Fracture risk prediction method based on bone trabecula extraction features

Abstract

The invention relates to the technical field of medical diagnosis, and discloses a fracture risk prediction method based on bone trabecula extraction features, and the method comprises the steps: obtaining two-dimensional image data of a bone of a patient, and carrying out the preprocessing; separating the bone trabecula from the background by using an improved image segmentation algorithm; reconstructing a three-dimensional structure of the bone trabecula, extracting features of the bone trabecula and collecting clinical index data of a patient; screening bone trabecula characteristics and clinical indexes; constructing a fracture risk prediction model, and evaluating and verifying the model; scanning to obtain a normal position spine image of the patient and obtaining a TBS value; and combining the TBS value with a probability value output by the fracture risk prediction model, performing comprehensive evaluation on the patient, and determining a fracture risk level. According to the method, the bone trabecula image features and clinical indexes can be effectively combined, and more accurate and more comprehensive fracture risk prediction is realized through an improved image processing technology and an optimized machine learning model.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis technology, and particularly to a fracture risk prediction method based on trabecular bone feature extraction. Background Art

[0002] Currently, the assessment of fracture risk mainly relies on bone density measurement and the assessment of clinical risk factors. Although these methods can provide a preliminary judgment of fracture risk to a certain extent, they have obvious limitations. For example, bone density measurement can only reflect the mineral content of bones and cannot directly reflect the microscopic structural characteristics of bones. These microscopic structural characteristics are crucial for the assessment of fracture risk because they directly affect the mechanical properties and fracture resistance of bones. In addition, existing assessment methods often ignore the comprehensive association between trabecular bone structural characteristics and clinical indicators, resulting in inaccurate and incomplete assessment results.

[0003] In the field of image processing, although various image segmentation and feature extraction methods have been applied to medical image analysis, their application in fracture risk assessment still has deficiencies. Traditional image segmentation methods, such as fixed threshold segmentation, often cannot adapt to the gray-scale distribution differences of bone images of different patients, resulting in unsatisfactory segmentation effects. In addition, most existing feature extraction methods only focus on a single type of feature, such as morphological features or texture features, and ignore the comprehensive analysis of multiple features. This makes it impossible to fully utilize the rich information in trabecular bone images in fracture risk assessment, thus limiting the accuracy of the assessment.

[0004] In the field of machine learning, although models such as support vector machines have been used for medical image classification and risk prediction, existing methods still face challenges in dealing with fracture risk prediction. For example, existing support vector machine models perform poorly in dealing with class imbalance problems, resulting in insufficient recognition ability of the model for high-risk samples. In addition, the output of existing models usually lacks an intuitive risk level classification, making it difficult for clinicians to give effective intervention and treatment suggestions based on the model results.

[0005] Generally speaking, existing fracture risk assessment methods have deficiencies in image processing, feature extraction, and machine learning model construction. These deficiencies lead to the inability of existing assessment methods to comprehensively and accurately reflect fracture risk. Therefore, the present invention proposes a fracture risk prediction method based on trabecular bone feature extraction. Summary of the Invention

[0006] The purpose of the present invention is to propose a fracture risk prediction method based on trabecular bone feature extraction to solve the problems of insufficient accuracy, incomplete feature extraction, and poor image segmentation effect in the prior art.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A fracture risk prediction method based on trabecular bone feature extraction, comprising the following steps: Step S1, obtaining two-dimensional image data of a patient's bone through an X-ray imaging device, and preprocessing the obtained two-dimensional image data; Step S2, applying an improved image segmentation algorithm, combining local threshold and global threshold, and separating trabecular bone from the background through an adaptive threshold segmentation method; Step S3, applying a multi-view image reconstruction technique to reconstruct the three-dimensional structure of trabecular bone from the two-dimensional image, and extracting trabecular bone features and collecting clinical index data of the patient; Step S4, combining a feature selection method based on statistics and a feature selection method based on a model to screen trabecular bone features and clinical indexes; Step S5, using an improved support vector machine to construct a fracture risk prediction model, and evaluating and validating the model by using a nested cross-validation method; Step S6, obtaining a frontal spine image of the patient by dual-energy X-ray absorptiometry scanning, inputting it into a TBS evaluation system for TBS analysis, and obtaining a TBS value; Step S7, combining the TBS value with the probability value output by the fracture risk prediction model to comprehensively evaluate the patient and determine the fracture risk level.

[0008] Further, in step S1, the following sub-steps are further included: S1-1, scanning bone tissue through an X-ray imaging device to obtain a two-dimensional projection image of the bone tissue; S1-2, performing denoising processing on the obtained two-dimensional image data, and using a Gaussian filter to reduce noise in the image; S1-3, enhancing the contrast of the denoised image, adjusting the gray value range of the image through an adaptive contrast enhancement method, increasing the contrast of the image, and highlighting the structural features of trabecular bone; S1-4, performing edge sharpening on the contrast-enhanced image, highlighting the edge information in the image through an edge enhancement algorithm, and enhancing the boundary features of trabecular bone.

