A method for predicting the risk of neck, back, and waist pain based on muscle ultrasound imaging analysis
By constructing a deep neural network model based on muscle ultrasound images, and combining multi-dimensional data and standardized processing, the problems of low diagnostic accuracy and subjective pain grading in existing technologies for myofascial pain syndromes are solved, and efficient myofascial pain risk prediction and quantification are achieved.
Patent Information
- Application Number
- CN202611533156.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
Current technologies cannot effectively predict and quantify the risk of systemic myofascial pain, especially myofascial pain syndrome, which has a low diagnostic accuracy, cannot achieve objective grading of pain severity, and relies on patient subjective pain scores which are prone to error.
By collecting multi-dimensional data from subjects, combining ultrasound images and elastography to extract morphological and elastic mechanical parameters, a fully connected deep neural network model is constructed, feature selection and encoding are performed, a synthetic minority class oversampling algorithm is used for class balancing, the model is trained using standardized parameters, and blind prediction is performed in an independent queue.
It improved the accuracy of myofascial pain risk prediction, achieved stable identification of level 4 pain risk, and enhanced the model's generalization ability in real clinical scenarios, with an accuracy of 87.2% and a multi-class AUC of 0.937.
Smart Images

Figure CN123120274A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence-assisted diagnosis technology in medical ultrasound imaging, and includes, but is not limited to, a method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis. Background Technology
[0002] Myofascial pain syndrome (MPS) is a chronic soft tissue pain disorder characterized by the presence of myofascial trigger points (MTrPs) within skeletal muscles. It commonly affects the neck, shoulder, lower back, scapula, and the areas surrounding the joints of the limbs. Typical clinical manifestations include localized tension and spasm of muscle fibers, and localized pain and referred pain induced by pressure. Clinically, MTrPs are usually classified into active trigger points (A-MTrPs) that cause spontaneous pain and latent trigger points (L-MTrPs) that cause pain only upon pressure.
[0003] In existing technologies, musculoskeletal ultrasound, which involves no ionizing radiation and supports real-time dynamic scanning at the bedside, combined with shear wave elastography (SWE), can quantitatively assess the stiffness of fascia and muscle tissue, making it the preferred imaging method for non-invasive and objective assessment of myofascial pain syndrome (MTrPs). However, existing technologies have significant limitations: Firstly, current research integrates factors such as shear wave velocity, anal right angle morphology, pelvic floor hiatus area, and weight gain to construct a predictive model for postpartum pelvic floor myofascial pain. This model is only suitable for diagnosing myofascial pain in a single postpartum pelvic floor area and cannot be universally applied to myofascial lesions of the entire trunk and limbs. Secondly, the accuracy of pure ultrasound texture feature recognition is low. Existing methods only extract texture features such as local binary pattern (LBP) and Gabor filtering from B-mode grayscale images to distinguish between A-MTrPs and L-MTrPs, without combining basic human indicators, quantitative elastography, and clinical palpation information. Its predictive accuracy is only 53.33%, and it cannot perform pain severity grading. Furthermore, clinically used pain rating scales (such as the Numerical Rating Scale (NRS) and the Visual Analogue Scale (VAS) rely entirely on patients' subjective feelings. However, individual tolerance varies greatly, and pain fluctuates significantly between day and night. Relying solely on these scales cannot achieve an objective quantitative classification of pain severity. Mild cases are easily missed, and the pain level of severe cases is easily underestimated, resulting in a lack of objective and reliable stratification criteria for clinical diagnosis and treatment decisions.
[0004] Therefore, there is an urgent need for a more comprehensive method for predicting the risk of neck, back and waist pain, in order to solve the problems existing in the current technology, thereby achieving accurate prediction and objective quantitative classification of the risk of myofascial pain throughout the body, and improving the clinical generalization ability. Summary of the Invention
[0005] This application provides a method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis.
[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for predicting the risk of neck, back, and lumbar pain based on muscle ultrasound imaging analysis. The method includes: collecting multi-dimensional data of subjects; extracting morphological and elastic parameters based on ultrasound images and elastography data; converting digital pain scores into pain risk labels; associating multi-dimensional data of the same subject with pain risk labels to construct an original dataset; performing feature screening and categorical feature encoding on the original dataset; dividing the dataset into training and validation sets based on pain risk labels; performing class balancing on the training set only using a synthetic minority class oversampling algorithm; calculating feature standardization parameters based on the balanced training set; and standardizing the balanced training and validation sets using the same standardization parameters; constructing a fully connected deep neural network; training the fully connected deep neural network based on the standardized balanced training and validation sets to obtain the optimal model; enrolling an independent prospective patient cohort after the modeling time window; processing the independent cohort data using the same standardization parameters as in the training phase; and having a third party independent of the model development process load the optimal model in a blinded manner to predict the standardized independent cohort data and output the prediction results.
