KOA risk prediction method based on automated measurement of body composition by CT examination

Abstract

The present invention provides a KOA risk prediction method based on CT examination to automatically measure physical constitution, including: obtaining CT data sets of healthy people and KOA patients respectively; preprocessing the CT data sets using a mobile image enhancement window to obtain a preprocessed data set; using the preprocessed data set to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information; automatically obtaining physical parameters using the segmentation information; establishing a sample data set according to the physical parameters and preset clinical data, and extracting features from the sample data set; training the preset model using feature values ​​to obtain a KOA risk prediction model; using the KOA risk prediction model to predict the KOA risk of the population to be tested to obtain a prediction result. The present invention can construct a risk prediction system for KOA that integrates human physical constitution and clinical factors through a machine learning prediction model, achieve accurate prediction of the risk of KOA, and provide important data reference for the prevention and treatment of clinical KOA.

Description

Technical Field

[0001] The present invention relates to the technical field of knee osteoarthritis prediction, and in particular to a KOA risk prediction method based on CT examination and automatic measurement of body composition. Background Art

[0002] As the global population ages, age-related musculoskeletal diseases have become a hot topic of research at home and abroad. Knee osteoarthritis (KOA) is a multifactorial disease of the whole joint, and its occurrence is related to age, obesity, inflammation, trauma, and genetic factors. The main pathological characteristics of KOA are cartilage erosion, subchondral bone remodeling, osteophyte formation, and chronic low-grade inflammation of the synovium. It mainly involves articular cartilage, subchondral bone, joint capsule and ligaments, synovium, and surrounding muscles. It is one of the most common degenerative diseases that cause disability in the elderly. KOA not only brings great pain and inconvenience to patients, but also brings huge medical and economic burdens to patients' families and society. At present, existing medical technologies have not yet been able to cure KOA. The main treatment principle is to alleviate symptoms and delay the progression of the disease, but more severe KOA can only be treated with joint replacement. However, it also faces greater surgical risks, the life of the prosthesis is limited, and patients will still have a certain degree of joint stiffness, limited activity, and discomfort after surgery. Therefore, it is still crucial to identify high-risk groups for KOA as early as possible and carry out early prevention to reduce the incidence or progression of KOA. Skeletal muscle, as the basis of joint function, plays a key role in the pathogenesis of KOA. A deeper understanding of the changes in skeletal muscle structure and function and its mechanism of action in KOA patients will help to carry out targeted intervention measures at an early stage. The research team has previously constructed an automated segmentation technology for skeletal muscle CT images based on deep learning, which provides an effective tool for the quantitative analysis of skeletal muscle.

[0003] Physical constitution is closely related to the occurrence and development of KOA. How to accurately evaluate physical constitution such as the structure and function of skeletal muscle and fat ratio will provide important value for the prevention and treatment of KOA. As a joint inflammatory disease caused by multiple factors, physical constitution has received more and more attention in the occurrence and development of KOA. Muscle is the main component of physical constitution. A large number of studies have confirmed that changes in muscle structure and function, that is, the decline in muscle mass and strength, play a key role in the onset of KOA. Previous studies have mainly focused on the muscles around the knee joint. Among them, the quadriceps femoris, as a shock absorber and joint stabilizer of the knee joint, can protect the joint surface during weight bearing and exercise. Its functional impairment can lead to the onset of KOA. With the deepening of research, the structure and function of core skeletal muscles in the waist and abdomen have also been confirmed to play an important role in the occurrence and development of KOA. Azuma et al. analyzed the association between the size and quality of abdominal trunk muscles and KOA. The results showed that the cross-sectional area of ​​paraspinal muscles in KOA patients was significantly reduced, and the muscle attenuation value was higher, suggesting that the size and quality of paraspinal muscles are related to the high risk of KOA. Therefore, strengthening the core muscles such as the waist and abdomen is also considered an important strategy to prevent and improve KOA.

[0004] In addition to muscle tissue, the role of body fat components in KOA cannot be ignored. A large number of previous studies have confirmed that obesity is a high-risk factor for KOA. First, overweight caused by excessive fat can increase the load on the knee joint and accelerate the wear of cartilage, leading to the onset and progression of KOA. In addition, fat is not only a substance that stores energy, but also an active endocrine tissue that can secrete a variety of hormones and cytokines to regulate the body's metabolism and immune response. Excessive or abnormal distribution of fat may lead to endocrine imbalance and inflammatory response, leading to loss of knee cartilage and inflammatory response, further damaging the structure and function of the joint. In addition, among systemic adipose tissue, subcutaneous adipose tissue is significantly negatively correlated with muscle mass and strength, which may also be a related mechanism for excessive fat to cause KOA. The abdomen, as the main site of fat accumulation (including subcutaneous and visceral fat), has been proven to be closely related to the occurrence and severity of KOA.