[0009] Further, in step S2, the following sub-steps are further included: S2-1, calculating the global threshold of the image, determining the basic segmentation threshold based on the overall gray distribution of the image, and the specific formula is: ; wherein, is the global threshold, is the gray value of the i-th pixel in the image, and N is the total number of pixels in the image; S2-2. Calculate the local threshold for each pixel. Based on the neighborhood information of the pixel, adjust the threshold through an adaptive method to adapt to the gray-scale changes in different regions of the image. The specific formula is: ; where is the local threshold for each pixel, is the weight factor, is the pixel the average gray-scale value within the neighborhood; S2-3. Combine the global threshold and the local threshold to generate the final segmentation threshold for distinguishing the trabecular bone region and the background region. The specific formula is: ; where is the final segmentation threshold, is the weight factor used to balance the influence of the local and global thresholds; S2-4. Binarize the image using the generated segmentation threshold to separate the trabecular bone region from the background region; S2-5. Perform morphological processing on the segmented trabecular bone region to refine the boundary of the trabecular bone region. The morphological processing includes dilation and erosion.

[0010] Furthermore, in step S3, the following sub-steps are also included: S3-1. Reconstruct the three-dimensional structure of the trabecular bone from the two-dimensional X-ray image through multi-view image reconstruction technology. The specific formula is: ; where represents the pixel value at the coordinate in the three-dimensional space, E represents the number of views, represents the pixel value of the two-dimensional X-ray image of the f-th view at the coordinate , represents the weight of the f-th view used to balance the contributions of different views; S3-2. Extract the morphological features and structural features of the trabecular bone from the three-dimensional structure image of the trabecular bone. The morphological features include trabecular bone thickness, trabecular bone number, and trabecular bone separation, and the structural features include trabecular bone connectivity density and anisotropy; S3-3. Collect the clinical index data of the patient. The clinical index data includes age, gender, body mass index, vitamin D level, blood glucose level, blood lipid level, thyroid function index, renal function index, liver function index, and inflammatory markers; S3-4. Standardize the extracted trabecular bone features and clinical index data to convert all feature values to the same dimension range.

[0011] Furthermore, in step S4, the following sub-steps are further included: S4-1. Apply Pearson correlation analysis to calculate the correlation coefficients between each feature and the fracture risk, and screen out the trabecular bone features and clinical indicators with the absolute value of the correlation coefficient greater than 0.5. The calculation formula for the correlation coefficient between each feature and the fracture risk is: ; where r represents the correlation coefficient between each feature and the fracture risk, is the feature value of the u-th sample, is the fracture risk value of the u-th sample, and are the average values of the feature value and the fracture risk value; S4-2. Through the Lasso regression algorithm, screen out the trabecular bone features and clinical indicators corresponding to the non-zero coefficients in the Lasso regression model. The objective function of the Lasso regression is: ; where, () represents taking the minimum value of, is the regression coefficient vector, m is the number of samples, is the fracture risk value of the u-th sample, is the feature value of the u-th sample, is the transpose of, is the regularization parameter, p is the number of features, is the sum of squared residuals, representing the difference between the model prediction value and the actual value, is the sum of the absolute values of the regression coefficients, used to achieve feature selection.

[0012] Furthermore, in step S5, the following sub-steps are further included: S5-1. Select the radial basis function as the kernel function and determine the parameters of the kernel function through the grid search method. The parameters include the value and the regularization parameter C; S5-2. Introduce a class weight adjustment mechanism to assign weights to samples of different classes to solve the class imbalance problem; S5-3. Build a fracture risk prediction model, and use the improved support vector machine model for the prediction of fracture risk. The output of the model is the probability of fracture risk; S5-4. Randomly divide the training data set into K subsets of equal size. For each outer cross-validation, select one subset as the outer validation set, and the remaining K-1 subsets as the inner training set; S5-5. In the inner training set, use M-fold cross-validation to optimize the model parameters, where the model parameters include the parameters of the kernel function and the regularization parameter C; S5-6. Retrain the model on the inner training set using the optimized parameters, and evaluate the model performance by calculating the performance metrics of the model on the current outer validation set. The performance metrics include accuracy, recall, and F1 value. The specific formulas are as follows: ; ; ; where Accuracy represents accuracy, Recall represents recall, F1 represents the F1 value, TP is the number of true positive samples, which is the number of high-risk samples correctly identified by the model, TN is the number of true negative samples, which is the number of low-risk samples correctly identified by the model, FP is the number of false positive samples, which is the number of low-risk samples misidentified as high-risk by the model, and FN is the number of false negative samples, which is the number of high-risk samples misidentified as low-risk by the model; S5-7. Repeat the above steps S5-4 to S5-6 for K times. Each time, select a different subset as the outer validation set and calculate the average performance metrics of the model on all outer validation sets.

[0013] Furthermore, in step S6, the following sub-steps are also included: S6-1. Use a dual-energy X-ray absorptiometry device to scan the patient and obtain a frontal spine image; S6-2. Input the frontal spine image into the TBS evaluation system, perform multi-scale texture analysis on the frontal spine image, and extract texture features. The texture features include gray-level co-occurrence matrix, wavelet transform features, and local binary pattern; S6-3. Fusion the extracted texture features to generate a comprehensive texture feature vector and calculate the TBS value.

[0014] Furthermore, in step S7, the following sub-steps are also included: S7-1. Input the extracted trabecular bone features and clinical index data into the fracture risk prediction model for prediction and output the fracture risk probability value; S7-2. Normalize the obtained TBS value and fracture risk probability value to convert them to the same dimension range; S7-3. Use Bayesian fusion to combine the TBS value and fracture risk probability value and calculate the comprehensive risk score. The specific formula is as follows: ; where, represents the comprehensive risk score, P(FR) represents the probability value output by the fracture risk prediction model, and P(TBS) represents the probability value corresponding to the TBS value. and are the weights of P(FR) and P(TBS) respectively, and ; S7-4, convert the comprehensive risk score into a comprehensive probability value between 1 and 1.5 through a probability calibration method; S7-5, divide the fracture risk into high-risk level, medium-risk level and low-risk level according to the comprehensive probability value. The high-risk level is that the comprehensive probability value is less than or equal to 1.23, the medium-risk level is that the comprehensive probability value is greater than 1.23 but less than 1.31, and the low-risk level is that the comprehensive probability value is greater than or equal to 1.31; S7-6, generate a prediction report according to the comprehensive probability value and the fracture risk level. The prediction report includes patient basic information, comprehensive probability value, fracture risk level, feature analysis and preventive measure suggestions.

