An automatic assessment method for limb dysfunction in patients with low back pain based on multi-source features
U-Net neural network and support vector machine algorithm combined with MRI imaging and clinical scale characteristics to automatically evaluate limb dysfunction in patients with low back pain, solving the subjectivity and accuracy of the existing evaluation methods, and achieving high accuracy and Automatic evaluation of reliability.
Patent Information
- Application Number
- CN202210081133.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-24
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-01-24
AI Technical Summary
The existing methods for evaluating limb dysfunction in patients with low back pain have strong subjectivity and low reliability. Relying on doctors' subjective experience and manual segmentation leads to low accuracy and time-consuming and labor-intensive results.
U-Net neural network and support vector machine algorithm are used to combine MRI imaging features and clinical scale features to automatically segment multifission muscles and erection spinal muscles, and the cross-sectional area and fat infiltration characteristics are extracted, and support vector machine is used to recursively The feature elimination algorithm reduces feature redundancy and builds an automatic evaluation model.
Automatic and accurate assessment of limb dysfunction in patients with low back pain is realized, which improves the accuracy and reliability of the assessment, and reduces operational complexity and time cost.
Smart Images

Figure CN114533042B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of limb dysfunction assessment, and in particular relates to an automatic assessment method for limb dysfunction of low back pain patients based on multi-source features. Background Art
[0002] Low back pain is a common clinical disease that seriously affects patients' work and life, and brings a heavy burden to patients and their families. Accurate assessment of functional impairment is an important prerequisite for determining patient rehabilitation treatment and assistive device adaptation plans.
[0003] At present, the limb dysfunction of patients with low back pain is mainly assessed based on behavioral scales. Commonly used scales in clinical practice include the Visual Analogue Scale (VAS), the Oswestry Disability Index (ODI), the Japanese Orthopaedic Association Scores (JOA), the Roland-Morris Disability Questionnaire (RMDQ), the Quebec Back Pain Disability Scale (QBPDS), etc.
[0004] In addition, imaging methods such as computed tomography (CT), ultrasound (US) and magnetic resonance imaging (MRI) are used clinically to explore the causes of low back pain, but there is a lack of research on the extraction of MRI image features related to limb dysfunction.
[0005] Although behavioral scale examinations and MRI-based imaging examinations can provide multiple bases for the assessment of limb dysfunction in patients with low back pain, clinicians still need to make comprehensive judgments based on their own experience. The assessment results are heavily dependent on the subjective experience of doctors, and there are problems such as strong subjectivity, inconsistent assessment standards, and low reliability. Therefore, achieving automatic and accurate assessment of limb dysfunction in patients with low back pain has become an important clinical problem that needs to be solved urgently.
[0006] The lumbar paraspinal muscles are an important component of protecting the stability of the spine. Studies have shown that the reduction of the cross-sectional area (CSA) and the increase of the fat infiltration rate (FIR) of the lumbar paraspinal muscles can reflect the functional degeneration of the muscles. The present invention has selected the multifidus and erector spinae muscles in the lumbar paraspinal muscles through research and comparison. The study found that their functional degeneration is particularly obvious in patients with low back pain. Subsequently, the cross-sectional area and fat infiltration of the multifidus and erector spinae muscles will be extracted as MRI imaging features of patients with low back pain.
[0007] At present, the extraction of multifidus and erector spinae regions mainly relies on manual segmentation by doctors, which results in low accuracy, time-consuming and labor-intensive calculation of muscle cross-sectional area and fat infiltration imaging features. The U-Net model has a convolutional network structure and is based on a fully convolutional neural network (FCN). The model can achieve accurate image segmentation even with a small amount of image training.
[0008] Machine learning is the science of using computers to simulate or realize human learning activities. It is a common research hotspot in the field of artificial intelligence and pattern recognition. Machine learning can achieve intelligent classification and recognition of different samples by learning input data. Support Vector Machines (SVM) is a type of machine learning method with powerful prediction capabilities. Its core idea is to find an optimal discriminant hyperplane that can achieve the maximum boundary distance to sample points of different categories based on the structural risk minimization criterion. It has the advantages of good generalization ability, excellent small sample classification performance, and low computational complexity. Summary of the invention
[0009] In view of the defects of existing clinical limb dysfunction assessment methods for low back pain patients, such as strong subjectivity and low reliability, the purpose of the present invention is to provide an automatic limb dysfunction assessment method for low back pain patients based on multi-source features.