[0005] In summary, it can be seen that systematic evaluation of physical composition is of great significance for the prevention and treatment of KOA. The waist and abdomen, as an important component of the body's core skeletal muscle and the main site of fat accumulation, are ideal sites for analyzing the characteristics of human physical composition. Analyzing the physical composition characteristics of the waist and abdomen will provide important information for the assessment of the risk of KOA and its intervention.

[0006] CT imaging automatic segmentation technology based on deep learning algorithms provides an effective means for clinical quantification and systematic analysis of physical parameters such as skeletal muscle structure and function and fat composition. Previous body composition analysis techniques include bioelectrical impedance analysis, dual-energy X-ray absorptiometry, ultrasound, magnetic resonance imaging (MRI), computed tomography (CT) and other imaging technologies to complete the determination of muscle and tissue content. Among them, CT and MRI are regarded as the gold standard for evaluating muscle fat composition. However, traditional CT detection requires technicians to manually outline and perform image recognition and analysis based on experience. Not only is it time-consuming, but due to individual differences in physician skills, the results are often difficult to replicate, and the effectiveness and efficiency of health services are not high. With the rapid development of artificial intelligence (AI) technology, AI-based automated segmentation has become a hot spot and important technology for the analysis of body tissue and organ structure.

[0007] Deep learning (DL) is an AI algorithm based on neural networks. It uses multiple nonlinear processing layers to extract data information layer by layer. It analyzes high-dimensional data to detect subtle changes beyond human recognition capabilities and mine deeper features, thereby achieving image segmentation and classification prediction. Because deep learning can directly apply raw data to avoid information loss, and it shows higher efficiency when the data scale is large and the prediction task is complex, it has become a current research hotspot. In recent years, convolutional neural networks, as one of the representative algorithms of deep learning, have shown good performance in the automatic segmentation of human body components. The research of Song et al. shows that the heart image segmentation technology based on deep learning has good performance, which effectively improves the segmentation accuracy and thus improves the accuracy of disease diagnosis. It has also been proven to have significant performance in the automated segmentation of organ tissues such as brain tumors and liver diseases. At the same time, the automated segmentation technology of CT images of skeletal muscle constructed by the research team based on deep learning in the early stage also showed good performance. In both the training (0.963, 0.970, 0.941, and 0.968, respectively) and testing (0.950, 0.960, 0.929, and 0.961, respectively) datasets, the algorithm's automatic segmentation of paraspinal muscles had high Dice similarity coefficients, while the muscle cross-sectional area errors were low, at 1.9%, 1.6%, 3.1%, and 1.3%, respectively, in the training set and 3.4%, 3.0%, 4.6%, and 1.9%, respectively, in the testing set.

[0008] In addition to physical composition factors such as muscle and fat, clinical factors such as the patient's height, weight, smoking, and drinking are also closely related to the occurrence and development of KOA. Therefore, in order to scientifically and accurately predict the disease risk of KOA, it is necessary to systematically integrate and analyze relevant clinical factors and skeletal muscle-related factors, so as to build a prediction tool that can be widely used in clinical practice. This requires solving problems such as how to deal with large amounts of nonlinear high-dimensional data and how to avoid overfitting of prediction models. Machine learning (ML), as the main type of AI, provides a powerful tool for solving these problems. Compared with traditional models, machine learning has better performance, does not require any prior assumptions, and can handle large amounts of nonlinear high-dimensional data. It is an important means of big data analysis technology at present. Common algorithms for machine learning include support vector machines, decision trees, random forests, and neural networks, all of which have good predictive performance. It has great application potential in the diagnosis, treatment, and prediction of clinical diseases, and has also received attention in KOA. Leung et al.'s research showed that machine learning can better predict the risk of total knee replacement for osteoarthritis, with an area under the ROC curve (AUC) as high as 0.87. The study by Kwon et al. used machine learning to analyze gait characteristics and predict the severity of KOA, providing a feasible method for objectively evaluating the severity of KOA, which is expected to be used as a tool for KOA rehabilitation guidance and progress assessment. In addition, machine learning has also shown good results in the evaluation of KOA phenotype, cartilage volume loss, etc.