[0007] Secondly, embodiments of this application provide a neck, back, and lumbar pain risk prediction system based on muscle ultrasound image analysis, comprising: The data acquisition module is used to collect multi-dimensional data of subjects, extract morphological and elastic mechanical parameters based on ultrasound and elastography data, convert digital pain scores into pain risk labels, and associate multi-dimensional data of the same subject with pain risk labels to construct the original dataset. The data preprocessing module is used to perform feature filtering and categorical feature encoding on the original dataset, and to divide the training set and validation set according to the pain risk label; The model training module is used to perform class balancing on the training set only by using a synthetic minority class oversampling algorithm, calculate the feature standardization parameters based on the balanced training set, and use the same standardization parameters to standardize both the balanced training set and the validation set; construct a fully connected deep neural network, train the fully connected deep neural network based on the standardized balanced training set and the validation set, and obtain the optimal model; The prediction module is used to enroll independent prospective patient cohorts after the modeling time window. It processes the independent cohort data using the same standardized parameters as the training phase. A third party, independent of the model development process, loads the optimal model in a blinded manner to predict the standardized independent cohort data and outputs the prediction results.
[0008] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the above-mentioned method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis.
[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the above-described method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis.
[0010] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides a method for predicting the risk of neck, back, and waist pain based on muscle ultrasound imaging analysis. By collecting multi-dimensional data from subjects, morphological and elastic parameters are extracted from ultrasound and elastography data. The numerical pain score is converted into a four-level pain risk label, and a structured raw dataset covering four heterogeneous dimensions—anthropometry, ultrasound morphology, elastography, and clinical palpation—is constructed. Ablation experiments show that the model accuracy decreases significantly after any feature is removed. SHAP feature importance analysis shows that the elastic modulus of the control area is the primary predictive feature of the model, followed by height, lesion depth, BMI, and age, demonstrating that the fusion of the four heterogeneous features produces a synergistic effect. By sequentially performing outlier removal, missing value imputation, correlation analysis, feature selection for multicollinearity diagnosis, categorical feature encoding, and stratified partitioning on the original dataset, a standardized preprocessing workflow for clinically unbalanced pain samples was constructed. Among these methods, box plots combined with Z-Score thresholding were used for anomaly detection, median imputation was used for continuous feature missing values, mode imputation was used for categorical feature missing values, redundant features with VIF greater than 10 were progressively removed, and 7:3 stratification was used to ensure that the proportion of each category was consistent with the original distribution. Furthermore, validation set information was not involved in the determination of any preprocessing parameters, thus avoiding data leakage from the source and providing a high-quality, unbiased data foundation for model training. By employing a synthetic minority class oversampling algorithm to balance the number of risk samples across all four levels of the training set, the model addresses the bias caused by the scarcity of high-risk samples. Standardized parameters are calculated based on the balanced training set, and the same parameters are used to transform the validation set, ensuring that validation set statistical parameters are not involved in the calculation. A fully connected deep neural network combined with dual regularization, early stopping, and grid search achieves an optimal model accuracy of 88.1% on the validation set and 89.6% on the internal test set, with a multi-class AUC of 0.943. Compared to the pure ultrasonic texture recognition model's accuracy of 53.33%, this approach improves accuracy by 36.27 percentage points while maintaining stable recognition accuracy across all four risk levels, preventing missed detections of high-risk severe cases.
[0011] By enrolling independent prospective patient cohorts after the modeling time window, and processing independent data using standardized parameters saved during the training phase, an independent third party loaded the optimal model for prediction in a blinded manner. This constructed an external validation system with triple isolation: time isolation, parameter locking, and operational blinding. The blinded test results of 247 independent prospective cohorts showed an overall accuracy of 87.2% and a multi-class AUC of 0.937, highly consistent with the internal validation, demonstrating the model's good generalization ability in real-world clinical scenarios. Most similar imaging AI models only use internal segmentation validation on a single dataset. This application added a new patient cohort blind test outside the modeling cycle, fully validating the model's generalization ability in real-world clinical scenarios. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis, provided in an embodiment of this application. Figure 2 This application provides a schematic diagram of ultrasonic quantitative parameter measurement of the hypoechoic region of a myofascial trigger point. Figure 3 This is a schematic diagram of the classification of affected myofascia provided in an embodiment of this application; Figure 4 This application provides a neck, back, and waist pain risk prediction system based on muscle ultrasound image analysis. Detailed Implementation
[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0014] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0015] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.
[0016] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.
[0017] The embodiments of this application will be further described below with reference to the accompanying drawings.
[0018] In view of the current problems in the research on risk prediction of neck, back and waist pain in the field of artificial intelligence-assisted diagnosis of medical ultrasound imaging, this application provides a method for predicting the risk of neck, back and waist pain based on muscle ultrasound image analysis.
[0019] The technical solution of this application is described below, starting with the method embodiments.
[0020] Please refer to Figure 1 It illustrates a flowchart of a method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis, as provided in an embodiment of this application. Figure 1 As shown, the method includes at least the following steps S110 to S140.
[0021] Step S110: Collect multi-dimensional data of the subjects, extract morphological parameters and elastic mechanical parameters based on ultrasound images and elastography data, and convert digital pain scores into pain risk labels; associate the multi-dimensional data of the same subject with the pain risk labels to construct the original dataset.
[0022] In this embodiment, multidimensional raw data are collected for each enrolled subject. This includes basic human information, pain location, musculoskeletal ultrasound and shear wave elastography data, clinical palpation information, and digital pain score. Specifically, basic human information is collected, including recording the subject's age, sex, height, and weight, and calculating the body mass index based on height and weight. The location of myofascial pain is collected and recorded, including any one or more areas from the neck and shoulder region, scapular region, chest and back region, upper lumbar region, lateral abdominal muscle group region, lower lumbar region, and limb muscle group region. This data collection is performed and recorded by clinicians according to standardized physical examination procedures.