[0010] In order to achieve the above tasks, the present invention adopts the following technical solutions:
[0011] An automatic assessment method for limb dysfunction in patients with low back pain based on multi-source features, characterized in that the specific implementation steps are as follows:
[0012] Step 1: Use the U-Net neural network to automatically segment the multifidus and erector spinae muscles in the lumbar MRI images of patients with low back pain, and use the threshold method to extract imaging features such as muscle cross-sectional area and fat infiltration;
[0013] Step 2: Based on multiple clinical assessment scales for low back pain patients, screen out the scales with strong correlation with limb function and extract the scale characteristics;
[0014] Step 3: Normalize the scale features and imaging features to form a feature set;
[0015] Step 4: Use the support vector machine recursive feature elimination algorithm to screen all features, reduce feature redundancy, and determine the optimal feature combination;
[0016] Step 5: Select the support vector machine method in supervised machine learning to build an evaluation model, use the optimal feature combination as the model input, and use the visual analog scale score as the tutor supervision signal of the evaluation model;
[0017] Step 6: Optimize the model parameters using a grid optimization method, and complete the construction of a limb dysfunction assessment model for low back pain patients through model training, thereby achieving automatic assessment of the degree of limb dysfunction in low back pain patients;
[0018] The method for extracting MRI imaging features using the U-Net neural network and threshold method described in step 1 above is:
[0019] ① Select three MRI axial images of the patient's L3-L4 spinal disc;
[0020] ② Improve the contrast of MRI images based on contrast-limited adaptive histogram equalization, which limits noise amplification and enhances local contrast by reducing the height of the local histogram;
[0021] ③Build a U-Net network based on the Adam optimizer with a learning rate of 10 -4 , the attenuation factor is 10 -6 , the number of iterations is 100, the momentum coefficient is 0.9, and the graphics processing unit (GPU) is used for training and testing, and finally the automatic segmentation of multifidus and erector spinae muscles is achieved;
[0022] ④Use the maximum inter-class variance method to distinguish muscle and fat tissue. Determine a gray value as the threshold based on the image histogram, and then divide the image histogram into two different classes based on this threshold, so that the inter-class variance of the two classes reaches the maximum value:
[0023] σ t 2 =ω0*ω1*(μ0-μ1) 2
[0024] In the formula, ω0 and ω1 are the occurrence probabilities of the two categories, and μ0 and μ1 are the means of the two categories;
[0025] Traverse all the grayscale values of the image and find Get the maximum gray value t, then the optimal threshold is t;
[0026] ⑤ Based on the number of pixels of muscle and fat tissue, calculate the muscle cross-sectional area (CSA) and fat infiltration rate (FIR) features:
[0027] CSA=p total *K
[0028] In the formula, p total is the sum of the number of pixels in the ROI; K is the scale.
[0029]
[0030] In the formula, p fat is the total number of fat pixels after threshold segmentation; p total is the total number of pixels in the area.
[0031] Finally, the muscle cross-sectional area and fat infiltration characteristics of the multifidus and erector spinae muscles were used as MRI imaging features.
[0032] The results of low back pain scale screening and scale feature extraction described in step 2 above are:
[0033] After research and screening, the scales that are strongly correlated with limb dysfunction in patients with low back pain are selected, including: ODI scoring standard (The Oswestry Disability Index, ODI) and Japanese Orthopaedic Association Scores (Japanese Orthopaetic Association Scores, JOA). The sub-question scores and total scores of the ODI scale and the JOA scale are used as the behavioral characteristics of the scale.
[0034] The calculation method for the feature normalization process described in the above step 3 is:
[0035]
[0036] In the formula, x nor is the normalized feature, x max is the maximum value of the feature, x min is the minimum value of the feature, and x is the feature value.
