A method for predicting spinal muscle and ligament injuries, an electronic device, and a storage medium.
By combining medical imaging, surface electromyography signals, and symptom data through a multi-source data processing model, the subjectivity and efficiency issues of traditional spinal muscle and ligament injury diagnosis have been resolved, achieving efficient and accurate injury prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AIR FORCE MEDICAL CENT PLA
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional diagnosis of spinal muscle and ligament injuries suffers from high subjectivity, low efficiency, and insufficient data mining, making it difficult to meet peak demand and neglecting the relationship between symptoms and structure or function.
A multi-source data processing model is adopted, which combines medical images, surface electromyography signals and symptom data. Damage prediction is performed through multi-branch feature extraction and feature alignment network, including image feature extraction, sEMG feature extraction and symptom feature extraction. Multi-dimensional fusion prediction is performed using feature fusion and prediction network.
It achieves multi-dimensional data coverage of structure, function, and subjective symptoms, improving the accuracy and efficiency of diagnosis, eliminating the heterogeneity of feature distribution among different types of data, and improving the model's prediction accuracy and generalization ability.
Smart Images
Figure CN122135943A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical and artificial intelligence technologies, and in particular to a method for predicting spinal muscle and ligament injuries, an electronic device, and a storage medium. Background Technology
[0002] Magnetic resonance imaging (MRI) and computed tomography (CT) are both common medical procedures. Spinal muscle and ligament injuries are prevalent in orthopedics and rehabilitation medicine, encompassing injuries to the cervical, thoracic, and lumbar spine, as well as surrounding muscle and ligament groups. Spinal ligament sprains are classified as supraspinous ligament sprains, interspinous ligament sprains, and muscle sprains. For muscle sprains, the nature of the ligament, the location of the sprain, and its characteristics are crucial to determine whether it is a multifidus muscle sprain, rotatores muscle sprain, or sacrospinalis muscle sprain. Traditional medical diagnosis typically relies on the physician's experience and a step-by-step examination process.
[0003] Traditional medical diagnosis is greatly affected by subjective bias, resulting in a high rate of misdiagnosis and the potential for missing the optimal treatment period. Relying on manual image interpretation, functional analysis, and symptom analysis is time-consuming and cannot meet the needs of peak outpatient periods. With the development of artificial intelligence, it has been gradually applied to the medical field for assisted diagnosis. However, existing diagnostic methods rely on MRI or CT images and ignore the relationship between symptoms and structure or function. For example, lumbar disc herniation is usually associated with low back pain accompanied by radiating pain in the lower extremities. Summary of the Invention
[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method, electronic device and storage medium for predicting spinal muscle and ligament injuries, and to solve the problems of strong subjectivity, low efficiency and insufficient data mining in traditional spinal muscle and ligament injury diagnosis.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for predicting spinal muscle and ligament injuries includes: Medical images and surface electromyography signals of the spinal region of the target patient were collected, and symptom data of the target patient were collected using a pre-defined symptom statistics table. The medical images and the surface electromyography signals are preprocessed. The symptom data, preprocessed medical images, and surface electromyography signals are input into a pre-trained muscle and ligament injury prediction model for injury identification to obtain muscle and ligament injury prediction results; the construction process of the muscle and ligament injury prediction model includes: Collect and study the patient's symptom data, as well as the preprocessed medical images and surface electromyography signals; A multi-source data processing model is constructed; the multi-source data processing model includes: a multi-source feature extraction network, a feature space alignment network, and a feature fusion and prediction network connected in sequence; the multi-source feature extraction network includes: a parallel image feature extraction branch, an sEMG feature extraction branch, and a symptom feature extraction branch; The image feature extraction branch is used to extract features from the preprocessed medical image to obtain an image feature vector. The sEMG feature extraction branch is used to extract features from the preprocessed surface electromyography signal to obtain the sEMG feature vector. Based on the pre-constructed symptom atlas, entity matching is performed on the symptom data using the symptom feature extraction branch to obtain symptom feature vectors; The feature space alignment network is used to perform feature alignment processing on the image feature vector, the sEMG feature vector, and the symptom feature vector. The feature fusion and prediction network is used to fuse the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features. The multi-output DNN network is then used to predict and calculate the multi-dimensional fused features to obtain the network output. The multi-source data processing model is iteratively trained using a pre-constructed multidimensional loss function based on the network output to obtain the muscle and ligament injury prediction model.
[0006] Preferably, medical images and surface electromyography signals of the spinal region of the target patient are acquired, and symptom data of the target patient are collected using a pre-defined symptom statistics table, including: The target patient's spinal region was scanned using an MRI scanner to obtain the medical images; Several auxiliary movements are set up, and surface electrodes are attached to the muscle areas of the spinal region of the target patient. The target patient is guided to perform each of the auxiliary movements, and electrical signals are collected using the surface electrodes during the movement to obtain the surface electromyography signal. A symptom statistics table is set up, and information is collected from the target patient based on the symptom statistics table to obtain the symptom data; the symptom statistics table includes: pain index, functional impairment index, and accompanying symptoms; the pain index includes: location and intensity score; the functional impairment index includes: activity limitation level; the accompanying symptoms include: numbness, weakness, and spasm.