[0023] Furthermore, musculoskeletal ultrasound scanning and shear wave elastography were performed on the painful lesion areas of the subjects. Specifically, a high-frequency linear array ultrasound probe was used to perform multi-section, standardized B-mode ultrasound scanning of the skeletal muscles at the site of the subject's reported pain. Simultaneously, shear wave elastography was activated to acquire elastograms of the lesion area in real time. Based on the ultrasound imaging characteristics of the myofascial trigger points, the ultrasound physician identified and marked the hypoechoic lesion areas of the myofascial trigger points on the B-mode images, and marked a control area in the adjacent normal tissue area of the same skeletal muscle, located 3 mm from the edge of the lesion. Morphological and elastic mechanical parameters were quantitatively extracted from the ultrasound images and elastography data. Specifically, morphological parameters of the hypoechoic area at the myofascial trigger point were measured and recorded on B-mode ultrasound images. These parameters included the major and minor axes of the hypoechoic area, lesion depth, and target skeletal muscle thickness. The major axis was the maximum diameter along the direction of muscle fiber travel, the minor axis was the maximum diameter perpendicular to the direction of muscle fiber travel, the lesion depth was the vertical distance from the skin surface to the center of the hypoechoic area, and the target skeletal muscle thickness was the full-thickness thickness of the skeletal muscle measured at the location of the myofascial trigger point. Each parameter was measured multiple times and the average was taken to reduce random error. Regions of interest were set in the lesion area and the control area on the shear wave elastography elastogram. Elastic parameters were extracted, including the elastic modulus of the lesion area and the elastic modulus of the control area 3 mm from the lesion edge. During extraction, obvious blood vessels and skeletal muscle fascia were avoided. Each parameter was measured multiple times and the average was taken.
[0024] Further, clinical palpation information of the subjects' muscles was collected, including the fascial classification of the affected myofascia. This information was determined and recorded by clinicians using standardized palpation techniques. The subjects' numerical pain scores were collected, ranging from 0 to 10. Subjects completed the scoring independently in a quiet environment, with 0 representing no pain and 10 representing the most severe pain. The numerical pain scores were converted into four levels of pain risk labels according to a pre-defined grading rule: a numerical pain score of 0 corresponds to no risk; a score of 1, 2, or 3 corresponds to low risk; a score of 4, 5, or 6 corresponds to medium risk; and a score of 7, 8, 9, or 10 corresponds to high risk.
[0025] In this embodiment, all collected data from the same subject are associated with pain risk labels to construct a structured raw dataset. Specifically, using the subject ID as the primary key, all the data collected and calculated above are horizontally concatenated and associated by subject to form a complete structured raw sample record. All data includes basic human information such as age, gender, height, weight, and body mass index (BMI); pain location; morphological parameters such as the long and short diameters of the hypoechoic area of the myofascial trigger point, lesion depth, and target skeletal muscle thickness; elastic parameters such as the elastic modulus of the lesion area and the elastic modulus of the control area 3 mm from the lesion edge; clinical palpation information such as myofascial classification; and a four-level pain risk label. The above process is repeated until all enrolled subjects have completed data collection and feature construction, resulting in a structured raw dataset.
[0026] Step S120: Perform feature filtering and categorical feature encoding on the original dataset, and divide the training set and validation set according to the pain risk label.
[0027] In this embodiment, outlier removal is performed on continuous features in the structured raw dataset. Specifically, a box plot method combined with a Z-Score thresholding method is used to detect anomalies in all continuous features. The box plot method calculates the interquartile range (IMR) of each continuous feature and identifies values less than the first quartile minus 1.5 IMR or greater than the third quartile plus 1.5 IMR as outliers. The Z-Score thresholding method calculates the standardized score of each continuous feature and identifies values with an absolute value greater than 3 as outliers. In this embodiment, if any continuous feature of a subject is identified as outlier by either of the above two methods, all data samples of that subject are removed.
[0028] Furthermore, missing values in the dataset after outlier removal are imputed. Specifically, missing values are imputed according to feature type: for missing values of continuous features, the median of the feature is used; for missing values of categorical features, the mode of the feature is used.
[0029] Furthermore, feature filtering is performed on the padded dataset to retain core input features. Specifically, this feature filtering includes two dimensions: correlation analysis and multicollinearity diagnosis. In the correlation analysis dimension, the correlation coefficient between each candidate feature in the padded dataset and the four-level pain risk label is calculated, and a correlation heatmap is plotted to preliminarily observe the strength of the association between each feature and the label, as well as the pairwise correlation patterns between features. In this embodiment, Spearman's rank correlation coefficient is used for the correlation analysis between continuous features and categorical labels; Pearson's correlation coefficient is used for the correlation analysis between continuous features. In the multicollinearity diagnosis dimension, the variance inflation factor among each candidate feature is calculated. This variance inflation factor is used to quantify the degree to which each feature is linearly represented by all other features, and is independent of and complementary to the correlation analysis. Redundant features with a variance inflation factor greater than 10 are gradually eliminated. Specifically, in each round, the feature with the largest variance inflation factor value is eliminated, and the variance inflation factor of the remaining features is recalculated. This process is repeated until the variance inflation factor of all remaining features is less than or equal to 10. After the above feature screening, a total of 12 core input features were obtained, including age, gender, height, weight, body mass index (BMI), long diameter of the hypoechoic area of the myofascial trigger point, short diameter of the hypoechoic area of the myofascial trigger point, lesion depth, skeletal muscle thickness, elastic modulus of the control area, elastic modulus of the lesion area, and fascial classification.