[0037] The specific implementation method of the SVM-RFE algorithm for feature screening described in step 4 above is:
[0038] ① Use the data set to train the SVM model and obtain the trained SVM model parameters;
[0039] ② Select an appropriate ranking criterion and calculate the scores of the behavioral characteristics of the scale and the MRI imaging characteristics under the criterion respectively;
[0040] ③ Eliminate the feature with the lowest score in the current data set and perform the above process until the last feature remains.
[0041] Commonly used sorting criteria are: R c =|W 2 -W -(P)2 |
[0042] Where W 2 and W -(P)2 They represent the weight of the complete SVM and the weight of the SVM after assuming that the Pth feature is removed.
[0043] W 2 The representation is:
[0044] In the formula, i and j represent loop variables; y represents the category label; N represents the number of samples; K(x i ,x j ) represents the kernel function; and It is obtained by solving the SVM dual optimization problem.
[0045] The SVM-RFE algorithm sorts the scores of each feature, removes the feature with the smallest score, and then uses the remaining features to train the model again for the next iteration. Finally, the optimal feature combination is obtained.
[0046] The specific method of selecting the supervisory signal and constructing the evaluation model in step 5 above is:
[0047] The optimal feature combination is used as the model input, and the VAS score is used as the model's mentor supervision signal. According to the VAS score range, low back pain patients are defined as mild disorder, moderate disorder, and severe disorder. Among them, VAS scores of 0-3 points are mild disorders, 4-6 points are moderate disorders, and 7-10 points are severe disorders.
[0048] The evaluation model was constructed using the multi-classification support vector machine method, in which the kernel function was the radial basis function. The model output was divided into three levels, defined as "1", "2", and "3", corresponding to mild impairment, moderate impairment, and severe impairment, respectively.
[0049] In the above step 6, the specific method of optimizing the SVM model parameters and automatically evaluating the degree of limb dysfunction is:
[0050] ① The kernel function in the support vector machine model uses the radial basis kernel function, and its definition expression is:
[0051]
[0052] Where σ is the radial base radius;
[0053] ②Definition Penalty factor C;
[0054] ③ Set the initial model parameters C, g optimization range and step size, traverse each set of C, g values through grid search, and calculate the classification accuracy of the model;
[0055] ④ Select the model parameters C and g with the highest classification accuracy as the optimal parameters of the model.
[0056] Based on the trained model, automatic assessment of limb dysfunction in patients with low back pain can be achieved.
[0057] The present invention proposes for the first time an automatic assessment method for limb dysfunction in patients with low back pain, using the SVM algorithm combined with clinical scale features and MRI imaging features to achieve comprehensive judgment and assessment of limb dysfunction in patients with low back pain; further, a deep learning method based on the U-Net network is used to achieve automatic segmentation of MRI images of bilateral multifidus muscles and erector spinae muscles, thereby improving the accuracy of segmentation. For the first time, a method of combining behavioral detection-based scale features with MRI-based imaging features to form a multi-source feature set is proposed, and feature optimization and redundancy reduction are performed; finally, a machine learning method is used to achieve the assessment of limb dysfunction in patients with low back pain. This method has the advantages of high accuracy, strong reliability, convenient operation, and easy clinical promotion. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 Technical roadmap for an automated assessment method of limb dysfunction in patients with low back pain.