[0007] Preferably, the medical images and the surface electromyography signals are preprocessed, including: A grayscale standard deviation threshold is set, and images in the medical images whose grayscale standard deviation is less than the grayscale standard deviation threshold are identified as blurred images. The blurred images are removed, and the blurred images are filled by mean interpolation using two sets of images adjacent to the blurred images. The interpolated medical images are then subjected to Gaussian filtering. The filtered medical images were processed to remove artifacts using a frequency-selective saturation method. Data corresponding to the transition segment of the action switching in the surface electromyography signal is removed.
[0008] Preferably, the image feature extraction branch is used to extract features from the preprocessed medical image to obtain an image feature vector, including: The medical image is subjected to size mapping, grayscale normalization, and spinal region segmentation to obtain a standard input image. The standard input image is progressively feature-extracted using a pre-trained ResNet50 network to obtain the original extracted features. These original extracted features include: edge texture features, local structural features, and global semantic features. The edge texture features correspond to the outputs of layers 1 to 7 of the ResNet50 network; the local structural features correspond to the outputs of layers 8 to 35 of the ResNet50 network; and the global semantic features correspond to the outputs of layers 36 to 49 of the ResNet50 network. The original extracted features are processed sequentially using two convolutional layers and one global average pooling layer to obtain comprehensive image features; The integrated image features are subjected to L1 regularization to obtain the image feature vector.
[0009] Preferably, the sEMG feature extraction branch is used to extract features from the preprocessed surface electromyography signal to obtain an sEMG feature vector, including: Each segment of the surface electromyography signal is divided using a fixed time window to obtain several time segments; Extract the basic features of each time segment; the basic features include: root mean square, mean absolute value, peak value, peak power spectrum, spectral centroid, spectral bandwidth, wavelet entropy, and sample entropy; The extracted basic features are input into a bidirectional LSTM network in temporal order for forward and backward computation to obtain a temporal feature matrix; The time-series feature matrix is subjected to global average pooling and BatchNorm2d normalization in sequence to obtain the sEMG feature vector.
[0010] Preferably, entity matching is performed on the symptom data using the symptom feature extraction branch based on the pre-constructed symptom atlas to obtain a symptom feature vector, including: Define the symptom atlas; the symptom atlas includes: core entities, entity attributes, entity relationships, and an entity vector library; the core entities include: symptom entities, attribute entities, and relationship entities; the entity vector library includes: a vector generated for each core entity using the Node2Vec algorithm; The symptom data is converted into structured key-value pairs; The structured key-value pairs are matched with entities in the symptom map to obtain the symptom feature vector.
[0011] Preferably, the pre-training process of the feature space alignment network includes: The pre-collected image feature vectors, sEMG feature vectors, and symptom feature vectors are projected onto the same dimension; A bidirectional cross-modal attention mechanism is used to perform cross-modal interactive processing on the projected image feature vector, the sEMG feature vector, and the symptom feature vector; The image feature vector, sEMG feature vector, and symptom feature vector after each patient interaction are set as the basic triplet; Set any one of the data in the basic triplet as a positive sample pair, and set the data in the basic triplet excluding the positive sample pair as a negative sample pair; Construct supervised contrastive loss and center loss; the expression for the supervised contrastive loss is: The expression for the center loss is: ;in, , These are the supervised comparison loss and the center loss, respectively. , These are the first vector and the second vector, respectively. For scaling parameters; This represents the amount of negative sample data. The feature of the i-th negative sample; It is the set of pairs of the image feature vector, the sEMG feature vector, and the symptom feature vector; express Any combination of the following; This represents the total number of samples. , , These represent the imaging modality, sEMG modality, and symptom modality, respectively. It can be any one of the image modality, the sEMG modality, and the symptom modality; The output of the j-th sample and its t-th mode after cross-modal interaction processing; Let be the class center of the t-th modality; The supervised contrast loss and the center loss are calculated based on the positive sample pairs and the negative sample pairs, and the calculated supervised contrast loss and center loss are weighted and fused to obtain the joint total loss; The network layers of the bidirectional cross-modal attention mechanism are iteratively optimized using the joint total loss to obtain the feature space alignment network after iteration.
