateral, is the vertex coordinate of the vertebra quadrilateral.

[0077] 3) Multi-scale feature calculation: for the 24 feature points, the least square method is used to minimize the sum of squared errors ; is the longitudinal coordinate of the feature point, is the predicted longitudinal coordinate of the fitting curve, to perform 12-order polynomial curve fitting.

[0078] 4) Calculate multi-scale feature parameters based on the fitting curve, a total of 72 initial features are generated, including: a First derivative:

[0079] Reflects the rate of change of the position of the key points of the spine, the physical meaning is the dynamic mechanical transmission characteristics of the vertebral body, the larger the absolute value, the higher the relative displacement rate of adjacent vertebral bodies;

[0080] b Second derivative:

[0081] Reflects the bending degree and concave-convex shape of the spinal joint, the second derivative of the lower convex region is > 0, and the second derivative of the upper convex region is < 0, the larger the absolute value, the more significant the bending degree;

[0082] c Curvature:

[0083] Reflects the degree of deviation from a straight line, which is directly related to the mechanical load-bearing of the spine, the larger the curvature, the more concentrated the local load.

[0084] Second, key feature screening: based on Spearman rank correlation analysis, screening the feature parameters with strong correlation with the type of spine, and eliminating redundant information.

[0085] Based on the Spearman rank correlation analysis, the key features are screened, and the specific steps are as follows:

[0086] 1) Correlation calculation:

[0087] Calculate the Spearman rank correlation coefficient of 72 initial features and the type of spine (based on the improved Lee classification TypeI-TypeIV), the formula is:

[0088] ;

[0089] Where, is the rank correlation coefficient (value range [-1, 1]), is the rank difference of the pair of samples, is the sample size (in this study ).

[0090] 2) Feature screening:

[0091] Set the screening threshold: the absolute value of the correlation coefficient is ≥0.6, finally retain 12 key features, as shown in Table 1:

[0092] Table 1: 12 key features retained by feature screening

[0093]

[0094] Third, data balancing and optimization: using SMOTE-ENN algorithm to handle the data class imbalance problem, combined with GridSearchCV to optimize the model hyperparameters.

[0095] The data balancing and optimization module specifically includes:

[0096] 1) Data partitioning and class balancing: 439 samples were divided into training set (351 cases) and test set (88 cases) by 8:2 stratified sampling, ensuring consistent class distribution in both data sets (Type I 20 cases, Type II 211 cases, Type III 171 cases, Type IV 37 cases); SMOTE algorithm was used to oversample the minority class samples (Type I, Type IV) in the training set (sampling ratio 1:3), and ENN algorithm was used to remove noise samples introduced by oversampling (samples with different class labels from 3 nearest neighbors were removed), finally 390 training samples were obtained.

[0097] 2) Model hyperparameter optimization: GridSearchCV method was used to traverse the random forest hyperparameter combination, and the hyperparameter search range was set as:

[0098] Decision tree number (n_estimators): 100, 200, 300;

[0099] Maximum feature number (max_features): sqrt, log2;

[0100] Minimum sample number for node splitting (min_samples_split): 2, 5;

[0101] Minimum sample number for leaf node (min_samples_leaf): 1, 2;

[0102] Maximum depth of tree (max_depth): None, 10, 20.

[0103] 3) The average F1 score of five-fold cross-validation was used as the evaluation index to determine the optimal hyperparameter combination: n_estimators=200, max_features=sqrt, min_samples_split=2, min_samples_leaf=1, max_depth=None, bootstrap=False, class_weight=balanced.

[0104] Fourth, random forest classification: build an optimized random forest model to realize accurate classification of spinal type (Type I-Type IV), and output the classification results and performance indicators.

[0105] Model construction and spine type classification, 12 key features as input, build a random forest classification model: based on Bootstrap sampling to generate 200 decision trees, using Gini coefficient as node splitting criterion (measure of data set purity), through majority voting to determine the final classification result.

[0106] The data set is divided into training set and test set, 80% of the data is used for training, 20% of the data is used for testing, to perform random forest classification prediction, and the selected parameter indicators are applied to the random forest classification prediction model.