[0009] In summary, physical constitution is closely related to the occurrence and development of KOA, and the waist and abdomen, as an important component of the body's core skeletal muscle and the main site of fat accumulation, play an important role in the occurrence and development of KOA. Therefore, by systematically analyzing the physical characteristics of this part of muscle and adipose tissue, it will provide important information for the assessment of the risk of KOA and its intervention. How to efficiently, systematically and accurately assess physical constitution is one of the key issues that needs to be solved. Summary of the invention

[0010] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a KOA risk prediction method based on CT examination to automatically measure the physical composition, which can efficiently, systematically and accurately evaluate the physical composition, thereby improving the accuracy of predicting the risk of KOA. To achieve the above purpose, the present invention provides the following scheme:

[0011] A KOA risk prediction method based on automated measurement of body composition by CT examination, comprising:

[0012] Obtain CT datasets of healthy people and KOA patients respectively;

[0013] Preprocessing the CT data set using a moving image enhancement window to obtain a preprocessed data set;

[0014] Using the preprocessed data set to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information;

[0015] calculating physical parameters using the segmentation information;

[0016] Establishing a sample data set according to the physical parameters and preset clinical data, and performing feature extraction on the sample data set to obtain feature values;

[0017] Using the characteristic values ​​to train a preset initial model to obtain a trained KOA risk prediction model;

[0018] The KOA risk prediction model is used to predict the KOA risk of the tested population to obtain a prediction result.

[0019] Preferably, the segmentation information includes: skeletal muscle data, subcutaneous fat data and visceral fat data.

[0020] Preferably, the physical parameters include: the area of ​​each muscle or muscle group, the subcutaneous fat area, the visceral fat area and muscle fat infiltration degree parameters, the overall lean tissue content, the fat to lean tissue ratio, the bone mass to lean tissue ratio, the total abdominal fat content and other parameters.

[0021] Preferably, the CT data set is preprocessed using a moving image enhancement window to obtain a preprocessed data set, including:

[0022] Filtering the CT data set based on a Gaussian filtering method to obtain filtered data;

[0023] Constructing a moving image enhancement window according to the scale of the filtering data;

[0024] According to formula F ij =|x ij -y ij |Get the pixel difference of each pixel under each image enhancement window; where F ij Represents the pixel difference of the filtered data at position (i, j), x ij represents the pixel value of the original knee joint image at position (i, j), y ij Represents the pixel value of the filtered data at position (i, j);

[0025] According to the formula Perform brightness uniformity correction on the pixel difference to obtain a corrected difference; wherein, D ij is the correction difference of the pixel point, D darkis the dark signal value of the pixel point, K(i) is the correction coefficient; β is the preset ratio;

[0026] Pixel points whose correction difference is greater than the feature threshold are taken as feature points and the number is recorded;

[0027] The ratio of the feature points to all the pixels in the image enhancement window is taken as a ratio value;

[0028] When the proportion value in the moving image enhancement window is greater than a preset threshold, median denoising is performed on the filtered data in the corresponding moving image enhancement window to obtain denoised data;

[0029] The moving image enhancement window is slid until the entire denoised data is traversed to obtain the preprocessed data set.

[0030] Preferably, when the pixel point is within the skeletal muscle range, β is the ratio of the muscle density of the person to which the CT data set belongs to and the average muscle density threshold of the age of the person to which the CT data set belongs; when the pixel point is within the fat range, β is the ratio of the body fat percentage of the person to which the CT data set belongs to and the average body fat percentage threshold of the age of the person to which the CT data set belongs.

[0031] Preferably, the preprocessed data set is used to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information, including:

[0032] Obtaining a preset skeletal muscle dataset, subcutaneous fat dataset, and visceral fat dataset;

[0033] Construct a convolutional neural network deep learning model for multi-target segmentation;

[0034] The convolutional neural network deep learning model for multi-target segmentation is trained according to the skeletal muscle data set, the subcutaneous fat data set, and the visceral fat data set to obtain a trained first classifier, a second classifier, and a third classifier;

[0035] Using the first classifier to identify skeletal muscles in the skeletal muscle data set to obtain a skeletal muscle confidence set, using the second classifier to identify subcutaneous fat in the subcutaneous fat data set to obtain a subcutaneous fat confidence set, and using the third classifier to identify visceral fat in the visceral fat data set to obtain a visceral fat confidence set;

[0036] Summing the confidence values ​​of the various parts in the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, to obtain the object confidence sums of the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, respectively;

[0037] Select the data with the largest sum of object confidence as the training threshold data, and retrain the classifiers corresponding to the other object confidence sums so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network;

[0038] Connecting the trained LSTM neural network after the classification network to obtain a CNN-LSTM combined neural network model;

[0039] The preprocessed data set is segmented according to the CNN-LSTM combined neural network model to obtain the segmentation information.