[0015] The beneficial effects brought by the technical solution provided by the present invention at least include: By combining trabecular bone image features and clinical index data, and using an improved image segmentation algorithm and an optimized feature selection method, the present invention can more comprehensively reflect the fracture risk and improve the accuracy of fracture risk prediction.

[0016] The present invention has made multiple improvements in image processing and feature extraction. By combining an adaptive segmentation method of local and global thresholds, it can better adapt to the gray distribution differences of bone images of different patients, and significantly improve the accuracy and robustness of image segmentation.

[0017] The present invention uses an improved support vector machine to construct a fracture risk prediction model and introduces a class weight adjustment mechanism, effectively solving the class imbalance problem, enabling clinicians to more intuitively understand the prediction results, and thus formulating more effective intervention measures.

[0018] The present invention can not only output the fracture risk probability and risk level, but also generate a detailed prediction report, which can help clinicians better understand the fracture risk of patients and provide personalized preventive measure suggestions according to the risk level and key features. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 This is the flowchart of the method provided by the embodiment of the present invention. Detailed implementation manners

[0021] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, elaborate in detail on a fracture risk prediction method based on trabecular bone feature extraction proposed according to the present invention, its specific implementation manners, structures, features and effects. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0023] The following embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0024] The following will specifically describe the specific solution of a fracture risk prediction method based on trabecular bone feature extraction provided by the present invention with reference to the accompanying drawings.

[0025] Please refer to Figure 1 , which shows the flowchart of a fracture risk prediction method based on trabecular bone feature extraction provided by an embodiment of the present invention. The method includes the following steps: Step S1, obtaining two-dimensional image data of a patient's bone through an X-ray imaging device and preprocessing the obtained two-dimensional image data; Among them, in step S1, the following sub-steps are further included: S1-1, scanning the bone tissue through an X-ray imaging device to obtain a two-dimensional projection image of the bone tissue; S1-2, performing denoising processing on the obtained two-dimensional image data and using a Gaussian filter to reduce the noise in the image; S1-3, enhancing the contrast of the denoised image, adjusting the gray value range of the image through an adaptive contrast enhancement method, increasing the contrast of the image, and highlighting the structural features of the trabecular bone; S1-4, performing edge sharpening on the contrast-enhanced image, highlighting the edge information in the image through an edge enhancement algorithm, and enhancing the boundary features of the trabecular bone.

[0026] It should be noted that an X-ray imaging device is a medical imaging tool that uses X-rays to penetrate an object and record its projection image. By emitting an X-ray beam, it penetrates the human body or other objects and then forms a two-dimensional projection image on the detector.

[0027] The Gaussian filter is a linear filter used for image denoising. It reduces noise by performing a weighted average of each pixel in the image with the pixels in its neighborhood. The weights of the Gaussian filter are determined by the Gaussian function, with the weight of the central pixel being the largest and the weights of the neighboring pixels decreasing as the distance increases.

[0028] The adaptive contrast enhancement method is a technique for dynamically adjusting the contrast of an image. It automatically adjusts the grayscale value range according to the local features of the image to enhance the contrast of the image. It is particularly suitable for processing local regions in the image and can better highlight the structural features of interest. For trabecular bone images, special attention can be paid to enhancing the contrast of the trabecular bone region.

[0029] The edge enhancement algorithm is a technique for highlighting edge information in an image, including the Laplacian operator or the Sobel operator. The algorithm enhances edges by calculating the gradient of the image. The Laplacian operator calculates the second derivative of the image and can highlight the edges in the image; the Sobel operator enhances the edge information by calculating the gradient magnitude of the image.

[0030] Step S2: Apply an improved image segmentation algorithm, combine local threshold and global threshold, and separate the trabecular bone from the background through an adaptive threshold segmentation method. Among them, in step S2, the following sub-steps are also included: S2-1: Calculate the global threshold of the image and determine the basic segmentation threshold based on the overall grayscale distribution of the image. The specific formula is: ; where, is the global threshold, is the grayscale value of the i-th pixel in the image, and N is the total number of pixels in the image; S2-2: Calculate the local threshold of each pixel, and adjust the threshold adaptively based on the neighborhood information of the pixel to adapt to the grayscale changes in different regions of the image. The specific formula is: ; where, is the local threshold of each pixel, is the weight factor, is the pixel the average grayscale value within the neighborhood; S2-3: Combine the global threshold and the local threshold to generate the final segmentation threshold for distinguishing the trabecular bone region and the background region. The specific formula is: ; where, is the final segmentation threshold, is a weight factor used to balance the influence of local and global thresholds; S2-4, Binarize the image using the generated segmentation threshold to separate the trabecular bone region from the background region; S2-5, Perform morphological processing on the segmented trabecular bone region to refine the boundary of the trabecular bone region. The morphological processing includes dilation and erosion.

[0031] It should be noted that the global threshold is calculated based on the overall gray level distribution of the image, which reflects the average gray level of the image. By calculating the average value of all pixel gray values, a basic segmentation threshold can be obtained. This threshold can initially distinguish the target region and the background region in the image, but may not be able to adapt to the gray level changes in different regions of the image.