[0030] Furthermore, one-hot encoding is performed on the categorical features in the core input features. Specifically, the categorical features include gender and fascial type. One-hot encoding expands each value of the categorical feature into an independent binary numerical feature column, thereby transforming all core input features into numerical values.
[0031] In this embodiment, the dataset after one-hot encoding is randomly divided into training and validation sets in a 7:3 ratio based on a four-level pain risk label. Specifically, all samples are grouped according to the four risk categories. Within each risk category group, 70% and 30% of the samples are randomly sampled independently. The 70% samples from each group are combined to form the training set, and the 30% samples are combined to form the validation set. This stratified random partitioning ensures that the sample ratios of the four risk categories in the training and validation sets are completely consistent with the category distribution of the original dataset. In this embodiment, the stratified random partitioning uses the StratifiedShuffleSplit algorithm. This algorithm shuffles the sample order before each partition and then performs stratified sampling by category to ensure that the partitioning result is both random and maintains the stability of the category ratio. The resulting training set is used for parameter learning and weight updates during subsequent model training, while the validation set is used for hyperparameter tuning, early stopping monitoring, and generalization ability evaluation during subsequent model training. The two sets are isolated from each other after construction.
[0032] Step S130: Perform class balancing on the training set only by synthesizing minority class oversampling algorithm, calculate feature standardization parameters based on the balanced training set, and use the same standardization parameters to standardize the balanced training set and validation set respectively; construct a fully connected deep neural network, train the fully connected deep neural network based on the standardized balanced training set and validation set, and obtain the optimal model.
[0033] In this embodiment, a synthetic minority oversampling algorithm is used to perform class balancing on the training set. Specifically, the synthetic minority oversampling algorithm is the SMOTE algorithm. The SMOTE algorithm synthesizes virtual samples only for high-risk samples in the training set, thus balancing the number of samples at all four risk levels in the training set. The specific operation is as follows: For each original sample in the high-risk category, find its K nearest neighbors in the feature space (K is set to 5 in this embodiment). Randomly select one nearest neighbor from the K nearest neighbors and perform random linear interpolation on the line connecting the original sample and the selected nearest neighbor to generate a new virtual sample. The mathematical expression of this interpolation process is as follows:
[0034] In the formula, This is the feature vector of an original sample in the high-risk category. Let be the feature vector of a randomly selected nearest neighbor sample from the K nearest neighbors of this sample. A random number between 0 and 1 This is the feature vector of the synthesized new sample. The above process is repeated until the number of samples in the four categories (no risk, low risk, medium risk, and high risk) in the training set is balanced. In this embodiment, the validation set does not participate in any oversampling operation and maintains its original category distribution.
[0035] In this embodiment, feature standardization parameters are calculated based on the balanced training set. Specifically, these feature standardization parameters are either MinMaxScaler normalization parameters or StandardScaler normalization parameters. When using MinMaxScaler normalization parameters, for each core input feature in the balanced training set, the minimum and maximum values of that feature are calculated, and the normalization formula is as follows:
[0036] In the formula, This refers to the original value of a core input feature. This is the minimum value of the feature in the balanced training set. This is the maximum value of the feature in the balanced training set. These are the normalized values. When using the StandardScaler standardization parameters, for each core input feature in the balanced training set, the mean and standard deviation of that feature are calculated. The standardization formula is:
[0037] In the formula, This refers to the original value of a core input feature. This is the mean of the feature in the balanced training set. The standard deviation of this feature in the balanced training set. These are the standardized values. In this embodiment, the standardized parameters are entirely derived from the training set, and the validation set uses the same parameter transformation. The statistical parameters of the validation set itself do not participate in the standardization calculation.
[0038] Furthermore, the same standardization parameters are used to transform the training set and the validation set respectively. Specifically, based on the minimum, maximum or mean, and standard deviation of each feature calculated from the balanced training set, the balanced training set is standardized to obtain a standardized balanced training set; at the same time, the validation set is standardized using the exact same standardization parameters to obtain a standardized validation set.
[0039] In this embodiment, a fully connected deep neural network is constructed, comprising an input layer, hidden layers, and an output layer. The input layer has 12 neurons, corresponding one-to-one with 12 core input features. The hidden layers consist of 3 to 5 fully connected layers, with the number of neurons in each hidden layer decreasing progressively after the input layer. In a preferred embodiment, there are 4 hidden layers with 256, 128, 64, and 32 neurons respectively, and all hidden layers use the ReLU activation function. The output layer has 4 neurons, corresponding to four levels of pain risk, and the output value is transformed into a probability distribution of the four risk levels using the Softmax activation function.
[0040] In this embodiment, a fully connected deep neural network is iteratively trained based on a standardized, balanced training set. Specifically, the loss function is set as classification cross-entropy, and the loss function is... The mathematical expression is:
[0041] In the formula, For the real label in the first The value on the category, when the sample belongs to the category The value is 1 if there is a specific category, otherwise it is 0; The model predicts that the sample belongs to the first... The probability of each class is calculated. The optimizer is set to Adam, with an initial learning rate selected between 0.0005 and 0.002. Dropout regularization and L2 weight regularization are configured simultaneously to prevent overfitting, with the Dropout inactivation rate selected between 0.3 and 0.5. The training batch size is set to 32 or 64, and the maximum number of training epochs is 100 to 300. The standardized, balanced training set is input into the fully connected deep neural network in batches according to the set batch size. The prediction loss is calculated using the classification cross-entropy loss function, and the network weight parameters are updated along the gradient descent direction using the Adam optimizer to complete one round of iterative training.