[0040] Preferably, after the trained LSTM neural network is connected after the classification network to obtain the CNN-LSTM combined neural network model, the method further includes:

[0041] The Dice coefficient and Jaccard coefficient are used to verify the constructed CNN-LSTM combined neural network model.

[0042] Preferably, a sample data set is established according to the physical parameters and preset clinical data, and feature extraction is performed on the sample data set to obtain feature values, including:

[0043] The Spearman correlation coefficient is used to reduce the dimension of the data in the sample data set to remove highly correlated features and obtain reduced dimension data.

[0044] Based on the ridge regression algorithm, the features of the dimension reduction data were screened through five-fold cross validation to obtain the initial features;

[0045] Non-numeric features in the initial features are converted into numeric features, and missing values ​​of the converted features are preprocessed to obtain the feature values.

[0046] Preferably, the KOA risk prediction model is any one of support vector machine, extreme gradient boosting, random forest, linear discriminant analysis, logistic regression, naive Bayes model, k nearest neighbor, multilayer perceptron, long short-term memory and convolutional neural network.

[0047] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] The present invention provides a KOA risk prediction method based on CT examination to automatically measure physical constitution, including: obtaining CT data sets of healthy people and KOA patients respectively; preprocessing the CT data sets using a moving image enhancement window to obtain a preprocessed data set; using the preprocessed data set to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information; using the segmentation information to calculate physical parameters; establishing a sample data set according to the physical parameters and preset clinical data, and extracting features from the sample data set to obtain characteristic values; using the characteristic values ​​to train a preset initial model to obtain a trained KOA risk prediction model; using the KOA risk prediction model to predict the KOA risk of the population to be tested to obtain a prediction result. The present invention can construct a KOA risk prediction system that integrates KOA-related skeletal muscle physical constitution and clinical factors through a machine learning prediction model, and can accurately predict the risk of KOA, providing an important reference for the prevention and treatment of clinical KOA. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of clinical application research provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a KOA risk prediction method based on CT examination to automatically measure body composition, comprising:

[0055] Step 100: Acquire CT data sets of healthy people and KOA patients respectively;

[0056] Step 200: preprocessing the CT data set using a moving image enhancement window to obtain a preprocessed data set;

[0057] Step 300: Using the preprocessed data set to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information;

[0058] Step 400: Calculating physical parameters using the segmentation information;

[0059] Step 500: establishing a sample data set according to the physical parameters and preset clinical data, and performing feature extraction on the sample data set to obtain feature values;

[0060] Step 600: training a preset initial model using the characteristic values ​​to obtain a trained KOA risk prediction model;

[0061] Step 700: Use the KOA risk prediction model to perform KOA risk prediction on the population to be tested to obtain a prediction result.

[0062] Specifically, the clinical data collection in this embodiment is as follows:

[0063] ① Baseline indicators and information collection: age, gender, height, weight, smoking, drinking history, past history, family history, medication history, etc.

[0064] ②Imaging data: CT examination data, knee joint X-ray, etc.

[0065] ③The radiological diagnostic criteria for osteoarthritis (Kellgren-Lawrence method) are divided into 5 grades: Grade 0, normal; Grade I, suspected narrowing of the joint space, possible osteophytes; Grade II, obvious osteophytes, suspected narrowing of the joint space; Grade HI, moderate osteophytes, obvious narrowing of the joint space, and sclerotic changes; Grade IV, a large number of osteophytes, obvious narrowing of the joint space, severe sclerotic lesions and obvious deformities.

[0066] ④ Knee symptom / function scoring system: Western Ontario and McMaster University Osteoarthritis Index (WOMAC). In this scoring system, WOMAC ADL, WOMAC PK and WOMAC STF are used to evaluate the patient's knee physiological function, pain and stiffness, respectively. In addition, WOMAC Ts is the total score of the above three scores. A score of 0 for each indicator in this project means that the corresponding knee symptom / function is normal, and a higher score means that the corresponding symptom is more obvious or the function is worse.

[0067] ⑤Others: including course of disease, site of onset, range of pain, associated symptoms, etc.

[0068] Preferably, the segmentation information includes: skeletal muscle data, subcutaneous fat data and visceral fat data.

[0069] Preferably, the physical parameters include: the area of ​​each muscle or muscle group, the subcutaneous fat area, the visceral fat area and muscle fat infiltration degree parameters, the overall lean tissue content, the fat to lean tissue ratio, the bone mass to lean tissue ratio, the total abdominal fat content and other parameters.