[0107] The definition of four types of spine and sample ratio is as follows:

[0108] Type I: horizontal lumbar line (HLL) is located below L4 / 5 intervertebral space, PI=34.7°, SS=20.2°, LL=26.8°, accounting for 3.6%;

[0109] Type II: HLL passes through L4 vertebral body and L4 / 5 intervertebral space, PI=38.1°, SS=30.0°, accounting for 39.3%;

[0110] Type III: HLL passes through L3 vertebral body and L3 / 4 intervertebral space, PI=46.9°, SS=36.8°, accounting for 45.7%;

[0111] Type IV: HLL is located above the upper endplate of L3, LL=47.0°, thoracic kyphosis (TK) significantly increases, accounting for 11.4%.

[0112] Performance verification and result output, using stratified five-fold cross-validation to evaluate the generalization ability of the model, outputting classification results (spine type) and performance indicators (accuracy, precision, recall, F1 score), ensuring that the model test set accuracy is ≥99%, and the average accuracy of cross-validation is ≥96.5%.

[0113] Example five, a spine sagittal plane shape intelligent classification method based on multi-scale feature extraction, specifically comprising:

[0114] Data preparation: 439 cases of Chinese Han nationality adult spine-pelvis sagittal plane X-ray images (18-49 years old, male and female ratio 1:1.37) are selected, which are exempted from ethical review by Xuzhou Central Hospital (IRB approval number: XZCH-2024-032);

[0115] Feature extraction: 30 feature points are extracted according to the above steps, 12 order polynomial curves are fitted, 72 initial features are calculated, and 12 key features are selected;

[0116] Data optimization: 390 samples were included in the training set after SMOTE-ENN processing, and the optimal hyperparameters were determined by GridSearchCV.

[0117] Model validation: The test set validation results are shown in Table 2, with an overall accuracy of 99% and an average accuracy of 96.46% in five-fold cross-validation, meeting the clinical requirements.

[0118] Table 2: Test set validation results

[0119]

[0120] Adult sagittal spine-pelvis X-ray images were obtained, and 24 key points of the spine and 6 key points of the sacrococcyx were detected after preprocessing. The centroid coordinates were calculated by the polygon centroid formula. A 12th order polynomial was used to fit the centroid points, and the first derivative (reflecting the displacement rate of key points), the second derivative (reflecting the joint bending degree), and the curvature (reflecting the load concentration degree) of the fitted curve were extracted, generating 72 initial features. Based on Spearman rank correlation analysis, 12 key features (absolute value of correlation coefficient ≥0.6) were selected. The SMOTE-ENN algorithm was used to process the data class imbalance, and the GridSearchCV was used to optimize the random forest hyperparameters (optimal combination: n_estimators=200, max_features=sqrt, etc.), and a classification model was constructed. Finally, the Type I-Type IV spine type based on the improved Lee classification was output as shown in Table 2, and the performance indicators such as accuracy and F1 score were also provided. Figure 8

[0121] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be considered as limiting the claims involved.

[0122] In addition, it should be understood that although the present specification is described in terms of embodiments, each embodiment does not contain only one independent technical solution, and the description of the specification is only for clarity. The skilled person should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art.​