[0042] The training process is monitored using a standardized validation set, and early stopping is implemented. Specifically, after each training round, the loss value is calculated on the current standardized validation set, and an early stopping mechanism is configured to monitor the validation set loss. Training is automatically terminated when the validation set loss no longer decreases within a predetermined number of consecutive rounds. In a preferred embodiment of this application, the predetermined number of consecutive rounds is 20 rounds, meaning that training is automatically terminated and the model weights are restored to the state where the validation set loss was lowest when the validation set loss did not decrease for 20 consecutive rounds.
[0043] The hyperparameters are systematically optimized using either grid search or random search. Specifically, grid search or random search is performed on the number of hidden layers, Dropout inactivation rate, training batch size, and initial learning rate. The search range for the number of hidden layers is 3, 4, and 5 layers; the search range for the Dropout inactivation rate is 0.3, 0.4, and 0.5; the search range for the training batch size is 32 and 64; and the search range for the initial learning rate is 0.0005, 0.001, and 0.002. The entire process of network construction, training, and early stopping is repeated for each hyperparameter combination.
[0044] The optimal model is selected based on validation set performance metrics, and its weights are preserved. Specifically, validation set accuracy and multi-class AUC are the core selection metrics, supplemented by confusion matrix and ROC curve for comprehensive evaluation. Accuracy, confusion matrix, ROC curve, and multi-class AUC are calculated independently on the training and validation sets, respectively. The multi-class AUC is calculated using a one-to-many strategy, i.e., the AUC value is calculated separately for each risk class, and then a macro-average or weighted average is taken for the AUC values of all risk classes. The model corresponding to the hyperparameter combination with the highest validation set accuracy and optimal multi-class AUC is selected from all hyperparameter combinations as the optimal model.
[0045] Step S140: After the modeling time window, an independent prospective patient cohort is enrolled, and the independent cohort data is processed using the same standardized parameters as in the training phase; a third party independent of the model development process loads the optimal model in a blinded manner to predict the standardized independent cohort data and outputs the prediction results.
[0046] In this embodiment, an independent prospective patient cohort is enrolled after the modeling sample enrollment window ends. This independent cohort is completely isolated from the modeling sample in time; all subjects in the independent cohort are enrolled after the modeling sample enrollment window ends, ensuring that the data from this independent cohort is not used in the training process in any way. The inclusion criteria for this independent cohort are consistent with those for the modeling sample, including patients meeting the clinical diagnostic criteria for myofascial pain syndrome, and covering myofascial lesions in any one or more of the following areas: neck and shoulder, scapular region, chest and back, upper lumbar region, lateral abdominal muscles, lower lumbar region, and limb muscles.
[0047] In this embodiment, for independent prospective patient cohorts, the same scheme as in the training phase is used to acquire raw data and perform preprocessing. The independent cohort data is then standardized based on the standardized parameters of the training set to obtain a standardized independent cohort dataset. In this embodiment, the standardized parameters are completely locked to the parameters calculated from the training set; the independent cohorts do not recalculate their own standardized parameters.
[0048] In this embodiment, a third party independent of the model development process loads the optimal model in a blinded manner to predict the independent cohort data. Specifically, the optimal model is encapsulated as an independent inference module, which includes the network structure and network weight file of the optimal model. An independent third-party operator, not involved in the entire model development process, loads the inference module and performs risk level prediction on the standardized independent cohort dataset. During the prediction process, the operator remains blinded to the true level 4 pain risk labels of the independent cohorts; that is, the operator does not know the true risk level attribution of each sample in the independent cohorts.
[0049] In this embodiment, the predicted results are compared with the true labels to evaluate the model's generalization ability. Specifically, after the prediction is completed, the true level 4 pain risk labels of the independent cohort are obtained through unblinding, and the predicted risk level output by the model is compared with the true labels sample by sample. The overall accuracy of the model on the independent cohort is calculated, and the multi-class AUC value is calculated using the same multi-class AUC calculation strategy as described above. In one specific embodiment of this application, the independent cohort contains 247 samples, and the overall accuracy of the optimal model on the independent cohort is 87.2%, with a multi-class AUC of 0.937, demonstrating that the optimal model has good real-world clinical generalization ability.
[0050] After confirming the generalization ability through external validation, the optimal model can be deployed for practical applications such as bedside rapid screening and clinical auxiliary diagnosis. During deployment, for any subject to be tested, multi-dimensional raw data is obtained, 12 core input features are extracted, and these 12 features are standardized based on the training set's standardized parameters. The standardized features are then input into the optimal model, which outputs a probability distribution of four levels of pain risk. The risk category with the highest probability is taken as the subject's predicted risk level, and the grading result is output to assist clinical decision-making.
[0051] The following describes a specific embodiment of the above-mentioned method for predicting the risk of neck, back, and waist pain based on muscle ultrasound image analysis. However, it is worth noting that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.