[0070] Specifically, the original data set in this application is labeled with physical composition, among which the lumbar spine CT volume data obtained from the picture archiving and communication system (PACS) is saved in DICOM format. One attending physician in the Department of Radiology, who has 8 and 9 years of experience in musculoskeletal system diagnosis and has received professional training, used Materialise Mimics Medical 20.0 medical imaging software to manually mark various parts of muscle, fat tissue and other tissues. When there is a dispute between the two doctors, another deputy chief physician with 15 years of work experience is asked to make a judgment. The standard data set finally obtained is used to train the network model and evaluate the segmentation performance.

[0071] Preferably, the CT data set is preprocessed using a moving image enhancement window to obtain a preprocessed data set, including:

[0072] Filtering the CT data set based on a Gaussian filtering method to obtain filtered data;

[0073] Constructing a moving image enhancement window according to the scale of the filtering data;

[0074] According to formula F ij =|x ij -y ij |Get the pixel difference of each pixel under each image enhancement window; where F ij Represents the pixel difference of the filtered data at position (i, j), x ij represents the pixel value of the original knee joint image at position (i, j), y ij Represents the pixel value of the filtered data at position (i, j);

[0075] According to the formula Perform brightness uniformity correction on the pixel difference to obtain a corrected difference; wherein, D ij is the correction difference of the pixel point, D darkis the dark signal value of the pixel point, K(i) is the correction coefficient; β is the preset ratio;

[0076] Pixel points whose correction difference is greater than the feature threshold are taken as feature points and the number is recorded;

[0077] The ratio of the feature points to all the pixels in the image enhancement window is taken as a ratio value;

[0078] When the proportion value in the moving image enhancement window is greater than a preset threshold, median denoising is performed on the filtered data in the corresponding moving image enhancement window to obtain denoised data;

[0079] The moving image enhancement window is slid until the entire denoised data is traversed to obtain the preprocessed data set.

[0080] Optionally, in this embodiment, based on the image enhancement window, median denoising is performed on the image in the corresponding window, which can make certain features in the image disappear. This can retain the original pixel information of the image to the greatest extent, further improving the image restoration degree.

[0081] Preferably, when the pixel point is within the skeletal muscle range, β is the ratio of the muscle density of the person to which the CT data set belongs to and the average muscle density threshold of the age of the person to which the CT data set belongs; when the pixel point is within the fat range, β is the ratio of the body fat percentage of the person to which the CT data set belongs to and the average body fat percentage threshold of the age of the person to which the CT data set belongs.

[0082] Furthermore, in this embodiment, a brightness uniformity correction formula is used to perform brightness correction on the filtered data. In this application, the differences between the skeletal muscle range and the fat range are taken into consideration, thereby correcting the original data set, that is, β appears as a correction parameter, thereby making the result of the brightness uniformity correction more accurate and reliable.

[0083] Furthermore, in this embodiment, the color in the image is adjusted by a brightness uniformity correction formula so that white objects in the data can appear pure white regardless of the lighting conditions, and other colors will also be more accurate.

[0084] Preferably, the preprocessed data set is used to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information, including:

[0085] Obtaining a preset skeletal muscle dataset, subcutaneous fat dataset, and visceral fat dataset;

[0086] Construct a convolutional neural network deep learning model for multi-target segmentation;

[0087] The convolutional neural network deep learning model for multi-target segmentation is trained according to the skeletal muscle data set, the subcutaneous fat data set, and the visceral fat data set to obtain a trained first classifier, a second classifier, and a third classifier;

[0088] Using the first classifier to identify skeletal muscles in the skeletal muscle data set to obtain a skeletal muscle confidence set, using the second classifier to identify subcutaneous fat in the subcutaneous fat data set to obtain a subcutaneous fat confidence set, and using the third classifier to identify visceral fat in the visceral fat data set to obtain a visceral fat confidence set;

[0089] Summing the confidence values ​​of the various parts in the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, to obtain the object confidence sums of the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, respectively;

[0090] Select the data with the largest sum of object confidence as the training threshold data, and retrain the classifiers corresponding to the other object confidence sums so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network;

[0091] Connecting the trained LSTM neural network after the classification network to obtain a CNN-LSTM combined neural network model;

[0092] The preprocessed data set is segmented according to the CNN-LSTM combined neural network model to obtain the segmentation information.

[0093] Specifically, this embodiment performs deep learning automatic segmentation training for each part of muscle and fat tissue respectively, and then automatically calculates parameters such as each muscle of skeletal muscle or the area of ​​skeletal muscle, subcutaneous fat area, visceral fat area and muscle fat infiltration degree parameters, overall lean tissue content, fat to lean tissue ratio, bone mass to lean tissue ratio, and total abdominal fat content.