[0012] Preferably, the feature fusion and prediction network is used to fuse the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features. A multi-output DNN network is then used to predict and calculate the multi-dimensional fused features to obtain the network output, including: The multi-output DNN network is constructed as follows: the multi-output DNN network includes: an input layer, a first hidden layer, a second hidden layer, and a two-branch output layer connected in sequence; the two-branch output layer includes: a parallel damage type output layer and a damage location output layer. The input layer is used to sequentially perform outlier filtering and feature standardization on the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain the input matrix; The input matrix is subjected to feature abstraction processing using the first hidden layer to obtain initial hidden features; the expression of the first hidden layer is: ;in, The initial hidden feature; This is the first weight matrix; The input matrix; This is the first bias; The initial hidden features are processed by feature association using the second hidden layer to obtain optimized hidden features; the expression for the second hidden layer is: ;in, For the optimized hidden features; This is the second weight matrix; This is the second bias; The regularization coefficient; The optimized hidden features are synchronously input into the damage type output layer and the damage location output layer for probability mapping calculation, so as to obtain the damage type prediction output and the damage location prediction output.
[0013] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned method for predicting spinal muscle and ligament injuries.
[0014] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned method for predicting spinal muscle and ligament injuries.
[0015] The present invention discloses the following technical effects: This invention provides a method, electronic device, and storage medium for predicting spinal muscle and ligament injuries. By employing medical images, surface electromyography (EMG) signals, and symptom data, it addresses the problems of traditional diagnostic methods having a single data source and not considering EMG signals, achieving multi-dimensional data coverage of structure, function, and subjective symptoms. By using three-branch channel feature extraction and feature alignment processing, it solves the problems of insufficient data mining and the reduction in model prediction accuracy caused by the distribution of different types of data features, thus ensuring the effectiveness of feature extraction while eliminating the heterogeneity of different types of output feature distributions. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the spinal muscle and ligament injury prediction process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-source data acquisition process provided in an embodiment of the present invention; Figure 3 A schematic diagram of the data preprocessing process is provided for embodiments of the present invention; Figure 4 This is a schematic diagram of the image feature vector extraction process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the sEMG feature vector extraction process provided in an embodiment of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] The purpose of this invention is to provide a method, electronic device, and storage medium for predicting spinal muscle and ligament injuries, thereby solving the problems of high subjectivity, low efficiency, and insufficient data mining in traditional spinal muscle and ligament injury diagnosis.
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0021] Figure 1 This is a schematic diagram of the spinal muscle and ligament injury prediction process provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the present invention provides a method for predicting spinal muscle and ligament injuries, comprising: Step 100: Acquire medical images and surface electromyography signals of the spinal region of the target patient, and collect symptom data of the target patient using a pre-defined symptom statistics table; Step 200: Preprocess the medical images and the surface electromyography signals; Step 300: Input the symptom data, the preprocessed medical images, and the surface electromyography signals into a pre-trained muscle and ligament injury prediction model for injury identification, and obtain muscle and ligament injury prediction results; the construction process of the muscle and ligament injury prediction model includes: Step 301: Collect the symptom data of the study patients, as well as the preprocessed medical images and surface electromyography signals; Step 302: Construct a multi-source data processing model; the multi-source data processing model includes: a multi-source feature extraction network, a feature space alignment network, and a feature fusion and prediction network connected in sequence; the multi-source feature extraction network includes: a parallel image feature extraction branch, an sEMG feature extraction branch, and a symptom feature extraction branch; Step 303: Use the image feature extraction branch to extract features from the preprocessed medical image to obtain an image feature vector; Step 304: Use the sEMG feature extraction branch to extract features from the preprocessed surface electromyography signal to obtain the sEMG feature vector; Step 305: Based on the pre-constructed symptom atlas, perform entity matching on the symptom data using the symptom feature extraction branch to obtain the symptom feature vector; Step 306: Use the feature space alignment network to perform feature alignment processing on the image feature vector, the sEMG feature vector, and the symptom feature vector; Step 307: Use the feature fusion and prediction network to perform feature fusion on the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features, and use a multi-output DNN network to predict and calculate the multi-dimensional fused features to obtain the network output; Step 308: Based on the network output, the multi-source data processing model is iteratively trained using a pre-constructed multidimensional loss function to obtain the muscle and ligament injury prediction model.
[0022] refer to Figure 2 Medical images and surface electromyography signals of the spinal region of the target patient were acquired, and symptom data of the target patient were collected using a pre-defined symptom statistics table, including: Step 101: Use an MRI scanner to scan the spinal region of the target patient to obtain the medical images; Step 102: Set several auxiliary movements, attach the surface electrodes to the muscle areas of the spinal region of the target patient, guide the target patient to perform each auxiliary movement, and collect electrical signals using the surface electrodes during the movement to obtain the surface electromyography signal; Step 103: Set up a symptom statistics table and collect information from the target patient according to the symptom statistics table to obtain the symptom data; the symptom statistics table includes: pain index, functional impairment index, and accompanying symptoms; the pain index includes: location and intensity score; the functional impairment index includes: activity limitation level; the accompanying symptoms include: numbness, weakness, and spasm.