[0059] The present invention is further described in detail below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION
[0060] See also Figure 1 This embodiment provides an automatic assessment method for limb dysfunction in patients with low back pain, and the specific implementation steps are as follows:
[0061] Step 1: Use the U-Net neural network to automatically segment the multifidus and erector spinae muscles in the lumbar MRI images of patients with low back pain, and use the threshold method to extract imaging features such as muscle cross-sectional area and fat infiltration;
[0062] The method of extracting imaging features using U-Net neural network and threshold method is as follows:
[0063] ① Select three MRI axial images of the patient's L3-L4 spinal disc;
[0064] ② Improve the contrast of MRI images based on contrast-limited adaptive histogram equalization, which limits noise amplification and enhances local contrast by reducing the height of the local histogram;
[0065] ③Build a U-Net network based on the Adam optimizer with a learning rate of 10 -4 , the attenuation factor is 10 -6 , the number of iterations is 100, the momentum coefficient is 0.9, and the graphics processing unit (GPU) is used for training and testing, and finally the automatic segmentation of multifidus and erector spinae muscles is achieved;
[0066] ④Use the maximum inter-class variance method to distinguish muscle and fat tissue. Determine a gray value as the threshold based on the image histogram, and then divide the image histogram into two different classes based on this threshold, so that the inter-class variance of the two classes reaches the maximum value:
[0067] σ t 2 =ω0*ω1*(μ0-μ1) 2
[0068] In the formula, ω0 and ω1 are the occurrence probabilities of the two categories, and μ0 and μ1 are the means of the two categories;
[0069] Traverse all the grayscale values of the image and find Get the maximum gray value t, then the optimal threshold is t;
[0070] ⑤ Based on the number of pixels of muscle and fat tissue, calculate the muscle cross-sectional area (CSA) and fat infiltration rate (FIR) features:
[0071] CSA=p total *K
[0072] In the formula, p total is the sum of the number of pixels in the ROI; K is the scale.
[0073]
[0074] In the formula, p fat is the total number of fat pixels after threshold segmentation; p total is the total number of pixels in the area.
[0075] Finally, the muscle cross-sectional area and fat infiltration characteristics of the multifidus and erector spinae muscles were used as MRI imaging features.
[0076] Step 2: Based on multiple clinical assessment scales for low back pain patients, screen out the scales with strong correlation with limb function and extract the scale characteristics;
[0077] The results of screening and feature extraction of low back pain scales are as follows:
[0078] After research and screening, the scales that are strongly correlated with limb dysfunction in patients with low back pain are selected, including: ODI scoring standard (The Oswestry Disability Index, ODI) and Japanese Orthopaedic Association Scores (Japanese Orthopaetic Association Scores, JOA). The sub-question scores and total scores of the ODI scale and the JOA scale are used as the behavioral characteristics of the scale.
[0079] Step 3: Normalize the scale features and imaging features to form a feature set;
[0080] The method of feature normalization is as follows:
[0081]
[0082] In the formula, x nor is the normalized feature, x max is the maximum value of the feature, x min is the minimum value of the feature, and x is the feature value.
[0083] Step 4: Use the support vector machine recursive feature elimination algorithm (SVM-RFE) to screen all features, reduce feature redundancy, and determine the optimal feature combination;
[0084] The specific implementation method of the SVM-RFE algorithm is as follows:
[0085] ① Use the training set to train SVM and obtain the trained SVM parameters;
[0086] ② Select an appropriate ranking criterion and calculate the scores of the behavioral characteristics of the scale and the MRI imaging characteristics under the criterion respectively;
[0087] ③ Eliminate the feature with the lowest score in the current data set and perform the above process until the last feature remains.
[0088] Commonly used sorting criteria are: R c =|W 2 -W -(P)2 |
[0089] Where W 2 and W -(P)2 They represent the weight of the complete SVM and the weight of the SVM after assuming that the Pth feature is removed.
[0090] W 2 The representation is:
[0091] In the formula, i and j represent loop variables; y represents the category label; N represents the number of samples; K(xi ,x j ) represents the kernel function; and It is obtained by solving the SVM dual optimization problem.
[0092] The SVM-RFE algorithm sorts the scores of each feature, removes the feature with the smallest score, and then uses the remaining features to train the model again for the next iteration. Finally, the optimal feature combination is obtained.
[0093] In this embodiment, the inventors compared the accuracy of the SVM-RFE method before and after use and found that the SVM-RFE algorithm can effectively remove feature redundancy and improve the classification accuracy of the model by 5%.
[0094] Step 5: Select the support vector machine method in supervised machine learning to build an evaluation model, use the optimal feature combination as the model input, and use the visual analog scale (VAS) score as the tutor supervision signal of the evaluation model;
[0095] The method of selecting supervisory signals and building evaluation models is as follows:
[0096] The optimal feature combination is used as the model input, and the VAS score is used as the model's mentor supervision signal. According to the VAS score range, low back pain patients are defined as: mild disorder, moderate disorder, and severe disorder. Among them, VAS scores of 0-3 points are mild disorders, 4-6 points are moderate disorders, and 7-10 points are severe disorders.