[0032] The local threshold is calculated based on the neighborhood information of each pixel and can adapt to the gray level changes in different regions of the image. By considering the neighborhood information of the pixel, the local characteristics in the image can be better processed. The weight factor is used to adjust the influence degree of the neighborhood information to ensure that the local threshold can effectively reflect the local gray level changes.

[0033] The final segmentation threshold is the weighted sum of the global threshold and the local threshold. The weight factor is used to balance the influence of the global and local thresholds to ensure that the segmentation threshold can comprehensively consider the overall and local characteristics of the image. In this way, the trabecular bone region can be more accurately segmented, and good results can be obtained even in images with uneven gray level distributions.

[0034] Binarization is the process of converting the pixel gray values in the image into binary values (0 or 1). According to the final segmentation threshold, the pixels with gray values higher than the threshold are classified into the trabecular bone region, and the pixels with gray values lower than the threshold are classified into the background region. Through binarization, the trabecular bone region can be clearly separated from the background.

[0035] Morphological processing is a processing method based on the shape of the image and is used to refine and optimize the segmentation result. In the present invention, dilation and erosion operations are used to refine the boundary of the trabecular bone region. The dilation operation can fill the small holes in the trabecular bone region, and the erosion operation can remove the small burrs and noises on the boundary. Through morphological processing, the boundary of the trabecular bone region can be further optimized to make it smoother and more accurate.

[0036] Step S3, Apply multi-view image reconstruction technology to reconstruct the three-dimensional structure of the trabecular bone from the two-dimensional image, and extract trabecular bone features and collect clinical index data of the patient; Among them, in step S3, the following sub-steps are further included: S3-1. Reconstruct the three-dimensional structure of trabecular bone from two-dimensional X-ray images through multi-view image reconstruction technology. The specific formula is as follows: ; where represents the pixel value at the coordinate in the three-dimensional space, E represents the number of views, represents the pixel value of the two-dimensional X-ray image of the f-th view at the coordinate , and represents the weight of the f-th view, which is used to balance the contributions of different views; S3-2. Extract the morphological and structural features of trabecular bone from the three-dimensional structure image of trabecular bone. The morphological features include trabecular bone thickness, trabecular bone number, and trabecular bone separation. The structural features include trabecular bone connectivity density and anisotropy; S3-3. Collect the clinical index data of the patient. The clinical index data includes age, gender, body mass index, vitamin D level, blood glucose level, blood lipid level, thyroid function index, renal function index, liver function index, and inflammatory markers; S3-4. Standardize the extracted trabecular bone features and clinical index data to convert all feature values into the same dimension range.

[0037] It should be noted that the multi-view image reconstruction technology is a technology that converts two-dimensional image data from multiple views into a three-dimensional model and has a wide range of applications in the field of medical imaging.

[0038] Trabecular bone thickness: Reflects the thickness of trabecular bone and is an important indicator for measuring bone strength. Thicker trabecular bone usually has higher anti-fracture ability.

[0039] Trabecular bone number: Refers to the number of trabecular bones per unit area and reflects the density of the bone. The more trabecular bones, the stronger the supporting ability of the bone.

[0040] Trabecular bone separation: Describes the spacing between trabecular bones and reflects the distribution uniformity of trabecular bones. Higher separation may mean osteoporosis and increase the risk of fracture.

[0041] Trabecular bone connectivity density: Reflects the connection degree between trabecular bones and is a key indicator for measuring the overall stability of the bone. Higher connectivity density means a stronger bone structure.

[0042] Anisotropy: Describes the distribution difference of trabecular bones in different directions and reflects the mechanical properties of the bone. Higher anisotropy may mean that the bone is more likely to fracture in certain directions.

[0043] Age: Age is an important factor in fracture risk. As people age, the fracture risk increases significantly.

[0044] Gender: Gender also has an impact on fracture risk. Women are generally more prone to fractures than men, especially postmenopausal women.

[0045] Body Mass Index (BMI): BMI reflects the ratio of weight to height. A lower BMI may be associated with osteoporosis and an increased risk of fractures.

[0046] Vitamin D level: Vitamin D is crucial for calcium absorption and bone health. Low levels of vitamin D may increase the fracture risk.

[0047] Blood glucose level: High blood glucose may affect bone health and increase the fracture risk.

[0048] Blood lipid level: High blood lipids may affect the microvascular supply of bones and indirectly influence the fracture risk.

[0049] Thyroid function indicators: Abnormal thyroid function may affect bone metabolism and increase the fracture risk.

[0050] Renal function indicators: Renal insufficiency may affect the metabolism of calcium and phosphorus, and thus affect bone health.

[0051] Liver function indicators: Liver diseases may affect the metabolism of vitamin D and indirectly influence the fracture risk.

[0052] Inflammatory markers: Chronic inflammation may affect bone health and increase the fracture risk.

[0053] Standardization is to adjust the value ranges of different features to a unified standard range; because different features have different dimensions, directly using these features may cause some features to dominate numerically, and the dimensions and ranges of feature values may vary greatly, which will affect the performance and training efficiency of the model.