[0023] refer to Figure 3 The medical images and surface electromyography signals are preprocessed, including: Step 201: Set a grayscale standard deviation threshold, determine images in the medical images whose grayscale standard deviation is less than the grayscale standard deviation threshold as blurred images, remove the blurred images, and use two sets of images adjacent to the blurred image to perform mean interpolation to complete the blurred image. Step 202: Perform Gaussian filtering on the interpolated medical image; Step 203: Use the frequency-selective saturation method to remove artifacts from the filtered medical image; Step 204: Remove the data corresponding to the action switching transition segment in the surface electromyography signal.
[0024] refer to Figure 4The image feature extraction branch is used to extract features from the preprocessed medical image to obtain an image feature vector, including: Step 30301: Perform size mapping, grayscale normalization, and spinal region segmentation on the medical image to obtain a standard input image; Step 30302: Use a pre-trained ResNet50 network to perform stepwise feature extraction on the standard input image to obtain the original extracted features; the original extracted features include: edge texture features, local structural features, and global semantic features; the edge texture features correspond to the outputs of the first to seventh layers of the ResNet50 network; the local structural features correspond to the outputs of the eighth to thirty-fifth layers of the ResNet50 network; the global semantic features correspond to the outputs of the thirty-sixth to forty-ninth layers of the ResNet50 network; Step 30303: The original extracted features are processed sequentially using two convolutional layers and one global average pooling layer to obtain comprehensive image features; Step 30304: Perform L1 regularization on the integrated image features to obtain the image feature vector.
[0025] refer to Figure 5 The sEMG feature extraction branch is used to extract features from the preprocessed surface electromyography signal to obtain the sEMG feature vector, including: Step 30401: Segment each segment of the surface electromyography signal using a fixed time window to obtain several time segments; Step 30402: Extract the basic features of each time segment; the basic features include: root mean square, mean absolute value, peak value, peak power spectrum, spectral centroid, spectral bandwidth, wavelet entropy, and sample entropy; Step 30403: Input the extracted basic features into the bidirectional LSTM network in temporal order for forward and backward computation to obtain the temporal feature matrix; Step 30404: Perform global average pooling and BatchNorm2d normalization on the time-series feature matrix in sequence to obtain the sEMG feature vector.
[0026] Furthermore, based on the pre-constructed symptom atlas, entity matching is performed on the symptom data using the symptom feature extraction branch to obtain a symptom feature vector, including: Define the symptom atlas; the symptom atlas includes: core entities, entity attributes, entity relationships, and an entity vector library; the core entities include: symptom entities, attribute entities, and relationship entities; the entity vector library includes: a vector generated for each core entity using the Node2Vec algorithm; The symptom data is converted into structured key-value pairs; The structured key-value pairs are matched with entities in the symptom map to obtain the symptom feature vector.
[0027] Specifically, the pre-training process of the feature space alignment network includes: The pre-collected image feature vectors, sEMG feature vectors, and symptom feature vectors are projected onto the same dimension; A bidirectional cross-modal attention mechanism is used to perform cross-modal interactive processing on the projected image feature vector, the sEMG feature vector, and the symptom feature vector; The image feature vector, sEMG feature vector, and symptom feature vector after each patient interaction are set as the basic triplet; Set any one of the data in the basic triplet as a positive sample pair, and set the data in the basic triplet excluding the positive sample pair as a negative sample pair; Construct supervised contrastive loss and center loss; the expression for the supervised contrastive loss is: The expression for the center loss is: ;in, , These are the supervised comparison loss and the center loss, respectively. , These are the first vector and the second vector, respectively. For scaling parameters; This represents the amount of negative sample data. The feature of the i-th negative sample; It is the set of pairs of the image feature vector, the sEMG feature vector, and the symptom feature vector; express Any combination of the following; This represents the total number of samples. , , These represent the imaging modality, sEMG modality, and symptom modality, respectively. It can be any one of the image modality, the sEMG modality, and the symptom modality; The output of the j-th sample and its t-th mode after cross-modal interaction processing; Let be the class center of the t-th modality; The supervised contrast loss and the center loss are calculated based on the positive sample pairs and the negative sample pairs, and the calculated supervised contrast loss and center loss are weighted and fused to obtain the joint total loss; The network layers of the bidirectional cross-modal attention mechanism are iteratively optimized using the joint total loss to obtain the feature space alignment network after iteration.
[0028] Furthermore, the feature fusion and prediction network is used to fuse the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features. A multi-output DNN network is then used to predict and calculate the multi-dimensional fused features to obtain the network output, including: The multi-output DNN network is constructed as follows: the multi-output DNN network includes: an input layer, a first hidden layer, a second hidden layer, and a two-branch output layer connected in sequence; the two-branch output layer includes: a parallel damage type output layer and a damage location output layer. The input layer is used to sequentially perform outlier filtering and feature standardization on the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain the input matrix; The input matrix is subjected to feature abstraction processing using the first hidden layer to obtain initial hidden features; the expression of the first hidden layer is: ;in, The initial hidden feature; This is the first weight matrix; The input matrix; This is the first bias; The initial hidden features are processed by feature association using the second hidden layer to obtain optimized hidden features; the expression for the second hidden layer is: ;in, For the optimized hidden features; This is the second weight matrix; This is the second bias; The regularization coefficient; The optimized hidden features are synchronously input into the damage type output layer and the damage location output layer for probability mapping calculation, so as to obtain the damage type prediction output and the damage location prediction output.