[0052] In a specific embodiment, the total enrolled modeling sample consisted of 1105 patients, covering myofascial pain syndrome patients in the neck and shoulder region, scapular region, chest and back region, and lateral abdominal muscle group region. The risk stratification distribution of these 1105 samples was as follows: no risk 289 cases, low risk 512 cases, intermediate risk 210 cases, and high risk 94 cases, with the high-risk sample being a minority. Basic human information of the subjects was collected, including age, sex, height, weight, and body mass index (BMI). Musculoskeletal ultrasound scanning and shear wave elastography were performed on the painful lesion areas of the subjects. The ultrasound physician identified and marked the hypoechoic areas of the myofascial trigger points on the B-mode ultrasound images, and a control area was marked 3mm from the lesion edge. The long axis, short axis, lesion depth, and target skeletal muscle thickness of the hypoechoic areas of the myofascial trigger points were measured on the B-mode ultrasound images. Please refer to [reference needed]. Figure 2 This illustration shows a schematic diagram of ultrasound quantitative parameter measurement of the hypoechoic area of a myofascial trigger point according to an embodiment of this application. On the elastogram of shear wave elastography, the elastic modulus of the lesion area and the control area are extracted, respectively. The myofascial classification of the affected myofascia is determined and recorded by clinicians using standardized palpation techniques. Please refer to [reference needed]. Figure 3 The illustration shows a schematic diagram of an affected myofascia classification provided in an embodiment of this application. The subject's digital pain score is collected and converted into a four-level pain risk label: 0 points indicate no risk, 1 to 3 points indicate low risk, 4 to 6 points indicate medium risk, and 7 to 10 points indicate high risk.
[0053] During data preprocessing, a box plot method combined with the Z-Score thresholding method was used to detect anomalies in all continuous features. The absolute value of the Z-Score threshold was set to 3, and samples with any continuous feature deemed abnormal were removed entirely, leaving 1000 valid samples after cleaning. Missing values in the dataset after outlier removal were imputed. Missing values for continuous features were imputed using the median, including age, body mass index (BMI), and elastic modulus; missing values for categorical features were imputed using the mode, including gender and myofascial type. Categorical features were uniformly encoded using one-hot encoding before being input into the model. The correlation coefficients between each candidate feature and the four-level pain risk label were calculated and heatmaps were plotted. The variance inflation factor (VIF) among each candidate feature was calculated, and redundant variables with a VIF greater than 10 were gradually removed, ultimately retaining 12 core input features: age, gender, height, weight, body mass index (BMI), long and short diameters of the hypoechoic area of the myofascial trigger point, lesion depth, skeletal muscle thickness, elastic modulus of the control area, elastic modulus of the lesion area, and myofascial type.
[0054] During dataset splitting and sample balancing, the dataset after one-hot encoding was randomly divided into training and validation sets in a 7:3 ratio, based on the four-level pain risk label. The resulting training set had 700 cases and the validation set had 300 cases. The SMOTE synthetic minority class oversampling algorithm was used to synthesize dummy samples only for high-risk samples in the training set, thus balancing the number of four-level risk samples in the training set. After balancing, the total number of samples in the training set was 2080.
[0055] When building and training the deep neural network model, a fully connected deep neural network was built based on the TensorFlow 2.8 framework. The input layer had 12 neurons, and the hidden layers consisted of 4 fully connected layers with 256, 128, 64, and 32 neurons respectively. All hidden layers used the ReLU activation function, and the output layer had 4 neurons using the Softmax activation function. The loss function was classification cross-entropy, the optimizer was Adam, the initial learning rate was 0.001, the Dropout inactivation rate was 0.4, the L2 weight regularization coefficient was 0.0001, the training batch size was 32, and the maximum number of training epochs was 200. An early stopping mechanism was configured to monitor the validation set loss, and training was automatically terminated when the validation set loss no longer decreased within 20 consecutive epochs. A grid search was performed on the number of hidden layers, Dropout inactivation rate, training batch size, and initial learning rate, ranging from 3 to 5 layers, 0.3 to 0.5 layers, 32 or 64 layers, and 0.0005 to 0.002 layers, respectively. Validation accuracy and multi-class AUC were used as the core selection metrics. The optimal hyperparameter combination was 4 hidden layers, a Dropout inactivation rate of 0.4, a training batch size of 32, and an initial learning rate of 0.001. The optimal model achieved an accuracy of 88.1% on the validation set, 89.6% on the internal test set, and a multi-class AUC of 0.943.
[0056] In the external blinded validation, a new cohort of 247 independent prospective patients was enrolled after the modeling sample enrollment window ended. Raw data for this independent cohort was acquired using the same collection protocol and preprocessing procedures as the modeling sample. The data from the independent cohort was standardized using the training set parameters saved during the training phase. The optimal model was then encapsulated as an independent inference module, which was loaded by an independent third-party operator not involved in the model development process and used for blinded prediction on the independent cohort data. After the predictions were completed and unblinded, the prediction results were compared with the true labels. The blinded test results of the optimal model on the 247 independent prospective cohort showed an overall accuracy of 87.2% and a multi-class AUC of 0.937.