[0054] Step S4, combine the feature selection method based on statistics and the feature selection method based on the model to screen the trabecular bone features and clinical indicators; Among them, in step S4, the following sub-steps are also included: S4-1, apply Pearson correlation analysis to calculate the correlation coefficients between each feature and the fracture risk, and screen out the trabecular bone features and clinical indicators whose absolute values of the correlation coefficients are greater than 0.5. The calculation formula for the correlation coefficient between each feature and the fracture risk is: ; where r represents the correlation coefficient between each feature and the fracture risk, is the feature value of the u-th sample, is the fracture risk value of the u-th sample, and is the average of the eigenvalue and the fracture risk value; S4-2, through the Lasso regression algorithm, screen out the trabecular bone features and clinical indicators corresponding to the non-zero coefficients in the Lasso regression model. The objective function of the Lasso regression is: ; Among them, () represents taking the minimum value of, is the regression coefficient vector, m is the number of samples, is the fracture risk value of the u-th sample, is the eigenvalue of the u-th sample, is the transpose of, is the regularization parameter, p is the number of features, is the sum of squared residuals, representing the difference between the model predicted value and the actual value, is the sum of the absolute values of the regression coefficients, used to achieve feature selection.

[0055] It should be noted that Pearson correlation analysis is a feature selection method based on statistics, used to measure the linear correlation between two variables. By calculating the correlation coefficients between each feature and the fracture risk, features highly correlated with the fracture risk can be screened out. Features with an absolute value of the correlation coefficient greater than 0.5 are considered to have a strong linear relationship with the fracture risk and are therefore retained for subsequent analysis.

[0056] Lasso regression is a model-based feature selection method. By introducing an L1 regularization term in the regression model, the regression coefficients of some unimportant features can be compressed to zero, thus achieving feature selection; the objective function of Lasso regression includes two parts: one part is the sum of squared residuals, used to measure the fitting effect of the model; the other part is the regularization term, used to control the number of features. By adjusting the regularization parameter , the complexity and fitting effect of the model can be balanced. Only those features with non-zero corresponding regression coefficients are considered to have a significant impact on the fracture risk and are retained for subsequent analysis. This method can not only select important features but also improve the sparsity and interpretability of the model.

[0057] Step S5, use the improved support vector machine to construct a fracture risk prediction model, and use the nested cross-validation method to evaluate and validate the model; Among them, in step S5, the following sub-steps are also included: S5-1, select the radial basis function as the kernel function, and determine the parameters of the kernel function through the grid search method. The parameters include The value and the regularization parameter C; S5-2. Introduce a class weight adjustment mechanism to assign weights to samples of different classes to solve the problem of class imbalance; S5-3. Build a fracture risk prediction model, and use the improved support vector machine model for fracture risk prediction. The output of the model is the probability of fracture risk; S5-4. Randomly divide the training data set into K subsets of equal size. For each outer cross-validation, select one subset as the outer validation set, and the remaining K-1 subsets as the inner training set; S5-5. In the inner training set, use M-fold cross-validation to optimize the model parameters. The model parameters include the parameters of the kernel function and the regularization parameter C; S5-6. Retrain the model on the inner training set using the optimized parameters, and evaluate the model performance on the current outer validation set by calculating the performance metrics of the model. The performance metrics include accuracy, recall, and F1 value. The specific formulas are as follows: ; ; ; where Accuracy represents accuracy, Recall represents recall, F1 represents the F1 value, TP is the number of true positive samples, which is the number of high-risk samples correctly identified by the model, TN is the number of true negative samples, which is the number of low-risk samples correctly identified by the model, FP is the number of false positive samples, which is the number of low-risk samples misidentified as high-risk by the model, and FN is the number of false negative samples, which is the number of high-risk samples misidentified as low-risk by the model; S5-7. Repeat the above steps S5-4 to S5-6 for K times, each time selecting a different subset as the outer validation set, and calculate the average performance metrics of the model on all outer validation sets.

[0058] It should be noted that the support vector machine (SVM) is a powerful classification and regression algorithm, which is widely used in medical image analysis and risk prediction. In fracture risk prediction, SVM can effectively process high-dimensional feature data and handle non-linear relationships through kernel functions. In this invention, by improving the SVM model and introducing a class weight adjustment mechanism and a probability calibration method, the performance and reliability of the model are further improved.

[0059] The radial basis function (RBF) kernel function is a commonly used non-linear kernel function, which can effectively handle non-linear relationships in the feature space. By selecting the RBF kernel function, the SVM model can better capture the complex relationships between trabecular bone features and clinical indicators.

[0060] The present invention adopts the grid search method to determine the parameters of the kernel function and regularization parameter C. The grid search method finds the parameter combination that optimizes the model performance by performing an exhaustive search within the predefined parameter range. This method can ensure that the model achieves the best performance on the training data.

[0061] In actual medical data, there is often an imbalance in the categories of fracture risk, that is, the number of high-risk samples is far less than that of low-risk samples. This category imbalance will cause the model to favor the majority class, thereby reducing the ability to recognize the minority class. The present invention introduces a category weight adjustment mechanism. By assigning weights to samples of different categories, the model will pay more attention to minority class samples during training, thereby improving the ability to recognize high-risk samples.

[0062] After determining the kernel function parameters and category weights, the improved SVM model was used to construct a fracture risk prediction model. The improved SVM model can predict fracture risk more accurately. The output of the model is a probability value, which indicates the probability that the sample belongs to the high-risk category. The output of the probability value makes the prediction results of the model more intuitive and easy to understand and apply.

[0063] Nested cross-validation is a model evaluation method that is widely used in machine learning and statistical modeling. By dividing the dataset into multiple subsets and performing training and validation on different subsets, it can effectively evaluate the performance and generalization ability of the model. Nested cross-validation is particularly suitable for model parameter optimization and model selection. It can simultaneously evaluate the training and validation effects of the model to avoid overfitting and underfitting problems.