[0029] Specifically, a standardized scan of the target patient's spinal region was performed using a 3T high-field MRI scanner. The procedure was as follows: The patient's position was adjusted according to the target spinal location, with the patient lying supine with knees flexed and a soft pillow underneath to relax the back muscles and reduce interference from the physiological curvature of the vertebrae on the images. For the diagnosis of muscle and ligament injuries, a multi-sequence combination scan was used. T1-weighted images were used to clearly display the anatomical boundaries of the vertebral bodies, intervertebral discs, and ligaments; T2-weighted images were used to identify lesions such as muscle and ligament edema and tears; and fat suppression sequences were used to suppress adipose tissue signals and eliminate fat interference to accurately locate the area of muscle and ligament injury.
[0030] Optionally, electrical signals from the target patient's spinal muscle groups are acquired using a multi-channel surface electromyography (EMG) acquisition system. The skin surface of the target muscle groups is wiped with 75% medical alcohol to remove oil and keratin. After the alcohol evaporates, a small amount of conductive gel is applied to the electrode attachment sites to reduce skin impedance. Silver-silver chloride surface electrodes are used, with the attachment muscles determined according to the spinal location: for the cervical spine, the sternocleidomastoid, upper trapezius, and splenius capitis muscles; for the lumbar spine, the erector spinae, multifidus, and psoas major muscles. The long axis of the electrode aligns with the muscle fiber direction, and reference electrodes are attached to adjacent non-exerting muscle groups. Six standardized auxiliary movements are designed for different spinal locations, each designed to activate a specific target muscle group, as detailed in Table 1.
[0031] Table 1
[0032] The acquisition system's sampling frequency is set to 1000Hz, with a bandwidth of 200Hz to 500Hz. It filters out 50Hz power frequency interference and high-frequency noise. During acquisition, the signal waveform is monitored in real time. If abnormal spikes occur, acquisition must be paused, and the electrode positions readjusted before resuming. After acquisition, data from the transitional phase of each action signal is discarded, retaining only the stable sEMG signal during the action hold phase for subsequent processing. The amplitude change rate of the sEMG signal is calculated. If the amplitude change rate of a data segment exceeds the normal range, it is determined to be a transitional segment and needs further removal to ensure the remaining signal is completely stable.
[0033] Preferably, a structured symptom statistics table is designed for clinicians to collect symptom data. The table includes three main categories of indicators and detailed dimensions: pain indicators, functional impairment indicators, and accompanying symptoms. Pain indicators include the location of pain, pain intensity score, pain nature, and triggering factors. Functional impairment indicators use a 5-point grading system to measure the degree of activity limitation. Accompanying symptoms are used to record related discomfort symptoms, such as the presence of numbness, weakness, or muscle spasms; if present, the time and frequency of these episodes are recorded.
[0034] Furthermore, a grayscale standard deviation threshold is set, and the grayscale standard deviation of each of the acquired MRI images is calculated. Images with a standard deviation less than the grayscale standard deviation threshold are marked as blurred images. For blurred image completion, one set of consecutive clear images before and after the blurred image is selected as the interpolation data source. For each pixel of the blurred image, the grayscale mean of the corresponding pixels in the preceding and following clear images is calculated, and this mean value is used to replace the pixel value of the blurred image, thus achieving completion. A 3×3 pixel Gaussian kernel is used, with a Gaussian standard deviation of 1. The Gaussian kernel is traversed pixel by pixel through the interpolated MRI image. For each pixel, the sum of the products of its neighboring 3×3 pixels and the weights of the Gaussian kernel is calculated as the new grayscale value for that pixel. Motion artifacts and chemical shift artifacts are prone to appear in MRI images. A two-dimensional fast Fourier transform is performed on the Gaussian filtered image to convert the image from the spatial domain to the frequency domain, obtaining the frequency spectrum. The frequency ranges of motion artifacts and chemical shift artifacts are located. For the identified artifact frequency ranges, the signal amplitude is multiplied by an attenuation factor of 0.1. A two-dimensional inverse fast Fourier transform is then performed on the processed frequency spectrum to convert the image back to the spatial domain, resulting in a clear image without artifacts.