[0097] The evaluation model was constructed using the multi-classification support vector machine method, in which the radial basis function was selected as the kernel function. The model output was divided into three levels, defined as "1", "2", and "3", corresponding to mild obstacles, moderate obstacles, and severe obstacles respectively.
[0098] Step 6: Optimize the model parameters using a grid optimization method, and complete the construction of a limb dysfunction assessment model for low back pain patients through model training, thereby achieving automatic assessment of the degree of limb dysfunction in low back pain patients;
[0099] The specific methods for optimizing the parameters of the SVM model and automatically evaluating the level of limb dysfunction are as follows:
[0100] ① The kernel function in the support vector machine model uses the radial basis kernel function, and its definition expression is:
[0101]
[0102] Where σ is the radial base radius;
[0103] ②Definition Penalty factor C;
[0104] ③ Set the initial model parameters C, g optimization range and step size, traverse each set of C, g values through grid search, and calculate the classification accuracy of the model;
[0105] ④ Select the model parameters C and g with the highest classification accuracy as the optimal parameters of the model.
[0106] Finally, based on the trained model, automatic assessment of limb dysfunction in patients with low back pain can be achieved.
[0107] In summary, the automatic assessment method for limb dysfunction of patients with low back pain based on multi-source features provided in this embodiment uses machine learning methods and combines multi-source features such as behavior and imaging to achieve automatic assessment of limb dysfunction of patients with low back pain. It has the characteristics of high accuracy, strong reliability, convenient use, and easy clinical promotion, and is of great significance for the formulation of clinical treatment plans for low back pain, efficacy evaluation, rehabilitation training, and assistive device adaptation.
Claims
1. An automatic assessment method for limb dysfunction in patients with low back pain based on multi-source features, characterized in that: The specific implementation steps are as follows: Step 1: Use the U-Net neural network to automatically segment the multifidus and erector spinae muscles in the lumbar MRI images of patients with low back pain, and use the threshold method to extract the imaging features of muscle cross-sectional area and fat infiltration; Step 2: Based on multiple clinical assessment scales for low back pain patients, screen out the scales with strong correlation with limb function and extract the scale characteristics; The screening of low back pain scales and the extraction of scale characteristics are as follows: after research and screening, the scales with strong correlation with limb dysfunction in patients with low back pain are selected, mainly including: ODI scoring standard and Japanese Orthopedic Association evaluation treatment score, and the sub-question scores and total scores of the above scales are used as the behavioral characteristics of the scales; Step 3: Normalize the scale features and imaging features to form a feature set; Step 4: Use the support vector machine recursive feature elimination algorithm to screen all features, reduce feature redundancy, and determine the optimal feature combination; Step 5: Select the support vector machine method in supervised machine learning to build an evaluation model, use the optimal feature combination as the model input, and use the visual analog scale score as the tutor supervision signal of the evaluation model; Step 6: Use the grid optimization method to optimize the model parameters, and through model training, complete the construction of the limb dysfunction assessment model for low back pain patients, and realize the automatic assessment of the degree of limb dysfunction in low back pain patients.