[0064] The first step of nested cross-validation is to randomly divide the training dataset into K subsets of equal size. The subsets are called "folds". The choice of K depends on the size of the dataset and the computing resources. By randomly dividing, it can be ensured that each subset is statistically representative, thereby improving the reliability of model evaluation.

[0065] In nested cross-validation, the outer cross-validation is used to evaluate the final performance of the model, while the inner cross-validation is used to optimize the model parameters. For each outer cross-validation, one subset is selected as the outer validation set, and the remaining K-1 subsets are used as the inner training set. This ensures that in each iteration, the model is trained and validated on different data sets, thereby improving the stability and reliability of model evaluation.

[0066] Inner cross-validation is used to optimize model parameters, including kernel function parameters in support vector machines With the regularization parameter C, in the inner training set, the data is further divided into M subsets, and M-fold cross-validation is used to evaluate the model performance under different parameter combinations. Through the grid search method, the parameter combination that optimizes the model performance can be found. This method can effectively avoid overfitting and improve the generalization ability of the model.

[0067] Accuracy represents the proportion of samples correctly classified by the model and is a basic indicator for evaluating model performance.

[0068] Recall represents the proportion of high-risk samples correctly identified by the model and is particularly suitable for class imbalance problems.

[0069] The F1 value is the harmonic mean of precision and recall, which can comprehensively evaluate the model performance and is particularly suitable for class imbalance problems.

[0070] To ensure the stability and reliability of model evaluation, the above steps from S5-4 to S5-6 need to be repeated K times. Each time, a different subset is selected as the outer validation set. By calculating the average performance metrics of the model on all outer validation sets, a more comprehensive and reliable model performance evaluation result can be obtained, which can effectively reduce the fluctuations in evaluation results caused by different data partitions and improve the credibility of model evaluation.

[0071] Step S6: Obtain the anteroposterior spine image of the patient through dual-energy X-ray absorptiometry scanning, input it into the TBS evaluation system for TBS analysis, and obtain the TBS value; Among them, in step S6, the following sub-steps are also included: S6-1: Use the dual-energy X-ray absorptiometry device to scan the patient to obtain the anteroposterior spine image; S6-2: Input the anteroposterior spine image into the TBS evaluation system, perform multi-scale texture analysis on the anteroposterior spine image, and extract texture features. The texture features include gray-level co-occurrence matrix, wavelet transform features, and local binary pattern; S6-3: Fuse the extracted texture features to generate a comprehensive texture feature vector and calculate the TBS value.

[0072] It should be noted that TBS is the trabecular bone score, a unitless indicator used to quantify the heterogeneity of trabecular bone microstructure in the lateral X-ray image of the spine. By analyzing the texture features of the image, the density and structure of trabecular bone are evaluated to predict the fracture risk; the lower the TBS value, the worse the trabecular bone microstructure and the higher the fracture risk.

[0073] Dual-energy X-ray absorptiometry (DXA) is a commonly used bone density measurement technique that can generate high-quality spinal images. It can be used to measure bone density. In this step, the anteroposterior spinal image can clearly show the trabecular bone structure of the spine, providing basic data for subsequent TBS analysis.

[0074] The TBS assessment system is a software tool specifically used to analyze the trabecular bone structure. In this step, the anteroposterior spinal image is input into the TBS assessment system for multi-scale texture analysis to extract various texture features that can reflect the microstructure and quality of the trabecular bone, which is the basis for calculating the TBS value.

[0075] Multi-scale texture analysis: Through multi-scale analysis, texture features of the image are extracted from different scales to more comprehensively reflect the microstructure of the trabecular bone. Multi-scale analysis can capture texture changes at different scales, improving the accuracy and reliability of the TBS value.

[0076] Gray-level co-occurrence matrix (GLCM): It reflects the spatial distribution relationship of pixel gray values in the image and can extract texture features of the image, including contrast, uniformity, and correlation.

[0077] Wavelet transform features: Multi-scale texture features of the image are extracted through wavelet transform, which can capture details and texture changes in the image.

[0078] Local binary pattern (LBP): Texture features of the image are extracted through local binary pattern, which can reflect the local texture structure of the image.

[0079] After extracting various texture features, these features need to be fused to generate a comprehensive texture feature vector that can more comprehensively reflect the microstructure and quality of the trabecular bone. Through the comprehensive texture feature vector, the TBS value can be calculated.

[0080] Feature fusion: The gray-level co-occurrence matrix, wavelet transform features, and local binary pattern are fused to generate a comprehensive texture feature vector. Feature fusion can be achieved through weighted average, feature splicing, or deep learning models.

[0081] TBS value calculation: Machine learning algorithms are used to analyze the comprehensive texture feature vector to calculate the TBS value. The TBS value is an index that comprehensively reflects the quality of the trabecular bone structure and can provide additional information about fracture risk.

[0082] In step S7, the TBS value is combined with the probability value output by the fracture risk prediction model to comprehensively evaluate the patient and determine the fracture risk level. Among them, in step S7, the following sub-steps are also included: S7-1. Input the extracted trabecular bone features and clinical index data into the fracture risk prediction model for prediction, and output the fracture risk probability value. S7-2. Normalize the obtained TBS value and fracture risk probability value to convert them to the same dimension range. S7-3. Use Bayesian fusion to combine the TBS value and fracture risk probability value to calculate the comprehensive risk score. The specific formula is: ; where, represents the comprehensive risk score, P(FR) represents the probability value output by the fracture risk prediction model, P(TBS) represents the probability value corresponding to the TBS value, and are the weights of P(FR) and P(TBS) respectively, and ; S7-4. Convert the comprehensive risk score to a comprehensive probability value between 1 and 1.5 through the probability calibration method. S7-5. Divide the fracture risk into high-risk level, medium-risk level and low-risk level according to the comprehensive probability value. The high-risk level is that the comprehensive probability value is less than or equal to 1.23, the medium-risk level is that the comprehensive probability value is greater than 1.23 but less than 1.31, and the low-risk level is that the comprehensive probability value is greater than or equal to 1.31. S7-6. Generate a prediction report according to the comprehensive probability value and fracture risk level. The prediction report includes the patient's basic information, comprehensive probability value, fracture risk level, feature analysis and prevention measure suggestions.