[0035] Specifically, bilinear interpolation is used to uniformly map the preprocessed MRI images to the standard input size of the ResNet50 network; the Min-Max normalization algorithm is used to map the image grayscale values from the original range to a unit interval to eliminate grayscale deviations from different scanning devices; the U-Net++ semantic segmentation model is used to automatically segment the vertebral bodies, intervertebral discs, and surrounding muscle and ligament regions, removing the background; after segmentation, the minimum bounding rectangle of the target region is retained to ensure that subsequent feature extraction focuses on injury-related structures. The ResNet50 network structure includes one convolutional stem layer, four residual block groups, one global average pooling layer, and a fully connected layer; layers one through seven output edge texture features, layers eight through thirty-five output local structural features, and layers thirty-six through forty-nine output global semantic features; the first convolutional layer is used to fuse the multi-level features output by ResNet50, and the second convolutional layer is used to compress the feature dimension and reduce the computational cost; L1 regularization is applied to the comprehensive image features to sparsify the feature vector, remove redundant non-injury-related features, and improve the model's generalization ability.
[0036] Furthermore, a sliding time window is set, and baseline drift removal is performed on the sEMG signal before segmentation to avoid interference from resting potential fluctuations on feature extraction. The sliding time window is used to segment the surface electromyography signal. Eight basic features are extracted from the amplitude, frequency, and complexity of the sEMG signal, namely root mean square, mean absolute value, peak value, peak power spectrum, spectral centroid, spectral bandwidth, wavelet entropy, and sample entropy. The feature sequence of the eight basic features is concatenated according to the time window and input into a bidirectional LSTM network. The forward LSTM captures historical time-series information, and the backward LSTM captures future time-series information, outputting a time-series feature matrix that reflects the dynamic changes of muscle contraction. The time-series feature matrix is standardized in the channel dimension, and the mean and variance of each channel are calculated to eliminate the influence of differences in electromyography signal intensity among different patients and improve the feature universality.
[0037] Preferably, symptom entities represent patients' subjective feelings or objective signs, attribute entities represent quantitative or categorical attributes describing symptom entities, and relation entities represent the logical associations connecting symptom entities with each other and with attribute entities. A three-level relational structure is used to ensure the logical consistency of the graph: symptom-attribute association, symptom-symptom accompaniment, and symptom-injury association. The Node2Vec algorithm is used to generate entity vectors with the following parameters: 100 node sequences are generated per round of walk, the window size is 10, and each node is walked 20 times. Based on pre-collected clinical records of spinal injury patients, entities and relationships are manually labeled to construct a graph training set. The final generated entity vector library can be used to calculate the entity association degree using cosine similarity. The information collected from the symptom statistics table is converted into structured key-value pairs. Entities to be matched are extracted from these structured key-value pairs. For each entity to be matched, the cosine similarity is calculated with all entities in the entity vector library of the symptom atlas. A similarity threshold of 0.8 is set. If the similarity between the entity to be matched and an entity in the atlas is ≥0.8, it is considered a successful match. If the similarity is <0.8, it is considered a non-match, and the entity is marked as a new entity for subsequent updates to the atlas. All successfully matched entity vectors are concatenated in the order of pain symptoms, accompanying symptoms, and functional impairment.
[0038] Specifically, the original feature dimensions of different modalities differ significantly, and direct interaction can lead to an imbalance in feature weights. Therefore, it is necessary to first project all three into the same dimensional space and perform cross-modal interaction. A separate fully connected network (FC layer) is constructed for each modality as a projector, with the original feature vectors of each modality as input. Traditional unidirectional attention can only achieve unidirectional information transmission and cannot capture bidirectional correlations between modalities. Therefore, this embodiment adopts a bidirectional cross-modal attention mechanism. Attention weights are calculated using ScaledDot-ProductAttention. In the first stage, the attention weights of image features are used to weight the sEMG features to obtain sEMG-image related features. In the second stage, the attention weights of sEMG features are used to weight the image features to obtain image-sEMG related features. The above process is repeated to complete the bidirectional interaction between image-symptom and sEMG-symptom in sequence, finally obtaining three sets of cross-modal enhanced features, denoted as set S.
[0039] Furthermore, the image enhancement features, sEMG enhancement features, and symptom enhancement features of the same patient essentially correspond to the same injury state, and the three have a natural homology relationship. Therefore, they are combined into a basic triplet. Any two sets of features within the same basic triplet are positive sample pairs. From the basic triplets of other patients, features that are of the same modality or different modality as the positive sample pairs are randomly selected as negative samples. This embodiment designs a supervised contrastive loss, the expression of which is as follows: The expression for center loss is: In the formula, , These are supervised comparison loss and center loss, respectively. , These are the first vector and the second vector, respectively. For scaling parameters; This represents the amount of negative sample data. The feature of the i-th negative sample; express Any combination of the following; This represents the total number of samples. , , These represent the imaging modality, sEMG modality, and symptom modality, respectively. It can be any one of the following: image modality, sEMG modality, or symptom modality; The output of the j-th sample and its t-th mode after cross-modal interaction processing; Let be the class center of the t-th modality.
[0040] Using the Adam optimizer, the model parameters for each round are saved during training, and the final model is selected as the feature space alignment network.
[0041] A pre-trained feature space alignment network can be used to convert input images, sEMGs, and symptom features into aligned features.