[0094] Preferably, after the trained LSTM neural network is connected after the classification network to obtain the CNN-LSTM combined neural network model, the method further includes:

[0095] The Dice coefficient and Jaccard coefficient are used to verify the constructed CNN-LSTM combined neural network model.

[0096] Furthermore, the Dice coefficient is a metric for evaluating the similarity or overlap of two sets. It calculates the ratio of twice the area of ​​the overlapping area of ​​the predicted value and the true value to the sum of the areas of their respective areas. Its range is [0, 1]. The closer it is to 1, the higher the accuracy of the segmented image, and the closer it is to 0, the lower the accuracy of the image segmentation.

[0097] The Jaccard coefficient is used to compare the similarities and differences between finite sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.

[0098] Preferably, a sample data set is established according to the physical parameters and preset clinical data, and feature extraction is performed on the sample data set to obtain feature values, including:

[0099] The Spearman correlation coefficient is used to reduce the dimension of the data in the sample data set to remove highly correlated features and obtain reduced dimension data.

[0100] Based on the ridge regression algorithm, the features of the dimension reduction data were screened through five-fold cross validation to obtain the initial features;

[0101] Non-numeric features in the initial features are converted into numeric features, and missing values ​​of the converted features are preprocessed to obtain the feature values.

[0102] Preferably, the KOA risk prediction model is any one of support vector machine, extreme gradient boosting, random forest, linear discriminant analysis, logistic regression, naive Bayes model, k nearest neighbor, multilayer perceptron, long short-term memory and convolutional neural network.

[0103] Specifically, the method for constructing the KOA risk prediction model in this embodiment includes:

[0104] (1) Machine training samples and model selection

[0105] First, the relevant clinical data of the subjects in the above cohort and the automatic measurement data of CT physical composition were used to establish a sample database, which was divided into 70% training samples and 30% test samples, of which the training samples were used for model construction and the test samples were used for model evaluation. Ten algorithms were used for modeling: support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), linear discriminant analysis (LDA), logistic regression (LR), naive Bayes model (NB), k nearest neighbor (KNN), multilayer perceptron (MLP), long short-term memory (LSTM) and convolutional neural network (CNN) models.

[0106] (2) Variable screening

[0107] First, the Spearman Correlation Coefficients (PCC) are used to reduce the dimension of the data in the training set. All features are traversed and the correlation coefficients are calculated pairwise. When the coefficient is greater than or equal to 0.8 (|r|≥0.8, highly correlated), one of them is randomly removed. This method can make the features after dimensionality reduction not have high similarity. This method can filter out some eigenvalues, but the number of eigenvalues ​​obtained is still large, and overfitting is prone to occur in the subsequent modeling process. In order to alleviate the occurrence of overfitting, ridge regression is continued to be used through the five-fold cross-validation method to further filter features. The ridge regression algorithm makes the coefficient values ​​average, and the model of the filtered features established is more stable, unlike the coefficients of the Lasso regression algorithm, which will fluctuate significantly due to slight data changes.

[0108] (3) Processing of predictor variables

[0109] Machine learning algorithms require feature attributes to be numerical, so non-numerical feature attributes such as gender and family history need to be changed to numerical (0, 1). Missing values ​​are handled in three types: A. Jump variables: If the primary variable is selected as "no", the secondary variable is automatically filled with 0; B. Delete variables with more than 30% missing values; C. For the remaining missing values, categorical variables are filled with 99 missing values; continuous variables are filled with the mean missing values. The ROC curve is used to determine the optimal critical value of each continuous variable, and the Youden index of each indicator is calculated. When the Youden index is the largest, the corresponding value is the optimal critical value, and all data are converted into binary variables.

[0110] (4) Machine Learning Modeling

[0111] According to the corresponding model function provided by the R package, XGBoost and several common models are modeled respectively, and the parameter configuration of the model follows the default settings provided by the function. Ten types of models are constructed, including support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), linear discriminant analysis (LDA), logistic regression (LR), naive Bayes model (NB), k nearest neighbor (KNN), multilayer perceptron (MLP), long short-term memory (LSTM) and convolutional neural network (CNN).

[0112] (5) Screening of machine learning models

[0113] The accuracy of each model is determined through ten-fold cross-validation, and the best prediction model is screened based on the "horse racing mechanism", and the prediction model with the highest prediction accuracy is selected as the final prediction model.

[0114] (6) Verification of machine learning models

[0115] In the validation set, the ROC curve was used to analyze the predictive efficiency of the prediction model. The larger the AUC value, the higher its predictive value. The prediction accuracy, sensitivity, specificity, negative prediction rate, and positive prediction rate of the prediction model were analyzed. The accuracy rate (accuracy) and precision (precision error) were calculated, and the recall rate (recall rate) and F1 score were calculated.