[0083] It should be noted that the input data: The input data includes the features extracted from the trabecular bone image and the patient's clinical indexes.

[0084] Model output: The model outputs a probability value between 0 and 1, indicating the risk of the patient having a fracture. The higher the probability value, the greater the fracture risk.

[0085] Normalization processing is a data preprocessing method used to convert data with different dimensions or ranges to the same numerical range, which can eliminate the dimension differences between different features and improve the accuracy of subsequent fusion. Through normalization processing, both the TBS value and the fracture risk probability value are between 0 and 1, facilitating subsequent fusion calculations.

[0086] Bayesian fusion is a method based on Bayes' theorem used to combine multiple independent evidences (probability values) to improve the accuracy of decision-making; this method assumes that each input probability value is independent and Bayes' theorem can be used to update the prior probability.

[0087] The probability calibration method is a method that adjusts the probability values output by the model to a more accurate range. The calibrated probability values are closer to the actual occurrence probability, thereby improving the prediction accuracy of the model.

[0088] Basic patient information: including patient name, age, gender, and medical record number.

[0089] Comprehensive probability value: The comprehensive probability value predicted by the model for the patient to have a fracture, which is between 1 and 1.5, accurate to two decimal places, indicating the risk of the patient having a fracture. The lower the comprehensive probability value, the greater the fracture risk.

[0090] Fracture risk level: The risk level divided according to the comprehensive probability value, including "high risk", "medium risk", and "low risk".

[0091] Feature analysis: List the top five features that contribute the most to risk prediction and their values, and explain how these features affect the fracture risk.

[0092] Suggestions for preventive measures: Provide personalized suggestions for preventive measures based on the risk level and key features. For patients with medium to high risks, the suggestions include increasing the intake of calcium and vitamin D, performing weight-bearing exercises, and avoiding high-risk activities; for patients with low risks, the suggestions include maintaining a healthy lifestyle and having regular check-ups.

[0093] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A fracture risk prediction method based on trabecular bone extraction features, characterized in that: The method includes: Step S1, obtaining two-dimensional image data of a patient's bones through an X-ray imaging device, and preprocessing the obtained two-dimensional image data; Step S2, applying an improved image segmentation algorithm, combining local threshold and global threshold, and separating trabeculae from background by an adaptive threshold segmentation method; Step S3, applying multi-view image reconstruction technology to reconstruct the three-dimensional structure of trabeculae from the two-dimensional image, extracting the characteristics of trabeculae and collecting clinical index data of patients; Step S4, combining a statistical-based feature selection method and a model-based feature selection method to screen trabecular bone characteristics and clinical indicators; Step S5, using the improved support vector machine to construct a fracture risk prediction model, and using the nested cross-validation method to evaluate and validate the model; Step S6, scanning the patient with dual-energy X-ray absorptiometry to obtain an AP spinal image, and inputting the image into a TBS evaluation system for TBS analysis to obtain a TBS value; Step S7, combining the TBS value with the probability value output by the fracture risk prediction model, comprehensively assessing the patient and determining the fracture risk level.

2. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein step S1 also includes the following sub-steps: S1-1, scanning the bone tissue by an X-ray imaging device to obtain a two-dimensional projection image of the bone tissue; S1-2, performing denoising processing on the acquired two-dimensional image data, using a Gaussian filter to reduce the noise in the image; S1-3, contrast enhancement is performed on the denoised image, and the grayscale value range of the image is adjusted by an adaptive contrast enhancement method to increase the contrast of the image and highlight the structural characteristics of the trabecular bone; S1-4, edge sharpening is performed on the contrast-enhanced image, and the edge information in the image is highlighted by an edge enhancement algorithm to enhance the boundary features of the trabeculae.

3. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein, in step S2, the following sub-steps are also included: S2-1, calculate the global threshold of the image and determine the basic segmentation threshold based on the overall grayscale distribution of the image. The specific formula is: ； in, is the global threshold, is the gray value of the i-th pixel in the image, and N is the total number of pixels in the image; S2-2, calculate the local threshold of each pixel, and adjust the threshold through an adaptive method based on the neighborhood information of the pixel to adapt to the grayscale changes in different areas of the image. The specific formula is: ； in, is the local threshold for each pixel, is the weight factor, It's pixels The average gray value in the neighborhood; S2-3, combining the global threshold and the local threshold, generates the final segmentation threshold, which is used to distinguish the trabecular area from the background area. The specific formula is: ； in, is the final segmentation threshold, is a weight factor used to balance the influence of local and global thresholds; S2-4, binarization of the image is performed using the generated segmentation threshold to separate the trabecular bone area from the background area; S2-5, performing morphological processing on the segmented trabecular bone region to refine the boundary of the trabecular bone region, wherein the morphological processing includes expansion and erosion.

4. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein step S3 also includes the following sub-steps: S3-1, the three-dimensional structure of trabecular bone is reconstructed from the two-dimensional X-ray image through multi-view image reconstruction technology. The specific formula is: ； in, Represents coordinates in three-dimensional space The pixel value at , E represents the number of viewing angles, The two-dimensional X-ray image of the f-th viewing angle is in coordinate The pixel value at represents the weight of the f-th perspective, which is used to balance the contributions of different perspectives; S3-2, extracting morphological and structural features of trabeculae from the three-dimensional structural image of trabeculae, wherein the morphological features include trabeculae thickness, trabeculae number and trabeculae separation, and the structural features include trabeculae connection density and anisotropy; S3-3, collecting clinical indicator data of the patient, wherein the clinical indicator data include age, gender, body mass index, vitamin D level, blood sugar level, blood lipid level, thyroid function index, kidney function index, liver function index and inflammatory markers; S3-4, standardize the extracted trabecular bone characteristics and clinical index data, and convert all characteristic values ​​into the same dimension range.

5. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein step S4 also includes the following sub-steps: S4-1, Pearson correlation analysis was used to calculate the correlation coefficient between each feature and fracture risk, and trabecular bone features and clinical indicators with an absolute value of correlation coefficient greater than 0.5 were screened out. The calculation formula for the correlation coefficient between each feature and fracture risk was: ； Among them, r represents the correlation coefficient between each characteristic and fracture risk, is the eigenvalue of the u-th sample, is the fracture risk value of the u-th sample, and is the average of the characteristic value and the fracture risk value; S4-2, using the Lasso regression algorithm, the trabecular characteristics and clinical indicators corresponding to the non-zero coefficients in the Lasso regression model are screened out, and the objective function of the Lasso regression is: ； in, () indicates taking The minimum value of is the regression coefficient vector, m is the sample size, is the fracture risk value of the u-th sample, is the eigenvalue of the u-th sample, Yes, transpose. is the regularization parameter, p is the number of features, is the residual sum of squares, which represents the difference between the model prediction and the actual value, It is the sum of the absolute values ​​of the regression coefficients and is used to implement feature selection.

6. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein, in step S5, the following sub-steps are also included: S5-1, select radial basis function as kernel function, and determine the parameters of kernel function by grid search method, the parameters include value and regularization parameter C; S5-2, introduces a category weight adjustment mechanism to assign weights to samples of different categories to solve the category imbalance problem; S5-3, construct a fracture risk prediction model, use the improved support vector machine model to predict fracture risk, and the output of the model is the probability of fracture risk; S5-4, randomly divide the training data set into K subsets of equal size. For each outer cross-validation, select one subset as the outer validation set, and the remaining K-1 subsets as the inner training sets; S5-5, in the inner training set, M-fold cross validation is used to optimize the model parameters, wherein the model parameters include the parameters of the kernel function and regularization parameter C; S5-6, retrain the model on the inner training set using the optimized parameters, and evaluate the model performance by calculating the model's performance indicators on the current outer validation set. The performance indicators include accuracy, recall, and F1 value. The specific formula is: ； ； ； Among them, Accuracy represents accuracy, Recall represents recall, F1 represents F1 value, TP is the number of true positive samples, which is the number of high-risk samples correctly identified by the model, TN is the number of true negative samples, which is the number of low-risk samples correctly identified by the model, FP is the number of false positive samples, which is the number of low-risk samples mistakenly identified as high-risk by the model, and FN is the number of false negative samples, which is the number of high-risk samples mistakenly identified as low-risk by the model; S5-7, repeat the above steps S5-4 to S5-6, K times, each time selecting a different subset as the outer validation set, and calculating the average performance index of the model on all outer validation sets.

7. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein, in step S6, the following sub-steps are also included: S6-1, scan the patient using dual-energy X-ray absorptiometry equipment to obtain anteroposterior spinal images; S6-2, inputting the anteroposterior spinal image into the TBS evaluation system, performing multi-scale texture analysis on the anteroposterior spinal image, and extracting texture features, wherein the texture features include gray-level co-occurrence matrix, wavelet transform features, and local binary patterns; S6-3, the extracted texture features are fused to generate a comprehensive texture feature vector and the TBS value is calculated.

8. The fracture risk prediction method based on trabecular bone extraction feature according to claim 1, characterized in that: Wherein, in step S7, the following sub-steps are also included: S7-1, inputting the extracted trabecular bone characteristics and clinical index data into the fracture risk prediction model for prediction, and outputting the fracture risk probability value; S7-2, normalize the obtained TBS values ​​and fracture risk probability values ​​and convert them into the same dimension range; S7-3, Bayesian fusion was used to combine TBS value and fracture risk probability value to calculate the comprehensive risk score. The specific formula is: ； in, represents the comprehensive risk score, P(FR) represents the probability value output by the fracture risk prediction model, and P(TBS) represents the probability value corresponding to the TBS value. and are the weights of P(FR) and P(TBS), respectively, and ; S7-4, convert the comprehensive risk score into a comprehensive probability value between 1 and 1.5 through a probability calibration method; S7-5, dividing the fracture risk into high risk, medium risk and low risk levels according to the comprehensive probability value, wherein the high risk level is a comprehensive probability value less than or equal to 1.23, the medium risk level is a comprehensive probability value greater than 1.23 but less than 1.31, and the low risk level is a comprehensive probability value greater than or equal to 1.31; S7-6, generating a prediction report based on the comprehensive probability value and the fracture risk level, wherein the prediction report includes basic information of the patient, the comprehensive probability value, the fracture risk level, feature analysis and preventive measures recommendations.

Images