2. The method according to claim 1, characterized in that The method of extracting MRI imaging features using U-Net neural network and threshold method in step 1 is as follows: ① Select three MRI axial images of the patient's L3-L4 spinal disc; ② Improve the contrast of MRI images based on contrast-limited adaptive histogram equalization, which limits noise amplification and enhances local contrast by reducing the height of the local histogram; ③Build a U-Net network based on the Adam optimizer with a learning rate of 10 -4 , the attenuation factor is 10 -6 , the number of iterations is 100, the momentum coefficient is 0.9, and the graphics processor is used for training and testing, and finally the automatic segmentation of multifidus and erector spinae muscles is achieved; ④Use the maximum inter-class variance method to distinguish muscle and fat tissue; determine a gray value as the threshold based on the histogram of the image, and then divide the histogram of the image into two different classes based on this threshold, so that the inter-class variance of the two classes reaches the maximum value: s t 2 =ω0*ω1*(μ0-μ1) 2 In the formula, σ t 2 is the between-class variance of the two classes, ω0 and ω1 are the occurrence probabilities of the two classes, and μ0 and μ1 are the means of the two classes; Traverse all the grayscale values of the image and find Get the maximum gray value t, then the optimal threshold is t; ⑤ Based on the number of pixels of muscle and fat tissue, calculate the muscle cross-sectional area and fat infiltration characteristics: CSA=p total *K Where CSA is the cross-sectional area of muscle; p total is the sum of the number of pixels in ROI; K is the scale; Where, FIR is the fat infiltration characteristic; p fat is the total number of fat pixels after threshold segmentation; p total is the total number of pixels in the area; Finally, the muscle cross-sectional area and fat infiltration characteristics of the multifidus and erector spinae muscles were used as MRI imaging features.
3. The method according to claim 1, characterized in that The calculation method for feature normalization in step 3 is: In order to avoid the influence of different eigenvalues on the subsequent model construction, all scales and imaging features were normalized according to the following formula; In the formula, x nor is the normalized feature, x max is the maximum value of the feature, x min is the minimum value of the feature, and x is the feature value.
4. The method according to claim 1, characterized in that The specific implementation method of the SVM-RFE algorithm used for feature screening in step 4 is: ① Use the data set to train the SVM model and obtain the trained SVM parameters; ② Select an appropriate ranking criterion and calculate the scores of the behavioral characteristics of the scale and the MRI imaging characteristics under the criterion respectively; ③ Eliminate the feature with the lowest score in the current data set and perform the above process until the last feature remains; Commonly used sorting criteria are: R c =|W 2 -W -(P)2 | Where W 2 and W -(P)2 They represent the weight of the complete SVM model and the weight of the SVM model after assuming that the Pth feature is removed; W 2 The representation is: In the formula, i and j represent loop variables; y represents the category label; N represents the number of samples; K(x i ,x j ) represents the kernel function; and Obtained by solving the SVM dual optimization problem; The SVM-RFE algorithm sorts the scores of each feature, removes the feature with the smallest score, and then uses the remaining features to train the model again for the next iteration; finally, the optimal feature combination is obtained.
5. The method according to claim 1, characterized in that The specific method of selecting supervisory signals and building evaluation models in step 5 is as follows: The optimal feature combination was used as the model input, and the VAS score was used as the model's mentor supervision signal. According to the VAS score range, low back pain patients were defined as: mild disorder, moderate disorder, and severe disorder. Among them, VAS scores of 0-3 were mild disorders, 4-6 were moderate disorders, and 7-10 were severe disorders. The evaluation model was constructed using the multi-classification support vector machine method, in which the radial basis function was selected as the kernel function; the model output was divided into three levels, defined as "1", "2", and "3", corresponding to mild disability, moderate disability, and severe disability, respectively.
6. The method according to claim 1, characterized in that The specific method for optimizing the SVM model parameters in step 6 is: In order to improve the evaluation accuracy of the model, the model parameters are optimized. The process is as follows: ① The kernel function in the support vector machine model uses the radial basis kernel function, and its definition expression is: Where σ is the radial base radius; x i 、x j are two eigenvectors; ②Definition Penalty factor C; ③ Set the initial model parameters C, g optimization range and step size, traverse each set of C, g values through grid search, and calculate the classification accuracy of the model; ④ Select the model parameters C and g with the highest classification accuracy as the optimal parameters of the model; Based on the trained model, automatic assessment of limb dysfunction in patients with low back pain can be achieved.
Citation Information
Patent Citations
Muscle amount increasing method
JP2006061687A
Method for evaluating and treating musculoskeletal dysfunctions
US20090156968A1
Mr spectroscopy system and method for diagnosing painful and non-painful intervertebral discs
US20110087087A1
Assessment of Spinal Anatomy
US20120143090A1
Method and system for managing pain
US20160213276A1