[0116] (7) Statistical methods

[0117] PASS software was used to calculate the sample size, and SPSS25.0 statistical software was used to analyze the data. Count data were expressed as "n (%)", and chi-square test was performed in parallel. The test level was set to bilateral α = 0.05; measurement data that did not conform to normal distribution were expressed as median (interquartile range) "M (QR)", and Mann-Whitney U test was used for inter-group comparison. Measurement data that conformed to normal distribution were expressed as "x ± s", and independent sample t test was used for inter-group comparison. Influencing factors were analyzed by logistic regression, and correlation was analyzed by Spearman or Pearson correlation analysis. The test level was set to bilateral α = 0.05. P < 0.05 was considered statistically significant. The processing of machine learning models was completed using R4.0.4 software, and the main R packages used were 'mlr', 'nnet', 'rpart', 'e1071', 'catboost', 'stats', 'pcalg', 'ggplot2', 'caret', etc.

[0118] In addition, if Figure 2 As shown, the specific process of the KOA risk prediction model of this embodiment for clinical application research is as follows:

[0119] (1) Research subjects

[0120] KOA patients treated in our hospital and healthy subjects were prospectively selected as research subjects. Inclusion criteria: ① met the diagnostic criteria of KOA in the "Guidelines for the Diagnosis and Treatment of Osteoarthritis" and were confirmed by magnetic resonance imaging; ② the patient or family member informed consent and signed the informed consent form. Exclusion criteria: patients with severe heart, liver and other organ diseases; patients with rheumatic and gouty arthritis; patients with a history of knee surgery.

[0121] The calculation method of sample size is the same as above, and the corresponding sample size is calculated based on the indicators finally included in the prediction model.

[0122] (2) Clinical data collection: the same as the data collection in the above steps.

[0123] (3) Diagnostic analysis of KOA risk prediction system

[0124] The diagnostic results of the KOA risk prediction system were analyzed using MRI examination results as the gold standard. The AUC value, sensitivity, specificity, and accuracy in the ROC curve were used as the predictive efficiency evaluation indicators of the model. The decision curve analysis (DCA) diagram was used to demonstrate the clinical value of the model.

[0125] (4) SHAP visualization analysis and platform construction

[0126] SHAP can calculate the Shapley value of each feature to measure the contribution of each feature to the prediction result: when the value is greater than 0, it means that the feature improves the result value and has an improvement effect on the prediction value; when the value is less than 0, it means that the feature reduces the result value and has a negative impact. The feature visualization toolkit SHAP (version 0.40.0) was used to analyze the influence of each indicator of the prediction model. The interpretation method can output the feature importance ranking and the relationship between the feature and the prediction result, so as to explain how the indicator affects the result. It includes single sample SHAP value analysis, feature importance distribution analysis, and feature overall analysis. Finally, the "Shiny" package in R was used to build a visual data analysis platform.

[0127] The beneficial effects of the present invention are as follows:

[0128] (1) The present invention uses machine learning to integrate physical parameters such as KOA skeletal muscle and related clinical parameters to construct a more scientific, complete and accurate KOA risk prediction model.

[0129] (2) The present invention combines the significant advantages of machine learning in the diagnosis, treatment and prediction of diseases, and can play an important role in the risk prediction and management of KOA by integrating KOA-related skeletal muscle constitution components and clinical factors.

[0130] (3) The present invention constructs CT physical composition automated measurement technology based on deep learning, and constructs a KOA risk prediction system based on machine learning based on CT physical composition automated measurement data and clinical related data, forming a visual analysis platform. The research will provide scientific guidance and powerful tools for the screening of high-risk groups for KOA and early intervention of KOA. It is of great significance to prevent the onset and progression of KOA.