[0057] In summary, the neck, back, and waist pain risk prediction method based on muscle ultrasound image analysis provided in this application collects multi-dimensional data from subjects, extracts morphological and elastic parameters from ultrasound images and elastography data, and converts digital pain scores into four-level pain risk labels. This constructs a structured raw dataset covering four heterogeneous dimensions: anthropometry, ultrasound morphology, elastography, and clinical palpation. Ablation experiments demonstrate that the model accuracy significantly decreases after any feature is removed. SHAP feature importance analysis shows that the elastic modulus of the control area is the primary predictive feature of the model, followed by height, lesion depth, BMI, and age, proving that the fusion of the four heterogeneous features produces a synergistic effect. By sequentially performing outlier removal, missing value imputation, correlation analysis, feature selection for multicollinearity diagnosis, categorical feature encoding, and stratified partitioning on the original dataset, a standardized preprocessing workflow for clinically unbalanced pain samples was constructed. Among these methods, box plots combined with Z-Score thresholding were used for anomaly detection, median imputation was used for continuous feature missing values, mode imputation was used for categorical feature missing values, redundant features with VIF greater than 10 were progressively removed, and 7:3 stratification was used to ensure that the proportion of each category was consistent with the original distribution. Furthermore, validation set information was not involved in the determination of any preprocessing parameters, thus avoiding data leakage from the source and providing a high-quality, unbiased data foundation for model training. By employing a synthetic minority class oversampling algorithm to balance the number of risk samples across all four risk levels by performing class balancing on the training set, the model's identification bias caused by the scarcity of high-risk samples is resolved. Standardized parameters are calculated based on the balanced training set, and the same parameters are used to transform the validation set, ensuring that the validation set statistical parameters are not involved in the calculation. Based on a fully connected deep neural network combined with dual regularization, early stopping, and grid search, the optimal model achieves an accuracy of 88.1% on the validation set and 89.6% on the internal test set, with a multi-class AUC of 0.943. Compared to the pure ultrasonic texture recognition model with an accuracy of 53.33%, this approach improves accuracy by 36.27 percentage points while maintaining stable identification accuracy across all four risk levels, without missing high-risk severe cases.
[0058] By enrolling independent prospective patient cohorts after the modeling time window, and processing independent data using standardized parameters saved during the training phase, an independent third party loaded the optimal model for prediction in a blinded manner. This constructed an external validation system with triple isolation: time isolation, parameter locking, and operational blinding. The blinded test results of 247 independent prospective cohorts showed an overall accuracy of 87.2% and a multi-class AUC of 0.937, highly consistent with the internal validation, demonstrating the model's good generalization ability in real-world clinical scenarios. Most similar imaging AI models only use internal segmentation validation on a single dataset. This application added a new patient cohort blind test outside the modeling cycle, fully validating the model's generalization ability in real-world clinical scenarios.
[0059] The above describes the method embodiments of this application. Based on the foregoing embodiments, the system embodiments of this application are described below. The system of this application is used to execute the method of this application. For details not disclosed in the system embodiments, please refer to the method embodiments. This application embodiment further provides a neck, back, and lumbar pain risk prediction system based on muscle ultrasound image analysis. The system includes various modules, which can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.
[0060] Please refer to Figure 4 This illustration shows a schematic diagram of a neck, back, and lumbar pain risk prediction system 400 based on muscle ultrasound image analysis, provided in an embodiment of this application. This neck, back, and lumbar pain risk prediction system 400 can achieve the following: Figure 1 The method for predicting the risk of neck, back, and lower back pain is shown. Figure 4 As shown, the neck, back, and waist pain risk prediction system 400 includes: The data acquisition module 410 is used to collect multi-dimensional data of the subjects, extract morphological parameters and elastic mechanical parameters based on ultrasound images and elastography data, convert digital pain scores into pain risk labels, and associate multi-dimensional data of the same subject with pain risk labels to construct the original dataset. The data preprocessing module 420 is used to perform feature filtering and categorical feature encoding on the original dataset, and to divide the training set and validation set according to the pain risk label; The model training module 430 is used to perform class balancing on the training set only by using a synthetic minority class oversampling algorithm, calculate the feature standardization parameters based on the balanced training set, and use the same standardization parameters to standardize the balanced training set and the validation set respectively; construct a fully connected deep neural network, train the fully connected deep neural network based on the standardized balanced training set and the validation set, and obtain the optimal model; The prediction module 440 is used to enroll independent prospective patient cohorts after the modeling time window, and processes the independent cohort data using the same standardized parameters as in the training phase; a third party independent of the model development process loads the optimal model in a blinded manner to predict the standardized independent cohort data and outputs the prediction results.
[0061] It should be noted that the description of the system embodiments above is similar to the description of the method embodiments above, and has similar beneficial effects. For technical details not disclosed in the system embodiments of this application, please refer to the description of the method embodiments of this application for understanding.
[0062] In one embodiment of this application, a musculoskeletal ultrasound-assisted diagnostic device is provided. The device includes an ultrasound probe, a memory, and a processor. The ultrasound probe is used to acquire musculoskeletal ultrasound imaging data and shear wave elastography data of a subject. The memory stores a computer program that can run on the processor; when the processor executes the computer program, it implements the steps in the aforementioned method. This musculoskeletal ultrasound-assisted diagnostic device can directly output a four-level pain risk grading result for rapid bedside screening.
[0063] In one specific embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the steps of the above-described method. This storage medium can be embedded in a musculoskeletal ultrasound-assisted diagnostic device or exist independently, and is used to store the computer program implementing the method of this application, facilitating program distribution, installation, and updates.