[0042] Furthermore, this embodiment employs a multi-output DNN network, which consists of an input layer, a first hidden layer, a second hidden layer, and a two-branch output layer. The input layer is used to eliminate feature heterogeneity and extreme value interference, specifically by performing outlier filtering and feature standardization. The standardized image feature vector, sEMG feature vector, and symptom feature vector are concatenated in the order of image, sEMG, and symptom. The first hidden layer is used to perform preliminary nonlinear abstraction on the fused multidimensional features, extracting basic correlation features related to the injury, as shown in the following expression: In the formula These are the initial hidden features; This is the first weight matrix; The input matrix; This is the first bias.
[0043] The second hidden layer enhances cross-modal feature interactions and suppresses overfitting through L2 regularization, specifically: In the formula To optimize hidden features; This is the second weight matrix; This is the second bias; is the regularization coefficient.
[0044] The two-branch output layer is used for dual-task probability prediction, employing activation functions to map the probability of injury type and location. The cross-loss function is used as a multi-dimensional loss function to calculate the network output, yielding the total network loss. With the goal of minimizing the total loss, backpropagation is performed to update the model parameters until iteration is complete, outputting the muscle and ligament injury prediction model. In subsequent work, newly acquired multi-source data from target patients can be directly input into this muscle and ligament injury prediction model to obtain the predicted results.
[0045] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned method for predicting spinal muscle and ligament injuries.
[0046] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned method for predicting spinal muscle and ligament injuries.
[0047] The beneficial effects of this invention are as follows: This invention achieves multi-dimensional data coverage of structure, function, and subjective symptoms by using medical imaging, surface electromyography signals, and symptom data, thus compensating for the information deficiencies of single data. By employing three-branch channel feature extraction and feature alignment processing, it ensures the effectiveness of feature extraction while eliminating the heterogeneity of different types of output feature distributions, thereby improving the model's prediction accuracy and generalization ability.
[0048] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0049] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for predicting spinal muscle and ligament injuries, characterized in that, include: Medical images and surface electromyography signals of the spinal region of the target patient were collected, and symptom data of the target patient were collected using a pre-defined symptom statistics table. The medical images and the surface electromyography signals are preprocessed. The symptom data, the preprocessed medical images, and the surface electromyography signals are input into a pre-trained muscle and ligament injury prediction model for injury identification, and muscle and ligament injury prediction results are obtained. The process of constructing the muscle and ligament injury prediction model includes: Collect and study the patient's symptom data, as well as the preprocessed medical images and surface electromyography signals; A multi-source data processing model is constructed; the multi-source data processing model includes: a multi-source feature extraction network, a feature space alignment network, and a feature fusion and prediction network connected in sequence; the multi-source feature extraction network includes: a parallel image feature extraction branch, an sEMG feature extraction branch, and a symptom feature extraction branch; The image feature extraction branch is used to extract features from the preprocessed medical image to obtain an image feature vector. The sEMG feature extraction branch is used to extract features from the preprocessed surface electromyography signal to obtain the sEMG feature vector. Based on the pre-constructed symptom atlas, entity matching is performed on the symptom data using the symptom feature extraction branch to obtain symptom feature vectors; The feature space alignment network is used to perform feature alignment processing on the image feature vector, the sEMG feature vector, and the symptom feature vector. The feature fusion and prediction network is used to fuse the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features. The multi-output DNN network is then used to predict and calculate the multi-dimensional fused features to obtain the network output. The multi-source data processing model is iteratively trained using a pre-constructed multidimensional loss function based on the network output to obtain the muscle and ligament injury prediction model.
2. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, Medical images and surface electromyography signals of the spinal region of the target patient were acquired. Symptom data of the target patient were collected using a pre-defined symptom statistics table, including: The target patient's spinal region was scanned using an MRI scanner to obtain the medical images; Several auxiliary movements are set up, and surface electrodes are attached to the muscle areas of the spinal region of the target patient. The target patient is guided to perform each of the auxiliary movements, and electrical signals are collected using the surface electrodes during the movement to obtain the surface electromyography signal. A symptom statistics table is set up, and information is collected from the target patient based on the symptom statistics table to obtain the symptom data; the symptom statistics table includes: pain index, functional impairment index, and accompanying symptoms; the pain index includes: location and intensity score; the functional impairment index includes: activity limitation level; the accompanying symptoms include: numbness, weakness, and spasm.
3. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, Preprocessing of the medical images and the surface electromyography signals includes: A grayscale standard deviation threshold is set, and images in the medical images whose grayscale standard deviation is less than the grayscale standard deviation threshold are identified as blurred images. The blurred images are removed, and the blurred images are filled by mean interpolation using two sets of images adjacent to the blurred images. The interpolated medical images are then subjected to Gaussian filtering. The filtered medical images were processed to remove artifacts using a frequency-selective saturation method. Data corresponding to the transition segment of the action switching in the surface electromyography signal is removed.
4. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, The image feature extraction branch is used to extract features from the preprocessed medical image to obtain an image feature vector, including: The medical image is subjected to size mapping, grayscale normalization, and spinal region segmentation to obtain a standard input image. The standard input image is progressively feature-extracted using a pre-trained ResNet50 network to obtain the original extracted features. These original extracted features include: edge texture features, local structural features, and global semantic features. The edge texture features correspond to the outputs of layers 1 to 7 of the ResNet50 network; the local structural features correspond to the outputs of layers 8 to 35 of the ResNet50 network; and the global semantic features correspond to the outputs of layers 36 to 49 of the ResNet50 network. The original extracted features are processed sequentially using two convolutional layers and one global average pooling layer to obtain comprehensive image features; The integrated image features are subjected to L1 regularization to obtain the image feature vector.
5. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, The sEMG feature extraction branch is used to extract features from the preprocessed surface electromyography signal to obtain an sEMG feature vector, including: Each segment of the surface electromyography signal is divided using a fixed time window to obtain several time segments; Extract the basic features of each time segment; the basic features include: root mean square, mean absolute value, peak value, peak power spectrum, spectral centroid, spectral bandwidth, wavelet entropy, and sample entropy; The extracted basic features are input into a bidirectional LSTM network in temporal order for forward and backward computation to obtain a temporal feature matrix; The time-series feature matrix is subjected to global average pooling and BatchNorm2d normalization in sequence to obtain the sEMG feature vector.
6. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, Based on the pre-constructed symptom atlas, entity matching is performed on the symptom data using the symptom feature extraction branch to obtain a symptom feature vector, including: Define the symptom atlas; the symptom atlas includes: core entities, entity attributes, entity relationships, and an entity vector library; the core entities include: symptom entities, attribute entities, and relationship entities; the entity vector library includes: a vector generated for each core entity using the Node2Vec algorithm; The symptom data is converted into structured key-value pairs; The structured key-value pairs are matched with entities in the symptom map to obtain the symptom feature vector.
7. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, The pre-training process of the feature space alignment network includes: The pre-collected image feature vectors, sEMG feature vectors, and symptom feature vectors are projected onto the same dimension; A bidirectional cross-modal attention mechanism is used to perform cross-modal interactive processing on the projected image feature vector, the sEMG feature vector, and the symptom feature vector; The image feature vector, sEMG feature vector, and symptom feature vector after each patient interaction are set as the basic triplet; Set any one of the data in the basic triplet as a positive sample pair, and set the data in the basic triplet excluding the positive sample pair as a negative sample pair; Construct supervised contrastive loss and center loss; the expression for the supervised contrastive loss is: The expression for the center loss is: ;in, , These are the supervised comparison loss and the center loss, respectively. , These are the first vector and the second vector, respectively. For scaling parameters; This represents the amount of negative sample data. The feature of the i-th negative sample; It is the set of pairs of the image feature vector, the sEMG feature vector, and the symptom feature vector; express Any combination of the following; This represents the total number of samples. , , These represent the imaging modality, sEMG modality, and symptom modality, respectively. It can be any one of the image modality, the sEMG modality, and the symptom modality; The output of the j-th sample and its t-th mode after cross-modal interaction processing; Let be the class center of the t-th modality; The supervised contrast loss and the center loss are calculated based on the positive sample pairs and the negative sample pairs, and the calculated supervised contrast loss and center loss are weighted and fused to obtain the joint total loss; The network layers of the bidirectional cross-modal attention mechanism are iteratively optimized using the joint total loss to obtain the feature space alignment network after iteration.
8. The method for predicting spinal muscle and ligament injuries according to claim 1, characterized in that, The feature fusion and prediction network is used to fuse the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain multi-dimensional fused features. A multi-output DNN network is then used to predict and calculate these multi-dimensional fused features to obtain the network output, including: The multi-output DNN network is constructed as follows: the multi-output DNN network includes: an input layer, a first hidden layer, a second hidden layer, and a two-branch output layer connected in sequence; the two-branch output layer includes: a parallel damage type output layer and a damage location output layer. The input layer is used to sequentially perform outlier filtering and feature standardization on the feature-aligned image feature vector, the sEMG feature vector, and the symptom feature vector to obtain the input matrix; The input matrix is subjected to feature abstraction processing using the first hidden layer to obtain initial hidden features; the expression of the first hidden layer is: ;in, The initial hidden feature; This is the first weight matrix; The input matrix; This is the first bias; The initial hidden features are processed by feature association using the second hidden layer to obtain optimized hidden features; the expression for the second hidden layer is: ;in, For the optimized hidden features; This is the second weight matrix; This is the second bias; The regularization coefficient; The optimized hidden features are synchronously input into the damage type output layer and the damage location output layer for probability mapping calculation, so as to obtain the damage type prediction output and the damage location prediction output.
9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor to enable the processor to perform a method for predicting spinal muscle and ligament injuries according to any one of claims 1 to 8.
10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute a method for predicting spinal muscle and ligament injuries according to any one of claims 1 to 8.