[0131] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0132] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A KOA risk prediction method based on automated measurement of body constitution by CT examination, characterized in that: include: Obtain CT datasets of healthy people and KOA patients respectively; Preprocessing the CT data set using a moving image enhancement window to obtain a preprocessed data set; Using the preprocessed data set to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information; Automatically measuring physical parameters using the segmented information; Establishing a sample data set according to the physical parameters and preset clinical data, and performing feature extraction on the sample data set to obtain feature values; Using the characteristic values ​​to train a preset initial model to obtain a trained KOA risk prediction model; Using the KOA risk prediction model to predict the KOA risk of the tested population, and obtaining a prediction result; The segmentation information includes: skeletal muscle data, subcutaneous fat data and visceral fat data; The preprocessed data set is used to train a preset multi-target segmentation convolutional neural network deep learning model to obtain segmentation information, including: Obtaining a preset skeletal muscle dataset, subcutaneous fat dataset, and visceral fat dataset; Construct a convolutional neural network deep learning model for multi-target segmentation; The convolutional neural network deep learning model for multi-target segmentation is trained according to the skeletal muscle data set, the subcutaneous fat data set, and the visceral fat data set to obtain a trained first classifier, a second classifier, and a third classifier; Using the first classifier to identify skeletal muscles in the skeletal muscle data set to obtain a skeletal muscle confidence set, using the second classifier to identify subcutaneous fat in the subcutaneous fat data set to obtain a subcutaneous fat confidence set, and using the third classifier to identify visceral fat in the visceral fat data set to obtain a visceral fat confidence set; Summing the confidence values ​​of the various parts in the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, to obtain the object confidence sums of the skeletal muscle confidence set, the subcutaneous fat confidence set, and the visceral fat confidence set, respectively; Select the data with the largest sum of object confidence as the training threshold data, and retrain the classifiers corresponding to the other object confidence sums so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network; Connecting the trained LSTM neural network after the classification network to obtain a CNN-LSTM combined neural network model; Segmenting the preprocessed data set according to the CNN-LSTM combined neural network model to obtain the segmentation information; The physical parameters include: the area of ​​each muscle or muscle group, the subcutaneous fat area, the visceral fat area and the muscle fat infiltration degree parameter, the overall lean tissue content, the fat to lean tissue ratio, the bone mass to lean tissue ratio, and the total abdominal fat content parameter.

2. The KOA risk prediction method based on CT examination and automatic measurement of body composition according to claim 1, characterized in that: The CT data set is preprocessed using a moving image enhancement window to obtain a preprocessed data set, including: Filtering the CT data set based on a Gaussian filtering method to obtain filtered data; Constructing a moving image enhancement window according to the scale of the filtering data; According to formula F ij =|x ij -y ij |Get the pixel difference of each pixel under each image enhancement window; where F ij Represents the pixel difference of the filtered data at position (i, j), x ij represents the pixel value of the original knee joint image at position (i, j), y ij Represents the pixel value of the filtered data at position (i, j); According to the formula Perform brightness uniformity correction on the pixel difference to obtain a corrected difference; wherein, D ij is the correction difference of the pixel point, D dark is the dark signal value of the pixel point, K(i) is the correction coefficient; β is the preset ratio; Pixel points whose correction difference is greater than the feature threshold are taken as feature points and the number is recorded; The ratio of the feature points to all the pixels in the image enhancement window is taken as a ratio value; When the proportion value in the moving image enhancement window is greater than a preset threshold, median denoising is performed on the filtered data in the corresponding moving image enhancement window to obtain denoised data; The moving image enhancement window is slid until the entire denoised data is traversed to obtain the preprocessed data set.

3. The KOA risk prediction method based on CT examination and automatic measurement of body composition according to claim 2, characterized in that: When the pixel point is within the skeletal muscle range, β is the ratio of the muscle density of the person to which the CT data set belongs to and the average muscle density threshold of the age to which the person belongs; when the pixel point is within the fat range, β is the ratio of the body fat percentage of the person to which the CT data set belongs to and the average body fat percentage threshold of the age to which the person belongs.

4. The KOA risk prediction method based on CT examination and automatic measurement of body composition according to claim 1, characterized in that: After the trained LSTM neural network is connected to the classification network to obtain the CNN-LSTM combined neural network model, the following further includes: The Dice coefficient and Jaccard coefficient are used to verify the constructed CNN-LSTM combined neural network model.

5. The KOA risk prediction method based on CT examination and automatic measurement of body composition according to claim 1, characterized in that: A sample data set is established according to the physical parameters and the preset clinical data, and feature extraction is performed on the sample data set to obtain feature values, including: The Spearman correlation coefficient is used to reduce the dimension of the data in the sample data set to remove highly correlated features and obtain reduced dimension data. Based on the ridge regression algorithm, the features of the dimension reduction data were screened through five-fold cross validation to obtain the initial features; Non-numeric features in the initial features are converted into numeric features, and missing values ​​of the converted features are preprocessed to obtain the feature values.

6. The KOA risk prediction method based on CT examination and automatic measurement of body composition according to claim 1, characterized in that: The KOA risk prediction model is any one of support vector machine, extreme gradient boosting, random forest, linear discriminant analysis, logistic regression, naive Bayes model, k nearest neighbor, multilayer perceptron, long short-term memory and convolutional neural network.

Images