[0064] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the risk of neck, back, and lumbar pain based on muscle ultrasound image analysis, characterized in that, The method includes: Multidimensional data of subjects were collected, and morphological and elastic mechanical parameters were extracted based on ultrasound and elastography data. Digital pain scores were converted into pain risk labels. Multidimensional data of the same subject were associated with pain risk labels to construct the original dataset. The original dataset was subjected to feature selection and categorical feature encoding, and the training set and validation set were divided according to the pain risk label; By using a synthetic minority class oversampling algorithm to perform class balancing only on the training set, calculating feature standardization parameters based on the balanced training set, and using the same standardization parameters to standardize both the balanced training set and the validation set respectively; constructing a fully connected deep neural network, training the fully connected deep neural network based on the standardized balanced training set and validation set, and obtaining the optimal model; After the modeling time window, an independent prospective patient cohort was enrolled, and the independent cohort data were processed using the same standardized parameters as in the training phase. A third party, independent of the model development process, loaded the optimal model in a blinded manner to predict the standardized independent cohort data and output the prediction results.
2. The method according to claim 1, characterized in that, The multidimensional data includes basic human information, pain location, musculoskeletal ultrasound images and shear wave elastography data, clinical palpation information, and digital pain score. The basic human information includes age, sex, height, weight, and body mass index. The pain location includes any one or more areas among the neck and shoulder region, scapular region, chest and back region, upper lumbar region, lateral abdominal muscle group region, lower lumbar region, and limb muscle group region. The clinical palpation information includes the fascial classification of the affected myofascia. The morphological parameters include the long axis, short axis, lesion depth, and target skeletal muscle thickness of the hypoechoic area at the myofascial trigger point. The elastic parameters include the elastic modulus of the lesion area and the elastic modulus of the control area 3 mm from the lesion edge.
3. The method according to claim 1, characterized in that, The digital pain score uses a scale of 0 to 10. When the digital pain score is 0, the pain risk label is no risk; when the digital pain score is 1 to 3, the pain risk label is low risk; when the digital pain score is 4 to 6, the pain risk label is medium risk; and when the digital pain score is 7 to 10, the pain risk label is high risk.
4. The method according to claim 1, characterized in that, The process of feature selection and categorical feature encoding of the original dataset, and dividing the training and validation sets according to pain risk labels, includes: Perform outlier removal and missing value imputation on continuous features of the original dataset, and perform missing value imputation on categorical features; The inflated dataset is subjected to feature filtering to retain the core input features, and one-hot encoding is performed on the categorical features in the core input features. Based on pain risk labels, the dataset was randomly divided into training and validation sets in a 7:3 ratio, ensuring that the proportion of samples of each category in the training and validation sets was consistent with the original category distribution.
5. The method according to claim 4, characterized in that, The process involves outlier removal and missing value imputation for continuous features in the original dataset, and missing value imputation for categorical features. The imputed dataset is then subjected to feature filtering to retain core input features, including: Anomaly detection for continuous features is performed by combining box plot method with Z-Score thresholding method, and samples in which any continuous feature is judged to be abnormal are removed as a whole. The median was used to fill missing values for continuous features, and the mode was used to fill missing values for categorical features. The correlation coefficients between each candidate feature and the pain risk label in the filled dataset are calculated and a correlation heatmap is plotted. The variance inflation factor between each candidate feature is calculated, and redundant features with a variance inflation factor greater than 10 are gradually removed to obtain the core input features. The core input features include age, gender, height, weight, BMI, MTRPs hypoechoic area long axis, short axis, lesion depth, skeletal muscle thickness, elastic modulus of the control area, elastic modulus of the lesion area, and fascial classification.
6. The method according to claim 1, characterized in that, The synthetic minority class oversampling algorithm is the SMOTE algorithm, which synthesizes virtual samples only for high-risk samples in the training set, so that the number of risk samples at each level in the training set is balanced.
7. The method according to claim 1, characterized in that, The fully connected deep neural network includes an input layer, hidden layers, and an output layer. The input layer has 12 neurons, the hidden layers include 3 to 5 fully connected layers, the number of neurons in each hidden layer decreases layer by layer after the input layer, all hidden layers use the ReLU activation function, and the output layer has 4 neurons and uses the Softmax activation function.
8. A neck, back, and lumbar pain risk prediction system based on muscle ultrasound image analysis, characterized in that, The system includes: The data acquisition module is used to collect multi-dimensional data of subjects, extract morphological and elastic mechanical parameters based on ultrasound and elastography data, convert digital pain scores into pain risk labels, and associate multi-dimensional data of the same subject with pain risk labels to construct the original dataset. The data preprocessing module is used to perform feature filtering and categorical feature encoding on the original dataset, and to divide the training set and validation set according to the pain risk label; The model training module is used to perform class balancing on the training set only by using a synthetic minority class oversampling algorithm, calculate the feature standardization parameters based on the balanced training set, and use the same standardization parameters to standardize both the balanced training set and the validation set; construct a fully connected deep neural network, train the fully connected deep neural network based on the standardized balanced training set and the validation set, and obtain the optimal model; The prediction module is used to enroll independent prospective patient cohorts after the modeling time window. It processes the independent cohort data using the same standardized parameters as the training phase. A third party, independent of the model development process, loads the optimal model in a blinded manner to predict the standardized independent cohort data and outputs the prediction results.
9. A musculoskeletal ultrasound-assisted diagnostic device, comprising an ultrasound probe, a memory, and a processor, wherein the ultrasound probe is used to acquire musculoskeletal ultrasound image data and shear wave elastography data of a subject, and the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.