Claims

1. A method for intelligent classification of spinal sagittal plane morphology based on multi-scale feature extraction, characterized in that, Includes the following steps: Step 1: Image Preprocessing and Feature Extraction: For sagittal X-ray images of the spine and pelvis, a multi-task deep learning method using an existing hollow spatial pyramid pooling module and cross-task attention mechanism is employed to detect key points at the four vertebrae of each vertebra. After solving for the centroid coordinates of the vertebrae using the vertex coordinates, a 12th-order polynomial is used to fit the centroid coordinate sequence to generate a fitting curve. Based on this curve, multi-scale features that can characterize the dynamic changes in the morphology of the spine are extracted to form an initial feature set. In step one: The key points include 24 key points of the spinal vertebrae; The centroid coordinates of the spinal vertebral body are calculated using the polygon centroid formula; The 12th-order polynomial fitting employs the least squares method, using the sum of squared errors. The objective is to minimize the coefficients of the 12th-order polynomial, and the coefficients are solved using the Vandermonde matrix. The ordinate of the feature point is... Predict the ordinate of the fitted curve; The multi-scale features include the first derivative, second derivative, and curvature of the fitted curve corresponding to the centroids of the 24 vertebral bodies of the spine; Step 2: Key Feature Screening: Based on statistical correlation analysis, a subset of key features that are strongly correlated with the spine classification results is screened from the initial feature set; In step two: The statistical correlation analysis method is Spearman rank correlation analysis, and features with an absolute correlation coefficient of not less than 0.6 are selected. The key feature subset includes the first derivative of L4, the first derivative of L3, the second derivative of T12, the first derivative of L2, the second derivative of L1, the second derivative of L3, the second derivative of L2, the second derivative of T11, the first derivative of T10, the first derivative of T9, the curvature of L3, and the curvature of T11. The formula for calculating the Spearman rank correlation analysis is as follows: ； in, This is the rank correlation coefficient, with a value range of [-1, 1]. The rank difference is the pairwise difference between samples, and n is the sample size. Step 3, Data Balancing and Optimization: Divide the key feature subset into training and testing sets, perform class balancing, and optimize the hyperparameters of the random forest. Step 4: Random Forest Classification: Based on the aforementioned key feature subset, construct an optimized random forest model to classify the sagittal morphology of the spine, and output the classification results and performance indicators.

2. The intelligent classification method for spinal sagittal plane morphology based on multi-scale feature extraction according to claim 1, characterized in that: The first derivative reflects the displacement rate of key points in the vertebral body; the second derivative reflects the degree of joint curvature; and the curvature reflects the degree of load concentration on the spine.

3. The intelligent classification method for spinal sagittal plane morphology based on multi-scale feature extraction according to claim 1, characterized in that, In step three: The category balancing process uses the SMOTE-ENN algorithm, where the SMOTE oversampling ratio is 1:3 and the ENN algorithm removes noisy samples based on the three nearest neighbor samples. The hyperparameters of the random forest are obtained by performing grid search using the GridSearchCV method, and the hyperparameters of the random forest include at least one of the following: number of decision trees, maximum depth of decision trees, maximum number of features, minimum number of samples for node splits, minimum number of samples for leaf nodes, autonomous sampling, and class weights.

4. The intelligent classification method for spinal sagittal plane morphology based on multi-scale feature extraction according to claim 3, characterized in that, In step four: The optimized random forest classification model has the following hyperparameters: number of decision trees n_estimators=200, maximum number of features max_features=sqrt, minimum number of samples for node splits min_samples_split=2, bootstrap=False, class weight=balanced, maximum depth of decision trees max_depth=None, and minimum number of samples for leaf nodes min_samples_leaf=2.

5. The intelligent classification method for sagittal plane morphology of the spine according to claim 1, characterized in that, In step four: The sagittal morphology of the spine is classified into four categories based on the position of the horizontal lumbar spine line relative to the lumbar vertebral body and intervertebral space: Type I, HLL is located below the L4 / 5 intervertebral space; Type II, HLL passes through the L4 vertebral body and the L4 / 5 intervertebral space; Type III, HLL passes through the L3 vertebral body and the L3 / 4 intervertebral space; Type IV, HLL is located above the L3 upper endplate.

6. The intelligent classification method for spinal sagittal plane morphology based on multi-scale feature extraction according to claim 5, characterized in that: The training process of the random forest model uses a weighted cross-entropy loss function.

7. A system for implementing the intelligent classification method for spinal sagittal plane morphology based on multi-scale feature extraction as claimed in claim 1, characterized in that: The intelligent classification system for spinal sagittal morphology includes an image preprocessing module, a feature extraction module, a feature filtering module, a data optimization module, and a random forest classification module, which are connected in sequence and work together to complete the classification of spinal sagittal morphology. The image preprocessing module and the feature extraction module are used to perform image preprocessing and feature extraction in step one; The feature filtering module is used to perform the key feature filtering in step two. The data optimization module is used to perform data balancing and optimization in step three. The random forest classification module is used to perform the random forest classification in step four.

Images